linux/net/ipv4/tcp_output.c

3844 lines
111 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
/*
* INET An implementation of the TCP/IP protocol suite for the LINUX
* operating system. INET is implemented using the BSD Socket
* interface as the means of communication with the user level.
*
* Implementation of the Transmission Control Protocol(TCP).
*
* Authors: Ross Biro
* Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
* Mark Evans, <evansmp@uhura.aston.ac.uk>
* Corey Minyard <wf-rch!minyard@relay.EU.net>
* Florian La Roche, <flla@stud.uni-sb.de>
* Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
* Linus Torvalds, <torvalds@cs.helsinki.fi>
* Alan Cox, <gw4pts@gw4pts.ampr.org>
* Matthew Dillon, <dillon@apollo.west.oic.com>
* Arnt Gulbrandsen, <agulbra@nvg.unit.no>
* Jorge Cwik, <jorge@laser.satlink.net>
*/
/*
* Changes: Pedro Roque : Retransmit queue handled by TCP.
* : Fragmentation on mtu decrease
* : Segment collapse on retransmit
* : AF independence
*
* Linus Torvalds : send_delayed_ack
* David S. Miller : Charge memory using the right skb
* during syn/ack processing.
* David S. Miller : Output engine completely rewritten.
* Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
* Cacophonix Gaul : draft-minshall-nagle-01
* J Hadi Salim : ECN support
*
*/
#define pr_fmt(fmt) "TCP: " fmt
#include <net/tcp.h>
#include <linux/compiler.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 16:04:11 +08:00
#include <linux/gfp.h>
#include <linux/module.h>
#include <linux/static_key.h>
#include <trace/events/tcp.h>
/* Refresh clocks of a TCP socket,
* ensuring monotically increasing values.
*/
void tcp_mstamp_refresh(struct tcp_sock *tp)
{
u64 val = tcp_clock_ns();
tp->tcp_clock_cache = val;
tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
}
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
int push_one, gfp_t gfp);
/* Account for new data that has been sent to the network. */
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
{
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
struct inet_connection_sock *icsk = inet_csk(sk);
[TCP]: Sed magic converts func(sk, tp, ...) -> func(sk, ...) This is (mostly) automated change using magic: sed -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e 's|struct sock \*sk,[\n\t ]*struct tcp_sock \*tp\([^{]*\n{\n\)| struct sock \*sk\1\tstruct tcp_sock *tp = tcp_sk(sk);\n|g' -e 's|struct sock \*sk, struct tcp_sock \*tp| struct sock \*sk|g' -e 's|sk, tp\([^-]\)|sk\1|g' Fixed four unused variable (tp) warnings that were introduced. In addition, manually added newlines after local variables and tweaked function arguments positioning. $ gcc --version gcc (GCC) 4.1.1 20060525 (Red Hat 4.1.1-1) ... $ codiff -fV built-in.o.old built-in.o.new net/ipv4/route.c: rt_cache_flush | +14 1 function changed, 14 bytes added net/ipv4/tcp.c: tcp_setsockopt | -5 tcp_sendpage | -25 tcp_sendmsg | -16 3 functions changed, 46 bytes removed net/ipv4/tcp_input.c: tcp_try_undo_recovery | +3 tcp_try_undo_dsack | +2 tcp_mark_head_lost | -12 tcp_ack | -15 tcp_event_data_recv | -32 tcp_rcv_state_process | -10 tcp_rcv_established | +1 7 functions changed, 6 bytes added, 69 bytes removed, diff: -63 net/ipv4/tcp_output.c: update_send_head | -9 tcp_transmit_skb | +19 tcp_cwnd_validate | +1 tcp_write_wakeup | -17 __tcp_push_pending_frames | -25 tcp_push_one | -8 tcp_send_fin | -4 7 functions changed, 20 bytes added, 63 bytes removed, diff: -43 built-in.o.new: 18 functions changed, 40 bytes added, 178 bytes removed, diff: -138 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-04-21 13:18:02 +08:00
struct tcp_sock *tp = tcp_sk(sk);
unsigned int prior_packets = tp->packets_out;
[TCP]: Sed magic converts func(sk, tp, ...) -> func(sk, ...) This is (mostly) automated change using magic: sed -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e 's|struct sock \*sk,[\n\t ]*struct tcp_sock \*tp\([^{]*\n{\n\)| struct sock \*sk\1\tstruct tcp_sock *tp = tcp_sk(sk);\n|g' -e 's|struct sock \*sk, struct tcp_sock \*tp| struct sock \*sk|g' -e 's|sk, tp\([^-]\)|sk\1|g' Fixed four unused variable (tp) warnings that were introduced. In addition, manually added newlines after local variables and tweaked function arguments positioning. $ gcc --version gcc (GCC) 4.1.1 20060525 (Red Hat 4.1.1-1) ... $ codiff -fV built-in.o.old built-in.o.new net/ipv4/route.c: rt_cache_flush | +14 1 function changed, 14 bytes added net/ipv4/tcp.c: tcp_setsockopt | -5 tcp_sendpage | -25 tcp_sendmsg | -16 3 functions changed, 46 bytes removed net/ipv4/tcp_input.c: tcp_try_undo_recovery | +3 tcp_try_undo_dsack | +2 tcp_mark_head_lost | -12 tcp_ack | -15 tcp_event_data_recv | -32 tcp_rcv_state_process | -10 tcp_rcv_established | +1 7 functions changed, 6 bytes added, 69 bytes removed, diff: -63 net/ipv4/tcp_output.c: update_send_head | -9 tcp_transmit_skb | +19 tcp_cwnd_validate | +1 tcp_write_wakeup | -17 __tcp_push_pending_frames | -25 tcp_push_one | -8 tcp_send_fin | -4 7 functions changed, 20 bytes added, 63 bytes removed, diff: -43 built-in.o.new: 18 functions changed, 40 bytes added, 178 bytes removed, diff: -138 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-04-21 13:18:02 +08:00
WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
__skb_unlink(skb, &sk->sk_write_queue);
tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
tp->packets_out += tcp_skb_pcount(skb);
if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
tcp_rearm_rto(sk);
NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
tcp_skb_pcount(skb));
}
/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
* window scaling factor due to loss of precision.
* If window has been shrunk, what should we make? It is not clear at all.
* Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
* Anything in between SND.UNA...SND.UNA+SND.WND also can be already
* invalid. OK, let's make this for now:
*/
static inline __u32 tcp_acceptable_seq(const struct sock *sk)
{
const struct tcp_sock *tp = tcp_sk(sk);
[TCP]: Sed magic converts func(sk, tp, ...) -> func(sk, ...) This is (mostly) automated change using magic: sed -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e 's|struct sock \*sk,[\n\t ]*struct tcp_sock \*tp\([^{]*\n{\n\)| struct sock \*sk\1\tstruct tcp_sock *tp = tcp_sk(sk);\n|g' -e 's|struct sock \*sk, struct tcp_sock \*tp| struct sock \*sk|g' -e 's|sk, tp\([^-]\)|sk\1|g' Fixed four unused variable (tp) warnings that were introduced. In addition, manually added newlines after local variables and tweaked function arguments positioning. $ gcc --version gcc (GCC) 4.1.1 20060525 (Red Hat 4.1.1-1) ... $ codiff -fV built-in.o.old built-in.o.new net/ipv4/route.c: rt_cache_flush | +14 1 function changed, 14 bytes added net/ipv4/tcp.c: tcp_setsockopt | -5 tcp_sendpage | -25 tcp_sendmsg | -16 3 functions changed, 46 bytes removed net/ipv4/tcp_input.c: tcp_try_undo_recovery | +3 tcp_try_undo_dsack | +2 tcp_mark_head_lost | -12 tcp_ack | -15 tcp_event_data_recv | -32 tcp_rcv_state_process | -10 tcp_rcv_established | +1 7 functions changed, 6 bytes added, 69 bytes removed, diff: -63 net/ipv4/tcp_output.c: update_send_head | -9 tcp_transmit_skb | +19 tcp_cwnd_validate | +1 tcp_write_wakeup | -17 __tcp_push_pending_frames | -25 tcp_push_one | -8 tcp_send_fin | -4 7 functions changed, 20 bytes added, 63 bytes removed, diff: -43 built-in.o.new: 18 functions changed, 40 bytes added, 178 bytes removed, diff: -138 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-04-21 13:18:02 +08:00
if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
(tp->rx_opt.wscale_ok &&
((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
return tp->snd_nxt;
else
return tcp_wnd_end(tp);
}
/* Calculate mss to advertise in SYN segment.
* RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
*
* 1. It is independent of path mtu.
* 2. Ideally, it is maximal possible segment size i.e. 65535-40.
* 3. For IPv4 it is reasonable to calculate it from maximal MTU of
* attached devices, because some buggy hosts are confused by
* large MSS.
* 4. We do not make 3, we advertise MSS, calculated from first
* hop device mtu, but allow to raise it to ip_rt_min_advmss.
* This may be overridden via information stored in routing table.
* 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
* probably even Jumbo".
*/
static __u16 tcp_advertise_mss(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
const struct dst_entry *dst = __sk_dst_get(sk);
int mss = tp->advmss;
if (dst) {
unsigned int metric = dst_metric_advmss(dst);
if (metric < mss) {
mss = metric;
tp->advmss = mss;
}
}
return (__u16)mss;
}
/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
tcp: fix slow start after idle vs TSO/GSO slow start after idle might reduce cwnd, but we perform this after first packet was cooked and sent. With TSO/GSO, it means that we might send a full TSO packet even if cwnd should have been reduced to IW10. Moving the SSAI check in skb_entail() makes sense, because we slightly reduce number of times this check is done, especially for large send() and TCP Small queue callbacks from softirq context. As Neal pointed out, we also need to perform the check if/when receive window opens. Tested: Following packetdrill test demonstrates the problem // Test of slow start after idle `sysctl -q net.ipv4.tcp_slow_start_after_idle=1` 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 65535 <mss 1000,sackOK,nop,nop,nop,wscale 7> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6> +.100 < . 1:1(0) ack 1 win 511 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_SOCKET, SO_SNDBUF, [200000], 4) = 0 +0 write(4, ..., 26000) = 26000 +0 > . 1:5001(5000) ack 1 +0 > . 5001:10001(5000) ack 1 +0 %{ assert tcpi_snd_cwnd == 10 }% +.100 < . 1:1(0) ack 10001 win 511 +0 %{ assert tcpi_snd_cwnd == 20, tcpi_snd_cwnd }% +0 > . 10001:20001(10000) ack 1 +0 > P. 20001:26001(6000) ack 1 +.100 < . 1:1(0) ack 26001 win 511 +0 %{ assert tcpi_snd_cwnd == 36, tcpi_snd_cwnd }% +4 write(4, ..., 20000) = 20000 // If slow start after idle works properly, we should send 5 MSS here (cwnd/2) +0 > . 26001:31001(5000) ack 1 +0 %{ assert tcpi_snd_cwnd == 10, tcpi_snd_cwnd }% +0 > . 31001:36001(5000) ack 1 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-08-22 03:30:00 +08:00
* This is the first part of cwnd validation mechanism.
*/
void tcp_cwnd_restart(struct sock *sk, s32 delta)
{
struct tcp_sock *tp = tcp_sk(sk);
tcp: fix slow start after idle vs TSO/GSO slow start after idle might reduce cwnd, but we perform this after first packet was cooked and sent. With TSO/GSO, it means that we might send a full TSO packet even if cwnd should have been reduced to IW10. Moving the SSAI check in skb_entail() makes sense, because we slightly reduce number of times this check is done, especially for large send() and TCP Small queue callbacks from softirq context. As Neal pointed out, we also need to perform the check if/when receive window opens. Tested: Following packetdrill test demonstrates the problem // Test of slow start after idle `sysctl -q net.ipv4.tcp_slow_start_after_idle=1` 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 65535 <mss 1000,sackOK,nop,nop,nop,wscale 7> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6> +.100 < . 1:1(0) ack 1 win 511 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_SOCKET, SO_SNDBUF, [200000], 4) = 0 +0 write(4, ..., 26000) = 26000 +0 > . 1:5001(5000) ack 1 +0 > . 5001:10001(5000) ack 1 +0 %{ assert tcpi_snd_cwnd == 10 }% +.100 < . 1:1(0) ack 10001 win 511 +0 %{ assert tcpi_snd_cwnd == 20, tcpi_snd_cwnd }% +0 > . 10001:20001(10000) ack 1 +0 > P. 20001:26001(6000) ack 1 +.100 < . 1:1(0) ack 26001 win 511 +0 %{ assert tcpi_snd_cwnd == 36, tcpi_snd_cwnd }% +4 write(4, ..., 20000) = 20000 // If slow start after idle works properly, we should send 5 MSS here (cwnd/2) +0 > . 26001:31001(5000) ack 1 +0 %{ assert tcpi_snd_cwnd == 10, tcpi_snd_cwnd }% +0 > . 31001:36001(5000) ack 1 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-08-22 03:30:00 +08:00
u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
u32 cwnd = tp->snd_cwnd;
tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
tp->snd_ssthresh = tcp_current_ssthresh(sk);
restart_cwnd = min(restart_cwnd, cwnd);
while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
cwnd >>= 1;
tp->snd_cwnd = max(cwnd, restart_cwnd);
tp->snd_cwnd_stamp = tcp_jiffies32;
tp->snd_cwnd_used = 0;
}
/* Congestion state accounting after a packet has been sent. */
static void tcp_event_data_sent(struct tcp_sock *tp,
struct sock *sk)
{
struct inet_connection_sock *icsk = inet_csk(sk);
const u32 now = tcp_jiffies32;
if (tcp_packets_in_flight(tp) == 0)
tcp_ca_event(sk, CA_EVENT_TX_START);
/* If this is the first data packet sent in response to the
* previous received data,
* and it is a reply for ato after last received packet,
* increase pingpong count.
*/
if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
(u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
inet_csk_inc_pingpong_cnt(sk);
tp->lsndtime = now;
}
/* Account for an ACK we sent. */
static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
u32 rcv_nxt)
{
struct tcp_sock *tp = tcp_sk(sk);
tcp: defer SACK compression after DupThresh Jean-Louis reported a TCP regression and bisected to recent SACK compression. After a loss episode (receiver not able to keep up and dropping packets because its backlog is full), linux TCP stack is sending a single SACK (DUPACK). Sender waits a full RTO timer before recovering losses. While RFC 6675 says in section 5, "Algorithm Details", (2) If DupAcks < DupThresh but IsLost (HighACK + 1) returns true -- indicating at least three segments have arrived above the current cumulative acknowledgment point, which is taken to indicate loss -- go to step (4). ... (4) Invoke fast retransmit and enter loss recovery as follows: there are old TCP stacks not implementing this strategy, and still counting the dupacks before starting fast retransmit. While these stacks probably perform poorly when receivers implement LRO/GRO, we should be a little more gentle to them. This patch makes sure we do not enable SACK compression unless 3 dupacks have been sent since last rcv_nxt update. Ideally we should even rearm the timer to send one or two more DUPACK if no more packets are coming, but that will be work aiming for linux-4.21. Many thanks to Jean-Louis for bisecting the issue, providing packet captures and testing this patch. Fixes: 5d9f4262b7ea ("tcp: add SACK compression") Reported-by: Jean-Louis Dupond <jean-louis@dupond.be> Tested-by: Jean-Louis Dupond <jean-louis@dupond.be> Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-11-20 21:53:59 +08:00
if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
tcp: defer SACK compression after DupThresh Jean-Louis reported a TCP regression and bisected to recent SACK compression. After a loss episode (receiver not able to keep up and dropping packets because its backlog is full), linux TCP stack is sending a single SACK (DUPACK). Sender waits a full RTO timer before recovering losses. While RFC 6675 says in section 5, "Algorithm Details", (2) If DupAcks < DupThresh but IsLost (HighACK + 1) returns true -- indicating at least three segments have arrived above the current cumulative acknowledgment point, which is taken to indicate loss -- go to step (4). ... (4) Invoke fast retransmit and enter loss recovery as follows: there are old TCP stacks not implementing this strategy, and still counting the dupacks before starting fast retransmit. While these stacks probably perform poorly when receivers implement LRO/GRO, we should be a little more gentle to them. This patch makes sure we do not enable SACK compression unless 3 dupacks have been sent since last rcv_nxt update. Ideally we should even rearm the timer to send one or two more DUPACK if no more packets are coming, but that will be work aiming for linux-4.21. Many thanks to Jean-Louis for bisecting the issue, providing packet captures and testing this patch. Fixes: 5d9f4262b7ea ("tcp: add SACK compression") Reported-by: Jean-Louis Dupond <jean-louis@dupond.be> Tested-by: Jean-Louis Dupond <jean-louis@dupond.be> Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-11-20 21:53:59 +08:00
tp->compressed_ack - TCP_FASTRETRANS_THRESH);
tp->compressed_ack = TCP_FASTRETRANS_THRESH;
if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
__sock_put(sk);
}
if (unlikely(rcv_nxt != tp->rcv_nxt))
return; /* Special ACK sent by DCTCP to reflect ECN */
tcp_dec_quickack_mode(sk, pkts);
inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
}
/* Determine a window scaling and initial window to offer.
* Based on the assumption that the given amount of space
* will be offered. Store the results in the tp structure.
* NOTE: for smooth operation initial space offering should
* be a multiple of mss if possible. We assume here that mss >= 1.
* This MUST be enforced by all callers.
*/
void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
__u32 *rcv_wnd, __u32 *window_clamp,
int wscale_ok, __u8 *rcv_wscale,
__u32 init_rcv_wnd)
{
unsigned int space = (__space < 0 ? 0 : __space);
/* If no clamp set the clamp to the max possible scaled window */
if (*window_clamp == 0)
(*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
space = min(*window_clamp, space);
/* Quantize space offering to a multiple of mss if possible. */
if (space > mss)
space = rounddown(space, mss);
/* NOTE: offering an initial window larger than 32767
* will break some buggy TCP stacks. If the admin tells us
* it is likely we could be speaking with such a buggy stack
* we will truncate our initial window offering to 32K-1
* unless the remote has sent us a window scaling option,
* which we interpret as a sign the remote TCP is not
* misinterpreting the window field as a signed quantity.
*/
if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
else
tcp: up initial rmem to 128KB and SYN rwin to around 64KB Previously TCP initial receive buffer is ~87KB by default and the initial receive window is ~29KB (20 MSS). This patch changes the two numbers to 128KB and ~64KB (rounding down to the multiples of MSS) respectively. The patch also simplifies the calculations s.t. the two numbers are directly controlled by sysctl tcp_rmem[1]: 1) Initial receiver buffer budget (sk_rcvbuf): while this should be configured via sysctl tcp_rmem[1], previously tcp_fixup_rcvbuf() always override and set a larger size when a new connection establishes. 2) Initial receive window in SYN: previously it is set to 20 packets if MSS <= 1460. The number 20 was based on the initial congestion window of 10: the receiver needs twice amount to avoid being limited by the receive window upon out-of-order delivery in the first window burst. But since this only applies if the receiving MSS <= 1460, connection using large MTU (e.g. to utilize receiver zero-copy) may be limited by the receive window. With this patch TCP memory configuration is more straight-forward and more properly sized to modern high-speed networks by default. Several popular stacks have been announcing 64KB rwin in SYNs as well. Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Wei Wang <weiwan@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-09-28 02:21:19 +08:00
(*rcv_wnd) = min_t(u32, space, U16_MAX);
if (init_rcv_wnd)
*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
*rcv_wscale = 0;
if (wscale_ok) {
/* Set window scaling on max possible window */
space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
space = max_t(u32, space, sysctl_rmem_max);
space = min_t(u32, space, *window_clamp);
*rcv_wscale = clamp_t(int, ilog2(space) - 15,
0, TCP_MAX_WSCALE);
}
/* Set the clamp no higher than max representable value */
(*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
}
EXPORT_SYMBOL(tcp_select_initial_window);
/* Chose a new window to advertise, update state in tcp_sock for the
* socket, and return result with RFC1323 scaling applied. The return
* value can be stuffed directly into th->window for an outgoing
* frame.
*/
static u16 tcp_select_window(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
u32 old_win = tp->rcv_wnd;
u32 cur_win = tcp_receive_window(tp);
u32 new_win = __tcp_select_window(sk);
/* Never shrink the offered window */
if (new_win < cur_win) {
/* Danger Will Robinson!
* Don't update rcv_wup/rcv_wnd here or else
* we will not be able to advertise a zero
* window in time. --DaveM
*
* Relax Will Robinson.
*/
if (new_win == 0)
NET_INC_STATS(sock_net(sk),
LINUX_MIB_TCPWANTZEROWINDOWADV);
[TCP]: Fix shrinking windows with window scaling When selecting a new window, tcp_select_window() tries not to shrink the offered window by using the maximum of the remaining offered window size and the newly calculated window size. The newly calculated window size is always a multiple of the window scaling factor, the remaining window size however might not be since it depends on rcv_wup/rcv_nxt. This means we're effectively shrinking the window when scaling it down. The dump below shows the problem (scaling factor 2^7): - Window size of 557 (71296) is advertised, up to 3111907257: IP 172.2.2.3.33000 > 172.2.2.2.33000: . ack 3111835961 win 557 <...> - New window size of 514 (65792) is advertised, up to 3111907217, 40 bytes below the last end: IP 172.2.2.3.33000 > 172.2.2.2.33000: . 3113575668:3113577116(1448) ack 3111841425 win 514 <...> The number 40 results from downscaling the remaining window: 3111907257 - 3111841425 = 65832 65832 / 2^7 = 514 65832 % 2^7 = 40 If the sender uses up the entire window before it is shrunk, this can have chaotic effects on the connection. When sending ACKs, tcp_acceptable_seq() will notice that the window has been shrunk since tcp_wnd_end() is before tp->snd_nxt, which makes it choose tcp_wnd_end() as sequence number. This will fail the receivers checks in tcp_sequence() however since it is before it's tp->rcv_wup, making it respond with a dupack. If both sides are in this condition, this leads to a constant flood of ACKs until the connection times out. Make sure the window is never shrunk by aligning the remaining window to the window scaling factor. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-03-21 07:11:27 +08:00
new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
}
tp->rcv_wnd = new_win;
tp->rcv_wup = tp->rcv_nxt;
/* Make sure we do not exceed the maximum possible
* scaled window.
*/
if (!tp->rx_opt.rcv_wscale &&
sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
new_win = min(new_win, MAX_TCP_WINDOW);
else
new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
/* RFC1323 scaling applied */
new_win >>= tp->rx_opt.rcv_wscale;
/* If we advertise zero window, disable fast path. */
if (new_win == 0) {
tp->pred_flags = 0;
if (old_win)
NET_INC_STATS(sock_net(sk),
LINUX_MIB_TCPTOZEROWINDOWADV);
} else if (old_win == 0) {
NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
}
return new_win;
}
/* Packet ECN state for a SYN-ACK */
static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
{
const struct tcp_sock *tp = tcp_sk(sk);
TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
if (!(tp->ecn_flags & TCP_ECN_OK))
TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
else if (tcp_ca_needs_ecn(sk) ||
tcp_bpf_ca_needs_ecn(sk))
INET_ECN_xmit(sk);
}
/* Packet ECN state for a SYN. */
static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
{
struct tcp_sock *tp = tcp_sk(sk);
bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
net: allow setting ecn via routing table This patch allows to set ECN on a per-route basis in case the sysctl tcp_ecn is not set to 1. In other words, when ECN is set for specific routes, it provides a tcp_ecn=1 behaviour for that route while the rest of the stack acts according to the global settings. One can use 'ip route change dev $dev $net features ecn' to toggle this. Having a more fine-grained per-route setting can be beneficial for various reasons, for example, 1) within data centers, or 2) local ISPs may deploy ECN support for their own video/streaming services [1], etc. There was a recent measurement study/paper [2] which scanned the Alexa's publicly available top million websites list from a vantage point in US, Europe and Asia: Half of the Alexa list will now happily use ECN (tcp_ecn=2, most likely blamed to commit 255cac91c3 ("tcp: extend ECN sysctl to allow server-side only ECN") ;)); the break in connectivity on-path was found is about 1 in 10,000 cases. Timeouts rather than receiving back RSTs were much more common in the negotiation phase (and mostly seen in the Alexa middle band, ranks around 50k-150k): from 12-thousand hosts on which there _may_ be ECN-linked connection failures, only 79 failed with RST when _not_ failing with RST when ECN is not requested. It's unclear though, how much equipment in the wild actually marks CE when buffers start to fill up. We thought about a fallback to non-ECN for retransmitted SYNs as another global option (which could perhaps one day be made default), but as Eric points out, there's much more work needed to detect broken middleboxes. Two examples Eric mentioned are buggy firewalls that accept only a single SYN per flow, and middleboxes that successfully let an ECN flow establish, but later mark CE for all packets (so cwnd converges to 1). [1] http://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf, p.15 [2] http://ecn.ethz.ch/ Joint work with Daniel Borkmann. Reference: http://thread.gmane.org/gmane.linux.network/335797 Suggested-by: Hannes Frederic Sowa <hannes@stressinduktion.org> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-04 00:35:03 +08:00
bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
net: allow setting ecn via routing table This patch allows to set ECN on a per-route basis in case the sysctl tcp_ecn is not set to 1. In other words, when ECN is set for specific routes, it provides a tcp_ecn=1 behaviour for that route while the rest of the stack acts according to the global settings. One can use 'ip route change dev $dev $net features ecn' to toggle this. Having a more fine-grained per-route setting can be beneficial for various reasons, for example, 1) within data centers, or 2) local ISPs may deploy ECN support for their own video/streaming services [1], etc. There was a recent measurement study/paper [2] which scanned the Alexa's publicly available top million websites list from a vantage point in US, Europe and Asia: Half of the Alexa list will now happily use ECN (tcp_ecn=2, most likely blamed to commit 255cac91c3 ("tcp: extend ECN sysctl to allow server-side only ECN") ;)); the break in connectivity on-path was found is about 1 in 10,000 cases. Timeouts rather than receiving back RSTs were much more common in the negotiation phase (and mostly seen in the Alexa middle band, ranks around 50k-150k): from 12-thousand hosts on which there _may_ be ECN-linked connection failures, only 79 failed with RST when _not_ failing with RST when ECN is not requested. It's unclear though, how much equipment in the wild actually marks CE when buffers start to fill up. We thought about a fallback to non-ECN for retransmitted SYNs as another global option (which could perhaps one day be made default), but as Eric points out, there's much more work needed to detect broken middleboxes. Two examples Eric mentioned are buggy firewalls that accept only a single SYN per flow, and middleboxes that successfully let an ECN flow establish, but later mark CE for all packets (so cwnd converges to 1). [1] http://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf, p.15 [2] http://ecn.ethz.ch/ Joint work with Daniel Borkmann. Reference: http://thread.gmane.org/gmane.linux.network/335797 Suggested-by: Hannes Frederic Sowa <hannes@stressinduktion.org> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-04 00:35:03 +08:00
if (!use_ecn) {
const struct dst_entry *dst = __sk_dst_get(sk);
if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
use_ecn = true;
}
tp->ecn_flags = 0;
net: allow setting ecn via routing table This patch allows to set ECN on a per-route basis in case the sysctl tcp_ecn is not set to 1. In other words, when ECN is set for specific routes, it provides a tcp_ecn=1 behaviour for that route while the rest of the stack acts according to the global settings. One can use 'ip route change dev $dev $net features ecn' to toggle this. Having a more fine-grained per-route setting can be beneficial for various reasons, for example, 1) within data centers, or 2) local ISPs may deploy ECN support for their own video/streaming services [1], etc. There was a recent measurement study/paper [2] which scanned the Alexa's publicly available top million websites list from a vantage point in US, Europe and Asia: Half of the Alexa list will now happily use ECN (tcp_ecn=2, most likely blamed to commit 255cac91c3 ("tcp: extend ECN sysctl to allow server-side only ECN") ;)); the break in connectivity on-path was found is about 1 in 10,000 cases. Timeouts rather than receiving back RSTs were much more common in the negotiation phase (and mostly seen in the Alexa middle band, ranks around 50k-150k): from 12-thousand hosts on which there _may_ be ECN-linked connection failures, only 79 failed with RST when _not_ failing with RST when ECN is not requested. It's unclear though, how much equipment in the wild actually marks CE when buffers start to fill up. We thought about a fallback to non-ECN for retransmitted SYNs as another global option (which could perhaps one day be made default), but as Eric points out, there's much more work needed to detect broken middleboxes. Two examples Eric mentioned are buggy firewalls that accept only a single SYN per flow, and middleboxes that successfully let an ECN flow establish, but later mark CE for all packets (so cwnd converges to 1). [1] http://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf, p.15 [2] http://ecn.ethz.ch/ Joint work with Daniel Borkmann. Reference: http://thread.gmane.org/gmane.linux.network/335797 Suggested-by: Hannes Frederic Sowa <hannes@stressinduktion.org> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-04 00:35:03 +08:00
if (use_ecn) {
TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
tp->ecn_flags = TCP_ECN_OK;
if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
INET_ECN_xmit(sk);
}
}
tcp: add rfc3168, section 6.1.1.1. fallback This work as a follow-up of commit f7b3bec6f516 ("net: allow setting ecn via routing table") and adds RFC3168 section 6.1.1.1. fallback for outgoing ECN connections. In other words, this work adds a retry with a non-ECN setup SYN packet, as suggested from the RFC on the first timeout: [...] A host that receives no reply to an ECN-setup SYN within the normal SYN retransmission timeout interval MAY resend the SYN and any subsequent SYN retransmissions with CWR and ECE cleared. [...] Schematic client-side view when assuming the server is in tcp_ecn=2 mode, that is, Linux default since 2009 via commit 255cac91c3c9 ("tcp: extend ECN sysctl to allow server-side only ECN"): 1) Normal ECN-capable path: SYN ECE CWR -----> <----- SYN ACK ECE ACK -----> 2) Path with broken middlebox, when client has fallback: SYN ECE CWR ----X crappy middlebox drops packet (timeout, rtx) SYN -----> <----- SYN ACK ACK -----> In case we would not have the fallback implemented, the middlebox drop point would basically end up as: SYN ECE CWR ----X crappy middlebox drops packet (timeout, rtx) SYN ECE CWR ----X crappy middlebox drops packet (timeout, rtx) SYN ECE CWR ----X crappy middlebox drops packet (timeout, rtx) In any case, it's rather a smaller percentage of sites where there would occur such additional setup latency: it was found in end of 2014 that ~56% of IPv4 and 65% of IPv6 servers of Alexa 1 million list would negotiate ECN (aka tcp_ecn=2 default), 0.42% of these webservers will fail to connect when trying to negotiate with ECN (tcp_ecn=1) due to timeouts, which the fallback would mitigate with a slight latency trade-off. Recent related paper on this topic: Brian Trammell, Mirja Kühlewind, Damiano Boppart, Iain Learmonth, Gorry Fairhurst, and Richard Scheffenegger: "Enabling Internet-Wide Deployment of Explicit Congestion Notification." Proc. PAM 2015, New York. http://ecn.ethz.ch/ecn-pam15.pdf Thus, when net.ipv4.tcp_ecn=1 is being set, the patch will perform RFC3168, section 6.1.1.1. fallback on timeout. For users explicitly not wanting this which can be in DC use case, we add a net.ipv4.tcp_ecn_fallback knob that allows for disabling the fallback. tp->ecn_flags are not being cleared in tcp_ecn_clear_syn() on output, but rather we let tcp_ecn_rcv_synack() take that over on input path in case a SYN ACK ECE was delayed. Thus a spurious SYN retransmission will not prevent ECN being negotiated eventually in that case. Reference: https://www.ietf.org/proceedings/92/slides/slides-92-iccrg-1.pdf Reference: https://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: Mirja Kühlewind <mirja.kuehlewind@tik.ee.ethz.ch> Signed-off-by: Brian Trammell <trammell@tik.ee.ethz.ch> Cc: Eric Dumazet <edumazet@google.com> Cc: Dave That <dave.taht@gmail.com> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-20 03:04:22 +08:00
static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
{
if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
/* tp->ecn_flags are cleared at a later point in time when
* SYN ACK is ultimatively being received.
*/
TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
}
static void
tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
{
if (inet_rsk(req)->ecn_ok)
th->ece = 1;
}
/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
* be sent.
*/
static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
struct tcphdr *th, int tcp_header_len)
{
struct tcp_sock *tp = tcp_sk(sk);
if (tp->ecn_flags & TCP_ECN_OK) {
/* Not-retransmitted data segment: set ECT and inject CWR. */
if (skb->len != tcp_header_len &&
!before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
INET_ECN_xmit(sk);
if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
th->cwr = 1;
skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
}
} else if (!tcp_ca_needs_ecn(sk)) {
/* ACK or retransmitted segment: clear ECT|CE */
INET_ECN_dontxmit(sk);
}
if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
th->ece = 1;
}
}
/* Constructs common control bits of non-data skb. If SYN/FIN is present,
* auto increment end seqno.
*/
static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
{
skb->ip_summed = CHECKSUM_PARTIAL;
TCP_SKB_CB(skb)->tcp_flags = flags;
TCP_SKB_CB(skb)->sacked = 0;
tcp_skb_pcount_set(skb, 1);
TCP_SKB_CB(skb)->seq = seq;
if (flags & (TCPHDR_SYN | TCPHDR_FIN))
seq++;
TCP_SKB_CB(skb)->end_seq = seq;
}
static inline bool tcp_urg_mode(const struct tcp_sock *tp)
{
return tp->snd_una != tp->snd_up;
}
#define OPTION_SACK_ADVERTISE (1 << 0)
#define OPTION_TS (1 << 1)
#define OPTION_MD5 (1 << 2)
IPv4 TCP fails to send window scale option when window scale is zero Acknowledge TCP window scale support by inserting the proper option in SYN/ACK and SYN headers even if our window scale is zero. This fixes the following observed behavior: 1. Client sends a SYN with TCP window scaling option and non zero window scale value to a Linux box. 2. Linux box notes large receive window from client. 3. Linux decides on a zero value of window scale for its part. 4. Due to compare against requested window scale size option, Linux does not to send windows scale TCP option header on SYN/ACK at all. With the following result: Client box thinks TCP window scaling is not supported, since SYN/ACK had no TCP window scale option, while Linux thinks that TCP window scaling is supported (and scale might be non zero), since SYN had TCP window scale option and we have a mismatched idea between the client and server regarding window sizes. Probably it also fixes up the following bug (not observed in practice): 1. Linux box opens TCP connection to some server. 2. Linux decides on zero value of window scale. 3. Due to compare against computed window scale size option, Linux does not to set windows scale TCP option header on SYN. With the expected result that the server OS does not use window scale option due to not receiving such an option in the SYN headers, leading to suboptimal performance. Signed-off-by: Gilad Ben-Yossef <gilad@codefidence.com> Signed-off-by: Ori Finkelman <ori@comsleep.com> Acked-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-10-01 14:41:59 +08:00
#define OPTION_WSCALE (1 << 3)
#define OPTION_FAST_OPEN_COOKIE (1 << 8)
#define OPTION_SMC (1 << 9)
static void smc_options_write(__be32 *ptr, u16 *options)
{
#if IS_ENABLED(CONFIG_SMC)
if (static_branch_unlikely(&tcp_have_smc)) {
if (unlikely(OPTION_SMC & *options)) {
*ptr++ = htonl((TCPOPT_NOP << 24) |
(TCPOPT_NOP << 16) |
(TCPOPT_EXP << 8) |
(TCPOLEN_EXP_SMC_BASE));
*ptr++ = htonl(TCPOPT_SMC_MAGIC);
}
}
#endif
}
struct tcp_out_options {
u16 options; /* bit field of OPTION_* */
u16 mss; /* 0 to disable */
u8 ws; /* window scale, 0 to disable */
u8 num_sack_blocks; /* number of SACK blocks to include */
u8 hash_size; /* bytes in hash_location */
__u8 *hash_location; /* temporary pointer, overloaded */
__u32 tsval, tsecr; /* need to include OPTION_TS */
struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
};
/* Write previously computed TCP options to the packet.
*
* Beware: Something in the Internet is very sensitive to the ordering of
* TCP options, we learned this through the hard way, so be careful here.
* Luckily we can at least blame others for their non-compliance but from
* inter-operability perspective it seems that we're somewhat stuck with
* the ordering which we have been using if we want to keep working with
* those broken things (not that it currently hurts anybody as there isn't
* particular reason why the ordering would need to be changed).
*
* At least SACK_PERM as the first option is known to lead to a disaster
* (but it may well be that other scenarios fail similarly).
*/
static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
struct tcp_out_options *opts)
{
u16 options = opts->options; /* mungable copy */
if (unlikely(OPTION_MD5 & options)) {
*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
(TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
/* overload cookie hash location */
opts->hash_location = (__u8 *)ptr;
ptr += 4;
}
if (unlikely(opts->mss)) {
*ptr++ = htonl((TCPOPT_MSS << 24) |
(TCPOLEN_MSS << 16) |
opts->mss);
}
if (likely(OPTION_TS & options)) {
if (unlikely(OPTION_SACK_ADVERTISE & options)) {
*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
(TCPOLEN_SACK_PERM << 16) |
(TCPOPT_TIMESTAMP << 8) |
TCPOLEN_TIMESTAMP);
options &= ~OPTION_SACK_ADVERTISE;
} else {
*ptr++ = htonl((TCPOPT_NOP << 24) |
(TCPOPT_NOP << 16) |
(TCPOPT_TIMESTAMP << 8) |
TCPOLEN_TIMESTAMP);
}
*ptr++ = htonl(opts->tsval);
*ptr++ = htonl(opts->tsecr);
}
if (unlikely(OPTION_SACK_ADVERTISE & options)) {
*ptr++ = htonl((TCPOPT_NOP << 24) |
(TCPOPT_NOP << 16) |
(TCPOPT_SACK_PERM << 8) |
TCPOLEN_SACK_PERM);
}
if (unlikely(OPTION_WSCALE & options)) {
*ptr++ = htonl((TCPOPT_NOP << 24) |
(TCPOPT_WINDOW << 16) |
(TCPOLEN_WINDOW << 8) |
opts->ws);
}
if (unlikely(opts->num_sack_blocks)) {
struct tcp_sack_block *sp = tp->rx_opt.dsack ?
tp->duplicate_sack : tp->selective_acks;
int this_sack;
*ptr++ = htonl((TCPOPT_NOP << 24) |
(TCPOPT_NOP << 16) |
(TCPOPT_SACK << 8) |
(TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
TCPOLEN_SACK_PERBLOCK)));
for (this_sack = 0; this_sack < opts->num_sack_blocks;
++this_sack) {
*ptr++ = htonl(sp[this_sack].start_seq);
*ptr++ = htonl(sp[this_sack].end_seq);
}
tp->rx_opt.dsack = 0;
}
if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
u8 *p = (u8 *)ptr;
u32 len; /* Fast Open option length */
if (foc->exp) {
len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
TCPOPT_FASTOPEN_MAGIC);
p += TCPOLEN_EXP_FASTOPEN_BASE;
} else {
len = TCPOLEN_FASTOPEN_BASE + foc->len;
*p++ = TCPOPT_FASTOPEN;
*p++ = len;
}
memcpy(p, foc->val, foc->len);
if ((len & 3) == 2) {
p[foc->len] = TCPOPT_NOP;
p[foc->len + 1] = TCPOPT_NOP;
}
ptr += (len + 3) >> 2;
}
smc_options_write(ptr, &options);
}
static void smc_set_option(const struct tcp_sock *tp,
struct tcp_out_options *opts,
unsigned int *remaining)
{
#if IS_ENABLED(CONFIG_SMC)
if (static_branch_unlikely(&tcp_have_smc)) {
if (tp->syn_smc) {
if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
opts->options |= OPTION_SMC;
*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
}
}
}
#endif
}
static void smc_set_option_cond(const struct tcp_sock *tp,
const struct inet_request_sock *ireq,
struct tcp_out_options *opts,
unsigned int *remaining)
{
#if IS_ENABLED(CONFIG_SMC)
if (static_branch_unlikely(&tcp_have_smc)) {
if (tp->syn_smc && ireq->smc_ok) {
if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
opts->options |= OPTION_SMC;
*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
}
}
}
#endif
}
/* Compute TCP options for SYN packets. This is not the final
* network wire format yet.
*/
static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
struct tcp_out_options *opts,
struct tcp_md5sig_key **md5)
{
struct tcp_sock *tp = tcp_sk(sk);
unsigned int remaining = MAX_TCP_OPTION_SPACE;
struct tcp_fastopen_request *fastopen = tp->fastopen_req;
*md5 = NULL;
#ifdef CONFIG_TCP_MD5SIG
if (static_branch_unlikely(&tcp_md5_needed) &&
rcu_access_pointer(tp->md5sig_info)) {
*md5 = tp->af_specific->md5_lookup(sk, sk);
if (*md5) {
opts->options |= OPTION_MD5;
remaining -= TCPOLEN_MD5SIG_ALIGNED;
}
}
#endif
/* We always get an MSS option. The option bytes which will be seen in
* normal data packets should timestamps be used, must be in the MSS
* advertised. But we subtract them from tp->mss_cache so that
* calculations in tcp_sendmsg are simpler etc. So account for this
* fact here if necessary. If we don't do this correctly, as a
* receiver we won't recognize data packets as being full sized when we
* should, and thus we won't abide by the delayed ACK rules correctly.
* SACKs don't matter, we never delay an ACK when we have any of those
* going out. */
opts->mss = tcp_advertise_mss(sk);
remaining -= TCPOLEN_MSS_ALIGNED;
if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
opts->options |= OPTION_TS;
opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
opts->tsecr = tp->rx_opt.ts_recent;
remaining -= TCPOLEN_TSTAMP_ALIGNED;
}
if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
opts->ws = tp->rx_opt.rcv_wscale;
IPv4 TCP fails to send window scale option when window scale is zero Acknowledge TCP window scale support by inserting the proper option in SYN/ACK and SYN headers even if our window scale is zero. This fixes the following observed behavior: 1. Client sends a SYN with TCP window scaling option and non zero window scale value to a Linux box. 2. Linux box notes large receive window from client. 3. Linux decides on a zero value of window scale for its part. 4. Due to compare against requested window scale size option, Linux does not to send windows scale TCP option header on SYN/ACK at all. With the following result: Client box thinks TCP window scaling is not supported, since SYN/ACK had no TCP window scale option, while Linux thinks that TCP window scaling is supported (and scale might be non zero), since SYN had TCP window scale option and we have a mismatched idea between the client and server regarding window sizes. Probably it also fixes up the following bug (not observed in practice): 1. Linux box opens TCP connection to some server. 2. Linux decides on zero value of window scale. 3. Due to compare against computed window scale size option, Linux does not to set windows scale TCP option header on SYN. With the expected result that the server OS does not use window scale option due to not receiving such an option in the SYN headers, leading to suboptimal performance. Signed-off-by: Gilad Ben-Yossef <gilad@codefidence.com> Signed-off-by: Ori Finkelman <ori@comsleep.com> Acked-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-10-01 14:41:59 +08:00
opts->options |= OPTION_WSCALE;
remaining -= TCPOLEN_WSCALE_ALIGNED;
}
if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
opts->options |= OPTION_SACK_ADVERTISE;
if (unlikely(!(OPTION_TS & opts->options)))
remaining -= TCPOLEN_SACKPERM_ALIGNED;
}
if (fastopen && fastopen->cookie.len >= 0) {
u32 need = fastopen->cookie.len;
need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
TCPOLEN_FASTOPEN_BASE;
need = (need + 3) & ~3U; /* Align to 32 bits */
if (remaining >= need) {
opts->options |= OPTION_FAST_OPEN_COOKIE;
opts->fastopen_cookie = &fastopen->cookie;
remaining -= need;
tp->syn_fastopen = 1;
tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
}
}
smc_set_option(tp, opts, &remaining);
return MAX_TCP_OPTION_SPACE - remaining;
}
/* Set up TCP options for SYN-ACKs. */
static unsigned int tcp_synack_options(const struct sock *sk,
struct request_sock *req,
unsigned int mss, struct sk_buff *skb,
struct tcp_out_options *opts,
const struct tcp_md5sig_key *md5,
struct tcp_fastopen_cookie *foc)
{
struct inet_request_sock *ireq = inet_rsk(req);
unsigned int remaining = MAX_TCP_OPTION_SPACE;
#ifdef CONFIG_TCP_MD5SIG
if (md5) {
opts->options |= OPTION_MD5;
remaining -= TCPOLEN_MD5SIG_ALIGNED;
/* We can't fit any SACK blocks in a packet with MD5 + TS
* options. There was discussion about disabling SACK
* rather than TS in order to fit in better with old,
* buggy kernels, but that was deemed to be unnecessary.
*/
ireq->tstamp_ok &= !ireq->sack_ok;
}
#endif
/* We always send an MSS option. */
opts->mss = mss;
remaining -= TCPOLEN_MSS_ALIGNED;
if (likely(ireq->wscale_ok)) {
opts->ws = ireq->rcv_wscale;
IPv4 TCP fails to send window scale option when window scale is zero Acknowledge TCP window scale support by inserting the proper option in SYN/ACK and SYN headers even if our window scale is zero. This fixes the following observed behavior: 1. Client sends a SYN with TCP window scaling option and non zero window scale value to a Linux box. 2. Linux box notes large receive window from client. 3. Linux decides on a zero value of window scale for its part. 4. Due to compare against requested window scale size option, Linux does not to send windows scale TCP option header on SYN/ACK at all. With the following result: Client box thinks TCP window scaling is not supported, since SYN/ACK had no TCP window scale option, while Linux thinks that TCP window scaling is supported (and scale might be non zero), since SYN had TCP window scale option and we have a mismatched idea between the client and server regarding window sizes. Probably it also fixes up the following bug (not observed in practice): 1. Linux box opens TCP connection to some server. 2. Linux decides on zero value of window scale. 3. Due to compare against computed window scale size option, Linux does not to set windows scale TCP option header on SYN. With the expected result that the server OS does not use window scale option due to not receiving such an option in the SYN headers, leading to suboptimal performance. Signed-off-by: Gilad Ben-Yossef <gilad@codefidence.com> Signed-off-by: Ori Finkelman <ori@comsleep.com> Acked-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-10-01 14:41:59 +08:00
opts->options |= OPTION_WSCALE;
remaining -= TCPOLEN_WSCALE_ALIGNED;
}
if (likely(ireq->tstamp_ok)) {
opts->options |= OPTION_TS;
opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
opts->tsecr = req->ts_recent;
remaining -= TCPOLEN_TSTAMP_ALIGNED;
}
if (likely(ireq->sack_ok)) {
opts->options |= OPTION_SACK_ADVERTISE;
if (unlikely(!ireq->tstamp_ok))
remaining -= TCPOLEN_SACKPERM_ALIGNED;
}
if (foc != NULL && foc->len >= 0) {
u32 need = foc->len;
need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
TCPOLEN_FASTOPEN_BASE;
need = (need + 3) & ~3U; /* Align to 32 bits */
if (remaining >= need) {
opts->options |= OPTION_FAST_OPEN_COOKIE;
opts->fastopen_cookie = foc;
remaining -= need;
}
}
smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
return MAX_TCP_OPTION_SPACE - remaining;
}
/* Compute TCP options for ESTABLISHED sockets. This is not the
* final wire format yet.
*/
static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
struct tcp_out_options *opts,
struct tcp_md5sig_key **md5)
{
struct tcp_sock *tp = tcp_sk(sk);
unsigned int size = 0;
unsigned int eff_sacks;
opts->options = 0;
*md5 = NULL;
#ifdef CONFIG_TCP_MD5SIG
if (static_branch_unlikely(&tcp_md5_needed) &&
rcu_access_pointer(tp->md5sig_info)) {
*md5 = tp->af_specific->md5_lookup(sk, sk);
if (*md5) {
opts->options |= OPTION_MD5;
size += TCPOLEN_MD5SIG_ALIGNED;
}
}
#endif
if (likely(tp->rx_opt.tstamp_ok)) {
opts->options |= OPTION_TS;
opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
opts->tsecr = tp->rx_opt.ts_recent;
size += TCPOLEN_TSTAMP_ALIGNED;
}
eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
if (unlikely(eff_sacks)) {
const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
opts->num_sack_blocks =
min_t(unsigned int, eff_sacks,
(remaining - TCPOLEN_SACK_BASE_ALIGNED) /
TCPOLEN_SACK_PERBLOCK);
size += TCPOLEN_SACK_BASE_ALIGNED +
opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
}
return size;
}
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
/* TCP SMALL QUEUES (TSQ)
*
* TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
* to reduce RTT and bufferbloat.
* We do this using a special skb destructor (tcp_wfree).
*
* Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
* needs to be reallocated in a driver.
* The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
*
* Since transmit from skb destructor is forbidden, we use a tasklet
* to process all sockets that eventually need to send more skbs.
* We use one tasklet per cpu, with its own queue of sockets.
*/
struct tsq_tasklet {
struct tasklet_struct tasklet;
struct list_head head; /* queue of tcp sockets */
};
static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
static void tcp_tsq_write(struct sock *sk)
{
if ((1 << sk->sk_state) &
(TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
struct tcp_sock *tp = tcp_sk(sk);
if (tp->lost_out > tp->retrans_out &&
tp->snd_cwnd > tcp_packets_in_flight(tp)) {
tcp_mstamp_refresh(tp);
tcp_xmit_retransmit_queue(sk);
}
tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
0, GFP_ATOMIC);
}
}
static void tcp_tsq_handler(struct sock *sk)
{
bh_lock_sock(sk);
if (!sock_owned_by_user(sk))
tcp_tsq_write(sk);
else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
sock_hold(sk);
bh_unlock_sock(sk);
}
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
/*
* One tasklet per cpu tries to send more skbs.
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
* We run in tasklet context but need to disable irqs when
* transferring tsq->head because tcp_wfree() might
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
* interrupt us (non NAPI drivers)
*/
static void tcp_tasklet_func(unsigned long data)
{
struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
LIST_HEAD(list);
unsigned long flags;
struct list_head *q, *n;
struct tcp_sock *tp;
struct sock *sk;
local_irq_save(flags);
list_splice_init(&tsq->head, &list);
local_irq_restore(flags);
list_for_each_safe(q, n, &list) {
tp = list_entry(q, struct tcp_sock, tsq_node);
list_del(&tp->tsq_node);
sk = (struct sock *)tp;
smp_mb__before_atomic();
clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
tcp_tsq_handler(sk);
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
sk_free(sk);
}
}
#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
TCPF_WRITE_TIMER_DEFERRED | \
TCPF_DELACK_TIMER_DEFERRED | \
TCPF_MTU_REDUCED_DEFERRED)
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
/**
* tcp_release_cb - tcp release_sock() callback
* @sk: socket
*
* called from release_sock() to perform protocol dependent
* actions before socket release.
*/
void tcp_release_cb(struct sock *sk)
{
unsigned long flags, nflags;
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
/* perform an atomic operation only if at least one flag is set */
do {
flags = sk->sk_tsq_flags;
if (!(flags & TCP_DEFERRED_ALL))
return;
nflags = flags & ~TCP_DEFERRED_ALL;
} while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
if (flags & TCPF_TSQ_DEFERRED) {
tcp_tsq_write(sk);
__sock_put(sk);
}
tcp: tcp_release_cb() should release socket ownership Lars Persson reported following deadlock : -000 |M:0x0:0x802B6AF8(asm) <-- arch_spin_lock -001 |tcp_v4_rcv(skb = 0x8BD527A0) <-- sk = 0x8BE6B2A0 -002 |ip_local_deliver_finish(skb = 0x8BD527A0) -003 |__netif_receive_skb_core(skb = 0x8BD527A0, ?) -004 |netif_receive_skb(skb = 0x8BD527A0) -005 |elk_poll(napi = 0x8C770500, budget = 64) -006 |net_rx_action(?) -007 |__do_softirq() -008 |do_softirq() -009 |local_bh_enable() -010 |tcp_rcv_established(sk = 0x8BE6B2A0, skb = 0x87D3A9E0, th = 0x814EBE14, ?) -011 |tcp_v4_do_rcv(sk = 0x8BE6B2A0, skb = 0x87D3A9E0) -012 |tcp_delack_timer_handler(sk = 0x8BE6B2A0) -013 |tcp_release_cb(sk = 0x8BE6B2A0) -014 |release_sock(sk = 0x8BE6B2A0) -015 |tcp_sendmsg(?, sk = 0x8BE6B2A0, ?, ?) -016 |sock_sendmsg(sock = 0x8518C4C0, msg = 0x87D8DAA8, size = 4096) -017 |kernel_sendmsg(?, ?, ?, ?, size = 4096) -018 |smb_send_kvec() -019 |smb_send_rqst(server = 0x87C4D400, rqst = 0x87D8DBA0) -020 |cifs_call_async() -021 |cifs_async_writev(wdata = 0x87FD6580) -022 |cifs_writepages(mapping = 0x852096E4, wbc = 0x87D8DC88) -023 |__writeback_single_inode(inode = 0x852095D0, wbc = 0x87D8DC88) -024 |writeback_sb_inodes(sb = 0x87D6D800, wb = 0x87E4A9C0, work = 0x87D8DD88) -025 |__writeback_inodes_wb(wb = 0x87E4A9C0, work = 0x87D8DD88) -026 |wb_writeback(wb = 0x87E4A9C0, work = 0x87D8DD88) -027 |wb_do_writeback(wb = 0x87E4A9C0, force_wait = 0) -028 |bdi_writeback_workfn(work = 0x87E4A9CC) -029 |process_one_work(worker = 0x8B045880, work = 0x87E4A9CC) -030 |worker_thread(__worker = 0x8B045880) -031 |kthread(_create = 0x87CADD90) -032 |ret_from_kernel_thread(asm) Bug occurs because __tcp_checksum_complete_user() enables BH, assuming it is running from softirq context. Lars trace involved a NIC without RX checksum support but other points are problematic as well, like the prequeue stuff. Problem is triggered by a timer, that found socket being owned by user. tcp_release_cb() should call tcp_write_timer_handler() or tcp_delack_timer_handler() in the appropriate context : BH disabled and socket lock held, but 'owned' field cleared, as if they were running from timer handlers. Fixes: 6f458dfb4092 ("tcp: improve latencies of timer triggered events") Reported-by: Lars Persson <lars.persson@axis.com> Tested-by: Lars Persson <lars.persson@axis.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-03-11 00:50:11 +08:00
/* Here begins the tricky part :
* We are called from release_sock() with :
* 1) BH disabled
* 2) sk_lock.slock spinlock held
* 3) socket owned by us (sk->sk_lock.owned == 1)
*
* But following code is meant to be called from BH handlers,
* so we should keep BH disabled, but early release socket ownership
*/
sock_release_ownership(sk);
if (flags & TCPF_WRITE_TIMER_DEFERRED) {
tcp_write_timer_handler(sk);
tcp: fix possible socket refcount problem Commit 6f458dfb40 (tcp: improve latencies of timer triggered events) added bug leading to following trace : [ 2866.131281] IPv4: Attempt to release TCP socket in state 1 ffff880019ec0000 [ 2866.131726] [ 2866.132188] ========================= [ 2866.132281] [ BUG: held lock freed! ] [ 2866.132281] 3.6.0-rc1+ #622 Not tainted [ 2866.132281] ------------------------- [ 2866.132281] kworker/0:1/652 is freeing memory ffff880019ec0000-ffff880019ec0a1f, with a lock still held there! [ 2866.132281] (sk_lock-AF_INET-RPC){+.+...}, at: [<ffffffff81903619>] tcp_sendmsg+0x29/0xcc6 [ 2866.132281] 4 locks held by kworker/0:1/652: [ 2866.132281] #0: (rpciod){.+.+.+}, at: [<ffffffff81083567>] process_one_work+0x1de/0x47f [ 2866.132281] #1: ((&task->u.tk_work)){+.+.+.}, at: [<ffffffff81083567>] process_one_work+0x1de/0x47f [ 2866.132281] #2: (sk_lock-AF_INET-RPC){+.+...}, at: [<ffffffff81903619>] tcp_sendmsg+0x29/0xcc6 [ 2866.132281] #3: (&icsk->icsk_retransmit_timer){+.-...}, at: [<ffffffff81078017>] run_timer_softirq+0x1ad/0x35f [ 2866.132281] [ 2866.132281] stack backtrace: [ 2866.132281] Pid: 652, comm: kworker/0:1 Not tainted 3.6.0-rc1+ #622 [ 2866.132281] Call Trace: [ 2866.132281] <IRQ> [<ffffffff810bc527>] debug_check_no_locks_freed+0x112/0x159 [ 2866.132281] [<ffffffff818a0839>] ? __sk_free+0xfd/0x114 [ 2866.132281] [<ffffffff811549fa>] kmem_cache_free+0x6b/0x13a [ 2866.132281] [<ffffffff818a0839>] __sk_free+0xfd/0x114 [ 2866.132281] [<ffffffff818a08c0>] sk_free+0x1c/0x1e [ 2866.132281] [<ffffffff81911e1c>] tcp_write_timer+0x51/0x56 [ 2866.132281] [<ffffffff81078082>] run_timer_softirq+0x218/0x35f [ 2866.132281] [<ffffffff81078017>] ? run_timer_softirq+0x1ad/0x35f [ 2866.132281] [<ffffffff810f5831>] ? rb_commit+0x58/0x85 [ 2866.132281] [<ffffffff81911dcb>] ? tcp_write_timer_handler+0x148/0x148 [ 2866.132281] [<ffffffff81070bd6>] __do_softirq+0xcb/0x1f9 [ 2866.132281] [<ffffffff81a0a00c>] ? _raw_spin_unlock+0x29/0x2e [ 2866.132281] [<ffffffff81a1227c>] call_softirq+0x1c/0x30 [ 2866.132281] [<ffffffff81039f38>] do_softirq+0x4a/0xa6 [ 2866.132281] [<ffffffff81070f2b>] irq_exit+0x51/0xad [ 2866.132281] [<ffffffff81a129cd>] do_IRQ+0x9d/0xb4 [ 2866.132281] [<ffffffff81a0a3ef>] common_interrupt+0x6f/0x6f [ 2866.132281] <EOI> [<ffffffff8109d006>] ? sched_clock_cpu+0x58/0xd1 [ 2866.132281] [<ffffffff81a0a172>] ? _raw_spin_unlock_irqrestore+0x4c/0x56 [ 2866.132281] [<ffffffff81078692>] mod_timer+0x178/0x1a9 [ 2866.132281] [<ffffffff818a00aa>] sk_reset_timer+0x19/0x26 [ 2866.132281] [<ffffffff8190b2cc>] tcp_rearm_rto+0x99/0xa4 [ 2866.132281] [<ffffffff8190dfba>] tcp_event_new_data_sent+0x6e/0x70 [ 2866.132281] [<ffffffff8190f7ea>] tcp_write_xmit+0x7de/0x8e4 [ 2866.132281] [<ffffffff818a565d>] ? __alloc_skb+0xa0/0x1a1 [ 2866.132281] [<ffffffff8190f952>] __tcp_push_pending_frames+0x2e/0x8a [ 2866.132281] [<ffffffff81904122>] tcp_sendmsg+0xb32/0xcc6 [ 2866.132281] [<ffffffff819229c2>] inet_sendmsg+0xaa/0xd5 [ 2866.132281] [<ffffffff81922918>] ? inet_autobind+0x5f/0x5f [ 2866.132281] [<ffffffff810ee7f1>] ? trace_clock_local+0x9/0xb [ 2866.132281] [<ffffffff8189adab>] sock_sendmsg+0xa3/0xc4 [ 2866.132281] [<ffffffff810f5de6>] ? rb_reserve_next_event+0x26f/0x2d5 [ 2866.132281] [<ffffffff8103e6a9>] ? native_sched_clock+0x29/0x6f [ 2866.132281] [<ffffffff8103e6f8>] ? sched_clock+0x9/0xd [ 2866.132281] [<ffffffff810ee7f1>] ? trace_clock_local+0x9/0xb [ 2866.132281] [<ffffffff8189ae03>] kernel_sendmsg+0x37/0x43 [ 2866.132281] [<ffffffff8199ce49>] xs_send_kvec+0x77/0x80 [ 2866.132281] [<ffffffff8199cec1>] xs_sendpages+0x6f/0x1a0 [ 2866.132281] [<ffffffff8107826d>] ? try_to_del_timer_sync+0x55/0x61 [ 2866.132281] [<ffffffff8199d0d2>] xs_tcp_send_request+0x55/0xf1 [ 2866.132281] [<ffffffff8199bb90>] xprt_transmit+0x89/0x1db [ 2866.132281] [<ffffffff81999bcd>] ? call_connect+0x3c/0x3c [ 2866.132281] [<ffffffff81999d92>] call_transmit+0x1c5/0x20e [ 2866.132281] [<ffffffff819a0d55>] __rpc_execute+0x6f/0x225 [ 2866.132281] [<ffffffff81999bcd>] ? call_connect+0x3c/0x3c [ 2866.132281] [<ffffffff819a0f33>] rpc_async_schedule+0x28/0x34 [ 2866.132281] [<ffffffff810835d6>] process_one_work+0x24d/0x47f [ 2866.132281] [<ffffffff81083567>] ? process_one_work+0x1de/0x47f [ 2866.132281] [<ffffffff819a0f0b>] ? __rpc_execute+0x225/0x225 [ 2866.132281] [<ffffffff81083a6d>] worker_thread+0x236/0x317 [ 2866.132281] [<ffffffff81083837>] ? process_scheduled_works+0x2f/0x2f [ 2866.132281] [<ffffffff8108b7b8>] kthread+0x9a/0xa2 [ 2866.132281] [<ffffffff81a12184>] kernel_thread_helper+0x4/0x10 [ 2866.132281] [<ffffffff81a0a4b0>] ? retint_restore_args+0x13/0x13 [ 2866.132281] [<ffffffff8108b71e>] ? __init_kthread_worker+0x5a/0x5a [ 2866.132281] [<ffffffff81a12180>] ? gs_change+0x13/0x13 [ 2866.308506] IPv4: Attempt to release TCP socket in state 1 ffff880019ec0000 [ 2866.309689] ============================================================================= [ 2866.310254] BUG TCP (Not tainted): Object already free [ 2866.310254] ----------------------------------------------------------------------------- [ 2866.310254] The bug comes from the fact that timer set in sk_reset_timer() can run before we actually do the sock_hold(). socket refcount reaches zero and we free the socket too soon. timer handler is not allowed to reduce socket refcnt if socket is owned by the user, or we need to change sk_reset_timer() implementation. We should take a reference on the socket in case TCP_DELACK_TIMER_DEFERRED or TCP_DELACK_TIMER_DEFERRED bit are set in tsq_flags Also fix a typo in tcp_delack_timer(), where TCP_WRITE_TIMER_DEFERRED was used instead of TCP_DELACK_TIMER_DEFERRED. For consistency, use same socket refcount change for TCP_MTU_REDUCED_DEFERRED, even if not fired from a timer. Reported-by: Fengguang Wu <fengguang.wu@intel.com> Tested-by: Fengguang Wu <fengguang.wu@intel.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-08-20 08:22:46 +08:00
__sock_put(sk);
}
if (flags & TCPF_DELACK_TIMER_DEFERRED) {
tcp_delack_timer_handler(sk);
tcp: fix possible socket refcount problem Commit 6f458dfb40 (tcp: improve latencies of timer triggered events) added bug leading to following trace : [ 2866.131281] IPv4: Attempt to release TCP socket in state 1 ffff880019ec0000 [ 2866.131726] [ 2866.132188] ========================= [ 2866.132281] [ BUG: held lock freed! ] [ 2866.132281] 3.6.0-rc1+ #622 Not tainted [ 2866.132281] ------------------------- [ 2866.132281] kworker/0:1/652 is freeing memory ffff880019ec0000-ffff880019ec0a1f, with a lock still held there! [ 2866.132281] (sk_lock-AF_INET-RPC){+.+...}, at: [<ffffffff81903619>] tcp_sendmsg+0x29/0xcc6 [ 2866.132281] 4 locks held by kworker/0:1/652: [ 2866.132281] #0: (rpciod){.+.+.+}, at: [<ffffffff81083567>] process_one_work+0x1de/0x47f [ 2866.132281] #1: ((&task->u.tk_work)){+.+.+.}, at: [<ffffffff81083567>] process_one_work+0x1de/0x47f [ 2866.132281] #2: (sk_lock-AF_INET-RPC){+.+...}, at: [<ffffffff81903619>] tcp_sendmsg+0x29/0xcc6 [ 2866.132281] #3: (&icsk->icsk_retransmit_timer){+.-...}, at: [<ffffffff81078017>] run_timer_softirq+0x1ad/0x35f [ 2866.132281] [ 2866.132281] stack backtrace: [ 2866.132281] Pid: 652, comm: kworker/0:1 Not tainted 3.6.0-rc1+ #622 [ 2866.132281] Call Trace: [ 2866.132281] <IRQ> [<ffffffff810bc527>] debug_check_no_locks_freed+0x112/0x159 [ 2866.132281] [<ffffffff818a0839>] ? __sk_free+0xfd/0x114 [ 2866.132281] [<ffffffff811549fa>] kmem_cache_free+0x6b/0x13a [ 2866.132281] [<ffffffff818a0839>] __sk_free+0xfd/0x114 [ 2866.132281] [<ffffffff818a08c0>] sk_free+0x1c/0x1e [ 2866.132281] [<ffffffff81911e1c>] tcp_write_timer+0x51/0x56 [ 2866.132281] [<ffffffff81078082>] run_timer_softirq+0x218/0x35f [ 2866.132281] [<ffffffff81078017>] ? run_timer_softirq+0x1ad/0x35f [ 2866.132281] [<ffffffff810f5831>] ? rb_commit+0x58/0x85 [ 2866.132281] [<ffffffff81911dcb>] ? tcp_write_timer_handler+0x148/0x148 [ 2866.132281] [<ffffffff81070bd6>] __do_softirq+0xcb/0x1f9 [ 2866.132281] [<ffffffff81a0a00c>] ? _raw_spin_unlock+0x29/0x2e [ 2866.132281] [<ffffffff81a1227c>] call_softirq+0x1c/0x30 [ 2866.132281] [<ffffffff81039f38>] do_softirq+0x4a/0xa6 [ 2866.132281] [<ffffffff81070f2b>] irq_exit+0x51/0xad [ 2866.132281] [<ffffffff81a129cd>] do_IRQ+0x9d/0xb4 [ 2866.132281] [<ffffffff81a0a3ef>] common_interrupt+0x6f/0x6f [ 2866.132281] <EOI> [<ffffffff8109d006>] ? sched_clock_cpu+0x58/0xd1 [ 2866.132281] [<ffffffff81a0a172>] ? _raw_spin_unlock_irqrestore+0x4c/0x56 [ 2866.132281] [<ffffffff81078692>] mod_timer+0x178/0x1a9 [ 2866.132281] [<ffffffff818a00aa>] sk_reset_timer+0x19/0x26 [ 2866.132281] [<ffffffff8190b2cc>] tcp_rearm_rto+0x99/0xa4 [ 2866.132281] [<ffffffff8190dfba>] tcp_event_new_data_sent+0x6e/0x70 [ 2866.132281] [<ffffffff8190f7ea>] tcp_write_xmit+0x7de/0x8e4 [ 2866.132281] [<ffffffff818a565d>] ? __alloc_skb+0xa0/0x1a1 [ 2866.132281] [<ffffffff8190f952>] __tcp_push_pending_frames+0x2e/0x8a [ 2866.132281] [<ffffffff81904122>] tcp_sendmsg+0xb32/0xcc6 [ 2866.132281] [<ffffffff819229c2>] inet_sendmsg+0xaa/0xd5 [ 2866.132281] [<ffffffff81922918>] ? inet_autobind+0x5f/0x5f [ 2866.132281] [<ffffffff810ee7f1>] ? trace_clock_local+0x9/0xb [ 2866.132281] [<ffffffff8189adab>] sock_sendmsg+0xa3/0xc4 [ 2866.132281] [<ffffffff810f5de6>] ? rb_reserve_next_event+0x26f/0x2d5 [ 2866.132281] [<ffffffff8103e6a9>] ? native_sched_clock+0x29/0x6f [ 2866.132281] [<ffffffff8103e6f8>] ? sched_clock+0x9/0xd [ 2866.132281] [<ffffffff810ee7f1>] ? trace_clock_local+0x9/0xb [ 2866.132281] [<ffffffff8189ae03>] kernel_sendmsg+0x37/0x43 [ 2866.132281] [<ffffffff8199ce49>] xs_send_kvec+0x77/0x80 [ 2866.132281] [<ffffffff8199cec1>] xs_sendpages+0x6f/0x1a0 [ 2866.132281] [<ffffffff8107826d>] ? try_to_del_timer_sync+0x55/0x61 [ 2866.132281] [<ffffffff8199d0d2>] xs_tcp_send_request+0x55/0xf1 [ 2866.132281] [<ffffffff8199bb90>] xprt_transmit+0x89/0x1db [ 2866.132281] [<ffffffff81999bcd>] ? call_connect+0x3c/0x3c [ 2866.132281] [<ffffffff81999d92>] call_transmit+0x1c5/0x20e [ 2866.132281] [<ffffffff819a0d55>] __rpc_execute+0x6f/0x225 [ 2866.132281] [<ffffffff81999bcd>] ? call_connect+0x3c/0x3c [ 2866.132281] [<ffffffff819a0f33>] rpc_async_schedule+0x28/0x34 [ 2866.132281] [<ffffffff810835d6>] process_one_work+0x24d/0x47f [ 2866.132281] [<ffffffff81083567>] ? process_one_work+0x1de/0x47f [ 2866.132281] [<ffffffff819a0f0b>] ? __rpc_execute+0x225/0x225 [ 2866.132281] [<ffffffff81083a6d>] worker_thread+0x236/0x317 [ 2866.132281] [<ffffffff81083837>] ? process_scheduled_works+0x2f/0x2f [ 2866.132281] [<ffffffff8108b7b8>] kthread+0x9a/0xa2 [ 2866.132281] [<ffffffff81a12184>] kernel_thread_helper+0x4/0x10 [ 2866.132281] [<ffffffff81a0a4b0>] ? retint_restore_args+0x13/0x13 [ 2866.132281] [<ffffffff8108b71e>] ? __init_kthread_worker+0x5a/0x5a [ 2866.132281] [<ffffffff81a12180>] ? gs_change+0x13/0x13 [ 2866.308506] IPv4: Attempt to release TCP socket in state 1 ffff880019ec0000 [ 2866.309689] ============================================================================= [ 2866.310254] BUG TCP (Not tainted): Object already free [ 2866.310254] ----------------------------------------------------------------------------- [ 2866.310254] The bug comes from the fact that timer set in sk_reset_timer() can run before we actually do the sock_hold(). socket refcount reaches zero and we free the socket too soon. timer handler is not allowed to reduce socket refcnt if socket is owned by the user, or we need to change sk_reset_timer() implementation. We should take a reference on the socket in case TCP_DELACK_TIMER_DEFERRED or TCP_DELACK_TIMER_DEFERRED bit are set in tsq_flags Also fix a typo in tcp_delack_timer(), where TCP_WRITE_TIMER_DEFERRED was used instead of TCP_DELACK_TIMER_DEFERRED. For consistency, use same socket refcount change for TCP_MTU_REDUCED_DEFERRED, even if not fired from a timer. Reported-by: Fengguang Wu <fengguang.wu@intel.com> Tested-by: Fengguang Wu <fengguang.wu@intel.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-08-20 08:22:46 +08:00
__sock_put(sk);
}
if (flags & TCPF_MTU_REDUCED_DEFERRED) {
inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
tcp: fix possible socket refcount problem Commit 6f458dfb40 (tcp: improve latencies of timer triggered events) added bug leading to following trace : [ 2866.131281] IPv4: Attempt to release TCP socket in state 1 ffff880019ec0000 [ 2866.131726] [ 2866.132188] ========================= [ 2866.132281] [ BUG: held lock freed! ] [ 2866.132281] 3.6.0-rc1+ #622 Not tainted [ 2866.132281] ------------------------- [ 2866.132281] kworker/0:1/652 is freeing memory ffff880019ec0000-ffff880019ec0a1f, with a lock still held there! [ 2866.132281] (sk_lock-AF_INET-RPC){+.+...}, at: [<ffffffff81903619>] tcp_sendmsg+0x29/0xcc6 [ 2866.132281] 4 locks held by kworker/0:1/652: [ 2866.132281] #0: (rpciod){.+.+.+}, at: [<ffffffff81083567>] process_one_work+0x1de/0x47f [ 2866.132281] #1: ((&task->u.tk_work)){+.+.+.}, at: [<ffffffff81083567>] process_one_work+0x1de/0x47f [ 2866.132281] #2: (sk_lock-AF_INET-RPC){+.+...}, at: [<ffffffff81903619>] tcp_sendmsg+0x29/0xcc6 [ 2866.132281] #3: (&icsk->icsk_retransmit_timer){+.-...}, at: [<ffffffff81078017>] run_timer_softirq+0x1ad/0x35f [ 2866.132281] [ 2866.132281] stack backtrace: [ 2866.132281] Pid: 652, comm: kworker/0:1 Not tainted 3.6.0-rc1+ #622 [ 2866.132281] Call Trace: [ 2866.132281] <IRQ> [<ffffffff810bc527>] debug_check_no_locks_freed+0x112/0x159 [ 2866.132281] [<ffffffff818a0839>] ? __sk_free+0xfd/0x114 [ 2866.132281] [<ffffffff811549fa>] kmem_cache_free+0x6b/0x13a [ 2866.132281] [<ffffffff818a0839>] __sk_free+0xfd/0x114 [ 2866.132281] [<ffffffff818a08c0>] sk_free+0x1c/0x1e [ 2866.132281] [<ffffffff81911e1c>] tcp_write_timer+0x51/0x56 [ 2866.132281] [<ffffffff81078082>] run_timer_softirq+0x218/0x35f [ 2866.132281] [<ffffffff81078017>] ? run_timer_softirq+0x1ad/0x35f [ 2866.132281] [<ffffffff810f5831>] ? rb_commit+0x58/0x85 [ 2866.132281] [<ffffffff81911dcb>] ? tcp_write_timer_handler+0x148/0x148 [ 2866.132281] [<ffffffff81070bd6>] __do_softirq+0xcb/0x1f9 [ 2866.132281] [<ffffffff81a0a00c>] ? _raw_spin_unlock+0x29/0x2e [ 2866.132281] [<ffffffff81a1227c>] call_softirq+0x1c/0x30 [ 2866.132281] [<ffffffff81039f38>] do_softirq+0x4a/0xa6 [ 2866.132281] [<ffffffff81070f2b>] irq_exit+0x51/0xad [ 2866.132281] [<ffffffff81a129cd>] do_IRQ+0x9d/0xb4 [ 2866.132281] [<ffffffff81a0a3ef>] common_interrupt+0x6f/0x6f [ 2866.132281] <EOI> [<ffffffff8109d006>] ? sched_clock_cpu+0x58/0xd1 [ 2866.132281] [<ffffffff81a0a172>] ? _raw_spin_unlock_irqrestore+0x4c/0x56 [ 2866.132281] [<ffffffff81078692>] mod_timer+0x178/0x1a9 [ 2866.132281] [<ffffffff818a00aa>] sk_reset_timer+0x19/0x26 [ 2866.132281] [<ffffffff8190b2cc>] tcp_rearm_rto+0x99/0xa4 [ 2866.132281] [<ffffffff8190dfba>] tcp_event_new_data_sent+0x6e/0x70 [ 2866.132281] [<ffffffff8190f7ea>] tcp_write_xmit+0x7de/0x8e4 [ 2866.132281] [<ffffffff818a565d>] ? __alloc_skb+0xa0/0x1a1 [ 2866.132281] [<ffffffff8190f952>] __tcp_push_pending_frames+0x2e/0x8a [ 2866.132281] [<ffffffff81904122>] tcp_sendmsg+0xb32/0xcc6 [ 2866.132281] [<ffffffff819229c2>] inet_sendmsg+0xaa/0xd5 [ 2866.132281] [<ffffffff81922918>] ? inet_autobind+0x5f/0x5f [ 2866.132281] [<ffffffff810ee7f1>] ? trace_clock_local+0x9/0xb [ 2866.132281] [<ffffffff8189adab>] sock_sendmsg+0xa3/0xc4 [ 2866.132281] [<ffffffff810f5de6>] ? rb_reserve_next_event+0x26f/0x2d5 [ 2866.132281] [<ffffffff8103e6a9>] ? native_sched_clock+0x29/0x6f [ 2866.132281] [<ffffffff8103e6f8>] ? sched_clock+0x9/0xd [ 2866.132281] [<ffffffff810ee7f1>] ? trace_clock_local+0x9/0xb [ 2866.132281] [<ffffffff8189ae03>] kernel_sendmsg+0x37/0x43 [ 2866.132281] [<ffffffff8199ce49>] xs_send_kvec+0x77/0x80 [ 2866.132281] [<ffffffff8199cec1>] xs_sendpages+0x6f/0x1a0 [ 2866.132281] [<ffffffff8107826d>] ? try_to_del_timer_sync+0x55/0x61 [ 2866.132281] [<ffffffff8199d0d2>] xs_tcp_send_request+0x55/0xf1 [ 2866.132281] [<ffffffff8199bb90>] xprt_transmit+0x89/0x1db [ 2866.132281] [<ffffffff81999bcd>] ? call_connect+0x3c/0x3c [ 2866.132281] [<ffffffff81999d92>] call_transmit+0x1c5/0x20e [ 2866.132281] [<ffffffff819a0d55>] __rpc_execute+0x6f/0x225 [ 2866.132281] [<ffffffff81999bcd>] ? call_connect+0x3c/0x3c [ 2866.132281] [<ffffffff819a0f33>] rpc_async_schedule+0x28/0x34 [ 2866.132281] [<ffffffff810835d6>] process_one_work+0x24d/0x47f [ 2866.132281] [<ffffffff81083567>] ? process_one_work+0x1de/0x47f [ 2866.132281] [<ffffffff819a0f0b>] ? __rpc_execute+0x225/0x225 [ 2866.132281] [<ffffffff81083a6d>] worker_thread+0x236/0x317 [ 2866.132281] [<ffffffff81083837>] ? process_scheduled_works+0x2f/0x2f [ 2866.132281] [<ffffffff8108b7b8>] kthread+0x9a/0xa2 [ 2866.132281] [<ffffffff81a12184>] kernel_thread_helper+0x4/0x10 [ 2866.132281] [<ffffffff81a0a4b0>] ? retint_restore_args+0x13/0x13 [ 2866.132281] [<ffffffff8108b71e>] ? __init_kthread_worker+0x5a/0x5a [ 2866.132281] [<ffffffff81a12180>] ? gs_change+0x13/0x13 [ 2866.308506] IPv4: Attempt to release TCP socket in state 1 ffff880019ec0000 [ 2866.309689] ============================================================================= [ 2866.310254] BUG TCP (Not tainted): Object already free [ 2866.310254] ----------------------------------------------------------------------------- [ 2866.310254] The bug comes from the fact that timer set in sk_reset_timer() can run before we actually do the sock_hold(). socket refcount reaches zero and we free the socket too soon. timer handler is not allowed to reduce socket refcnt if socket is owned by the user, or we need to change sk_reset_timer() implementation. We should take a reference on the socket in case TCP_DELACK_TIMER_DEFERRED or TCP_DELACK_TIMER_DEFERRED bit are set in tsq_flags Also fix a typo in tcp_delack_timer(), where TCP_WRITE_TIMER_DEFERRED was used instead of TCP_DELACK_TIMER_DEFERRED. For consistency, use same socket refcount change for TCP_MTU_REDUCED_DEFERRED, even if not fired from a timer. Reported-by: Fengguang Wu <fengguang.wu@intel.com> Tested-by: Fengguang Wu <fengguang.wu@intel.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-08-20 08:22:46 +08:00
__sock_put(sk);
}
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
}
EXPORT_SYMBOL(tcp_release_cb);
void __init tcp_tasklet_init(void)
{
int i;
for_each_possible_cpu(i) {
struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
INIT_LIST_HEAD(&tsq->head);
tasklet_init(&tsq->tasklet,
tcp_tasklet_func,
(unsigned long)tsq);
}
}
/*
* Write buffer destructor automatically called from kfree_skb.
* We can't xmit new skbs from this context, as we might already
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
* hold qdisc lock.
*/
void tcp_wfree(struct sk_buff *skb)
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
{
struct sock *sk = skb->sk;
struct tcp_sock *tp = tcp_sk(sk);
unsigned long flags, nval, oval;
tcp: TCP Small Queues and strange attractors TCP Small queues tries to keep number of packets in qdisc as small as possible, and depends on a tasklet to feed following packets at TX completion time. Choice of tasklet was driven by latencies requirements. Then, TCP stack tries to avoid reorders, by locking flows with outstanding packets in qdisc in a given TX queue. What can happen is that many flows get attracted by a low performing TX queue, and cpu servicing TX completion has to feed packets for all of them, making this cpu 100% busy in softirq mode. This became particularly visible with latest skb->xmit_more support Strategy adopted in this patch is to detect when tcp_wfree() is called from ksoftirqd and let the outstanding queue for this flow being drained before feeding additional packets, so that skb->ooo_okay can be set to allow select_queue() to select the optimal queue : Incoming ACKS are normally handled by different cpus, so this patch gives more chance for these cpus to take over the burden of feeding qdisc with future packets. Tested: lpaa23:~# ./super_netperf 1400 --google-pacing-rate 3028000 -H lpaa24 -l 3600 & lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:18 AM eth1 595448.00 1190564.00 38381.09 1760253.12 0.00 0.00 1.00 06:16:19 AM eth1 594858.00 1189686.00 38340.76 1758952.72 0.00 0.00 0.00 06:16:20 AM eth1 597017.00 1194019.00 38480.79 1765370.29 0.00 0.00 1.00 06:16:21 AM eth1 595450.00 1190936.00 38380.19 1760805.05 0.00 0.00 0.00 06:16:22 AM eth1 596385.00 1193096.00 38442.56 1763976.29 0.00 0.00 1.00 06:16:23 AM eth1 598155.00 1195978.00 38552.97 1768264.60 0.00 0.00 0.00 06:16:24 AM eth1 594405.00 1188643.00 38312.57 1757414.89 0.00 0.00 1.00 06:16:25 AM eth1 593366.00 1187154.00 38252.16 1755195.83 0.00 0.00 0.00 06:16:26 AM eth1 593188.00 1186118.00 38232.88 1753682.57 0.00 0.00 1.00 06:16:27 AM eth1 596301.00 1192241.00 38440.94 1762733.09 0.00 0.00 0.00 Average: eth1 595457.30 1190843.50 38381.69 1760664.84 0.00 0.00 0.50 lpaa23:~# ./tc -s -d qd sh dev eth1 | grep backlog backlog 7606336b 2513p requeues 167982 backlog 224072b 74p requeues 566 backlog 581376b 192p requeues 5598 backlog 181680b 60p requeues 1070 backlog 5305056b 1753p requeues 110166 // Here, this TX queue is attracting flows backlog 157456b 52p requeues 1758 backlog 672216b 222p requeues 3025 backlog 60560b 20p requeues 24541 backlog 448144b 148p requeues 21258 lpaa23:~# echo 1 >/proc/sys/net/ipv4/tcp_tsq_enable_tcp_wfree_ksoftirqd_detect Immediate jump to full bandwidth, and traffic is properly shard on all tx queues. lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:46 AM eth1 1397632.00 2795397.00 90081.87 4133031.26 0.00 0.00 1.00 06:16:47 AM eth1 1396874.00 2793614.00 90032.99 4130385.46 0.00 0.00 0.00 06:16:48 AM eth1 1395842.00 2791600.00 89966.46 4127409.67 0.00 0.00 1.00 06:16:49 AM eth1 1395528.00 2791017.00 89946.17 4126551.24 0.00 0.00 0.00 06:16:50 AM eth1 1397891.00 2795716.00 90098.74 4133497.39 0.00 0.00 1.00 06:16:51 AM eth1 1394951.00 2789984.00 89908.96 4125022.51 0.00 0.00 0.00 06:16:52 AM eth1 1394608.00 2789190.00 89886.90 4123851.36 0.00 0.00 1.00 06:16:53 AM eth1 1395314.00 2790653.00 89934.33 4125983.09 0.00 0.00 0.00 06:16:54 AM eth1 1396115.00 2792276.00 89984.25 4128411.21 0.00 0.00 1.00 06:16:55 AM eth1 1396829.00 2793523.00 90030.19 4130250.28 0.00 0.00 0.00 Average: eth1 1396158.40 2792297.00 89987.09 4128439.35 0.00 0.00 0.50 lpaa23:~# tc -s -d qd sh dev eth1 | grep backlog backlog 7900052b 2609p requeues 173287 backlog 878120b 290p requeues 589 backlog 1068884b 354p requeues 5621 backlog 996212b 329p requeues 1088 backlog 984100b 325p requeues 115316 backlog 956848b 316p requeues 1781 backlog 1080996b 357p requeues 3047 backlog 975016b 322p requeues 24571 backlog 990156b 327p requeues 21274 (All 8 TX queues get a fair share of the traffic) Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-13 21:27:47 +08:00
/* Keep one reference on sk_wmem_alloc.
* Will be released by sk_free() from here or tcp_tasklet_func()
*/
WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
tcp: TCP Small Queues and strange attractors TCP Small queues tries to keep number of packets in qdisc as small as possible, and depends on a tasklet to feed following packets at TX completion time. Choice of tasklet was driven by latencies requirements. Then, TCP stack tries to avoid reorders, by locking flows with outstanding packets in qdisc in a given TX queue. What can happen is that many flows get attracted by a low performing TX queue, and cpu servicing TX completion has to feed packets for all of them, making this cpu 100% busy in softirq mode. This became particularly visible with latest skb->xmit_more support Strategy adopted in this patch is to detect when tcp_wfree() is called from ksoftirqd and let the outstanding queue for this flow being drained before feeding additional packets, so that skb->ooo_okay can be set to allow select_queue() to select the optimal queue : Incoming ACKS are normally handled by different cpus, so this patch gives more chance for these cpus to take over the burden of feeding qdisc with future packets. Tested: lpaa23:~# ./super_netperf 1400 --google-pacing-rate 3028000 -H lpaa24 -l 3600 & lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:18 AM eth1 595448.00 1190564.00 38381.09 1760253.12 0.00 0.00 1.00 06:16:19 AM eth1 594858.00 1189686.00 38340.76 1758952.72 0.00 0.00 0.00 06:16:20 AM eth1 597017.00 1194019.00 38480.79 1765370.29 0.00 0.00 1.00 06:16:21 AM eth1 595450.00 1190936.00 38380.19 1760805.05 0.00 0.00 0.00 06:16:22 AM eth1 596385.00 1193096.00 38442.56 1763976.29 0.00 0.00 1.00 06:16:23 AM eth1 598155.00 1195978.00 38552.97 1768264.60 0.00 0.00 0.00 06:16:24 AM eth1 594405.00 1188643.00 38312.57 1757414.89 0.00 0.00 1.00 06:16:25 AM eth1 593366.00 1187154.00 38252.16 1755195.83 0.00 0.00 0.00 06:16:26 AM eth1 593188.00 1186118.00 38232.88 1753682.57 0.00 0.00 1.00 06:16:27 AM eth1 596301.00 1192241.00 38440.94 1762733.09 0.00 0.00 0.00 Average: eth1 595457.30 1190843.50 38381.69 1760664.84 0.00 0.00 0.50 lpaa23:~# ./tc -s -d qd sh dev eth1 | grep backlog backlog 7606336b 2513p requeues 167982 backlog 224072b 74p requeues 566 backlog 581376b 192p requeues 5598 backlog 181680b 60p requeues 1070 backlog 5305056b 1753p requeues 110166 // Here, this TX queue is attracting flows backlog 157456b 52p requeues 1758 backlog 672216b 222p requeues 3025 backlog 60560b 20p requeues 24541 backlog 448144b 148p requeues 21258 lpaa23:~# echo 1 >/proc/sys/net/ipv4/tcp_tsq_enable_tcp_wfree_ksoftirqd_detect Immediate jump to full bandwidth, and traffic is properly shard on all tx queues. lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:46 AM eth1 1397632.00 2795397.00 90081.87 4133031.26 0.00 0.00 1.00 06:16:47 AM eth1 1396874.00 2793614.00 90032.99 4130385.46 0.00 0.00 0.00 06:16:48 AM eth1 1395842.00 2791600.00 89966.46 4127409.67 0.00 0.00 1.00 06:16:49 AM eth1 1395528.00 2791017.00 89946.17 4126551.24 0.00 0.00 0.00 06:16:50 AM eth1 1397891.00 2795716.00 90098.74 4133497.39 0.00 0.00 1.00 06:16:51 AM eth1 1394951.00 2789984.00 89908.96 4125022.51 0.00 0.00 0.00 06:16:52 AM eth1 1394608.00 2789190.00 89886.90 4123851.36 0.00 0.00 1.00 06:16:53 AM eth1 1395314.00 2790653.00 89934.33 4125983.09 0.00 0.00 0.00 06:16:54 AM eth1 1396115.00 2792276.00 89984.25 4128411.21 0.00 0.00 1.00 06:16:55 AM eth1 1396829.00 2793523.00 90030.19 4130250.28 0.00 0.00 0.00 Average: eth1 1396158.40 2792297.00 89987.09 4128439.35 0.00 0.00 0.50 lpaa23:~# tc -s -d qd sh dev eth1 | grep backlog backlog 7900052b 2609p requeues 173287 backlog 878120b 290p requeues 589 backlog 1068884b 354p requeues 5621 backlog 996212b 329p requeues 1088 backlog 984100b 325p requeues 115316 backlog 956848b 316p requeues 1781 backlog 1080996b 357p requeues 3047 backlog 975016b 322p requeues 24571 backlog 990156b 327p requeues 21274 (All 8 TX queues get a fair share of the traffic) Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-13 21:27:47 +08:00
/* If this softirq is serviced by ksoftirqd, we are likely under stress.
* Wait until our queues (qdisc + devices) are drained.
* This gives :
* - less callbacks to tcp_write_xmit(), reducing stress (batches)
* - chance for incoming ACK (processed by another cpu maybe)
* to migrate this flow (skb->ooo_okay will be eventually set)
*/
if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
tcp: TCP Small Queues and strange attractors TCP Small queues tries to keep number of packets in qdisc as small as possible, and depends on a tasklet to feed following packets at TX completion time. Choice of tasklet was driven by latencies requirements. Then, TCP stack tries to avoid reorders, by locking flows with outstanding packets in qdisc in a given TX queue. What can happen is that many flows get attracted by a low performing TX queue, and cpu servicing TX completion has to feed packets for all of them, making this cpu 100% busy in softirq mode. This became particularly visible with latest skb->xmit_more support Strategy adopted in this patch is to detect when tcp_wfree() is called from ksoftirqd and let the outstanding queue for this flow being drained before feeding additional packets, so that skb->ooo_okay can be set to allow select_queue() to select the optimal queue : Incoming ACKS are normally handled by different cpus, so this patch gives more chance for these cpus to take over the burden of feeding qdisc with future packets. Tested: lpaa23:~# ./super_netperf 1400 --google-pacing-rate 3028000 -H lpaa24 -l 3600 & lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:18 AM eth1 595448.00 1190564.00 38381.09 1760253.12 0.00 0.00 1.00 06:16:19 AM eth1 594858.00 1189686.00 38340.76 1758952.72 0.00 0.00 0.00 06:16:20 AM eth1 597017.00 1194019.00 38480.79 1765370.29 0.00 0.00 1.00 06:16:21 AM eth1 595450.00 1190936.00 38380.19 1760805.05 0.00 0.00 0.00 06:16:22 AM eth1 596385.00 1193096.00 38442.56 1763976.29 0.00 0.00 1.00 06:16:23 AM eth1 598155.00 1195978.00 38552.97 1768264.60 0.00 0.00 0.00 06:16:24 AM eth1 594405.00 1188643.00 38312.57 1757414.89 0.00 0.00 1.00 06:16:25 AM eth1 593366.00 1187154.00 38252.16 1755195.83 0.00 0.00 0.00 06:16:26 AM eth1 593188.00 1186118.00 38232.88 1753682.57 0.00 0.00 1.00 06:16:27 AM eth1 596301.00 1192241.00 38440.94 1762733.09 0.00 0.00 0.00 Average: eth1 595457.30 1190843.50 38381.69 1760664.84 0.00 0.00 0.50 lpaa23:~# ./tc -s -d qd sh dev eth1 | grep backlog backlog 7606336b 2513p requeues 167982 backlog 224072b 74p requeues 566 backlog 581376b 192p requeues 5598 backlog 181680b 60p requeues 1070 backlog 5305056b 1753p requeues 110166 // Here, this TX queue is attracting flows backlog 157456b 52p requeues 1758 backlog 672216b 222p requeues 3025 backlog 60560b 20p requeues 24541 backlog 448144b 148p requeues 21258 lpaa23:~# echo 1 >/proc/sys/net/ipv4/tcp_tsq_enable_tcp_wfree_ksoftirqd_detect Immediate jump to full bandwidth, and traffic is properly shard on all tx queues. lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:46 AM eth1 1397632.00 2795397.00 90081.87 4133031.26 0.00 0.00 1.00 06:16:47 AM eth1 1396874.00 2793614.00 90032.99 4130385.46 0.00 0.00 0.00 06:16:48 AM eth1 1395842.00 2791600.00 89966.46 4127409.67 0.00 0.00 1.00 06:16:49 AM eth1 1395528.00 2791017.00 89946.17 4126551.24 0.00 0.00 0.00 06:16:50 AM eth1 1397891.00 2795716.00 90098.74 4133497.39 0.00 0.00 1.00 06:16:51 AM eth1 1394951.00 2789984.00 89908.96 4125022.51 0.00 0.00 0.00 06:16:52 AM eth1 1394608.00 2789190.00 89886.90 4123851.36 0.00 0.00 1.00 06:16:53 AM eth1 1395314.00 2790653.00 89934.33 4125983.09 0.00 0.00 0.00 06:16:54 AM eth1 1396115.00 2792276.00 89984.25 4128411.21 0.00 0.00 1.00 06:16:55 AM eth1 1396829.00 2793523.00 90030.19 4130250.28 0.00 0.00 0.00 Average: eth1 1396158.40 2792297.00 89987.09 4128439.35 0.00 0.00 0.50 lpaa23:~# tc -s -d qd sh dev eth1 | grep backlog backlog 7900052b 2609p requeues 173287 backlog 878120b 290p requeues 589 backlog 1068884b 354p requeues 5621 backlog 996212b 329p requeues 1088 backlog 984100b 325p requeues 115316 backlog 956848b 316p requeues 1781 backlog 1080996b 357p requeues 3047 backlog 975016b 322p requeues 24571 backlog 990156b 327p requeues 21274 (All 8 TX queues get a fair share of the traffic) Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-13 21:27:47 +08:00
goto out;
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
struct tsq_tasklet *tsq;
bool empty;
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
goto out;
nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
if (nval != oval)
continue;
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
/* queue this socket to tasklet queue */
local_irq_save(flags);
tsq = this_cpu_ptr(&tsq_tasklet);
empty = list_empty(&tsq->head);
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
list_add(&tp->tsq_node, &tsq->head);
if (empty)
tasklet_schedule(&tsq->tasklet);
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
local_irq_restore(flags);
tcp: TCP Small Queues and strange attractors TCP Small queues tries to keep number of packets in qdisc as small as possible, and depends on a tasklet to feed following packets at TX completion time. Choice of tasklet was driven by latencies requirements. Then, TCP stack tries to avoid reorders, by locking flows with outstanding packets in qdisc in a given TX queue. What can happen is that many flows get attracted by a low performing TX queue, and cpu servicing TX completion has to feed packets for all of them, making this cpu 100% busy in softirq mode. This became particularly visible with latest skb->xmit_more support Strategy adopted in this patch is to detect when tcp_wfree() is called from ksoftirqd and let the outstanding queue for this flow being drained before feeding additional packets, so that skb->ooo_okay can be set to allow select_queue() to select the optimal queue : Incoming ACKS are normally handled by different cpus, so this patch gives more chance for these cpus to take over the burden of feeding qdisc with future packets. Tested: lpaa23:~# ./super_netperf 1400 --google-pacing-rate 3028000 -H lpaa24 -l 3600 & lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:18 AM eth1 595448.00 1190564.00 38381.09 1760253.12 0.00 0.00 1.00 06:16:19 AM eth1 594858.00 1189686.00 38340.76 1758952.72 0.00 0.00 0.00 06:16:20 AM eth1 597017.00 1194019.00 38480.79 1765370.29 0.00 0.00 1.00 06:16:21 AM eth1 595450.00 1190936.00 38380.19 1760805.05 0.00 0.00 0.00 06:16:22 AM eth1 596385.00 1193096.00 38442.56 1763976.29 0.00 0.00 1.00 06:16:23 AM eth1 598155.00 1195978.00 38552.97 1768264.60 0.00 0.00 0.00 06:16:24 AM eth1 594405.00 1188643.00 38312.57 1757414.89 0.00 0.00 1.00 06:16:25 AM eth1 593366.00 1187154.00 38252.16 1755195.83 0.00 0.00 0.00 06:16:26 AM eth1 593188.00 1186118.00 38232.88 1753682.57 0.00 0.00 1.00 06:16:27 AM eth1 596301.00 1192241.00 38440.94 1762733.09 0.00 0.00 0.00 Average: eth1 595457.30 1190843.50 38381.69 1760664.84 0.00 0.00 0.50 lpaa23:~# ./tc -s -d qd sh dev eth1 | grep backlog backlog 7606336b 2513p requeues 167982 backlog 224072b 74p requeues 566 backlog 581376b 192p requeues 5598 backlog 181680b 60p requeues 1070 backlog 5305056b 1753p requeues 110166 // Here, this TX queue is attracting flows backlog 157456b 52p requeues 1758 backlog 672216b 222p requeues 3025 backlog 60560b 20p requeues 24541 backlog 448144b 148p requeues 21258 lpaa23:~# echo 1 >/proc/sys/net/ipv4/tcp_tsq_enable_tcp_wfree_ksoftirqd_detect Immediate jump to full bandwidth, and traffic is properly shard on all tx queues. lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:46 AM eth1 1397632.00 2795397.00 90081.87 4133031.26 0.00 0.00 1.00 06:16:47 AM eth1 1396874.00 2793614.00 90032.99 4130385.46 0.00 0.00 0.00 06:16:48 AM eth1 1395842.00 2791600.00 89966.46 4127409.67 0.00 0.00 1.00 06:16:49 AM eth1 1395528.00 2791017.00 89946.17 4126551.24 0.00 0.00 0.00 06:16:50 AM eth1 1397891.00 2795716.00 90098.74 4133497.39 0.00 0.00 1.00 06:16:51 AM eth1 1394951.00 2789984.00 89908.96 4125022.51 0.00 0.00 0.00 06:16:52 AM eth1 1394608.00 2789190.00 89886.90 4123851.36 0.00 0.00 1.00 06:16:53 AM eth1 1395314.00 2790653.00 89934.33 4125983.09 0.00 0.00 0.00 06:16:54 AM eth1 1396115.00 2792276.00 89984.25 4128411.21 0.00 0.00 1.00 06:16:55 AM eth1 1396829.00 2793523.00 90030.19 4130250.28 0.00 0.00 0.00 Average: eth1 1396158.40 2792297.00 89987.09 4128439.35 0.00 0.00 0.50 lpaa23:~# tc -s -d qd sh dev eth1 | grep backlog backlog 7900052b 2609p requeues 173287 backlog 878120b 290p requeues 589 backlog 1068884b 354p requeues 5621 backlog 996212b 329p requeues 1088 backlog 984100b 325p requeues 115316 backlog 956848b 316p requeues 1781 backlog 1080996b 357p requeues 3047 backlog 975016b 322p requeues 24571 backlog 990156b 327p requeues 21274 (All 8 TX queues get a fair share of the traffic) Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-13 21:27:47 +08:00
return;
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
}
tcp: TCP Small Queues and strange attractors TCP Small queues tries to keep number of packets in qdisc as small as possible, and depends on a tasklet to feed following packets at TX completion time. Choice of tasklet was driven by latencies requirements. Then, TCP stack tries to avoid reorders, by locking flows with outstanding packets in qdisc in a given TX queue. What can happen is that many flows get attracted by a low performing TX queue, and cpu servicing TX completion has to feed packets for all of them, making this cpu 100% busy in softirq mode. This became particularly visible with latest skb->xmit_more support Strategy adopted in this patch is to detect when tcp_wfree() is called from ksoftirqd and let the outstanding queue for this flow being drained before feeding additional packets, so that skb->ooo_okay can be set to allow select_queue() to select the optimal queue : Incoming ACKS are normally handled by different cpus, so this patch gives more chance for these cpus to take over the burden of feeding qdisc with future packets. Tested: lpaa23:~# ./super_netperf 1400 --google-pacing-rate 3028000 -H lpaa24 -l 3600 & lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:18 AM eth1 595448.00 1190564.00 38381.09 1760253.12 0.00 0.00 1.00 06:16:19 AM eth1 594858.00 1189686.00 38340.76 1758952.72 0.00 0.00 0.00 06:16:20 AM eth1 597017.00 1194019.00 38480.79 1765370.29 0.00 0.00 1.00 06:16:21 AM eth1 595450.00 1190936.00 38380.19 1760805.05 0.00 0.00 0.00 06:16:22 AM eth1 596385.00 1193096.00 38442.56 1763976.29 0.00 0.00 1.00 06:16:23 AM eth1 598155.00 1195978.00 38552.97 1768264.60 0.00 0.00 0.00 06:16:24 AM eth1 594405.00 1188643.00 38312.57 1757414.89 0.00 0.00 1.00 06:16:25 AM eth1 593366.00 1187154.00 38252.16 1755195.83 0.00 0.00 0.00 06:16:26 AM eth1 593188.00 1186118.00 38232.88 1753682.57 0.00 0.00 1.00 06:16:27 AM eth1 596301.00 1192241.00 38440.94 1762733.09 0.00 0.00 0.00 Average: eth1 595457.30 1190843.50 38381.69 1760664.84 0.00 0.00 0.50 lpaa23:~# ./tc -s -d qd sh dev eth1 | grep backlog backlog 7606336b 2513p requeues 167982 backlog 224072b 74p requeues 566 backlog 581376b 192p requeues 5598 backlog 181680b 60p requeues 1070 backlog 5305056b 1753p requeues 110166 // Here, this TX queue is attracting flows backlog 157456b 52p requeues 1758 backlog 672216b 222p requeues 3025 backlog 60560b 20p requeues 24541 backlog 448144b 148p requeues 21258 lpaa23:~# echo 1 >/proc/sys/net/ipv4/tcp_tsq_enable_tcp_wfree_ksoftirqd_detect Immediate jump to full bandwidth, and traffic is properly shard on all tx queues. lpaa23:~# sar -n DEV 1 10 | grep eth1 06:16:46 AM eth1 1397632.00 2795397.00 90081.87 4133031.26 0.00 0.00 1.00 06:16:47 AM eth1 1396874.00 2793614.00 90032.99 4130385.46 0.00 0.00 0.00 06:16:48 AM eth1 1395842.00 2791600.00 89966.46 4127409.67 0.00 0.00 1.00 06:16:49 AM eth1 1395528.00 2791017.00 89946.17 4126551.24 0.00 0.00 0.00 06:16:50 AM eth1 1397891.00 2795716.00 90098.74 4133497.39 0.00 0.00 1.00 06:16:51 AM eth1 1394951.00 2789984.00 89908.96 4125022.51 0.00 0.00 0.00 06:16:52 AM eth1 1394608.00 2789190.00 89886.90 4123851.36 0.00 0.00 1.00 06:16:53 AM eth1 1395314.00 2790653.00 89934.33 4125983.09 0.00 0.00 0.00 06:16:54 AM eth1 1396115.00 2792276.00 89984.25 4128411.21 0.00 0.00 1.00 06:16:55 AM eth1 1396829.00 2793523.00 90030.19 4130250.28 0.00 0.00 0.00 Average: eth1 1396158.40 2792297.00 89987.09 4128439.35 0.00 0.00 0.50 lpaa23:~# tc -s -d qd sh dev eth1 | grep backlog backlog 7900052b 2609p requeues 173287 backlog 878120b 290p requeues 589 backlog 1068884b 354p requeues 5621 backlog 996212b 329p requeues 1088 backlog 984100b 325p requeues 115316 backlog 956848b 316p requeues 1781 backlog 1080996b 357p requeues 3047 backlog 975016b 322p requeues 24571 backlog 990156b 327p requeues 21274 (All 8 TX queues get a fair share of the traffic) Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-13 21:27:47 +08:00
out:
sk_free(sk);
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
}
/* Note: Called under soft irq.
* We can call TCP stack right away, unless socket is owned by user.
tcp: internal implementation for pacing BBR congestion control depends on pacing, and pacing is currently handled by sch_fq packet scheduler for performance reasons, and also because implemening pacing with FQ was convenient to truly avoid bursts. However there are many cases where this packet scheduler constraint is not practical. - Many linux hosts are not focusing on handling thousands of TCP flows in the most efficient way. - Some routers use fq_codel or other AQM, but still would like to use BBR for the few TCP flows they initiate/terminate. This patch implements an automatic fallback to internal pacing. Pacing is requested either by BBR or use of SO_MAX_PACING_RATE option. If sch_fq happens to be in the egress path, pacing is delegated to the qdisc, otherwise pacing is done by TCP itself. One advantage of pacing from TCP stack is to get more precise rtt estimations, and less work done from TX completion, since TCP Small queue limits are not generally hit. Setups with single TX queue but many cpus might even benefit from this. Note that unlike sch_fq, we do not take into account header sizes. Taking care of these headers would add additional complexity for no practical differences in behavior. Some performance numbers using 800 TCP_STREAM flows rate limited to ~48 Mbit per second on 40Gbit NIC. If MQ+pfifo_fast is used on the NIC : $ sar -n DEV 1 5 | grep eth 14:48:44 eth0 725743.00 2932134.00 46776.76 4335184.68 0.00 0.00 1.00 14:48:45 eth0 725349.00 2932112.00 46751.86 4335158.90 0.00 0.00 0.00 14:48:46 eth0 725101.00 2931153.00 46735.07 4333748.63 0.00 0.00 0.00 14:48:47 eth0 725099.00 2931161.00 46735.11 4333760.44 0.00 0.00 1.00 14:48:48 eth0 725160.00 2931731.00 46738.88 4334606.07 0.00 0.00 0.00 Average: eth0 725290.40 2931658.20 46747.54 4334491.74 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 4 0 0 259825920 45644 2708324 0 0 21 2 247 98 0 0 100 0 0 4 0 0 259823744 45644 2708356 0 0 0 0 2400825 159843 0 19 81 0 0 0 0 0 259824208 45644 2708072 0 0 0 0 2407351 159929 0 19 81 0 0 1 0 0 259824592 45644 2708128 0 0 0 0 2405183 160386 0 19 80 0 0 1 0 0 259824272 45644 2707868 0 0 0 32 2396361 158037 0 19 81 0 0 Now use MQ+FQ : lpaa23:~# echo fq >/proc/sys/net/core/default_qdisc lpaa23:~# tc qdisc replace dev eth0 root mq $ sar -n DEV 1 5 | grep eth 14:49:57 eth0 678614.00 2727930.00 43739.13 4033279.14 0.00 0.00 0.00 14:49:58 eth0 677620.00 2723971.00 43674.69 4027429.62 0.00 0.00 1.00 14:49:59 eth0 676396.00 2719050.00 43596.83 4020125.02 0.00 0.00 0.00 14:50:00 eth0 675197.00 2714173.00 43518.62 4012938.90 0.00 0.00 1.00 14:50:01 eth0 676388.00 2719063.00 43595.47 4020171.64 0.00 0.00 0.00 Average: eth0 676843.00 2720837.40 43624.95 4022788.86 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 2 0 0 259832240 46008 2710912 0 0 21 2 223 192 0 1 99 0 0 1 0 0 259832896 46008 2710744 0 0 0 0 1702206 198078 0 17 82 0 0 0 0 0 259830272 46008 2710596 0 0 0 0 1696340 197756 1 17 83 0 0 4 0 0 259829168 46024 2710584 0 0 16 0 1688472 197158 1 17 82 0 0 3 0 0 259830224 46024 2710408 0 0 0 0 1692450 197212 0 18 82 0 0 As expected, number of interrupts per second is very different. Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Van Jacobson <vanj@google.com> Cc: Jerry Chu <hkchu@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-16 19:24:36 +08:00
*/
enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
{
struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
struct sock *sk = (struct sock *)tp;
tcp_tsq_handler(sk);
sock_put(sk);
tcp: internal implementation for pacing BBR congestion control depends on pacing, and pacing is currently handled by sch_fq packet scheduler for performance reasons, and also because implemening pacing with FQ was convenient to truly avoid bursts. However there are many cases where this packet scheduler constraint is not practical. - Many linux hosts are not focusing on handling thousands of TCP flows in the most efficient way. - Some routers use fq_codel or other AQM, but still would like to use BBR for the few TCP flows they initiate/terminate. This patch implements an automatic fallback to internal pacing. Pacing is requested either by BBR or use of SO_MAX_PACING_RATE option. If sch_fq happens to be in the egress path, pacing is delegated to the qdisc, otherwise pacing is done by TCP itself. One advantage of pacing from TCP stack is to get more precise rtt estimations, and less work done from TX completion, since TCP Small queue limits are not generally hit. Setups with single TX queue but many cpus might even benefit from this. Note that unlike sch_fq, we do not take into account header sizes. Taking care of these headers would add additional complexity for no practical differences in behavior. Some performance numbers using 800 TCP_STREAM flows rate limited to ~48 Mbit per second on 40Gbit NIC. If MQ+pfifo_fast is used on the NIC : $ sar -n DEV 1 5 | grep eth 14:48:44 eth0 725743.00 2932134.00 46776.76 4335184.68 0.00 0.00 1.00 14:48:45 eth0 725349.00 2932112.00 46751.86 4335158.90 0.00 0.00 0.00 14:48:46 eth0 725101.00 2931153.00 46735.07 4333748.63 0.00 0.00 0.00 14:48:47 eth0 725099.00 2931161.00 46735.11 4333760.44 0.00 0.00 1.00 14:48:48 eth0 725160.00 2931731.00 46738.88 4334606.07 0.00 0.00 0.00 Average: eth0 725290.40 2931658.20 46747.54 4334491.74 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 4 0 0 259825920 45644 2708324 0 0 21 2 247 98 0 0 100 0 0 4 0 0 259823744 45644 2708356 0 0 0 0 2400825 159843 0 19 81 0 0 0 0 0 259824208 45644 2708072 0 0 0 0 2407351 159929 0 19 81 0 0 1 0 0 259824592 45644 2708128 0 0 0 0 2405183 160386 0 19 80 0 0 1 0 0 259824272 45644 2707868 0 0 0 32 2396361 158037 0 19 81 0 0 Now use MQ+FQ : lpaa23:~# echo fq >/proc/sys/net/core/default_qdisc lpaa23:~# tc qdisc replace dev eth0 root mq $ sar -n DEV 1 5 | grep eth 14:49:57 eth0 678614.00 2727930.00 43739.13 4033279.14 0.00 0.00 0.00 14:49:58 eth0 677620.00 2723971.00 43674.69 4027429.62 0.00 0.00 1.00 14:49:59 eth0 676396.00 2719050.00 43596.83 4020125.02 0.00 0.00 0.00 14:50:00 eth0 675197.00 2714173.00 43518.62 4012938.90 0.00 0.00 1.00 14:50:01 eth0 676388.00 2719063.00 43595.47 4020171.64 0.00 0.00 0.00 Average: eth0 676843.00 2720837.40 43624.95 4022788.86 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 2 0 0 259832240 46008 2710912 0 0 21 2 223 192 0 1 99 0 0 1 0 0 259832896 46008 2710744 0 0 0 0 1702206 198078 0 17 82 0 0 0 0 0 259830272 46008 2710596 0 0 0 0 1696340 197756 1 17 83 0 0 4 0 0 259829168 46024 2710584 0 0 16 0 1688472 197158 1 17 82 0 0 3 0 0 259830224 46024 2710408 0 0 0 0 1692450 197212 0 18 82 0 0 As expected, number of interrupts per second is very different. Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Van Jacobson <vanj@google.com> Cc: Jerry Chu <hkchu@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-16 19:24:36 +08:00
return HRTIMER_NORESTART;
}
static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
u64 prior_wstamp)
tcp: new list for sent but unacked skbs for RACK recovery This patch adds a new queue (list) that tracks the sent but not yet acked or SACKed skbs for a TCP connection. The list is chronologically ordered by skb->skb_mstamp (the head is the oldest sent skb). This list will be used to optimize TCP Rack recovery, which checks an skb's timestamp to judge if it has been lost and needs to be retransmitted. Since TCP write queue is ordered by sequence instead of sent time, RACK has to scan over the write queue to catch all eligible packets to detect lost retransmission, and iterates through SACKed skbs repeatedly. Special cares for rare events: 1. TCP repair fakes skb transmission so the send queue needs adjusted 2. SACK reneging would require re-inserting SACKed skbs into the send queue. For now I believe it's not worth the complexity to make RACK work perfectly on SACK reneging, so we do nothing here. 3. Fast Open: currently for non-TFO, send-queue correctly queues the pure SYN packet. For TFO which queues a pure SYN and then a data packet, send-queue only queues the data packet but not the pure SYN due to the structure of TFO code. This is okay because the SYN receiver would never respond with a SACK on a missing SYN (i.e. SYN is never fast-retransmitted by SACK/RACK). In order to not grow sk_buff, we use an union for the new list and _skb_refdst/destructor fields. This is a bit complicated because we need to make sure _skb_refdst and destructor are properly zeroed before skb is cloned/copied at transmit, and before being freed. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-05 03:59:58 +08:00
{
tcp: switch tcp and sch_fq to new earliest departure time model TCP keeps track of tcp_wstamp_ns by itself, meaning sch_fq no longer has to do it. Thanks to this model, TCP can get more accurate RTT samples, since pacing no longer inflates them. This has the nice effect of removing some delays caused by FQ quantum mechanism, causing inflated max/P99 latencies. Also we might relax TCP Small Queue tight limits in the future, since this new model allow TCP to build bigger batches, since sch_fq (or a device with earliest departure time offload) ensure these packets will be delivered on time. Note that other protocols are not converted (they will probably never be) so sch_fq has still support for SO_MAX_PACING_RATE Tested: Test showing FQ pacing quantum artifact for low-rate flows, adding unexpected throttles for RPC flows, inflating max and P99 latencies. The parameters chosen here are to show what happens typically when a TCP flow has a reduced pacing rate (this can be caused by a reduced cwin after few losses, or/and rtt above few ms) MIBS="MIN_LATENCY,MEAN_LATENCY,MAX_LATENCY,P99_LATENCY,STDDEV_LATENCY" Before : $ netperf -H 10.246.7.133 -t TCP_RR -Cc -T6,6 -- -q 2000000 -r 100,100 -o $MIBS MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.246.7.133 () port 0 AF_INET : first burst 0 : cpu bind Minimum Latency Microseconds,Mean Latency Microseconds,Maximum Latency Microseconds,99th Percentile Latency Microseconds,Stddev Latency Microseconds 19,82.78,5279,3825,482.02 After : $ netperf -H 10.246.7.133 -t TCP_RR -Cc -T6,6 -- -q 2000000 -r 100,100 -o $MIBS MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.246.7.133 () port 0 AF_INET : first burst 0 : cpu bind Minimum Latency Microseconds,Mean Latency Microseconds,Maximum Latency Microseconds,99th Percentile Latency Microseconds,Stddev Latency Microseconds 20,49.94,128,63,3.18 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-09-21 23:51:52 +08:00
struct tcp_sock *tp = tcp_sk(sk);
if (sk->sk_pacing_status != SK_PACING_NONE) {
unsigned long rate = sk->sk_pacing_rate;
tcp: switch tcp and sch_fq to new earliest departure time model TCP keeps track of tcp_wstamp_ns by itself, meaning sch_fq no longer has to do it. Thanks to this model, TCP can get more accurate RTT samples, since pacing no longer inflates them. This has the nice effect of removing some delays caused by FQ quantum mechanism, causing inflated max/P99 latencies. Also we might relax TCP Small Queue tight limits in the future, since this new model allow TCP to build bigger batches, since sch_fq (or a device with earliest departure time offload) ensure these packets will be delivered on time. Note that other protocols are not converted (they will probably never be) so sch_fq has still support for SO_MAX_PACING_RATE Tested: Test showing FQ pacing quantum artifact for low-rate flows, adding unexpected throttles for RPC flows, inflating max and P99 latencies. The parameters chosen here are to show what happens typically when a TCP flow has a reduced pacing rate (this can be caused by a reduced cwin after few losses, or/and rtt above few ms) MIBS="MIN_LATENCY,MEAN_LATENCY,MAX_LATENCY,P99_LATENCY,STDDEV_LATENCY" Before : $ netperf -H 10.246.7.133 -t TCP_RR -Cc -T6,6 -- -q 2000000 -r 100,100 -o $MIBS MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.246.7.133 () port 0 AF_INET : first burst 0 : cpu bind Minimum Latency Microseconds,Mean Latency Microseconds,Maximum Latency Microseconds,99th Percentile Latency Microseconds,Stddev Latency Microseconds 19,82.78,5279,3825,482.02 After : $ netperf -H 10.246.7.133 -t TCP_RR -Cc -T6,6 -- -q 2000000 -r 100,100 -o $MIBS MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.246.7.133 () port 0 AF_INET : first burst 0 : cpu bind Minimum Latency Microseconds,Mean Latency Microseconds,Maximum Latency Microseconds,99th Percentile Latency Microseconds,Stddev Latency Microseconds 20,49.94,128,63,3.18 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-09-21 23:51:52 +08:00
/* Original sch_fq does not pace first 10 MSS
* Note that tp->data_segs_out overflows after 2^32 packets,
* this is a minor annoyance.
*/
if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
/* take into account OS jitter */
len_ns -= min_t(u64, len_ns / 2, credit);
tp->tcp_wstamp_ns += len_ns;
tcp: switch tcp and sch_fq to new earliest departure time model TCP keeps track of tcp_wstamp_ns by itself, meaning sch_fq no longer has to do it. Thanks to this model, TCP can get more accurate RTT samples, since pacing no longer inflates them. This has the nice effect of removing some delays caused by FQ quantum mechanism, causing inflated max/P99 latencies. Also we might relax TCP Small Queue tight limits in the future, since this new model allow TCP to build bigger batches, since sch_fq (or a device with earliest departure time offload) ensure these packets will be delivered on time. Note that other protocols are not converted (they will probably never be) so sch_fq has still support for SO_MAX_PACING_RATE Tested: Test showing FQ pacing quantum artifact for low-rate flows, adding unexpected throttles for RPC flows, inflating max and P99 latencies. The parameters chosen here are to show what happens typically when a TCP flow has a reduced pacing rate (this can be caused by a reduced cwin after few losses, or/and rtt above few ms) MIBS="MIN_LATENCY,MEAN_LATENCY,MAX_LATENCY,P99_LATENCY,STDDEV_LATENCY" Before : $ netperf -H 10.246.7.133 -t TCP_RR -Cc -T6,6 -- -q 2000000 -r 100,100 -o $MIBS MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.246.7.133 () port 0 AF_INET : first burst 0 : cpu bind Minimum Latency Microseconds,Mean Latency Microseconds,Maximum Latency Microseconds,99th Percentile Latency Microseconds,Stddev Latency Microseconds 19,82.78,5279,3825,482.02 After : $ netperf -H 10.246.7.133 -t TCP_RR -Cc -T6,6 -- -q 2000000 -r 100,100 -o $MIBS MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.246.7.133 () port 0 AF_INET : first burst 0 : cpu bind Minimum Latency Microseconds,Mean Latency Microseconds,Maximum Latency Microseconds,99th Percentile Latency Microseconds,Stddev Latency Microseconds 20,49.94,128,63,3.18 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-09-21 23:51:52 +08:00
}
}
tcp: new list for sent but unacked skbs for RACK recovery This patch adds a new queue (list) that tracks the sent but not yet acked or SACKed skbs for a TCP connection. The list is chronologically ordered by skb->skb_mstamp (the head is the oldest sent skb). This list will be used to optimize TCP Rack recovery, which checks an skb's timestamp to judge if it has been lost and needs to be retransmitted. Since TCP write queue is ordered by sequence instead of sent time, RACK has to scan over the write queue to catch all eligible packets to detect lost retransmission, and iterates through SACKed skbs repeatedly. Special cares for rare events: 1. TCP repair fakes skb transmission so the send queue needs adjusted 2. SACK reneging would require re-inserting SACKed skbs into the send queue. For now I believe it's not worth the complexity to make RACK work perfectly on SACK reneging, so we do nothing here. 3. Fast Open: currently for non-TFO, send-queue correctly queues the pure SYN packet. For TFO which queues a pure SYN and then a data packet, send-queue only queues the data packet but not the pure SYN due to the structure of TFO code. This is okay because the SYN receiver would never respond with a SACK on a missing SYN (i.e. SYN is never fast-retransmitted by SACK/RACK). In order to not grow sk_buff, we use an union for the new list and _skb_refdst/destructor fields. This is a bit complicated because we need to make sure _skb_refdst and destructor are properly zeroed before skb is cloned/copied at transmit, and before being freed. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-05 03:59:58 +08:00
list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
}
/* This routine actually transmits TCP packets queued in by
* tcp_do_sendmsg(). This is used by both the initial
* transmission and possible later retransmissions.
* All SKB's seen here are completely headerless. It is our
* job to build the TCP header, and pass the packet down to
* IP so it can do the same plus pass the packet off to the
* device.
*
* We are working here with either a clone of the original
* SKB, or a fresh unique copy made by the retransmit engine.
*/
static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
{
const struct inet_connection_sock *icsk = inet_csk(sk);
struct inet_sock *inet;
struct tcp_sock *tp;
struct tcp_skb_cb *tcb;
struct tcp_out_options opts;
unsigned int tcp_options_size, tcp_header_size;
tcp: update skb->skb_mstamp more carefully liujian reported a problem in TCP_USER_TIMEOUT processing with a patch in tcp_probe_timer() : https://www.spinics.net/lists/netdev/msg454496.html After investigations, the root cause of the problem is that we update skb->skb_mstamp of skbs in write queue, even if the attempt to send a clone or copy of it failed. One reason being a routing problem. This patch prevents this, solving liujian issue. It also removes a potential RTT miscalculation, since __tcp_retransmit_skb() is not OR-ing TCP_SKB_CB(skb)->sacked with TCPCB_EVER_RETRANS if a failure happens, but skb->skb_mstamp has been changed. A future ACK would then lead to a very small RTT sample and min_rtt would then be lowered to this too small value. Tested: # cat user_timeout.pkt --local_ip=192.168.102.64 0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 `ifconfig tun0 192.168.102.64/16; ip ro add 192.0.2.1 dev tun0` +0 < S 0:0(0) win 0 <mss 1460> +0 > S. 0:0(0) ack 1 <mss 1460> +.1 < . 1:1(0) ack 1 win 65530 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_USER_TIMEOUT, [3000], 4) = 0 +0 write(4, ..., 24) = 24 +0 > P. 1:25(24) ack 1 win 29200 +.1 < . 1:1(0) ack 25 win 65530 //change the ipaddress +1 `ifconfig tun0 192.168.0.10/16` +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +0 `ifconfig tun0 192.168.102.64/16` +0 < . 1:2(1) ack 25 win 65530 +0 `ifconfig tun0 192.168.0.10/16` +3 write(4, ..., 24) = -1 # ./packetdrill user_timeout.pkt Signed-off-by: Eric Dumazet <edumazet@googl.com> Reported-by: liujian <liujian56@huawei.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-14 11:30:39 +08:00
struct sk_buff *oskb = NULL;
struct tcp_md5sig_key *md5;
struct tcphdr *th;
u64 prior_wstamp;
int err;
BUG_ON(!skb || !tcp_skb_pcount(skb));
tp = tcp_sk(sk);
prior_wstamp = tp->tcp_wstamp_ns;
tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
if (clone_it) {
TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
- tp->snd_una;
tcp: update skb->skb_mstamp more carefully liujian reported a problem in TCP_USER_TIMEOUT processing with a patch in tcp_probe_timer() : https://www.spinics.net/lists/netdev/msg454496.html After investigations, the root cause of the problem is that we update skb->skb_mstamp of skbs in write queue, even if the attempt to send a clone or copy of it failed. One reason being a routing problem. This patch prevents this, solving liujian issue. It also removes a potential RTT miscalculation, since __tcp_retransmit_skb() is not OR-ing TCP_SKB_CB(skb)->sacked with TCPCB_EVER_RETRANS if a failure happens, but skb->skb_mstamp has been changed. A future ACK would then lead to a very small RTT sample and min_rtt would then be lowered to this too small value. Tested: # cat user_timeout.pkt --local_ip=192.168.102.64 0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 `ifconfig tun0 192.168.102.64/16; ip ro add 192.0.2.1 dev tun0` +0 < S 0:0(0) win 0 <mss 1460> +0 > S. 0:0(0) ack 1 <mss 1460> +.1 < . 1:1(0) ack 1 win 65530 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_USER_TIMEOUT, [3000], 4) = 0 +0 write(4, ..., 24) = 24 +0 > P. 1:25(24) ack 1 win 29200 +.1 < . 1:1(0) ack 25 win 65530 //change the ipaddress +1 `ifconfig tun0 192.168.0.10/16` +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +0 `ifconfig tun0 192.168.102.64/16` +0 < . 1:2(1) ack 25 win 65530 +0 `ifconfig tun0 192.168.0.10/16` +3 write(4, ..., 24) = -1 # ./packetdrill user_timeout.pkt Signed-off-by: Eric Dumazet <edumazet@googl.com> Reported-by: liujian <liujian56@huawei.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-14 11:30:39 +08:00
oskb = skb;
tcp: new list for sent but unacked skbs for RACK recovery This patch adds a new queue (list) that tracks the sent but not yet acked or SACKed skbs for a TCP connection. The list is chronologically ordered by skb->skb_mstamp (the head is the oldest sent skb). This list will be used to optimize TCP Rack recovery, which checks an skb's timestamp to judge if it has been lost and needs to be retransmitted. Since TCP write queue is ordered by sequence instead of sent time, RACK has to scan over the write queue to catch all eligible packets to detect lost retransmission, and iterates through SACKed skbs repeatedly. Special cares for rare events: 1. TCP repair fakes skb transmission so the send queue needs adjusted 2. SACK reneging would require re-inserting SACKed skbs into the send queue. For now I believe it's not worth the complexity to make RACK work perfectly on SACK reneging, so we do nothing here. 3. Fast Open: currently for non-TFO, send-queue correctly queues the pure SYN packet. For TFO which queues a pure SYN and then a data packet, send-queue only queues the data packet but not the pure SYN due to the structure of TFO code. This is okay because the SYN receiver would never respond with a SACK on a missing SYN (i.e. SYN is never fast-retransmitted by SACK/RACK). In order to not grow sk_buff, we use an union for the new list and _skb_refdst/destructor fields. This is a bit complicated because we need to make sure _skb_refdst and destructor are properly zeroed before skb is cloned/copied at transmit, and before being freed. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-05 03:59:58 +08:00
tcp_skb_tsorted_save(oskb) {
if (unlikely(skb_cloned(oskb)))
skb = pskb_copy(oskb, gfp_mask);
else
skb = skb_clone(oskb, gfp_mask);
} tcp_skb_tsorted_restore(oskb);
if (unlikely(!skb))
return -ENOBUFS;
}
inet = inet_sk(sk);
tcb = TCP_SKB_CB(skb);
memset(&opts, 0, sizeof(opts));
tcp: force a PSH flag on TSO packets When tcp sends a TSO packet, adding a PSH flag on it reduces the sojourn time of GRO packet in GRO receivers. This is particularly the case under pressure, since RX queues receive packets for many concurrent flows. A sender can give a hint to GRO engines when it is appropriate to flush a super-packet, especially when pacing is in the picture, since next packet is probably delayed by one ms. Having less packets in GRO engine reduces chance of LRU eviction or inflated RTT, and reduces GRO cost. We found recently that we must not set the PSH flag on individual full-size MSS segments [1] : Under pressure (CWR state), we better let the packet sit for a small delay (depending on NAPI logic) so that the ACK packet is delayed, and thus next packet we send is also delayed a bit. Eventually the bottleneck queue can be drained. DCTCP flows with CWND=1 have demonstrated the issue. This patch allows to slowdown the aggregate traffic without involving high resolution timers on senders and/or receivers. It has been used at Google for about four years, and has been discussed at various networking conferences. [1] segments smaller than MSS already have PSH flag set by tcp_sendmsg() / tcp_mark_push(), unless MSG_MORE has been requested by the user. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Tariq Toukan <tariqt@mellanox.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-11 05:49:28 +08:00
if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
tcp: force a PSH flag on TSO packets When tcp sends a TSO packet, adding a PSH flag on it reduces the sojourn time of GRO packet in GRO receivers. This is particularly the case under pressure, since RX queues receive packets for many concurrent flows. A sender can give a hint to GRO engines when it is appropriate to flush a super-packet, especially when pacing is in the picture, since next packet is probably delayed by one ms. Having less packets in GRO engine reduces chance of LRU eviction or inflated RTT, and reduces GRO cost. We found recently that we must not set the PSH flag on individual full-size MSS segments [1] : Under pressure (CWR state), we better let the packet sit for a small delay (depending on NAPI logic) so that the ACK packet is delayed, and thus next packet we send is also delayed a bit. Eventually the bottleneck queue can be drained. DCTCP flows with CWND=1 have demonstrated the issue. This patch allows to slowdown the aggregate traffic without involving high resolution timers on senders and/or receivers. It has been used at Google for about four years, and has been discussed at various networking conferences. [1] segments smaller than MSS already have PSH flag set by tcp_sendmsg() / tcp_mark_push(), unless MSG_MORE has been requested by the user. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Tariq Toukan <tariqt@mellanox.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-11 05:49:28 +08:00
} else {
tcp_options_size = tcp_established_options(sk, skb, &opts,
&md5);
tcp: force a PSH flag on TSO packets When tcp sends a TSO packet, adding a PSH flag on it reduces the sojourn time of GRO packet in GRO receivers. This is particularly the case under pressure, since RX queues receive packets for many concurrent flows. A sender can give a hint to GRO engines when it is appropriate to flush a super-packet, especially when pacing is in the picture, since next packet is probably delayed by one ms. Having less packets in GRO engine reduces chance of LRU eviction or inflated RTT, and reduces GRO cost. We found recently that we must not set the PSH flag on individual full-size MSS segments [1] : Under pressure (CWR state), we better let the packet sit for a small delay (depending on NAPI logic) so that the ACK packet is delayed, and thus next packet we send is also delayed a bit. Eventually the bottleneck queue can be drained. DCTCP flows with CWND=1 have demonstrated the issue. This patch allows to slowdown the aggregate traffic without involving high resolution timers on senders and/or receivers. It has been used at Google for about four years, and has been discussed at various networking conferences. [1] segments smaller than MSS already have PSH flag set by tcp_sendmsg() / tcp_mark_push(), unless MSG_MORE has been requested by the user. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Tariq Toukan <tariqt@mellanox.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-11 05:49:28 +08:00
/* Force a PSH flag on all (GSO) packets to expedite GRO flush
* at receiver : This slightly improve GRO performance.
* Note that we do not force the PSH flag for non GSO packets,
* because they might be sent under high congestion events,
* and in this case it is better to delay the delivery of 1-MSS
* packets and thus the corresponding ACK packet that would
* release the following packet.
*/
if (tcp_skb_pcount(skb) > 1)
tcb->tcp_flags |= TCPHDR_PSH;
}
tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
/* if no packet is in qdisc/device queue, then allow XPS to select
* another queue. We can be called from tcp_tsq_handler()
* which holds one reference to sk.
*
* TODO: Ideally, in-flight pure ACK packets should not matter here.
* One way to get this would be to set skb->truesize = 2 on them.
*/
skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
/* If we had to use memory reserve to allocate this skb,
* this might cause drops if packet is looped back :
* Other socket might not have SOCK_MEMALLOC.
* Packets not looped back do not care about pfmemalloc.
*/
skb->pfmemalloc = 0;
skb_push(skb, tcp_header_size);
skb_reset_transport_header(skb);
tcp: TCP Small Queues This introduce TSQ (TCP Small Queues) TSQ goal is to reduce number of TCP packets in xmit queues (qdisc & device queues), to reduce RTT and cwnd bias, part of the bufferbloat problem. sk->sk_wmem_alloc not allowed to grow above a given limit, allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a given time. TSO packets are sized/capped to half the limit, so that we have two TSO packets in flight, allowing better bandwidth use. As a side effect, setting the limit to 40000 automatically reduces the standard gso max limit (65536) to 40000/2 : It can help to reduce latencies of high prio packets, having smaller TSO packets. This means we divert sock_wfree() to a tcp_wfree() handler, to queue/send following frames when skb_orphan() [2] is called for the already queued skbs. Results on my dev machines (tg3/ixgbe nics) are really impressive, using standard pfifo_fast, and with or without TSO/GSO. Without reduction of nominal bandwidth, we have reduction of buffering per bulk sender : < 1ms on Gbit (instead of 50ms with TSO) < 8ms on 100Mbit (instead of 132 ms) I no longer have 4 MBytes backlogged in qdisc by a single netperf session, and both side socket autotuning no longer use 4 Mbytes. As skb destructor cannot restart xmit itself ( as qdisc lock might be taken at this point ), we delegate the work to a tasklet. We use one tasklest per cpu for performance reasons. If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag. This flag is tested in a new protocol method called from release_sock(), to eventually send new segments. [1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable [2] skb_orphan() is usually called at TX completion time, but some drivers call it in their start_xmit() handler. These drivers should at least use BQL, or else a single TCP session can still fill the whole NIC TX ring, since TSQ will have no effect. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Dave Taht <dave.taht@bufferbloat.net> Cc: Tom Herbert <therbert@google.com> Cc: Matt Mathis <mattmathis@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-07-11 13:50:31 +08:00
skb_orphan(skb);
skb->sk = sk;
skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
skb_set_hash_from_sk(skb, sk);
refcount_add(skb->truesize, &sk->sk_wmem_alloc);
skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
/* Build TCP header and checksum it. */
th = (struct tcphdr *)skb->data;
th->source = inet->inet_sport;
th->dest = inet->inet_dport;
th->seq = htonl(tcb->seq);
th->ack_seq = htonl(rcv_nxt);
*(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
tcb->tcp_flags);
th->check = 0;
th->urg_ptr = 0;
/* The urg_mode check is necessary during a below snd_una win probe */
tcp: Always set urgent pointer if it's beyond snd_nxt Our TCP stack does not set the urgent flag if the urgent pointer does not fit in 16 bits, i.e., if it is more than 64K from the sequence number of a packet. This behaviour is different from the BSDs, and clearly contradicts the purpose of urgent mode, which is to send the notification (though not necessarily the associated data) as soon as possible. Our current behaviour may in fact delay the urgent notification indefinitely if the receiver window does not open up. Simply matching BSD however may break legacy applications which incorrectly rely on the out-of-band delivery of urgent data, and conversely the in-band delivery of non-urgent data. Alexey Kuznetsov suggested a safe solution of following BSD only if the urgent pointer itself has not yet been transmitted. This way we guarantee that when the remote end sees the packet with non-urgent data marked as urgent due to wrap-around we would have advanced the urgent pointer beyond, either to the actual urgent data or to an as-yet untransmitted packet. The only potential downside is that applications on the remote end may see multiple SIGURG notifications. However, this would occur anyway with other TCP stacks. More importantly, the outcome of such a duplicate notification is likely to be harmless since the signal itself does not carry any information other than the fact that we're in urgent mode. Thanks to Ilpo Järvinen for fixing a critical bug in this and Jeff Chua for reporting that bug. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Acked-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-02-22 15:52:29 +08:00
if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
if (before(tp->snd_up, tcb->seq + 0x10000)) {
th->urg_ptr = htons(tp->snd_up - tcb->seq);
th->urg = 1;
} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
th->urg_ptr = htons(0xFFFF);
tcp: Always set urgent pointer if it's beyond snd_nxt Our TCP stack does not set the urgent flag if the urgent pointer does not fit in 16 bits, i.e., if it is more than 64K from the sequence number of a packet. This behaviour is different from the BSDs, and clearly contradicts the purpose of urgent mode, which is to send the notification (though not necessarily the associated data) as soon as possible. Our current behaviour may in fact delay the urgent notification indefinitely if the receiver window does not open up. Simply matching BSD however may break legacy applications which incorrectly rely on the out-of-band delivery of urgent data, and conversely the in-band delivery of non-urgent data. Alexey Kuznetsov suggested a safe solution of following BSD only if the urgent pointer itself has not yet been transmitted. This way we guarantee that when the remote end sees the packet with non-urgent data marked as urgent due to wrap-around we would have advanced the urgent pointer beyond, either to the actual urgent data or to an as-yet untransmitted packet. The only potential downside is that applications on the remote end may see multiple SIGURG notifications. However, this would occur anyway with other TCP stacks. More importantly, the outcome of such a duplicate notification is likely to be harmless since the signal itself does not carry any information other than the fact that we're in urgent mode. Thanks to Ilpo Järvinen for fixing a critical bug in this and Jeff Chua for reporting that bug. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Acked-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-02-22 15:52:29 +08:00
th->urg = 1;
}
}
tcp_options_write((__be32 *)(th + 1), tp, &opts);
skb_shinfo(skb)->gso_type = sk->sk_gso_type;
if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
th->window = htons(tcp_select_window(sk));
tcp_ecn_send(sk, skb, th, tcp_header_size);
} else {
/* RFC1323: The window in SYN & SYN/ACK segments
* is never scaled.
*/
th->window = htons(min(tp->rcv_wnd, 65535U));
}
#ifdef CONFIG_TCP_MD5SIG
/* Calculate the MD5 hash, as we have all we need now */
if (md5) {
sk_nocaps_add(sk, NETIF_F_GSO_MASK);
tp->af_specific->calc_md5_hash(opts.hash_location,
md5, sk, skb);
}
#endif
icsk->icsk_af_ops->send_check(sk, skb);
if (likely(tcb->tcp_flags & TCPHDR_ACK))
tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
tcp: Add RFC4898 tcpEStatsPerfDataSegsOut/In Per RFC4898, they count segments sent/received containing a positive length data segment (that includes retransmission segments carrying data). Unlike tcpi_segs_out/in, tcpi_data_segs_out/in excludes segments carrying no data (e.g. pure ack). The patch also updates the segs_in in tcp_fastopen_add_skb() so that segs_in >= data_segs_in property is kept. Together with retransmission data, tcpi_data_segs_out gives a better signal on the rxmit rate. v6: Rebase on the latest net-next v5: Eric pointed out that checking skb->len is still needed in tcp_fastopen_add_skb() because skb can carry a FIN without data. Hence, instead of open coding segs_in and data_segs_in, tcp_segs_in() helper is used. Comment is added to the fastopen case to explain why segs_in has to be reset and tcp_segs_in() has to be called before __skb_pull(). v4: Add comment to the changes in tcp_fastopen_add_skb() and also add remark on this case in the commit message. v3: Add const modifier to the skb parameter in tcp_segs_in() v2: Rework based on recent fix by Eric: commit a9d99ce28ed3 ("tcp: fix tcpi_segs_in after connection establishment") Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Chris Rapier <rapier@psc.edu> Cc: Eric Dumazet <edumazet@google.com> Cc: Marcelo Ricardo Leitner <mleitner@redhat.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-15 01:52:15 +08:00
if (skb->len != tcp_header_size) {
tcp_event_data_sent(tp, sk);
tcp: Add RFC4898 tcpEStatsPerfDataSegsOut/In Per RFC4898, they count segments sent/received containing a positive length data segment (that includes retransmission segments carrying data). Unlike tcpi_segs_out/in, tcpi_data_segs_out/in excludes segments carrying no data (e.g. pure ack). The patch also updates the segs_in in tcp_fastopen_add_skb() so that segs_in >= data_segs_in property is kept. Together with retransmission data, tcpi_data_segs_out gives a better signal on the rxmit rate. v6: Rebase on the latest net-next v5: Eric pointed out that checking skb->len is still needed in tcp_fastopen_add_skb() because skb can carry a FIN without data. Hence, instead of open coding segs_in and data_segs_in, tcp_segs_in() helper is used. Comment is added to the fastopen case to explain why segs_in has to be reset and tcp_segs_in() has to be called before __skb_pull(). v4: Add comment to the changes in tcp_fastopen_add_skb() and also add remark on this case in the commit message. v3: Add const modifier to the skb parameter in tcp_segs_in() v2: Rework based on recent fix by Eric: commit a9d99ce28ed3 ("tcp: fix tcpi_segs_in after connection establishment") Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Chris Rapier <rapier@psc.edu> Cc: Eric Dumazet <edumazet@google.com> Cc: Marcelo Ricardo Leitner <mleitner@redhat.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-15 01:52:15 +08:00
tp->data_segs_out += tcp_skb_pcount(skb);
tp->bytes_sent += skb->len - tcp_header_size;
tcp: Add RFC4898 tcpEStatsPerfDataSegsOut/In Per RFC4898, they count segments sent/received containing a positive length data segment (that includes retransmission segments carrying data). Unlike tcpi_segs_out/in, tcpi_data_segs_out/in excludes segments carrying no data (e.g. pure ack). The patch also updates the segs_in in tcp_fastopen_add_skb() so that segs_in >= data_segs_in property is kept. Together with retransmission data, tcpi_data_segs_out gives a better signal on the rxmit rate. v6: Rebase on the latest net-next v5: Eric pointed out that checking skb->len is still needed in tcp_fastopen_add_skb() because skb can carry a FIN without data. Hence, instead of open coding segs_in and data_segs_in, tcp_segs_in() helper is used. Comment is added to the fastopen case to explain why segs_in has to be reset and tcp_segs_in() has to be called before __skb_pull(). v4: Add comment to the changes in tcp_fastopen_add_skb() and also add remark on this case in the commit message. v3: Add const modifier to the skb parameter in tcp_segs_in() v2: Rework based on recent fix by Eric: commit a9d99ce28ed3 ("tcp: fix tcpi_segs_in after connection establishment") Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Chris Rapier <rapier@psc.edu> Cc: Eric Dumazet <edumazet@google.com> Cc: Marcelo Ricardo Leitner <mleitner@redhat.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-15 01:52:15 +08:00
}
if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
tcp_skb_pcount(skb));
tp->segs_out += tcp_skb_pcount(skb);
/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
/* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
/* Cleanup our debris for IP stacks */
memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
sizeof(struct inet6_skb_parm)));
tcp: add optional per socket transmit delay Adding delays to TCP flows is crucial for studying behavior of TCP stacks, including congestion control modules. Linux offers netem module, but it has unpractical constraints : - Need root access to change qdisc - Hard to setup on egress if combined with non trivial qdisc like FQ - Single delay for all flows. EDT (Earliest Departure Time) adoption in TCP stack allows us to enable a per socket delay at a very small cost. Networking tools can now establish thousands of flows, each of them with a different delay, simulating real world conditions. This requires FQ packet scheduler or a EDT-enabled NIC. This patchs adds TCP_TX_DELAY socket option, to set a delay in usec units. unsigned int tx_delay = 10000; /* 10 msec */ setsockopt(fd, SOL_TCP, TCP_TX_DELAY, &tx_delay, sizeof(tx_delay)); Note that FQ packet scheduler limits might need some tweaking : man tc-fq PARAMETERS limit Hard limit on the real queue size. When this limit is reached, new packets are dropped. If the value is lowered, packets are dropped so that the new limit is met. Default is 10000 packets. flow_limit Hard limit on the maximum number of packets queued per flow. Default value is 100. Use of TCP_TX_DELAY option will increase number of skbs in FQ qdisc, so packets would be dropped if any of the previous limit is hit. Use of a jump label makes this support runtime-free, for hosts never using the option. Also note that TSQ (TCP Small Queues) limits are slightly changed with this patch : we need to account that skbs artificially delayed wont stop us providind more skbs to feed the pipe (netem uses skb_orphan_partial() for this purpose, but FQ can not use this trick) Because of that, using big delays might very well trigger old bugs in TSO auto defer logic and/or sndbuf limited detection. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-13 02:57:25 +08:00
tcp_add_tx_delay(skb, tp);
err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
tcp: update skb->skb_mstamp more carefully liujian reported a problem in TCP_USER_TIMEOUT processing with a patch in tcp_probe_timer() : https://www.spinics.net/lists/netdev/msg454496.html After investigations, the root cause of the problem is that we update skb->skb_mstamp of skbs in write queue, even if the attempt to send a clone or copy of it failed. One reason being a routing problem. This patch prevents this, solving liujian issue. It also removes a potential RTT miscalculation, since __tcp_retransmit_skb() is not OR-ing TCP_SKB_CB(skb)->sacked with TCPCB_EVER_RETRANS if a failure happens, but skb->skb_mstamp has been changed. A future ACK would then lead to a very small RTT sample and min_rtt would then be lowered to this too small value. Tested: # cat user_timeout.pkt --local_ip=192.168.102.64 0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 `ifconfig tun0 192.168.102.64/16; ip ro add 192.0.2.1 dev tun0` +0 < S 0:0(0) win 0 <mss 1460> +0 > S. 0:0(0) ack 1 <mss 1460> +.1 < . 1:1(0) ack 1 win 65530 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_USER_TIMEOUT, [3000], 4) = 0 +0 write(4, ..., 24) = 24 +0 > P. 1:25(24) ack 1 win 29200 +.1 < . 1:1(0) ack 25 win 65530 //change the ipaddress +1 `ifconfig tun0 192.168.0.10/16` +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +0 `ifconfig tun0 192.168.102.64/16` +0 < . 1:2(1) ack 25 win 65530 +0 `ifconfig tun0 192.168.0.10/16` +3 write(4, ..., 24) = -1 # ./packetdrill user_timeout.pkt Signed-off-by: Eric Dumazet <edumazet@googl.com> Reported-by: liujian <liujian56@huawei.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-14 11:30:39 +08:00
if (unlikely(err > 0)) {
tcp_enter_cwr(sk);
err = net_xmit_eval(err);
}
if (!err && oskb) {
tcp_update_skb_after_send(sk, oskb, prior_wstamp);
tcp_rate_skb_sent(sk, oskb);
}
tcp: update skb->skb_mstamp more carefully liujian reported a problem in TCP_USER_TIMEOUT processing with a patch in tcp_probe_timer() : https://www.spinics.net/lists/netdev/msg454496.html After investigations, the root cause of the problem is that we update skb->skb_mstamp of skbs in write queue, even if the attempt to send a clone or copy of it failed. One reason being a routing problem. This patch prevents this, solving liujian issue. It also removes a potential RTT miscalculation, since __tcp_retransmit_skb() is not OR-ing TCP_SKB_CB(skb)->sacked with TCPCB_EVER_RETRANS if a failure happens, but skb->skb_mstamp has been changed. A future ACK would then lead to a very small RTT sample and min_rtt would then be lowered to this too small value. Tested: # cat user_timeout.pkt --local_ip=192.168.102.64 0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 `ifconfig tun0 192.168.102.64/16; ip ro add 192.0.2.1 dev tun0` +0 < S 0:0(0) win 0 <mss 1460> +0 > S. 0:0(0) ack 1 <mss 1460> +.1 < . 1:1(0) ack 1 win 65530 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_USER_TIMEOUT, [3000], 4) = 0 +0 write(4, ..., 24) = 24 +0 > P. 1:25(24) ack 1 win 29200 +.1 < . 1:1(0) ack 25 win 65530 //change the ipaddress +1 `ifconfig tun0 192.168.0.10/16` +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +1 write(4, ..., 24) = 24 +0 `ifconfig tun0 192.168.102.64/16` +0 < . 1:2(1) ack 25 win 65530 +0 `ifconfig tun0 192.168.0.10/16` +3 write(4, ..., 24) = -1 # ./packetdrill user_timeout.pkt Signed-off-by: Eric Dumazet <edumazet@googl.com> Reported-by: liujian <liujian56@huawei.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-14 11:30:39 +08:00
return err;
}
static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
gfp_t gfp_mask)
{
return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
tcp_sk(sk)->rcv_nxt);
}
/* This routine just queues the buffer for sending.
*
* NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
* otherwise socket can stall.
*/
static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
{
struct tcp_sock *tp = tcp_sk(sk);
/* Advance write_seq and place onto the write_queue. */
WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
__skb_header_release(skb);
tcp_add_write_queue_tail(sk, skb);
sk_wmem_queued_add(sk, skb->truesize);
[NET] CORE: Introducing new memory accounting interface. This patch introduces new memory accounting functions for each network protocol. Most of them are renamed from memory accounting functions for stream protocols. At the same time, some stream memory accounting functions are removed since other functions do same thing. Renaming: sk_stream_free_skb() -> sk_wmem_free_skb() __sk_stream_mem_reclaim() -> __sk_mem_reclaim() sk_stream_mem_reclaim() -> sk_mem_reclaim() sk_stream_mem_schedule -> __sk_mem_schedule() sk_stream_pages() -> sk_mem_pages() sk_stream_rmem_schedule() -> sk_rmem_schedule() sk_stream_wmem_schedule() -> sk_wmem_schedule() sk_charge_skb() -> sk_mem_charge() Removeing sk_stream_rfree(): consolidates into sock_rfree() sk_stream_set_owner_r(): consolidates into skb_set_owner_r() sk_stream_mem_schedule() The following functions are added. sk_has_account(): check if the protocol supports accounting sk_mem_uncharge(): do the opposite of sk_mem_charge() In addition, to achieve consolidation, updating sk_wmem_queued is removed from sk_mem_charge(). Next, to consolidate memory accounting functions, this patch adds memory accounting calls to network core functions. Moreover, present memory accounting call is renamed to new accounting call. Finally we replace present memory accounting calls with new interface in TCP and SCTP. Signed-off-by: Takahiro Yasui <tyasui@redhat.com> Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-12-31 16:11:19 +08:00
sk_mem_charge(sk, skb->truesize);
}
/* Initialize TSO segments for a packet. */
static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
{
if (skb->len <= mss_now) {
/* Avoid the costly divide in the normal
* non-TSO case.
*/
tcp_skb_pcount_set(skb, 1);
TCP_SKB_CB(skb)->tcp_gso_size = 0;
} else {
tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
}
}
/* Pcount in the middle of the write queue got changed, we need to do various
* tweaks to fix counters
*/
static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
{
struct tcp_sock *tp = tcp_sk(sk);
tp->packets_out -= decr;
if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
tp->sacked_out -= decr;
if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
tp->retrans_out -= decr;
if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
tp->lost_out -= decr;
/* Reno case is special. Sigh... */
if (tcp_is_reno(tp) && decr > 0)
tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
if (tp->lost_skb_hint &&
before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
tp->lost_cnt_hint -= decr;
tcp_verify_left_out(tp);
}
static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
{
return TCP_SKB_CB(skb)->txstamp_ack ||
(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
}
static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
{
struct skb_shared_info *shinfo = skb_shinfo(skb);
if (unlikely(tcp_has_tx_tstamp(skb)) &&
!before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
shinfo->tx_flags &= ~tsflags;
shinfo2->tx_flags |= tsflags;
swap(shinfo->tskey, shinfo2->tskey);
tcp: Carry txstamp_ack in tcp_fragment_tstamp When a tcp skb is sliced into two smaller skbs (e.g. in tcp_fragment() and tso_fragment()), it does not carry the txstamp_ack bit to the newly created skb if it is needed. The end result is a timestamping event (SCM_TSTAMP_ACK) will be missing from the sk->sk_error_queue. This patch carries this bit to the new skb2 in tcp_fragment_tstamp(). BPF Output Before: ~~~~~~ <No output due to missing SCM_TSTAMP_ACK timestamp> BPF Output After: ~~~~~~ <...>-2050 [000] d.s. 100.928763: : ee_data:14599 Packetdrill Script: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 14600) = 14600 +0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0 0.200 > . 1:7301(7300) ack 1 0.200 > P. 7301:14601(7300) ack 1 0.300 < . 1:1(0) ack 14601 win 257 0.300 close(4) = 0 0.300 > F. 14601:14601(0) ack 1 0.400 < F. 1:1(0) ack 16062 win 257 0.400 > . 14602:14602(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Tested-by: Soheil Hassas Yeganeh <soheil@google.com> Acked-by: Willem de Bruijn <willemb@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-20 13:50:47 +08:00
TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
TCP_SKB_CB(skb)->txstamp_ack = 0;
}
}
tcp: Handle eor bit when fragmenting a skb When fragmenting a skb, the next_skb should carry the eor from prev_skb. The eor of prev_skb should also be reset. Packetdrill script for testing: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 sendto(4, ..., 15330, MSG_EOR, ..., ...) = 15330 0.200 sendto(4, ..., 730, 0, ..., ...) = 730 0.200 > . 1:7301(7300) ack 1 0.200 > . 7301:14601(7300) ack 1 0.300 < . 1:1(0) ack 14601 win 257 0.300 > P. 14601:15331(730) ack 1 0.300 > P. 15331:16061(730) ack 1 0.400 < . 1:1(0) ack 16061 win 257 0.400 close(4) = 0 0.400 > F. 16061:16061(0) ack 1 0.400 < F. 1:1(0) ack 16062 win 257 0.400 > . 16062:16062(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-26 05:44:50 +08:00
static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
{
TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
TCP_SKB_CB(skb)->eor = 0;
}
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
/* Insert buff after skb on the write or rtx queue of sk. */
static void tcp_insert_write_queue_after(struct sk_buff *skb,
struct sk_buff *buff,
struct sock *sk,
enum tcp_queue tcp_queue)
{
if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
__skb_queue_after(&sk->sk_write_queue, skb, buff);
else
tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
}
/* Function to create two new TCP segments. Shrinks the given segment
* to the specified size and appends a new segment with the rest of the
* packet to the list. This won't be called frequently, I hope.
* Remember, these are still headerless SKBs at this point.
*/
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
struct sk_buff *skb, u32 len,
unsigned int mss_now, gfp_t gfp)
{
struct tcp_sock *tp = tcp_sk(sk);
struct sk_buff *buff;
int nsize, old_factor;
long limit;
int nlen;
u8 flags;
if (WARN_ON(len > skb->len))
return -EINVAL;
nsize = skb_headlen(skb) - len;
if (nsize < 0)
nsize = 0;
/* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
* We need some allowance to not penalize applications setting small
* SO_SNDBUF values.
* Also allow first and last skb in retransmit queue to be split.
*/
limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
skb != tcp_rtx_queue_head(sk) &&
skb != tcp_rtx_queue_tail(sk))) {
NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
return -ENOMEM;
}
if (skb_unclone(skb, gfp))
return -ENOMEM;
/* Get a new skb... force flag on. */
buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
if (!buff)
return -ENOMEM; /* We'll just try again later. */
net/tls: prevent skb_orphan() from leaking TLS plain text with offload sk_validate_xmit_skb() and drivers depend on the sk member of struct sk_buff to identify segments requiring encryption. Any operation which removes or does not preserve the original TLS socket such as skb_orphan() or skb_clone() will cause clear text leaks. Make the TCP socket underlying an offloaded TLS connection mark all skbs as decrypted, if TLS TX is in offload mode. Then in sk_validate_xmit_skb() catch skbs which have no socket (or a socket with no validation) and decrypted flag set. Note that CONFIG_SOCK_VALIDATE_XMIT, CONFIG_TLS_DEVICE and sk->sk_validate_xmit_skb are slightly interchangeable right now, they all imply TLS offload. The new checks are guarded by CONFIG_TLS_DEVICE because that's the option guarding the sk_buff->decrypted member. Second, smaller issue with orphaning is that it breaks the guarantee that packets will be delivered to device queues in-order. All TLS offload drivers depend on that scheduling property. This means skb_orphan_partial()'s trick of preserving partial socket references will cause issues in the drivers. We need a full orphan, and as a result netem delay/throttling will cause all TLS offload skbs to be dropped. Reusing the sk_buff->decrypted flag also protects from leaking clear text when incoming, decrypted skb is redirected (e.g. by TC). See commit 0608c69c9a80 ("bpf: sk_msg, sock{map|hash} redirect through ULP") for justification why the internal flag is safe. The only location which could leak the flag in is tcp_bpf_sendmsg(), which is taken care of by clearing the previously unused bit. v2: - remove superfluous decrypted mark copy (Willem); - remove the stale doc entry (Boris); - rely entirely on EOR marking to prevent coalescing (Boris); - use an internal sendpages flag instead of marking the socket (Boris). v3 (Willem): - reorganize the can_skb_orphan_partial() condition; - fix the flag leak-in through tcp_bpf_sendmsg. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Acked-by: Willem de Bruijn <willemb@google.com> Reviewed-by: Boris Pismenny <borisp@mellanox.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-08 08:03:59 +08:00
skb_copy_decrypted(buff, skb);
sk_wmem_queued_add(sk, buff->truesize);
[NET] CORE: Introducing new memory accounting interface. This patch introduces new memory accounting functions for each network protocol. Most of them are renamed from memory accounting functions for stream protocols. At the same time, some stream memory accounting functions are removed since other functions do same thing. Renaming: sk_stream_free_skb() -> sk_wmem_free_skb() __sk_stream_mem_reclaim() -> __sk_mem_reclaim() sk_stream_mem_reclaim() -> sk_mem_reclaim() sk_stream_mem_schedule -> __sk_mem_schedule() sk_stream_pages() -> sk_mem_pages() sk_stream_rmem_schedule() -> sk_rmem_schedule() sk_stream_wmem_schedule() -> sk_wmem_schedule() sk_charge_skb() -> sk_mem_charge() Removeing sk_stream_rfree(): consolidates into sock_rfree() sk_stream_set_owner_r(): consolidates into skb_set_owner_r() sk_stream_mem_schedule() The following functions are added. sk_has_account(): check if the protocol supports accounting sk_mem_uncharge(): do the opposite of sk_mem_charge() In addition, to achieve consolidation, updating sk_wmem_queued is removed from sk_mem_charge(). Next, to consolidate memory accounting functions, this patch adds memory accounting calls to network core functions. Moreover, present memory accounting call is renamed to new accounting call. Finally we replace present memory accounting calls with new interface in TCP and SCTP. Signed-off-by: Takahiro Yasui <tyasui@redhat.com> Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-12-31 16:11:19 +08:00
sk_mem_charge(sk, buff->truesize);
nlen = skb->len - len - nsize;
buff->truesize += nlen;
skb->truesize -= nlen;
/* Correct the sequence numbers. */
TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
/* PSH and FIN should only be set in the second packet. */
flags = TCP_SKB_CB(skb)->tcp_flags;
TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
TCP_SKB_CB(buff)->tcp_flags = flags;
TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
tcp: Handle eor bit when fragmenting a skb When fragmenting a skb, the next_skb should carry the eor from prev_skb. The eor of prev_skb should also be reset. Packetdrill script for testing: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 sendto(4, ..., 15330, MSG_EOR, ..., ...) = 15330 0.200 sendto(4, ..., 730, 0, ..., ...) = 730 0.200 > . 1:7301(7300) ack 1 0.200 > . 7301:14601(7300) ack 1 0.300 < . 1:1(0) ack 14601 win 257 0.300 > P. 14601:15331(730) ack 1 0.300 > P. 15331:16061(730) ack 1 0.400 < . 1:1(0) ack 16061 win 257 0.400 close(4) = 0 0.400 > F. 16061:16061(0) ack 1 0.400 < F. 1:1(0) ack 16062 win 257 0.400 > . 16062:16062(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-26 05:44:50 +08:00
tcp_skb_fragment_eor(skb, buff);
skb_split(skb, buff, len);
buff->ip_summed = CHECKSUM_PARTIAL;
buff->tstamp = skb->tstamp;
tcp_fragment_tstamp(skb, buff);
old_factor = tcp_skb_pcount(skb);
/* Fix up tso_factor for both original and new SKB. */
tcp_set_skb_tso_segs(skb, mss_now);
tcp_set_skb_tso_segs(buff, mss_now);
tcp: track data delivery rate for a TCP connection This patch generates data delivery rate (throughput) samples on a per-ACK basis. These rate samples can be used by congestion control modules, and specifically will be used by TCP BBR in later patches in this series. Key state: tp->delivered: Tracks the total number of data packets (original or not) delivered so far. This is an already-existing field. tp->delivered_mstamp: the last time tp->delivered was updated. Algorithm: A rate sample is calculated as (d1 - d0)/(t1 - t0) on a per-ACK basis: d1: the current tp->delivered after processing the ACK t1: the current time after processing the ACK d0: the prior tp->delivered when the acked skb was transmitted t0: the prior tp->delivered_mstamp when the acked skb was transmitted When an skb is transmitted, we snapshot d0 and t0 in its control block in tcp_rate_skb_sent(). When an ACK arrives, it may SACK and ACK some skbs. For each SACKed or ACKed skb, tcp_rate_skb_delivered() updates the rate_sample struct to reflect the latest (d0, t0). Finally, tcp_rate_gen() generates a rate sample by storing (d1 - d0) in rs->delivered and (t1 - t0) in rs->interval_us. One caveat: if an skb was sent with no packets in flight, then tp->delivered_mstamp may be either invalid (if the connection is starting) or outdated (if the connection was idle). In that case, we'll re-stamp tp->delivered_mstamp. At first glance it seems t0 should always be the time when an skb was transmitted, but actually this could over-estimate the rate due to phase mismatch between transmit and ACK events. To track the delivery rate, we ensure that if packets are in flight then t0 and and t1 are times at which packets were marked delivered. If the initial and final RTTs are different then one may be corrupted by some sort of noise. The noise we see most often is sending gaps caused by delayed, compressed, or stretched acks. This either affects both RTTs equally or artificially reduces the final RTT. We approach this by recording the info we need to compute the initial RTT (duration of the "send phase" of the window) when we recorded the associated inflight. Then, for a filter to avoid bandwidth overestimates, we generalize the per-sample bandwidth computation from: bw = delivered / ack_phase_rtt to the following: bw = delivered / max(send_phase_rtt, ack_phase_rtt) In large-scale experiments, this filtering approach incorporating send_phase_rtt is effective at avoiding bandwidth overestimates due to ACK compression or stretched ACKs. Signed-off-by: Van Jacobson <vanj@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Nandita Dukkipati <nanditad@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-09-20 11:39:14 +08:00
/* Update delivered info for the new segment */
TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
/* If this packet has been sent out already, we must
* adjust the various packet counters.
*/
if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
int diff = old_factor - tcp_skb_pcount(skb) -
tcp_skb_pcount(buff);
if (diff)
tcp_adjust_pcount(sk, skb, diff);
}
/* Link BUFF into the send queue. */
__skb_header_release(buff);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
return 0;
}
/* This is similar to __pskb_pull_tail(). The difference is that pulled
* data is not copied, but immediately discarded.
*/
static int __pskb_trim_head(struct sk_buff *skb, int len)
{
struct skb_shared_info *shinfo;
int i, k, eat;
eat = min_t(int, len, skb_headlen(skb));
if (eat) {
__skb_pull(skb, eat);
len -= eat;
if (!len)
return 0;
}
eat = len;
k = 0;
shinfo = skb_shinfo(skb);
for (i = 0; i < shinfo->nr_frags; i++) {
int size = skb_frag_size(&shinfo->frags[i]);
if (size <= eat) {
skb_frag_unref(skb, i);
eat -= size;
} else {
shinfo->frags[k] = shinfo->frags[i];
if (eat) {
skb_frag_off_add(&shinfo->frags[k], eat);
skb_frag_size_sub(&shinfo->frags[k], eat);
eat = 0;
}
k++;
}
}
shinfo->nr_frags = k;
skb->data_len -= len;
skb->len = skb->data_len;
return len;
}
/* Remove acked data from a packet in the transmit queue. */
int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
{
u32 delta_truesize;
if (skb_unclone(skb, GFP_ATOMIC))
return -ENOMEM;
delta_truesize = __pskb_trim_head(skb, len);
TCP_SKB_CB(skb)->seq += len;
skb->ip_summed = CHECKSUM_PARTIAL;
if (delta_truesize) {
skb->truesize -= delta_truesize;
sk_wmem_queued_add(sk, -delta_truesize);
sk_mem_uncharge(sk, delta_truesize);
sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
}
/* Any change of skb->len requires recalculation of tso factor. */
if (tcp_skb_pcount(skb) > 1)
tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
return 0;
}
/* Calculate MSS not accounting any TCP options. */
static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
{
const struct tcp_sock *tp = tcp_sk(sk);
const struct inet_connection_sock *icsk = inet_csk(sk);
int mss_now;
/* Calculate base mss without TCP options:
It is MMS_S - sizeof(tcphdr) of rfc1122
*/
mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
ipv6: RTAX_FEATURE_ALLFRAG causes inefficient TCP segment sizing Quoting Tore Anderson from : https://bugzilla.kernel.org/show_bug.cgi?id=42572 When RTAX_FEATURE_ALLFRAG is set on a route, the effective TCP segment size does not take into account the size of the IPv6 Fragmentation header that needs to be included in outbound packets, causing every transmitted TCP segment to be fragmented across two IPv6 packets, the latter of which will only contain 8 bytes of actual payload. RTAX_FEATURE_ALLFRAG is typically set on a route in response to receving a ICMPv6 Packet Too Big message indicating a Path MTU of less than 1280 bytes. 1280 bytes is the minimum IPv6 MTU, however ICMPv6 PTBs with MTU < 1280 are still valid, in particular when an IPv6 packet is sent to an IPv4 destination through a stateless translator. Any ICMPv4 Need To Fragment packets originated from the IPv4 part of the path will be translated to ICMPv6 PTB which may then indicate an MTU of less than 1280. The Linux kernel refuses to reduce the effective MTU to anything below 1280 bytes, instead it sets it to exactly 1280 bytes, and RTAX_FEATURE_ALLFRAG is also set. However, the TCP segment size appears to be set to 1240 bytes (1280 Path MTU - 40 bytes of IPv6 header), instead of 1232 (additionally taking into account the 8 bytes required by the IPv6 Fragmentation extension header). This in turn results in rather inefficient transmission, as every transmitted TCP segment now is split in two fragments containing 1232+8 bytes of payload. After this patch, all the outgoing packets that includes a Fragmentation header all are "atomic" or "non-fragmented" fragments, i.e., they both have Offset=0 and More Fragments=0. With help from David S. Miller Reported-by: Tore Anderson <tore@fud.no> Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Maciej Żenczykowski <maze@google.com> Cc: Tom Herbert <therbert@google.com> Tested-by: Tore Anderson <tore@fud.no> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-24 15:37:38 +08:00
/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
if (icsk->icsk_af_ops->net_frag_header_len) {
const struct dst_entry *dst = __sk_dst_get(sk);
if (dst && dst_allfrag(dst))
mss_now -= icsk->icsk_af_ops->net_frag_header_len;
}
/* Clamp it (mss_clamp does not include tcp options) */
if (mss_now > tp->rx_opt.mss_clamp)
mss_now = tp->rx_opt.mss_clamp;
/* Now subtract optional transport overhead */
mss_now -= icsk->icsk_ext_hdr_len;
/* Then reserve room for full set of TCP options and 8 bytes of data */
mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
return mss_now;
}
/* Calculate MSS. Not accounting for SACKs here. */
int tcp_mtu_to_mss(struct sock *sk, int pmtu)
{
/* Subtract TCP options size, not including SACKs */
return __tcp_mtu_to_mss(sk, pmtu) -
(tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
}
/* Inverse of above */
ipv6: RTAX_FEATURE_ALLFRAG causes inefficient TCP segment sizing Quoting Tore Anderson from : https://bugzilla.kernel.org/show_bug.cgi?id=42572 When RTAX_FEATURE_ALLFRAG is set on a route, the effective TCP segment size does not take into account the size of the IPv6 Fragmentation header that needs to be included in outbound packets, causing every transmitted TCP segment to be fragmented across two IPv6 packets, the latter of which will only contain 8 bytes of actual payload. RTAX_FEATURE_ALLFRAG is typically set on a route in response to receving a ICMPv6 Packet Too Big message indicating a Path MTU of less than 1280 bytes. 1280 bytes is the minimum IPv6 MTU, however ICMPv6 PTBs with MTU < 1280 are still valid, in particular when an IPv6 packet is sent to an IPv4 destination through a stateless translator. Any ICMPv4 Need To Fragment packets originated from the IPv4 part of the path will be translated to ICMPv6 PTB which may then indicate an MTU of less than 1280. The Linux kernel refuses to reduce the effective MTU to anything below 1280 bytes, instead it sets it to exactly 1280 bytes, and RTAX_FEATURE_ALLFRAG is also set. However, the TCP segment size appears to be set to 1240 bytes (1280 Path MTU - 40 bytes of IPv6 header), instead of 1232 (additionally taking into account the 8 bytes required by the IPv6 Fragmentation extension header). This in turn results in rather inefficient transmission, as every transmitted TCP segment now is split in two fragments containing 1232+8 bytes of payload. After this patch, all the outgoing packets that includes a Fragmentation header all are "atomic" or "non-fragmented" fragments, i.e., they both have Offset=0 and More Fragments=0. With help from David S. Miller Reported-by: Tore Anderson <tore@fud.no> Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Maciej Żenczykowski <maze@google.com> Cc: Tom Herbert <therbert@google.com> Tested-by: Tore Anderson <tore@fud.no> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-24 15:37:38 +08:00
int tcp_mss_to_mtu(struct sock *sk, int mss)
{
const struct tcp_sock *tp = tcp_sk(sk);
const struct inet_connection_sock *icsk = inet_csk(sk);
int mtu;
mtu = mss +
tp->tcp_header_len +
icsk->icsk_ext_hdr_len +
icsk->icsk_af_ops->net_header_len;
ipv6: RTAX_FEATURE_ALLFRAG causes inefficient TCP segment sizing Quoting Tore Anderson from : https://bugzilla.kernel.org/show_bug.cgi?id=42572 When RTAX_FEATURE_ALLFRAG is set on a route, the effective TCP segment size does not take into account the size of the IPv6 Fragmentation header that needs to be included in outbound packets, causing every transmitted TCP segment to be fragmented across two IPv6 packets, the latter of which will only contain 8 bytes of actual payload. RTAX_FEATURE_ALLFRAG is typically set on a route in response to receving a ICMPv6 Packet Too Big message indicating a Path MTU of less than 1280 bytes. 1280 bytes is the minimum IPv6 MTU, however ICMPv6 PTBs with MTU < 1280 are still valid, in particular when an IPv6 packet is sent to an IPv4 destination through a stateless translator. Any ICMPv4 Need To Fragment packets originated from the IPv4 part of the path will be translated to ICMPv6 PTB which may then indicate an MTU of less than 1280. The Linux kernel refuses to reduce the effective MTU to anything below 1280 bytes, instead it sets it to exactly 1280 bytes, and RTAX_FEATURE_ALLFRAG is also set. However, the TCP segment size appears to be set to 1240 bytes (1280 Path MTU - 40 bytes of IPv6 header), instead of 1232 (additionally taking into account the 8 bytes required by the IPv6 Fragmentation extension header). This in turn results in rather inefficient transmission, as every transmitted TCP segment now is split in two fragments containing 1232+8 bytes of payload. After this patch, all the outgoing packets that includes a Fragmentation header all are "atomic" or "non-fragmented" fragments, i.e., they both have Offset=0 and More Fragments=0. With help from David S. Miller Reported-by: Tore Anderson <tore@fud.no> Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Maciej Żenczykowski <maze@google.com> Cc: Tom Herbert <therbert@google.com> Tested-by: Tore Anderson <tore@fud.no> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-04-24 15:37:38 +08:00
/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
if (icsk->icsk_af_ops->net_frag_header_len) {
const struct dst_entry *dst = __sk_dst_get(sk);
if (dst && dst_allfrag(dst))
mtu += icsk->icsk_af_ops->net_frag_header_len;
}
return mtu;
}
EXPORT_SYMBOL(tcp_mss_to_mtu);
/* MTU probing init per socket */
void tcp_mtup_init(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
struct inet_connection_sock *icsk = inet_csk(sk);
struct net *net = sock_net(sk);
icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
icsk->icsk_af_ops->net_header_len;
icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
icsk->icsk_mtup.probe_size = 0;
if (icsk->icsk_mtup.enabled)
icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
}
EXPORT_SYMBOL(tcp_mtup_init);
/* This function synchronize snd mss to current pmtu/exthdr set.
tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
for TCP options, but includes only bare TCP header.
tp->rx_opt.mss_clamp is mss negotiated at connection setup.
It is minimum of user_mss and mss received with SYN.
It also does not include TCP options.
inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
tp->mss_cache is current effective sending mss, including
all tcp options except for SACKs. It is evaluated,
taking into account current pmtu, but never exceeds
tp->rx_opt.mss_clamp.
NOTE1. rfc1122 clearly states that advertised MSS
DOES NOT include either tcp or ip options.
NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
are READ ONLY outside this function. --ANK (980731)
*/
unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
{
struct tcp_sock *tp = tcp_sk(sk);
struct inet_connection_sock *icsk = inet_csk(sk);
int mss_now;
if (icsk->icsk_mtup.search_high > pmtu)
icsk->icsk_mtup.search_high = pmtu;
mss_now = tcp_mtu_to_mss(sk, pmtu);
mss_now = tcp_bound_to_half_wnd(tp, mss_now);
/* And store cached results */
icsk->icsk_pmtu_cookie = pmtu;
if (icsk->icsk_mtup.enabled)
mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
tp->mss_cache = mss_now;
return mss_now;
}
EXPORT_SYMBOL(tcp_sync_mss);
/* Compute the current effective MSS, taking SACKs and IP options,
* and even PMTU discovery events into account.
*/
unsigned int tcp_current_mss(struct sock *sk)
{
const struct tcp_sock *tp = tcp_sk(sk);
const struct dst_entry *dst = __sk_dst_get(sk);
u32 mss_now;
unsigned int header_len;
struct tcp_out_options opts;
struct tcp_md5sig_key *md5;
mss_now = tp->mss_cache;
if (dst) {
u32 mtu = dst_mtu(dst);
if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
mss_now = tcp_sync_mss(sk, mtu);
}
header_len = tcp_established_options(sk, NULL, &opts, &md5) +
sizeof(struct tcphdr);
/* The mss_cache is sized based on tp->tcp_header_len, which assumes
* some common options. If this is an odd packet (because we have SACK
* blocks etc) then our calculated header_len will be different, and
* we have to adjust mss_now correspondingly */
if (header_len != tp->tcp_header_len) {
int delta = (int) header_len - tp->tcp_header_len;
mss_now -= delta;
}
return mss_now;
}
/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
* As additional protections, we do not touch cwnd in retransmission phases,
* and if application hit its sndbuf limit recently.
*/
static void tcp_cwnd_application_limited(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
/* Limited by application or receiver window. */
u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
u32 win_used = max(tp->snd_cwnd_used, init_win);
if (win_used < tp->snd_cwnd) {
tp->snd_ssthresh = tcp_current_ssthresh(sk);
tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
}
tp->snd_cwnd_used = 0;
}
tp->snd_cwnd_stamp = tcp_jiffies32;
}
static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
{
const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
[TCP]: Sed magic converts func(sk, tp, ...) -> func(sk, ...) This is (mostly) automated change using magic: sed -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e 's|struct sock \*sk,[\n\t ]*struct tcp_sock \*tp\([^{]*\n{\n\)| struct sock \*sk\1\tstruct tcp_sock *tp = tcp_sk(sk);\n|g' -e 's|struct sock \*sk, struct tcp_sock \*tp| struct sock \*sk|g' -e 's|sk, tp\([^-]\)|sk\1|g' Fixed four unused variable (tp) warnings that were introduced. In addition, manually added newlines after local variables and tweaked function arguments positioning. $ gcc --version gcc (GCC) 4.1.1 20060525 (Red Hat 4.1.1-1) ... $ codiff -fV built-in.o.old built-in.o.new net/ipv4/route.c: rt_cache_flush | +14 1 function changed, 14 bytes added net/ipv4/tcp.c: tcp_setsockopt | -5 tcp_sendpage | -25 tcp_sendmsg | -16 3 functions changed, 46 bytes removed net/ipv4/tcp_input.c: tcp_try_undo_recovery | +3 tcp_try_undo_dsack | +2 tcp_mark_head_lost | -12 tcp_ack | -15 tcp_event_data_recv | -32 tcp_rcv_state_process | -10 tcp_rcv_established | +1 7 functions changed, 6 bytes added, 69 bytes removed, diff: -63 net/ipv4/tcp_output.c: update_send_head | -9 tcp_transmit_skb | +19 tcp_cwnd_validate | +1 tcp_write_wakeup | -17 __tcp_push_pending_frames | -25 tcp_push_one | -8 tcp_send_fin | -4 7 functions changed, 20 bytes added, 63 bytes removed, diff: -43 built-in.o.new: 18 functions changed, 40 bytes added, 178 bytes removed, diff: -138 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-04-21 13:18:02 +08:00
struct tcp_sock *tp = tcp_sk(sk);
/* Track the maximum number of outstanding packets in each
* window, and remember whether we were cwnd-limited then.
*/
if (!before(tp->snd_una, tp->max_packets_seq) ||
tp->packets_out > tp->max_packets_out) {
tp->max_packets_out = tp->packets_out;
tp->max_packets_seq = tp->snd_nxt;
tp->is_cwnd_limited = is_cwnd_limited;
}
tcp: fix cwnd limited checking to improve congestion control Yuchung discovered tcp_is_cwnd_limited() was returning false in slow start phase even if the application filled the socket write queue. All congestion modules take into account tcp_is_cwnd_limited() before increasing cwnd, so this behavior limits slow start from probing the bandwidth at full speed. The problem is that even if write queue is full (aka we are _not_ application limited), cwnd can be under utilized if TSO should auto defer or TCP Small queues decided to hold packets. So the in_flight can be kept to smaller value, and we can get to the point tcp_is_cwnd_limited() returns false. With TCP Small Queues and FQ/pacing, this issue is more visible. We fix this by having tcp_cwnd_validate(), which is supposed to track such things, take into account unsent_segs, the number of segs that we are not sending at the moment due to TSO or TSQ, but intend to send real soon. Then when we are cwnd-limited, remember this fact while we are processing the window of ACKs that comes back. For example, suppose we have a brand new connection with cwnd=10; we are in slow start, and we send a flight of 9 packets. By the time we have received ACKs for all 9 packets we want our cwnd to be 18. We implement this by setting tp->lsnd_pending to 9, and considering ourselves to be cwnd-limited while cwnd is less than twice tp->lsnd_pending (2*9 -> 18). This makes tcp_is_cwnd_limited() more understandable, by removing the GSO/TSO kludge, that tried to work around the issue. Note the in_flight parameter can be removed in a followup cleanup patch. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-01 02:58:13 +08:00
if (tcp_is_cwnd_limited(sk)) {
/* Network is feed fully. */
tp->snd_cwnd_used = 0;
tp->snd_cwnd_stamp = tcp_jiffies32;
} else {
/* Network starves. */
if (tp->packets_out > tp->snd_cwnd_used)
tp->snd_cwnd_used = tp->packets_out;
if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
(s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
!ca_ops->cong_control)
tcp_cwnd_application_limited(sk);
/* The following conditions together indicate the starvation
* is caused by insufficient sender buffer:
* 1) just sent some data (see tcp_write_xmit)
* 2) not cwnd limited (this else condition)
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
* 3) no more data to send (tcp_write_queue_empty())
* 4) application is hitting buffer limit (SOCK_NOSPACE)
*/
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (tcp_write_queue_empty(sk) && sk->sk_socket &&
test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
(1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
}
}
tcp: refine TSO splits While investigating performance problems on small RPC workloads, I noticed linux TCP stack was always splitting the last TSO skb into two parts (skbs). One being a multiple of MSS, and a small one with the Push flag. This split is done even if TCP_NODELAY is set, or if no small packet is in flight. Example with request/response of 4K/4K IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> We can avoid this split by including the Nagle tests at the right place. Note : If some NIC had trouble sending TSO packets with a partial last segment, we would have hit the problem in GRO/forwarding workload already. tcp_minshall_update() is moved to tcp_output.c and is updated as we might feed a TSO packet with a partial last segment. This patch tremendously improves performance, as the traffic now looks like : IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> Before : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280774 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 205719.049006 task-clock # 9.278 CPUs utilized 8,449,968 context-switches # 0.041 M/sec 1,935,997 CPU-migrations # 0.009 M/sec 160,541 page-faults # 0.780 K/sec 548,478,722,290 cycles # 2.666 GHz [83.20%] 455,240,670,857 stalled-cycles-frontend # 83.00% frontend cycles idle [83.48%] 272,881,454,275 stalled-cycles-backend # 49.75% backend cycles idle [66.73%] 166,091,460,030 instructions # 0.30 insns per cycle # 2.74 stalled cycles per insn [83.39%] 29,150,229,399 branches # 141.699 M/sec [83.30%] 1,943,814,026 branch-misses # 6.67% of all branches [83.32%] 22.173517844 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 16851063 0.0 IpExtOutOctets 23878580777 0.0 After patch : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280877 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 107496.071918 task-clock # 4.847 CPUs utilized 5,635,458 context-switches # 0.052 M/sec 1,374,707 CPU-migrations # 0.013 M/sec 160,920 page-faults # 0.001 M/sec 281,500,010,924 cycles # 2.619 GHz [83.28%] 228,865,069,307 stalled-cycles-frontend # 81.30% frontend cycles idle [83.38%] 142,462,742,658 stalled-cycles-backend # 50.61% backend cycles idle [66.81%] 95,227,712,566 instructions # 0.34 insns per cycle # 2.40 stalled cycles per insn [83.43%] 16,209,868,171 branches # 150.795 M/sec [83.20%] 874,252,952 branch-misses # 5.39% of all branches [83.37%] 22.175821286 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 11239428 0.0 IpExtOutOctets 23595191035 0.0 Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher : 2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet scheduler. Many thanks to Neal for review and ideas. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Cc: Van Jacobson <vanj@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Tested-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-14 05:51:23 +08:00
/* Minshall's variant of the Nagle send check. */
static bool tcp_minshall_check(const struct tcp_sock *tp)
{
return after(tp->snd_sml, tp->snd_una) &&
!after(tp->snd_sml, tp->snd_nxt);
}
/* Update snd_sml if this skb is under mss
* Note that a TSO packet might end with a sub-mss segment
* The test is really :
* if ((skb->len % mss) != 0)
* tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
* But we can avoid doing the divide again given we already have
* skb_pcount = skb->len / mss_now
*/
tcp: refine TSO splits While investigating performance problems on small RPC workloads, I noticed linux TCP stack was always splitting the last TSO skb into two parts (skbs). One being a multiple of MSS, and a small one with the Push flag. This split is done even if TCP_NODELAY is set, or if no small packet is in flight. Example with request/response of 4K/4K IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> We can avoid this split by including the Nagle tests at the right place. Note : If some NIC had trouble sending TSO packets with a partial last segment, we would have hit the problem in GRO/forwarding workload already. tcp_minshall_update() is moved to tcp_output.c and is updated as we might feed a TSO packet with a partial last segment. This patch tremendously improves performance, as the traffic now looks like : IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> Before : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280774 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 205719.049006 task-clock # 9.278 CPUs utilized 8,449,968 context-switches # 0.041 M/sec 1,935,997 CPU-migrations # 0.009 M/sec 160,541 page-faults # 0.780 K/sec 548,478,722,290 cycles # 2.666 GHz [83.20%] 455,240,670,857 stalled-cycles-frontend # 83.00% frontend cycles idle [83.48%] 272,881,454,275 stalled-cycles-backend # 49.75% backend cycles idle [66.73%] 166,091,460,030 instructions # 0.30 insns per cycle # 2.74 stalled cycles per insn [83.39%] 29,150,229,399 branches # 141.699 M/sec [83.30%] 1,943,814,026 branch-misses # 6.67% of all branches [83.32%] 22.173517844 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 16851063 0.0 IpExtOutOctets 23878580777 0.0 After patch : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280877 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 107496.071918 task-clock # 4.847 CPUs utilized 5,635,458 context-switches # 0.052 M/sec 1,374,707 CPU-migrations # 0.013 M/sec 160,920 page-faults # 0.001 M/sec 281,500,010,924 cycles # 2.619 GHz [83.28%] 228,865,069,307 stalled-cycles-frontend # 81.30% frontend cycles idle [83.38%] 142,462,742,658 stalled-cycles-backend # 50.61% backend cycles idle [66.81%] 95,227,712,566 instructions # 0.34 insns per cycle # 2.40 stalled cycles per insn [83.43%] 16,209,868,171 branches # 150.795 M/sec [83.20%] 874,252,952 branch-misses # 5.39% of all branches [83.37%] 22.175821286 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 11239428 0.0 IpExtOutOctets 23595191035 0.0 Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher : 2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet scheduler. Many thanks to Neal for review and ideas. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Cc: Van Jacobson <vanj@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Tested-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-14 05:51:23 +08:00
static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
const struct sk_buff *skb)
{
if (skb->len < tcp_skb_pcount(skb) * mss_now)
tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
}
/* Return false, if packet can be sent now without violation Nagle's rules:
* 1. It is full sized. (provided by caller in %partial bool)
* 2. Or it contains FIN. (already checked by caller)
* 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
* 4. Or TCP_CORK is not set, and all sent packets are ACKed.
* With Minshall's modification: all sent small packets are ACKed.
*/
static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
int nonagle)
tcp: refine TSO splits While investigating performance problems on small RPC workloads, I noticed linux TCP stack was always splitting the last TSO skb into two parts (skbs). One being a multiple of MSS, and a small one with the Push flag. This split is done even if TCP_NODELAY is set, or if no small packet is in flight. Example with request/response of 4K/4K IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> We can avoid this split by including the Nagle tests at the right place. Note : If some NIC had trouble sending TSO packets with a partial last segment, we would have hit the problem in GRO/forwarding workload already. tcp_minshall_update() is moved to tcp_output.c and is updated as we might feed a TSO packet with a partial last segment. This patch tremendously improves performance, as the traffic now looks like : IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> Before : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280774 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 205719.049006 task-clock # 9.278 CPUs utilized 8,449,968 context-switches # 0.041 M/sec 1,935,997 CPU-migrations # 0.009 M/sec 160,541 page-faults # 0.780 K/sec 548,478,722,290 cycles # 2.666 GHz [83.20%] 455,240,670,857 stalled-cycles-frontend # 83.00% frontend cycles idle [83.48%] 272,881,454,275 stalled-cycles-backend # 49.75% backend cycles idle [66.73%] 166,091,460,030 instructions # 0.30 insns per cycle # 2.74 stalled cycles per insn [83.39%] 29,150,229,399 branches # 141.699 M/sec [83.30%] 1,943,814,026 branch-misses # 6.67% of all branches [83.32%] 22.173517844 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 16851063 0.0 IpExtOutOctets 23878580777 0.0 After patch : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280877 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 107496.071918 task-clock # 4.847 CPUs utilized 5,635,458 context-switches # 0.052 M/sec 1,374,707 CPU-migrations # 0.013 M/sec 160,920 page-faults # 0.001 M/sec 281,500,010,924 cycles # 2.619 GHz [83.28%] 228,865,069,307 stalled-cycles-frontend # 81.30% frontend cycles idle [83.38%] 142,462,742,658 stalled-cycles-backend # 50.61% backend cycles idle [66.81%] 95,227,712,566 instructions # 0.34 insns per cycle # 2.40 stalled cycles per insn [83.43%] 16,209,868,171 branches # 150.795 M/sec [83.20%] 874,252,952 branch-misses # 5.39% of all branches [83.37%] 22.175821286 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 11239428 0.0 IpExtOutOctets 23595191035 0.0 Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher : 2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet scheduler. Many thanks to Neal for review and ideas. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Cc: Van Jacobson <vanj@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Tested-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-14 05:51:23 +08:00
{
return partial &&
((nonagle & TCP_NAGLE_CORK) ||
(!nonagle && tp->packets_out && tcp_minshall_check(tp)));
}
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
/* Return how many segs we'd like on a TSO packet,
* to send one TSO packet per ms
*/
static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
int min_tso_segs)
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
{
u32 bytes, segs;
bytes = min_t(unsigned long,
sk->sk_pacing_rate >> sk->sk_pacing_shift,
sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
/* Goal is to send at least one packet per ms,
* not one big TSO packet every 100 ms.
* This preserves ACK clocking and is consistent
* with tcp_tso_should_defer() heuristic.
*/
segs = max_t(u32, bytes / mss_now, min_tso_segs);
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
tcp_bbr: better deal with suboptimal GSO BBR uses tcp_tso_autosize() in an attempt to probe what would be the burst sizes and to adjust cwnd in bbr_target_cwnd() with following gold formula : /* Allow enough full-sized skbs in flight to utilize end systems. */ cwnd += 3 * bbr->tso_segs_goal; But GSO can be lacking or be constrained to very small units (ip link set dev ... gso_max_segs 2) What we really want is to have enough packets in flight so that both GSO and GRO are efficient. So in the case GSO is off or downgraded, we still want to have the same number of packets in flight as if GSO/TSO was fully operational, so that GRO can hopefully be working efficiently. To fix this issue, we make tcp_tso_autosize() unaware of sk->sk_gso_max_segs Only tcp_tso_segs() has to enforce the gso_max_segs limit. Tested: ethtool -K eth0 tso off gso off tc qd replace dev eth0 root pfifo_fast Before patch: for f in {1..5}; do ./super_netperf 1 -H lpaa24 -- -K bbr; done     691  (ss -temoi shows cwnd is stuck around 6 )     667     651     631     517 After patch : # for f in {1..5}; do ./super_netperf 1 -H lpaa24 -- -K bbr; done    1733 (ss -temoi shows cwnd is around 386 )    1778    1746    1781    1718 Fixes: 0f8782ea1497 ("tcp_bbr: add BBR congestion control") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Oleksandr Natalenko <oleksandr@natalenko.name> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-02-21 22:43:03 +08:00
return segs;
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
}
/* Return the number of segments we want in the skb we are transmitting.
* See if congestion control module wants to decide; otherwise, autosize.
*/
static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
{
const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
u32 min_tso, tso_segs;
min_tso = ca_ops->min_tso_segs ?
ca_ops->min_tso_segs(sk) :
sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
tcp_bbr: better deal with suboptimal GSO BBR uses tcp_tso_autosize() in an attempt to probe what would be the burst sizes and to adjust cwnd in bbr_target_cwnd() with following gold formula : /* Allow enough full-sized skbs in flight to utilize end systems. */ cwnd += 3 * bbr->tso_segs_goal; But GSO can be lacking or be constrained to very small units (ip link set dev ... gso_max_segs 2) What we really want is to have enough packets in flight so that both GSO and GRO are efficient. So in the case GSO is off or downgraded, we still want to have the same number of packets in flight as if GSO/TSO was fully operational, so that GRO can hopefully be working efficiently. To fix this issue, we make tcp_tso_autosize() unaware of sk->sk_gso_max_segs Only tcp_tso_segs() has to enforce the gso_max_segs limit. Tested: ethtool -K eth0 tso off gso off tc qd replace dev eth0 root pfifo_fast Before patch: for f in {1..5}; do ./super_netperf 1 -H lpaa24 -- -K bbr; done     691  (ss -temoi shows cwnd is stuck around 6 )     667     651     631     517 After patch : # for f in {1..5}; do ./super_netperf 1 -H lpaa24 -- -K bbr; done    1733 (ss -temoi shows cwnd is around 386 )    1778    1746    1781    1718 Fixes: 0f8782ea1497 ("tcp_bbr: add BBR congestion control") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Oleksandr Natalenko <oleksandr@natalenko.name> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-02-21 22:43:03 +08:00
return min_t(u32, tso_segs, sk->sk_gso_max_segs);
}
tcp: refine TSO splits While investigating performance problems on small RPC workloads, I noticed linux TCP stack was always splitting the last TSO skb into two parts (skbs). One being a multiple of MSS, and a small one with the Push flag. This split is done even if TCP_NODELAY is set, or if no small packet is in flight. Example with request/response of 4K/4K IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> We can avoid this split by including the Nagle tests at the right place. Note : If some NIC had trouble sending TSO packets with a partial last segment, we would have hit the problem in GRO/forwarding workload already. tcp_minshall_update() is moved to tcp_output.c and is updated as we might feed a TSO packet with a partial last segment. This patch tremendously improves performance, as the traffic now looks like : IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> Before : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280774 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 205719.049006 task-clock # 9.278 CPUs utilized 8,449,968 context-switches # 0.041 M/sec 1,935,997 CPU-migrations # 0.009 M/sec 160,541 page-faults # 0.780 K/sec 548,478,722,290 cycles # 2.666 GHz [83.20%] 455,240,670,857 stalled-cycles-frontend # 83.00% frontend cycles idle [83.48%] 272,881,454,275 stalled-cycles-backend # 49.75% backend cycles idle [66.73%] 166,091,460,030 instructions # 0.30 insns per cycle # 2.74 stalled cycles per insn [83.39%] 29,150,229,399 branches # 141.699 M/sec [83.30%] 1,943,814,026 branch-misses # 6.67% of all branches [83.32%] 22.173517844 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 16851063 0.0 IpExtOutOctets 23878580777 0.0 After patch : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280877 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 107496.071918 task-clock # 4.847 CPUs utilized 5,635,458 context-switches # 0.052 M/sec 1,374,707 CPU-migrations # 0.013 M/sec 160,920 page-faults # 0.001 M/sec 281,500,010,924 cycles # 2.619 GHz [83.28%] 228,865,069,307 stalled-cycles-frontend # 81.30% frontend cycles idle [83.38%] 142,462,742,658 stalled-cycles-backend # 50.61% backend cycles idle [66.81%] 95,227,712,566 instructions # 0.34 insns per cycle # 2.40 stalled cycles per insn [83.43%] 16,209,868,171 branches # 150.795 M/sec [83.20%] 874,252,952 branch-misses # 5.39% of all branches [83.37%] 22.175821286 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 11239428 0.0 IpExtOutOctets 23595191035 0.0 Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher : 2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet scheduler. Many thanks to Neal for review and ideas. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Cc: Van Jacobson <vanj@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Tested-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-14 05:51:23 +08:00
/* Returns the portion of skb which can be sent right away */
static unsigned int tcp_mss_split_point(const struct sock *sk,
const struct sk_buff *skb,
unsigned int mss_now,
unsigned int max_segs,
int nonagle)
{
const struct tcp_sock *tp = tcp_sk(sk);
tcp: refine TSO splits While investigating performance problems on small RPC workloads, I noticed linux TCP stack was always splitting the last TSO skb into two parts (skbs). One being a multiple of MSS, and a small one with the Push flag. This split is done even if TCP_NODELAY is set, or if no small packet is in flight. Example with request/response of 4K/4K IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> We can avoid this split by including the Nagle tests at the right place. Note : If some NIC had trouble sending TSO packets with a partial last segment, we would have hit the problem in GRO/forwarding workload already. tcp_minshall_update() is moved to tcp_output.c and is updated as we might feed a TSO packet with a partial last segment. This patch tremendously improves performance, as the traffic now looks like : IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> Before : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280774 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 205719.049006 task-clock # 9.278 CPUs utilized 8,449,968 context-switches # 0.041 M/sec 1,935,997 CPU-migrations # 0.009 M/sec 160,541 page-faults # 0.780 K/sec 548,478,722,290 cycles # 2.666 GHz [83.20%] 455,240,670,857 stalled-cycles-frontend # 83.00% frontend cycles idle [83.48%] 272,881,454,275 stalled-cycles-backend # 49.75% backend cycles idle [66.73%] 166,091,460,030 instructions # 0.30 insns per cycle # 2.74 stalled cycles per insn [83.39%] 29,150,229,399 branches # 141.699 M/sec [83.30%] 1,943,814,026 branch-misses # 6.67% of all branches [83.32%] 22.173517844 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 16851063 0.0 IpExtOutOctets 23878580777 0.0 After patch : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280877 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 107496.071918 task-clock # 4.847 CPUs utilized 5,635,458 context-switches # 0.052 M/sec 1,374,707 CPU-migrations # 0.013 M/sec 160,920 page-faults # 0.001 M/sec 281,500,010,924 cycles # 2.619 GHz [83.28%] 228,865,069,307 stalled-cycles-frontend # 81.30% frontend cycles idle [83.38%] 142,462,742,658 stalled-cycles-backend # 50.61% backend cycles idle [66.81%] 95,227,712,566 instructions # 0.34 insns per cycle # 2.40 stalled cycles per insn [83.43%] 16,209,868,171 branches # 150.795 M/sec [83.20%] 874,252,952 branch-misses # 5.39% of all branches [83.37%] 22.175821286 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 11239428 0.0 IpExtOutOctets 23595191035 0.0 Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher : 2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet scheduler. Many thanks to Neal for review and ideas. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Cc: Van Jacobson <vanj@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Tested-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-14 05:51:23 +08:00
u32 partial, needed, window, max_len;
window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
max_len = mss_now * max_segs;
if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
return max_len;
needed = min(skb->len, window);
if (max_len <= needed)
return max_len;
tcp: refine TSO splits While investigating performance problems on small RPC workloads, I noticed linux TCP stack was always splitting the last TSO skb into two parts (skbs). One being a multiple of MSS, and a small one with the Push flag. This split is done even if TCP_NODELAY is set, or if no small packet is in flight. Example with request/response of 4K/4K IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> We can avoid this split by including the Nagle tests at the right place. Note : If some NIC had trouble sending TSO packets with a partial last segment, we would have hit the problem in GRO/forwarding workload already. tcp_minshall_update() is moved to tcp_output.c and is updated as we might feed a TSO packet with a partial last segment. This patch tremendously improves performance, as the traffic now looks like : IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> Before : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280774 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 205719.049006 task-clock # 9.278 CPUs utilized 8,449,968 context-switches # 0.041 M/sec 1,935,997 CPU-migrations # 0.009 M/sec 160,541 page-faults # 0.780 K/sec 548,478,722,290 cycles # 2.666 GHz [83.20%] 455,240,670,857 stalled-cycles-frontend # 83.00% frontend cycles idle [83.48%] 272,881,454,275 stalled-cycles-backend # 49.75% backend cycles idle [66.73%] 166,091,460,030 instructions # 0.30 insns per cycle # 2.74 stalled cycles per insn [83.39%] 29,150,229,399 branches # 141.699 M/sec [83.30%] 1,943,814,026 branch-misses # 6.67% of all branches [83.32%] 22.173517844 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 16851063 0.0 IpExtOutOctets 23878580777 0.0 After patch : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280877 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 107496.071918 task-clock # 4.847 CPUs utilized 5,635,458 context-switches # 0.052 M/sec 1,374,707 CPU-migrations # 0.013 M/sec 160,920 page-faults # 0.001 M/sec 281,500,010,924 cycles # 2.619 GHz [83.28%] 228,865,069,307 stalled-cycles-frontend # 81.30% frontend cycles idle [83.38%] 142,462,742,658 stalled-cycles-backend # 50.61% backend cycles idle [66.81%] 95,227,712,566 instructions # 0.34 insns per cycle # 2.40 stalled cycles per insn [83.43%] 16,209,868,171 branches # 150.795 M/sec [83.20%] 874,252,952 branch-misses # 5.39% of all branches [83.37%] 22.175821286 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 11239428 0.0 IpExtOutOctets 23595191035 0.0 Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher : 2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet scheduler. Many thanks to Neal for review and ideas. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Cc: Van Jacobson <vanj@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Tested-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-14 05:51:23 +08:00
partial = needed % mss_now;
/* If last segment is not a full MSS, check if Nagle rules allow us
* to include this last segment in this skb.
* Otherwise, we'll split the skb at last MSS boundary
*/
if (tcp_nagle_check(partial != 0, tp, nonagle))
tcp: refine TSO splits While investigating performance problems on small RPC workloads, I noticed linux TCP stack was always splitting the last TSO skb into two parts (skbs). One being a multiple of MSS, and a small one with the Push flag. This split is done even if TCP_NODELAY is set, or if no small packet is in flight. Example with request/response of 4K/4K IP A > B: . ack 68432 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 65537:68433(2896) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 68433:69633(1200) ack 69632 win 2783 <nop,nop,timestamp 6524593 6525001> IP B > A: . ack 68433 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: . 69632:72528(2896) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP B > A: P 72528:73728(1200) ack 69633 win 2768 <nop,nop,timestamp 6525001 6524593> IP A > B: . ack 72528 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: . 69633:72529(2896) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> IP A > B: P 72529:73729(1200) ack 73728 win 2783 <nop,nop,timestamp 6524593 6525001> We can avoid this split by including the Nagle tests at the right place. Note : If some NIC had trouble sending TSO packets with a partial last segment, we would have hit the problem in GRO/forwarding workload already. tcp_minshall_update() is moved to tcp_output.c and is updated as we might feed a TSO packet with a partial last segment. This patch tremendously improves performance, as the traffic now looks like : IP A > B: . ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP A > B: P 94209:98305(4096) ack 98304 win 2783 <nop,nop,timestamp 6834277 6834685> IP B > A: . ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP B > A: P 98304:102400(4096) ack 98305 win 2768 <nop,nop,timestamp 6834686 6834277> IP A > B: . ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP A > B: P 98305:102401(4096) ack 102400 win 2783 <nop,nop,timestamp 6834279 6834686> IP B > A: . ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP B > A: P 102400:106496(4096) ack 102401 win 2768 <nop,nop,timestamp 6834687 6834279> IP A > B: . ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP A > B: P 102401:106497(4096) ack 106496 win 2783 <nop,nop,timestamp 6834280 6834687> IP B > A: . ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> IP B > A: P 106496:110592(4096) ack 106497 win 2768 <nop,nop,timestamp 6834688 6834280> Before : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280774 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 205719.049006 task-clock # 9.278 CPUs utilized 8,449,968 context-switches # 0.041 M/sec 1,935,997 CPU-migrations # 0.009 M/sec 160,541 page-faults # 0.780 K/sec 548,478,722,290 cycles # 2.666 GHz [83.20%] 455,240,670,857 stalled-cycles-frontend # 83.00% frontend cycles idle [83.48%] 272,881,454,275 stalled-cycles-backend # 49.75% backend cycles idle [66.73%] 166,091,460,030 instructions # 0.30 insns per cycle # 2.74 stalled cycles per insn [83.39%] 29,150,229,399 branches # 141.699 M/sec [83.30%] 1,943,814,026 branch-misses # 6.67% of all branches [83.32%] 22.173517844 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 16851063 0.0 IpExtOutOctets 23878580777 0.0 After patch : lpq83:~# nstat >/dev/null;perf stat ./super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K 280877 Performance counter stats for './super_netperf 200 -t TCP_RR -H lpq84 -l 20 -- -r 4K,4K': 107496.071918 task-clock # 4.847 CPUs utilized 5,635,458 context-switches # 0.052 M/sec 1,374,707 CPU-migrations # 0.013 M/sec 160,920 page-faults # 0.001 M/sec 281,500,010,924 cycles # 2.619 GHz [83.28%] 228,865,069,307 stalled-cycles-frontend # 81.30% frontend cycles idle [83.38%] 142,462,742,658 stalled-cycles-backend # 50.61% backend cycles idle [66.81%] 95,227,712,566 instructions # 0.34 insns per cycle # 2.40 stalled cycles per insn [83.43%] 16,209,868,171 branches # 150.795 M/sec [83.20%] 874,252,952 branch-misses # 5.39% of all branches [83.37%] 22.175821286 seconds time elapsed lpq83:~# nstat | egrep "IpOutRequests|IpExtOutOctets" IpOutRequests 11239428 0.0 IpExtOutOctets 23595191035 0.0 Indeed, the occupancy of tx skbs (IpExtOutOctets/IpOutRequests) is higher : 2099 instead of 1417, thus helping GRO to be more efficient when using FQ packet scheduler. Many thanks to Neal for review and ideas. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Nandita Dukkipati <nanditad@google.com> Cc: Van Jacobson <vanj@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Tested-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-14 05:51:23 +08:00
return needed - partial;
return needed;
}
/* Can at least one segment of SKB be sent right now, according to the
* congestion window rules? If so, return how many segments are allowed.
*/
static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
const struct sk_buff *skb)
{
u32 in_flight, cwnd, halfcwnd;
/* Don't be strict about the congestion window for the final FIN. */
if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
tcp_skb_pcount(skb) == 1)
return 1;
in_flight = tcp_packets_in_flight(tp);
cwnd = tp->snd_cwnd;
if (in_flight >= cwnd)
return 0;
/* For better scheduling, ensure we have at least
* 2 GSO packets in flight.
*/
halfcwnd = max(cwnd >> 1, 1U);
return min(halfcwnd, cwnd - in_flight);
}
/* Initialize TSO state of a skb.
* This must be invoked the first time we consider transmitting
* SKB onto the wire.
*/
static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
{
int tso_segs = tcp_skb_pcount(skb);
if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
tcp_set_skb_tso_segs(skb, mss_now);
tso_segs = tcp_skb_pcount(skb);
}
return tso_segs;
}
/* Return true if the Nagle test allows this packet to be
* sent now.
*/
static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
unsigned int cur_mss, int nonagle)
{
/* Nagle rule does not apply to frames, which sit in the middle of the
* write_queue (they have no chances to get new data).
*
* This is implemented in the callers, where they modify the 'nonagle'
* argument based upon the location of SKB in the send queue.
*/
if (nonagle & TCP_NAGLE_PUSH)
return true;
tcp: refactor F-RTO The patch series refactor the F-RTO feature (RFC4138/5682). This is to simplify the loss recovery processing. Existing F-RTO was developed during the experimental stage (RFC4138) and has many experimental features. It takes a separate code path from the traditional timeout processing by overloading CA_Disorder instead of using CA_Loss state. This complicates CA_Disorder state handling because it's also used for handling dubious ACKs and undos. While the algorithm in the RFC does not change the congestion control, the implementation intercepts congestion control in various places (e.g., frto_cwnd in tcp_ack()). The new code implements newer F-RTO RFC5682 using CA_Loss processing path. F-RTO becomes a small extension in the timeout processing and interfaces with congestion control and Eifel undo modules. It lets congestion control (module) determines how many to send independently. F-RTO only chooses what to send in order to detect spurious retranmission. If timeout is found spurious it invokes existing Eifel undo algorithms like DSACK or TCP timestamp based detection. The first patch removes all F-RTO code except the sysctl_tcp_frto is left for the new implementation. Since CA_EVENT_FRTO is removed, TCP westwood now computes ssthresh on regular timeout CA_EVENT_LOSS event. Signed-off-by: Yuchung Cheng <ycheng@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-20 21:32:58 +08:00
/* Don't use the nagle rule for urgent data (or for the final FIN). */
if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
return true;
if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
return true;
return false;
}
/* Does at least the first segment of SKB fit into the send window? */
static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
const struct sk_buff *skb,
unsigned int cur_mss)
{
u32 end_seq = TCP_SKB_CB(skb)->end_seq;
if (skb->len > cur_mss)
end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
return !after(end_seq, tcp_wnd_end(tp));
}
/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
* which is put after SKB on the list. It is very much like
* tcp_fragment() except that it may make several kinds of assumptions
* in order to speed up the splitting operation. In particular, we
* know that all the data is in scatter-gather pages, and that the
* packet has never been sent out before (and thus is not cloned).
*/
static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
unsigned int mss_now, gfp_t gfp)
{
int nlen = skb->len - len;
struct sk_buff *buff;
u8 flags;
/* All of a TSO frame must be composed of paged data. */
if (skb->len != skb->data_len)
return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
skb, len, mss_now, gfp);
buff = sk_stream_alloc_skb(sk, 0, gfp, true);
if (unlikely(!buff))
return -ENOMEM;
net/tls: prevent skb_orphan() from leaking TLS plain text with offload sk_validate_xmit_skb() and drivers depend on the sk member of struct sk_buff to identify segments requiring encryption. Any operation which removes or does not preserve the original TLS socket such as skb_orphan() or skb_clone() will cause clear text leaks. Make the TCP socket underlying an offloaded TLS connection mark all skbs as decrypted, if TLS TX is in offload mode. Then in sk_validate_xmit_skb() catch skbs which have no socket (or a socket with no validation) and decrypted flag set. Note that CONFIG_SOCK_VALIDATE_XMIT, CONFIG_TLS_DEVICE and sk->sk_validate_xmit_skb are slightly interchangeable right now, they all imply TLS offload. The new checks are guarded by CONFIG_TLS_DEVICE because that's the option guarding the sk_buff->decrypted member. Second, smaller issue with orphaning is that it breaks the guarantee that packets will be delivered to device queues in-order. All TLS offload drivers depend on that scheduling property. This means skb_orphan_partial()'s trick of preserving partial socket references will cause issues in the drivers. We need a full orphan, and as a result netem delay/throttling will cause all TLS offload skbs to be dropped. Reusing the sk_buff->decrypted flag also protects from leaking clear text when incoming, decrypted skb is redirected (e.g. by TC). See commit 0608c69c9a80 ("bpf: sk_msg, sock{map|hash} redirect through ULP") for justification why the internal flag is safe. The only location which could leak the flag in is tcp_bpf_sendmsg(), which is taken care of by clearing the previously unused bit. v2: - remove superfluous decrypted mark copy (Willem); - remove the stale doc entry (Boris); - rely entirely on EOR marking to prevent coalescing (Boris); - use an internal sendpages flag instead of marking the socket (Boris). v3 (Willem): - reorganize the can_skb_orphan_partial() condition; - fix the flag leak-in through tcp_bpf_sendmsg. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Acked-by: Willem de Bruijn <willemb@google.com> Reviewed-by: Boris Pismenny <borisp@mellanox.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-08 08:03:59 +08:00
skb_copy_decrypted(buff, skb);
sk_wmem_queued_add(sk, buff->truesize);
[NET] CORE: Introducing new memory accounting interface. This patch introduces new memory accounting functions for each network protocol. Most of them are renamed from memory accounting functions for stream protocols. At the same time, some stream memory accounting functions are removed since other functions do same thing. Renaming: sk_stream_free_skb() -> sk_wmem_free_skb() __sk_stream_mem_reclaim() -> __sk_mem_reclaim() sk_stream_mem_reclaim() -> sk_mem_reclaim() sk_stream_mem_schedule -> __sk_mem_schedule() sk_stream_pages() -> sk_mem_pages() sk_stream_rmem_schedule() -> sk_rmem_schedule() sk_stream_wmem_schedule() -> sk_wmem_schedule() sk_charge_skb() -> sk_mem_charge() Removeing sk_stream_rfree(): consolidates into sock_rfree() sk_stream_set_owner_r(): consolidates into skb_set_owner_r() sk_stream_mem_schedule() The following functions are added. sk_has_account(): check if the protocol supports accounting sk_mem_uncharge(): do the opposite of sk_mem_charge() In addition, to achieve consolidation, updating sk_wmem_queued is removed from sk_mem_charge(). Next, to consolidate memory accounting functions, this patch adds memory accounting calls to network core functions. Moreover, present memory accounting call is renamed to new accounting call. Finally we replace present memory accounting calls with new interface in TCP and SCTP. Signed-off-by: Takahiro Yasui <tyasui@redhat.com> Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-12-31 16:11:19 +08:00
sk_mem_charge(sk, buff->truesize);
buff->truesize += nlen;
skb->truesize -= nlen;
/* Correct the sequence numbers. */
TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
/* PSH and FIN should only be set in the second packet. */
flags = TCP_SKB_CB(skb)->tcp_flags;
TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
TCP_SKB_CB(buff)->tcp_flags = flags;
/* This packet was never sent out yet, so no SACK bits. */
TCP_SKB_CB(buff)->sacked = 0;
tcp: Handle eor bit when fragmenting a skb When fragmenting a skb, the next_skb should carry the eor from prev_skb. The eor of prev_skb should also be reset. Packetdrill script for testing: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 sendto(4, ..., 15330, MSG_EOR, ..., ...) = 15330 0.200 sendto(4, ..., 730, 0, ..., ...) = 730 0.200 > . 1:7301(7300) ack 1 0.200 > . 7301:14601(7300) ack 1 0.300 < . 1:1(0) ack 14601 win 257 0.300 > P. 14601:15331(730) ack 1 0.300 > P. 15331:16061(730) ack 1 0.400 < . 1:1(0) ack 16061 win 257 0.400 close(4) = 0 0.400 > F. 16061:16061(0) ack 1 0.400 < F. 1:1(0) ack 16062 win 257 0.400 > . 16062:16062(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-26 05:44:50 +08:00
tcp_skb_fragment_eor(skb, buff);
buff->ip_summed = CHECKSUM_PARTIAL;
skb_split(skb, buff, len);
tcp_fragment_tstamp(skb, buff);
/* Fix up tso_factor for both original and new SKB. */
tcp_set_skb_tso_segs(skb, mss_now);
tcp_set_skb_tso_segs(buff, mss_now);
/* Link BUFF into the send queue. */
__skb_header_release(buff);
tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
return 0;
}
/* Try to defer sending, if possible, in order to minimize the amount
* of TSO splitting we do. View it as a kind of TSO Nagle test.
*
* This algorithm is from John Heffner.
*/
static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
bool *is_cwnd_limited,
bool *is_rwnd_limited,
u32 max_segs)
{
const struct inet_connection_sock *icsk = inet_csk(sk);
u32 send_win, cong_win, limit, in_flight;
tcp: tso: restore IW10 after TSO autosizing With sysctl_tcp_min_tso_segs being 4, it is very possible that tcp_tso_should_defer() decides not sending last 2 MSS of initial window of 10 packets. This also applies if autosizing decides to send X MSS per GSO packet, and cwnd is not a multiple of X. This patch implements an heuristic based on age of first skb in write queue : If it was sent very recently (less than half srtt), we can predict that no ACK packet will come in less than half rtt, so deferring might cause an under utilization of our window. This is visible on initial send (IW10) on web servers, but more generally on some RPC, as the last part of the message might need an extra RTT to get delivered. Tested: Ran following packetdrill test // A simple server-side test that sends exactly an initial window (IW10) // worth of packets. `sysctl -e -q net.ipv4.tcp_min_tso_segs=4` 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +.1 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6> +.1 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 write(4, ..., 14600) = 14600 +0 > . 1:5841(5840) ack 1 win 457 +0 > . 5841:11681(5840) ack 1 win 457 // Following packet should be sent right now. +0 > P. 11681:14601(2920) ack 1 win 457 +.1 < . 1:1(0) ack 14601 win 257 +0 close(4) = 0 +0 > F. 14601:14601(0) ack 1 +.1 < F. 1:1(0) ack 14602 win 257 +0 > . 14602:14602(0) ack 2 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-27 06:10:19 +08:00
struct tcp_sock *tp = tcp_sk(sk);
struct sk_buff *head;
int win_divisor;
s64 delta;
if (icsk->icsk_ca_state >= TCP_CA_Recovery)
goto send_now;
/* Avoid bursty behavior by allowing defer
* only if the last write was recent (1 ms).
* Note that tp->tcp_wstamp_ns can be in the future if we have
* packets waiting in a qdisc or device for EDT delivery.
*/
delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
if (delta > 0)
goto send_now;
in_flight = tcp_packets_in_flight(tp);
BUG_ON(tcp_skb_pcount(skb) <= 1);
BUG_ON(tp->snd_cwnd <= in_flight);
send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
/* From in_flight test above, we know that cwnd > in_flight. */
cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
limit = min(send_win, cong_win);
/* If a full-sized TSO skb can be sent, do it. */
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
if (limit >= max_segs * tp->mss_cache)
goto send_now;
/* Middle in queue won't get any more data, full sendable already? */
if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
goto send_now;
Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next Pull networking updates from David Miller: "Highlights: 1) Maintain the TCP retransmit queue using an rbtree, with 1GB windows at 100Gb this really has become necessary. From Eric Dumazet. 2) Multi-program support for cgroup+bpf, from Alexei Starovoitov. 3) Perform broadcast flooding in hardware in mv88e6xxx, from Andrew Lunn. 4) Add meter action support to openvswitch, from Andy Zhou. 5) Add a data meta pointer for BPF accessible packets, from Daniel Borkmann. 6) Namespace-ify almost all TCP sysctl knobs, from Eric Dumazet. 7) Turn on Broadcom Tags in b53 driver, from Florian Fainelli. 8) More work to move the RTNL mutex down, from Florian Westphal. 9) Add 'bpftool' utility, to help with bpf program introspection. From Jakub Kicinski. 10) Add new 'cpumap' type for XDP_REDIRECT action, from Jesper Dangaard Brouer. 11) Support 'blocks' of transformations in the packet scheduler which can span multiple network devices, from Jiri Pirko. 12) TC flower offload support in cxgb4, from Kumar Sanghvi. 13) Priority based stream scheduler for SCTP, from Marcelo Ricardo Leitner. 14) Thunderbolt networking driver, from Amir Levy and Mika Westerberg. 15) Add RED qdisc offloadability, and use it in mlxsw driver. From Nogah Frankel. 16) eBPF based device controller for cgroup v2, from Roman Gushchin. 17) Add some fundamental tracepoints for TCP, from Song Liu. 18) Remove garbage collection from ipv6 route layer, this is a significant accomplishment. From Wei Wang. 19) Add multicast route offload support to mlxsw, from Yotam Gigi" * git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (2177 commits) tcp: highest_sack fix geneve: fix fill_info when link down bpf: fix lockdep splat net: cdc_ncm: GetNtbFormat endian fix openvswitch: meter: fix NULL pointer dereference in ovs_meter_cmd_reply_start netem: remove unnecessary 64 bit modulus netem: use 64 bit divide by rate tcp: Namespace-ify sysctl_tcp_default_congestion_control net: Protect iterations over net::fib_notifier_ops in fib_seq_sum() ipv6: set all.accept_dad to 0 by default uapi: fix linux/tls.h userspace compilation error usbnet: ipheth: prevent TX queue timeouts when device not ready vhost_net: conditionally enable tx polling uapi: fix linux/rxrpc.h userspace compilation errors net: stmmac: fix LPI transitioning for dwmac4 atm: horizon: Fix irq release error net-sysfs: trigger netlink notification on ifalias change via sysfs openvswitch: Using kfree_rcu() to simplify the code openvswitch: Make local function ovs_nsh_key_attr_size() static openvswitch: Fix return value check in ovs_meter_cmd_features() ...
2017-11-16 03:56:19 +08:00
win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
if (win_divisor) {
u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
/* If at least some fraction of a window is available,
* just use it.
*/
chunk /= win_divisor;
if (limit >= chunk)
goto send_now;
} else {
/* Different approach, try not to defer past a single
* ACK. Receiver should ACK every other full sized
* frame, so if we have space for more than 3 frames
* then send now.
*/
if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
goto send_now;
}
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
/* TODO : use tsorted_sent_queue ? */
head = tcp_rtx_queue_head(sk);
if (!head)
goto send_now;
delta = tp->tcp_clock_cache - head->tstamp;
tcp: tso: restore IW10 after TSO autosizing With sysctl_tcp_min_tso_segs being 4, it is very possible that tcp_tso_should_defer() decides not sending last 2 MSS of initial window of 10 packets. This also applies if autosizing decides to send X MSS per GSO packet, and cwnd is not a multiple of X. This patch implements an heuristic based on age of first skb in write queue : If it was sent very recently (less than half srtt), we can predict that no ACK packet will come in less than half rtt, so deferring might cause an under utilization of our window. This is visible on initial send (IW10) on web servers, but more generally on some RPC, as the last part of the message might need an extra RTT to get delivered. Tested: Ran following packetdrill test // A simple server-side test that sends exactly an initial window (IW10) // worth of packets. `sysctl -e -q net.ipv4.tcp_min_tso_segs=4` 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +.1 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6> +.1 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 write(4, ..., 14600) = 14600 +0 > . 1:5841(5840) ack 1 win 457 +0 > . 5841:11681(5840) ack 1 win 457 // Following packet should be sent right now. +0 > P. 11681:14601(2920) ack 1 win 457 +.1 < . 1:1(0) ack 14601 win 257 +0 close(4) = 0 +0 > F. 14601:14601(0) ack 1 +.1 < F. 1:1(0) ack 14602 win 257 +0 > . 14602:14602(0) ack 2 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-27 06:10:19 +08:00
/* If next ACK is likely to come too late (half srtt), do not defer */
if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
tcp: tso: restore IW10 after TSO autosizing With sysctl_tcp_min_tso_segs being 4, it is very possible that tcp_tso_should_defer() decides not sending last 2 MSS of initial window of 10 packets. This also applies if autosizing decides to send X MSS per GSO packet, and cwnd is not a multiple of X. This patch implements an heuristic based on age of first skb in write queue : If it was sent very recently (less than half srtt), we can predict that no ACK packet will come in less than half rtt, so deferring might cause an under utilization of our window. This is visible on initial send (IW10) on web servers, but more generally on some RPC, as the last part of the message might need an extra RTT to get delivered. Tested: Ran following packetdrill test // A simple server-side test that sends exactly an initial window (IW10) // worth of packets. `sysctl -e -q net.ipv4.tcp_min_tso_segs=4` 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +.1 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6> +.1 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 write(4, ..., 14600) = 14600 +0 > . 1:5841(5840) ack 1 win 457 +0 > . 5841:11681(5840) ack 1 win 457 // Following packet should be sent right now. +0 > P. 11681:14601(2920) ack 1 win 457 +.1 < . 1:1(0) ack 14601 win 257 +0 close(4) = 0 +0 > F. 14601:14601(0) ack 1 +.1 < F. 1:1(0) ack 14602 win 257 +0 > . 14602:14602(0) ack 2 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-02-27 06:10:19 +08:00
goto send_now;
/* Ok, it looks like it is advisable to defer.
* Three cases are tracked :
* 1) We are cwnd-limited
* 2) We are rwnd-limited
* 3) We are application limited.
*/
if (cong_win < send_win) {
if (cong_win <= skb->len) {
*is_cwnd_limited = true;
return true;
}
} else {
if (send_win <= skb->len) {
*is_rwnd_limited = true;
return true;
}
}
/* If this packet won't get more data, do not wait. */
if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
TCP_SKB_CB(skb)->eor)
goto send_now;
return true;
send_now:
return false;
}
static inline void tcp_mtu_check_reprobe(struct sock *sk)
{
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
struct net *net = sock_net(sk);
u32 interval;
s32 delta;
interval = net->ipv4.sysctl_tcp_probe_interval;
delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
if (unlikely(delta >= interval * HZ)) {
int mss = tcp_current_mss(sk);
/* Update current search range */
icsk->icsk_mtup.probe_size = 0;
icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
sizeof(struct tcphdr) +
icsk->icsk_af_ops->net_header_len;
icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
/* Update probe time stamp */
icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
}
}
static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
{
struct sk_buff *skb, *next;
skb = tcp_send_head(sk);
tcp_for_write_queue_from_safe(skb, next, sk) {
if (len <= skb->len)
break;
if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
return false;
len -= skb->len;
}
return true;
}
/* Create a new MTU probe if we are ready.
* MTU probe is regularly attempting to increase the path MTU by
* deliberately sending larger packets. This discovers routing
* changes resulting in larger path MTUs.
*
* Returns 0 if we should wait to probe (no cwnd available),
* 1 if a probe was sent,
* -1 otherwise
*/
static int tcp_mtu_probe(struct sock *sk)
{
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
struct sk_buff *skb, *nskb, *next;
struct net *net = sock_net(sk);
int probe_size;
int size_needed;
int copy, len;
int mss_now;
int interval;
/* Not currently probing/verifying,
* not in recovery,
* have enough cwnd, and
* not SACKing (the variable headers throw things off)
*/
if (likely(!icsk->icsk_mtup.enabled ||
icsk->icsk_mtup.probe_size ||
inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
tp->snd_cwnd < 11 ||
tp->rx_opt.num_sacks || tp->rx_opt.dsack))
return -1;
/* Use binary search for probe_size between tcp_mss_base,
* and current mss_clamp. if (search_high - search_low)
* smaller than a threshold, backoff from probing.
*/
mss_now = tcp_current_mss(sk);
probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
icsk->icsk_mtup.search_low) >> 1);
size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
/* When misfortune happens, we are reprobing actively,
* and then reprobe timer has expired. We stick with current
* probing process by not resetting search range to its orignal.
*/
if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
interval < net->ipv4.sysctl_tcp_probe_threshold) {
/* Check whether enough time has elaplased for
* another round of probing.
*/
tcp_mtu_check_reprobe(sk);
return -1;
}
/* Have enough data in the send queue to probe? */
if (tp->write_seq - tp->snd_nxt < size_needed)
return -1;
if (tp->snd_wnd < size_needed)
return -1;
if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
return 0;
/* Do we need to wait to drain cwnd? With none in flight, don't stall */
if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
if (!tcp_packets_in_flight(tp))
return -1;
else
return 0;
}
if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
return -1;
/* We're allowed to probe. Build it now. */
nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
if (!nskb)
return -1;
sk_wmem_queued_add(sk, nskb->truesize);
[NET] CORE: Introducing new memory accounting interface. This patch introduces new memory accounting functions for each network protocol. Most of them are renamed from memory accounting functions for stream protocols. At the same time, some stream memory accounting functions are removed since other functions do same thing. Renaming: sk_stream_free_skb() -> sk_wmem_free_skb() __sk_stream_mem_reclaim() -> __sk_mem_reclaim() sk_stream_mem_reclaim() -> sk_mem_reclaim() sk_stream_mem_schedule -> __sk_mem_schedule() sk_stream_pages() -> sk_mem_pages() sk_stream_rmem_schedule() -> sk_rmem_schedule() sk_stream_wmem_schedule() -> sk_wmem_schedule() sk_charge_skb() -> sk_mem_charge() Removeing sk_stream_rfree(): consolidates into sock_rfree() sk_stream_set_owner_r(): consolidates into skb_set_owner_r() sk_stream_mem_schedule() The following functions are added. sk_has_account(): check if the protocol supports accounting sk_mem_uncharge(): do the opposite of sk_mem_charge() In addition, to achieve consolidation, updating sk_wmem_queued is removed from sk_mem_charge(). Next, to consolidate memory accounting functions, this patch adds memory accounting calls to network core functions. Moreover, present memory accounting call is renamed to new accounting call. Finally we replace present memory accounting calls with new interface in TCP and SCTP. Signed-off-by: Takahiro Yasui <tyasui@redhat.com> Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-12-31 16:11:19 +08:00
sk_mem_charge(sk, nskb->truesize);
skb = tcp_send_head(sk);
net/tls: prevent skb_orphan() from leaking TLS plain text with offload sk_validate_xmit_skb() and drivers depend on the sk member of struct sk_buff to identify segments requiring encryption. Any operation which removes or does not preserve the original TLS socket such as skb_orphan() or skb_clone() will cause clear text leaks. Make the TCP socket underlying an offloaded TLS connection mark all skbs as decrypted, if TLS TX is in offload mode. Then in sk_validate_xmit_skb() catch skbs which have no socket (or a socket with no validation) and decrypted flag set. Note that CONFIG_SOCK_VALIDATE_XMIT, CONFIG_TLS_DEVICE and sk->sk_validate_xmit_skb are slightly interchangeable right now, they all imply TLS offload. The new checks are guarded by CONFIG_TLS_DEVICE because that's the option guarding the sk_buff->decrypted member. Second, smaller issue with orphaning is that it breaks the guarantee that packets will be delivered to device queues in-order. All TLS offload drivers depend on that scheduling property. This means skb_orphan_partial()'s trick of preserving partial socket references will cause issues in the drivers. We need a full orphan, and as a result netem delay/throttling will cause all TLS offload skbs to be dropped. Reusing the sk_buff->decrypted flag also protects from leaking clear text when incoming, decrypted skb is redirected (e.g. by TC). See commit 0608c69c9a80 ("bpf: sk_msg, sock{map|hash} redirect through ULP") for justification why the internal flag is safe. The only location which could leak the flag in is tcp_bpf_sendmsg(), which is taken care of by clearing the previously unused bit. v2: - remove superfluous decrypted mark copy (Willem); - remove the stale doc entry (Boris); - rely entirely on EOR marking to prevent coalescing (Boris); - use an internal sendpages flag instead of marking the socket (Boris). v3 (Willem): - reorganize the can_skb_orphan_partial() condition; - fix the flag leak-in through tcp_bpf_sendmsg. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Acked-by: Willem de Bruijn <willemb@google.com> Reviewed-by: Boris Pismenny <borisp@mellanox.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-08 08:03:59 +08:00
skb_copy_decrypted(nskb, skb);
TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
TCP_SKB_CB(nskb)->sacked = 0;
nskb->csum = 0;
nskb->ip_summed = CHECKSUM_PARTIAL;
tcp_insert_write_queue_before(nskb, skb, sk);
tcp_highest_sack_replace(sk, skb, nskb);
len = 0;
tcp_for_write_queue_from_safe(skb, next, sk) {
copy = min_t(int, skb->len, probe_size - len);
skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
if (skb->len <= copy) {
/* We've eaten all the data from this skb.
* Throw it away. */
TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
/* If this is the last SKB we copy and eor is set
* we need to propagate it to the new skb.
*/
TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
tcp_skb_collapse_tstamp(nskb, skb);
tcp_unlink_write_queue(skb, sk);
[NET] CORE: Introducing new memory accounting interface. This patch introduces new memory accounting functions for each network protocol. Most of them are renamed from memory accounting functions for stream protocols. At the same time, some stream memory accounting functions are removed since other functions do same thing. Renaming: sk_stream_free_skb() -> sk_wmem_free_skb() __sk_stream_mem_reclaim() -> __sk_mem_reclaim() sk_stream_mem_reclaim() -> sk_mem_reclaim() sk_stream_mem_schedule -> __sk_mem_schedule() sk_stream_pages() -> sk_mem_pages() sk_stream_rmem_schedule() -> sk_rmem_schedule() sk_stream_wmem_schedule() -> sk_wmem_schedule() sk_charge_skb() -> sk_mem_charge() Removeing sk_stream_rfree(): consolidates into sock_rfree() sk_stream_set_owner_r(): consolidates into skb_set_owner_r() sk_stream_mem_schedule() The following functions are added. sk_has_account(): check if the protocol supports accounting sk_mem_uncharge(): do the opposite of sk_mem_charge() In addition, to achieve consolidation, updating sk_wmem_queued is removed from sk_mem_charge(). Next, to consolidate memory accounting functions, this patch adds memory accounting calls to network core functions. Moreover, present memory accounting call is renamed to new accounting call. Finally we replace present memory accounting calls with new interface in TCP and SCTP. Signed-off-by: Takahiro Yasui <tyasui@redhat.com> Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-12-31 16:11:19 +08:00
sk_wmem_free_skb(sk, skb);
} else {
TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
~(TCPHDR_FIN|TCPHDR_PSH);
if (!skb_shinfo(skb)->nr_frags) {
skb_pull(skb, copy);
} else {
__pskb_trim_head(skb, copy);
tcp_set_skb_tso_segs(skb, mss_now);
}
TCP_SKB_CB(skb)->seq += copy;
}
len += copy;
if (len >= probe_size)
break;
}
tcp_init_tso_segs(nskb, nskb->len);
/* We're ready to send. If this fails, the probe will
* be resegmented into mss-sized pieces by tcp_write_xmit().
*/
if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
/* Decrement cwnd here because we are sending
* effectively two packets. */
tp->snd_cwnd--;
tcp_event_new_data_sent(sk, nskb);
icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
return 1;
}
return -1;
}
static bool tcp_pacing_check(struct sock *sk)
tcp: internal implementation for pacing BBR congestion control depends on pacing, and pacing is currently handled by sch_fq packet scheduler for performance reasons, and also because implemening pacing with FQ was convenient to truly avoid bursts. However there are many cases where this packet scheduler constraint is not practical. - Many linux hosts are not focusing on handling thousands of TCP flows in the most efficient way. - Some routers use fq_codel or other AQM, but still would like to use BBR for the few TCP flows they initiate/terminate. This patch implements an automatic fallback to internal pacing. Pacing is requested either by BBR or use of SO_MAX_PACING_RATE option. If sch_fq happens to be in the egress path, pacing is delegated to the qdisc, otherwise pacing is done by TCP itself. One advantage of pacing from TCP stack is to get more precise rtt estimations, and less work done from TX completion, since TCP Small queue limits are not generally hit. Setups with single TX queue but many cpus might even benefit from this. Note that unlike sch_fq, we do not take into account header sizes. Taking care of these headers would add additional complexity for no practical differences in behavior. Some performance numbers using 800 TCP_STREAM flows rate limited to ~48 Mbit per second on 40Gbit NIC. If MQ+pfifo_fast is used on the NIC : $ sar -n DEV 1 5 | grep eth 14:48:44 eth0 725743.00 2932134.00 46776.76 4335184.68 0.00 0.00 1.00 14:48:45 eth0 725349.00 2932112.00 46751.86 4335158.90 0.00 0.00 0.00 14:48:46 eth0 725101.00 2931153.00 46735.07 4333748.63 0.00 0.00 0.00 14:48:47 eth0 725099.00 2931161.00 46735.11 4333760.44 0.00 0.00 1.00 14:48:48 eth0 725160.00 2931731.00 46738.88 4334606.07 0.00 0.00 0.00 Average: eth0 725290.40 2931658.20 46747.54 4334491.74 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 4 0 0 259825920 45644 2708324 0 0 21 2 247 98 0 0 100 0 0 4 0 0 259823744 45644 2708356 0 0 0 0 2400825 159843 0 19 81 0 0 0 0 0 259824208 45644 2708072 0 0 0 0 2407351 159929 0 19 81 0 0 1 0 0 259824592 45644 2708128 0 0 0 0 2405183 160386 0 19 80 0 0 1 0 0 259824272 45644 2707868 0 0 0 32 2396361 158037 0 19 81 0 0 Now use MQ+FQ : lpaa23:~# echo fq >/proc/sys/net/core/default_qdisc lpaa23:~# tc qdisc replace dev eth0 root mq $ sar -n DEV 1 5 | grep eth 14:49:57 eth0 678614.00 2727930.00 43739.13 4033279.14 0.00 0.00 0.00 14:49:58 eth0 677620.00 2723971.00 43674.69 4027429.62 0.00 0.00 1.00 14:49:59 eth0 676396.00 2719050.00 43596.83 4020125.02 0.00 0.00 0.00 14:50:00 eth0 675197.00 2714173.00 43518.62 4012938.90 0.00 0.00 1.00 14:50:01 eth0 676388.00 2719063.00 43595.47 4020171.64 0.00 0.00 0.00 Average: eth0 676843.00 2720837.40 43624.95 4022788.86 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 2 0 0 259832240 46008 2710912 0 0 21 2 223 192 0 1 99 0 0 1 0 0 259832896 46008 2710744 0 0 0 0 1702206 198078 0 17 82 0 0 0 0 0 259830272 46008 2710596 0 0 0 0 1696340 197756 1 17 83 0 0 4 0 0 259829168 46024 2710584 0 0 16 0 1688472 197158 1 17 82 0 0 3 0 0 259830224 46024 2710408 0 0 0 0 1692450 197212 0 18 82 0 0 As expected, number of interrupts per second is very different. Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Van Jacobson <vanj@google.com> Cc: Jerry Chu <hkchu@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-16 19:24:36 +08:00
{
struct tcp_sock *tp = tcp_sk(sk);
if (!tcp_needs_internal_pacing(sk))
return false;
if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
return false;
if (!hrtimer_is_queued(&tp->pacing_timer)) {
hrtimer_start(&tp->pacing_timer,
ns_to_ktime(tp->tcp_wstamp_ns),
HRTIMER_MODE_ABS_PINNED_SOFT);
sock_hold(sk);
}
return true;
tcp: internal implementation for pacing BBR congestion control depends on pacing, and pacing is currently handled by sch_fq packet scheduler for performance reasons, and also because implemening pacing with FQ was convenient to truly avoid bursts. However there are many cases where this packet scheduler constraint is not practical. - Many linux hosts are not focusing on handling thousands of TCP flows in the most efficient way. - Some routers use fq_codel or other AQM, but still would like to use BBR for the few TCP flows they initiate/terminate. This patch implements an automatic fallback to internal pacing. Pacing is requested either by BBR or use of SO_MAX_PACING_RATE option. If sch_fq happens to be in the egress path, pacing is delegated to the qdisc, otherwise pacing is done by TCP itself. One advantage of pacing from TCP stack is to get more precise rtt estimations, and less work done from TX completion, since TCP Small queue limits are not generally hit. Setups with single TX queue but many cpus might even benefit from this. Note that unlike sch_fq, we do not take into account header sizes. Taking care of these headers would add additional complexity for no practical differences in behavior. Some performance numbers using 800 TCP_STREAM flows rate limited to ~48 Mbit per second on 40Gbit NIC. If MQ+pfifo_fast is used on the NIC : $ sar -n DEV 1 5 | grep eth 14:48:44 eth0 725743.00 2932134.00 46776.76 4335184.68 0.00 0.00 1.00 14:48:45 eth0 725349.00 2932112.00 46751.86 4335158.90 0.00 0.00 0.00 14:48:46 eth0 725101.00 2931153.00 46735.07 4333748.63 0.00 0.00 0.00 14:48:47 eth0 725099.00 2931161.00 46735.11 4333760.44 0.00 0.00 1.00 14:48:48 eth0 725160.00 2931731.00 46738.88 4334606.07 0.00 0.00 0.00 Average: eth0 725290.40 2931658.20 46747.54 4334491.74 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 4 0 0 259825920 45644 2708324 0 0 21 2 247 98 0 0 100 0 0 4 0 0 259823744 45644 2708356 0 0 0 0 2400825 159843 0 19 81 0 0 0 0 0 259824208 45644 2708072 0 0 0 0 2407351 159929 0 19 81 0 0 1 0 0 259824592 45644 2708128 0 0 0 0 2405183 160386 0 19 80 0 0 1 0 0 259824272 45644 2707868 0 0 0 32 2396361 158037 0 19 81 0 0 Now use MQ+FQ : lpaa23:~# echo fq >/proc/sys/net/core/default_qdisc lpaa23:~# tc qdisc replace dev eth0 root mq $ sar -n DEV 1 5 | grep eth 14:49:57 eth0 678614.00 2727930.00 43739.13 4033279.14 0.00 0.00 0.00 14:49:58 eth0 677620.00 2723971.00 43674.69 4027429.62 0.00 0.00 1.00 14:49:59 eth0 676396.00 2719050.00 43596.83 4020125.02 0.00 0.00 0.00 14:50:00 eth0 675197.00 2714173.00 43518.62 4012938.90 0.00 0.00 1.00 14:50:01 eth0 676388.00 2719063.00 43595.47 4020171.64 0.00 0.00 0.00 Average: eth0 676843.00 2720837.40 43624.95 4022788.86 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 2 0 0 259832240 46008 2710912 0 0 21 2 223 192 0 1 99 0 0 1 0 0 259832896 46008 2710744 0 0 0 0 1702206 198078 0 17 82 0 0 0 0 0 259830272 46008 2710596 0 0 0 0 1696340 197756 1 17 83 0 0 4 0 0 259829168 46024 2710584 0 0 16 0 1688472 197158 1 17 82 0 0 3 0 0 259830224 46024 2710408 0 0 0 0 1692450 197212 0 18 82 0 0 As expected, number of interrupts per second is very different. Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Van Jacobson <vanj@google.com> Cc: Jerry Chu <hkchu@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-16 19:24:36 +08:00
}
/* TCP Small Queues :
* Control number of packets in qdisc/devices to two packets / or ~1 ms.
* (These limits are doubled for retransmits)
* This allows for :
* - better RTT estimation and ACK scheduling
* - faster recovery
* - high rates
* Alas, some drivers / subsystems require a fair amount
* of queued bytes to ensure line rate.
* One example is wifi aggregation (802.11 AMPDU)
*/
static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
unsigned int factor)
{
unsigned long limit;
limit = max_t(unsigned long,
2 * skb->truesize,
sk->sk_pacing_rate >> sk->sk_pacing_shift);
if (sk->sk_pacing_status == SK_PACING_NONE)
limit = min_t(unsigned long, limit,
sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
limit <<= factor;
tcp: add optional per socket transmit delay Adding delays to TCP flows is crucial for studying behavior of TCP stacks, including congestion control modules. Linux offers netem module, but it has unpractical constraints : - Need root access to change qdisc - Hard to setup on egress if combined with non trivial qdisc like FQ - Single delay for all flows. EDT (Earliest Departure Time) adoption in TCP stack allows us to enable a per socket delay at a very small cost. Networking tools can now establish thousands of flows, each of them with a different delay, simulating real world conditions. This requires FQ packet scheduler or a EDT-enabled NIC. This patchs adds TCP_TX_DELAY socket option, to set a delay in usec units. unsigned int tx_delay = 10000; /* 10 msec */ setsockopt(fd, SOL_TCP, TCP_TX_DELAY, &tx_delay, sizeof(tx_delay)); Note that FQ packet scheduler limits might need some tweaking : man tc-fq PARAMETERS limit Hard limit on the real queue size. When this limit is reached, new packets are dropped. If the value is lowered, packets are dropped so that the new limit is met. Default is 10000 packets. flow_limit Hard limit on the maximum number of packets queued per flow. Default value is 100. Use of TCP_TX_DELAY option will increase number of skbs in FQ qdisc, so packets would be dropped if any of the previous limit is hit. Use of a jump label makes this support runtime-free, for hosts never using the option. Also note that TSQ (TCP Small Queues) limits are slightly changed with this patch : we need to account that skbs artificially delayed wont stop us providind more skbs to feed the pipe (netem uses skb_orphan_partial() for this purpose, but FQ can not use this trick) Because of that, using big delays might very well trigger old bugs in TSO auto defer logic and/or sndbuf limited detection. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-13 02:57:25 +08:00
if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
tcp_sk(sk)->tcp_tx_delay) {
u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
/* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
* approximate our needs assuming an ~100% skb->truesize overhead.
* USEC_PER_SEC is approximated by 2^20.
* do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
*/
extra_bytes >>= (20 - 1);
limit += extra_bytes;
}
if (refcount_read(&sk->sk_wmem_alloc) > limit) {
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
/* Always send skb if rtx queue is empty.
* No need to wait for TX completion to call us back,
* after softirq/tasklet schedule.
* This helps when TX completions are delayed too much.
*/
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (tcp_rtx_queue_empty(sk))
return false;
set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
/* It is possible TX completion already happened
* before we set TSQ_THROTTLED, so we must
* test again the condition.
*/
smp_mb__after_atomic();
if (refcount_read(&sk->sk_wmem_alloc) > limit)
return true;
}
return false;
}
static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
{
const u32 now = tcp_jiffies32;
enum tcp_chrono old = tp->chrono_type;
if (old > TCP_CHRONO_UNSPEC)
tp->chrono_stat[old - 1] += now - tp->chrono_start;
tp->chrono_start = now;
tp->chrono_type = new;
}
void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
{
struct tcp_sock *tp = tcp_sk(sk);
/* If there are multiple conditions worthy of tracking in a
* chronograph then the highest priority enum takes precedence
* over the other conditions. So that if something "more interesting"
* starts happening, stop the previous chrono and start a new one.
*/
if (type > tp->chrono_type)
tcp_chrono_set(tp, type);
}
void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
{
struct tcp_sock *tp = tcp_sk(sk);
/* There are multiple conditions worthy of tracking in a
* chronograph, so that the highest priority enum takes
* precedence over the other conditions (see tcp_chrono_start).
* If a condition stops, we only stop chrono tracking if
* it's the "most interesting" or current chrono we are
* tracking and starts busy chrono if we have pending data.
*/
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (tcp_rtx_and_write_queues_empty(sk))
tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
else if (type == tp->chrono_type)
tcp_chrono_set(tp, TCP_CHRONO_BUSY);
}
/* This routine writes packets to the network. It advances the
* send_head. This happens as incoming acks open up the remote
* window for us.
*
* LARGESEND note: !tcp_urg_mode is overkill, only frames between
* snd_up-64k-mss .. snd_up cannot be large. However, taking into
* account rare use of URG, this is not a big flaw.
*
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
* Send at most one packet when push_one > 0. Temporarily ignore
* cwnd limit to force at most one packet out when push_one == 2.
* Returns true, if no segments are in flight and we have queued segments,
* but cannot send anything now because of SWS or another problem.
*/
static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
int push_one, gfp_t gfp)
{
struct tcp_sock *tp = tcp_sk(sk);
struct sk_buff *skb;
unsigned int tso_segs, sent_pkts;
int cwnd_quota;
int result;
bool is_cwnd_limited = false, is_rwnd_limited = false;
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
u32 max_segs;
sent_pkts = 0;
tcp_mstamp_refresh(tp);
if (!push_one) {
/* Do MTU probing. */
result = tcp_mtu_probe(sk);
if (!result) {
return false;
} else if (result > 0) {
sent_pkts = 1;
}
}
max_segs = tcp_tso_segs(sk, mss_now);
while ((skb = tcp_send_head(sk))) {
unsigned int limit;
if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
/* "skb_mstamp_ns" is used as a start point for the retransmit timer */
skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
tcp: repaired skbs must init their tso_segs syzbot reported a WARN_ON(!tcp_skb_pcount(skb)) in tcp_send_loss_probe() [1] This was caused by TCP_REPAIR sent skbs that inadvertenly were missing a call to tcp_init_tso_segs() [1] WARNING: CPU: 1 PID: 0 at net/ipv4/tcp_output.c:2534 tcp_send_loss_probe+0x771/0x8a0 net/ipv4/tcp_output.c:2534 Kernel panic - not syncing: panic_on_warn set ... CPU: 1 PID: 0 Comm: swapper/1 Not tainted 5.0.0-rc7+ #77 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: <IRQ> __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x172/0x1f0 lib/dump_stack.c:113 panic+0x2cb/0x65c kernel/panic.c:214 __warn.cold+0x20/0x45 kernel/panic.c:571 report_bug+0x263/0x2b0 lib/bug.c:186 fixup_bug arch/x86/kernel/traps.c:178 [inline] fixup_bug arch/x86/kernel/traps.c:173 [inline] do_error_trap+0x11b/0x200 arch/x86/kernel/traps.c:271 do_invalid_op+0x37/0x50 arch/x86/kernel/traps.c:290 invalid_op+0x14/0x20 arch/x86/entry/entry_64.S:973 RIP: 0010:tcp_send_loss_probe+0x771/0x8a0 net/ipv4/tcp_output.c:2534 Code: 88 fc ff ff 4c 89 ef e8 ed 75 c8 fb e9 c8 fc ff ff e8 43 76 c8 fb e9 63 fd ff ff e8 d9 75 c8 fb e9 94 f9 ff ff e8 bf 03 91 fb <0f> 0b e9 7d fa ff ff e8 b3 03 91 fb 0f b6 1d 37 43 7a 03 31 ff 89 RSP: 0018:ffff8880ae907c60 EFLAGS: 00010206 RAX: ffff8880a989c340 RBX: 0000000000000000 RCX: ffffffff85dedbdb RDX: 0000000000000100 RSI: ffffffff85dee0b1 RDI: 0000000000000005 RBP: ffff8880ae907c90 R08: ffff8880a989c340 R09: ffffed10147d1ae1 R10: ffffed10147d1ae0 R11: ffff8880a3e8d703 R12: ffff888091b90040 R13: ffff8880a3e8d540 R14: 0000000000008000 R15: ffff888091b90860 tcp_write_timer_handler+0x5c0/0x8a0 net/ipv4/tcp_timer.c:583 tcp_write_timer+0x10e/0x1d0 net/ipv4/tcp_timer.c:607 call_timer_fn+0x190/0x720 kernel/time/timer.c:1325 expire_timers kernel/time/timer.c:1362 [inline] __run_timers kernel/time/timer.c:1681 [inline] __run_timers kernel/time/timer.c:1649 [inline] run_timer_softirq+0x652/0x1700 kernel/time/timer.c:1694 __do_softirq+0x266/0x95a kernel/softirq.c:292 invoke_softirq kernel/softirq.c:373 [inline] irq_exit+0x180/0x1d0 kernel/softirq.c:413 exiting_irq arch/x86/include/asm/apic.h:536 [inline] smp_apic_timer_interrupt+0x14a/0x570 arch/x86/kernel/apic/apic.c:1062 apic_timer_interrupt+0xf/0x20 arch/x86/entry/entry_64.S:807 </IRQ> RIP: 0010:native_safe_halt+0x2/0x10 arch/x86/include/asm/irqflags.h:58 Code: ff ff ff 48 89 c7 48 89 45 d8 e8 59 0c a1 fa 48 8b 45 d8 e9 ce fe ff ff 48 89 df e8 48 0c a1 fa eb 82 90 90 90 90 90 90 fb f4 <c3> 0f 1f 00 66 2e 0f 1f 84 00 00 00 00 00 f4 c3 90 90 90 90 90 90 RSP: 0018:ffff8880a98afd78 EFLAGS: 00000286 ORIG_RAX: ffffffffffffff13 RAX: 1ffffffff1125061 RBX: ffff8880a989c340 RCX: 0000000000000000 RDX: dffffc0000000000 RSI: 0000000000000001 RDI: ffff8880a989cbbc RBP: ffff8880a98afda8 R08: ffff8880a989c340 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000001 R13: ffffffff889282f8 R14: 0000000000000001 R15: 0000000000000000 arch_cpu_idle+0x10/0x20 arch/x86/kernel/process.c:555 default_idle_call+0x36/0x90 kernel/sched/idle.c:93 cpuidle_idle_call kernel/sched/idle.c:153 [inline] do_idle+0x386/0x570 kernel/sched/idle.c:262 cpu_startup_entry+0x1b/0x20 kernel/sched/idle.c:353 start_secondary+0x404/0x5c0 arch/x86/kernel/smpboot.c:271 secondary_startup_64+0xa4/0xb0 arch/x86/kernel/head_64.S:243 Kernel Offset: disabled Rebooting in 86400 seconds.. Fixes: 79861919b889 ("tcp: fix TCP_REPAIR xmit queue setup") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: syzbot <syzkaller@googlegroups.com> Cc: Andrey Vagin <avagin@openvz.org> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-02-24 07:51:51 +08:00
tcp_init_tso_segs(skb, mss_now);
goto repair; /* Skip network transmission */
}
tcp: internal implementation for pacing BBR congestion control depends on pacing, and pacing is currently handled by sch_fq packet scheduler for performance reasons, and also because implemening pacing with FQ was convenient to truly avoid bursts. However there are many cases where this packet scheduler constraint is not practical. - Many linux hosts are not focusing on handling thousands of TCP flows in the most efficient way. - Some routers use fq_codel or other AQM, but still would like to use BBR for the few TCP flows they initiate/terminate. This patch implements an automatic fallback to internal pacing. Pacing is requested either by BBR or use of SO_MAX_PACING_RATE option. If sch_fq happens to be in the egress path, pacing is delegated to the qdisc, otherwise pacing is done by TCP itself. One advantage of pacing from TCP stack is to get more precise rtt estimations, and less work done from TX completion, since TCP Small queue limits are not generally hit. Setups with single TX queue but many cpus might even benefit from this. Note that unlike sch_fq, we do not take into account header sizes. Taking care of these headers would add additional complexity for no practical differences in behavior. Some performance numbers using 800 TCP_STREAM flows rate limited to ~48 Mbit per second on 40Gbit NIC. If MQ+pfifo_fast is used on the NIC : $ sar -n DEV 1 5 | grep eth 14:48:44 eth0 725743.00 2932134.00 46776.76 4335184.68 0.00 0.00 1.00 14:48:45 eth0 725349.00 2932112.00 46751.86 4335158.90 0.00 0.00 0.00 14:48:46 eth0 725101.00 2931153.00 46735.07 4333748.63 0.00 0.00 0.00 14:48:47 eth0 725099.00 2931161.00 46735.11 4333760.44 0.00 0.00 1.00 14:48:48 eth0 725160.00 2931731.00 46738.88 4334606.07 0.00 0.00 0.00 Average: eth0 725290.40 2931658.20 46747.54 4334491.74 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 4 0 0 259825920 45644 2708324 0 0 21 2 247 98 0 0 100 0 0 4 0 0 259823744 45644 2708356 0 0 0 0 2400825 159843 0 19 81 0 0 0 0 0 259824208 45644 2708072 0 0 0 0 2407351 159929 0 19 81 0 0 1 0 0 259824592 45644 2708128 0 0 0 0 2405183 160386 0 19 80 0 0 1 0 0 259824272 45644 2707868 0 0 0 32 2396361 158037 0 19 81 0 0 Now use MQ+FQ : lpaa23:~# echo fq >/proc/sys/net/core/default_qdisc lpaa23:~# tc qdisc replace dev eth0 root mq $ sar -n DEV 1 5 | grep eth 14:49:57 eth0 678614.00 2727930.00 43739.13 4033279.14 0.00 0.00 0.00 14:49:58 eth0 677620.00 2723971.00 43674.69 4027429.62 0.00 0.00 1.00 14:49:59 eth0 676396.00 2719050.00 43596.83 4020125.02 0.00 0.00 0.00 14:50:00 eth0 675197.00 2714173.00 43518.62 4012938.90 0.00 0.00 1.00 14:50:01 eth0 676388.00 2719063.00 43595.47 4020171.64 0.00 0.00 0.00 Average: eth0 676843.00 2720837.40 43624.95 4022788.86 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 2 0 0 259832240 46008 2710912 0 0 21 2 223 192 0 1 99 0 0 1 0 0 259832896 46008 2710744 0 0 0 0 1702206 198078 0 17 82 0 0 0 0 0 259830272 46008 2710596 0 0 0 0 1696340 197756 1 17 83 0 0 4 0 0 259829168 46024 2710584 0 0 16 0 1688472 197158 1 17 82 0 0 3 0 0 259830224 46024 2710408 0 0 0 0 1692450 197212 0 18 82 0 0 As expected, number of interrupts per second is very different. Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Van Jacobson <vanj@google.com> Cc: Jerry Chu <hkchu@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-16 19:24:36 +08:00
if (tcp_pacing_check(sk))
break;
tso_segs = tcp_init_tso_segs(skb, mss_now);
BUG_ON(!tso_segs);
cwnd_quota = tcp_cwnd_test(tp, skb);
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
if (!cwnd_quota) {
if (push_one == 2)
/* Force out a loss probe pkt. */
cwnd_quota = 1;
else
break;
}
if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
is_rwnd_limited = true;
break;
}
if (tso_segs == 1) {
if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
(tcp_skb_is_last(sk, skb) ?
nonagle : TCP_NAGLE_PUSH))))
break;
} else {
if (!push_one &&
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
&is_rwnd_limited, max_segs))
break;
}
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
limit = mss_now;
if (tso_segs > 1 && !tcp_urg_mode(tp))
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
limit = tcp_mss_split_point(sk, skb, mss_now,
min_t(unsigned int,
cwnd_quota,
max_segs),
nonagle);
if (skb->len > limit &&
unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
tcp: refine TSO autosizing Commit 95bd09eb2750 ("tcp: TSO packets automatic sizing") tried to control TSO size, but did this at the wrong place (sendmsg() time) At sendmsg() time, we might have a pessimistic view of flow rate, and we end up building very small skbs (with 2 MSS per skb). This is bad because : - It sends small TSO packets even in Slow Start where rate quickly increases. - It tends to make socket write queue very big, increasing tcp_ack() processing time, but also increasing memory needs, not necessarily accounted for, as fast clones overhead is currently ignored. - Lower GRO efficiency and more ACK packets. Servers with a lot of small lived connections suffer from this. Lets instead fill skbs as much as possible (64KB of payload), but split them at xmit time, when we have a precise idea of the flow rate. skb split is actually quite efficient. Patch looks bigger than necessary, because TCP Small Queue decision now has to take place after the eventual split. As Neal suggested, introduce a new tcp_tso_autosize() helper, so that tcp_tso_should_defer() can be synchronized on same goal. Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable contains number of mss that we can put in GSO packet, and is not related to the autosizing goal anymore. Tested: 40 ms rtt link nstat >/dev/null netperf -H remote -l -2000000 -- -s 1000000 nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets" Before patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/s 87380 2000000 2000000 0.36 44.22 IpInReceives 600 0.0 IpOutRequests 599 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2033249 0.0 After patch : Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 87380 2000000 2000000 0.36 44.27 IpInReceives 221 0.0 IpOutRequests 232 0.0 TcpOutSegs 1397 0.0 IpExtOutOctets 2013953 0.0 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-08 04:22:18 +08:00
break;
if (tcp_small_queue_check(sk, skb, 0))
break;
tcp: TSQ can use a dynamic limit When TCP Small Queues was added, we used a sysctl to limit amount of packets queues on Qdisc/device queues for a given TCP flow. Problem is this limit is either too big for low rates, or too small for high rates. Now TCP stack has rate estimation in sk->sk_pacing_rate, and TSO auto sizing, it can better control number of packets in Qdisc/device queues. New limit is two packets or at least 1 to 2 ms worth of packets. Low rates flows benefit from this patch by having even smaller number of packets in queues, allowing for faster recovery, better RTT estimations. High rates flows benefit from this patch by allowing more than 2 packets in flight as we had reports this was a limiting factor to reach line rate. [ In particular if TX completion is delayed because of coalescing parameters ] Example for a single flow on 10Gbp link controlled by FQ/pacing 14 packets in flight instead of 2 $ tc -s -d qd qdisc fq 8001: dev eth0 root refcnt 32 limit 10000p flow_limit 100p buckets 1024 quantum 3028 initial_quantum 15140 Sent 1168459366606 bytes 771822841 pkt (dropped 0, overlimits 0 requeues 6822476) rate 9346Mbit 771713pps backlog 953820b 14p requeues 6822476 2047 flow, 2046 inactive, 1 throttled, delay 15673 ns 2372 gc, 0 highprio, 0 retrans, 9739249 throttled, 0 flows_plimit Note that sk_pacing_rate is currently set to twice the actual rate, but this might be refined in the future when a flow is in congestion avoidance. Additional change : skb->destructor should be set to tcp_wfree(). A future patch (for linux 3.13+) might remove tcp_limit_output_bytes Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Wei Liu <wei.liu2@citrix.com> Cc: Cong Wang <xiyou.wangcong@gmail.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-27 18:28:54 +08:00
if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
break;
repair:
/* Advance the send_head. This one is sent out.
* This call will increment packets_out.
*/
tcp_event_new_data_sent(sk, skb);
tcp_minshall_update(tp, mss_now, skb);
Proportional Rate Reduction for TCP. This patch implements Proportional Rate Reduction (PRR) for TCP. PRR is an algorithm that determines TCP's sending rate in fast recovery. PRR avoids excessive window reductions and aims for the actual congestion window size at the end of recovery to be as close as possible to the window determined by the congestion control algorithm. PRR also improves accuracy of the amount of data sent during loss recovery. The patch implements the recommended flavor of PRR called PRR-SSRB (Proportional rate reduction with slow start reduction bound) and replaces the existing rate halving algorithm. PRR improves upon the existing Linux fast recovery under a number of conditions including: 1) burst losses where the losses implicitly reduce the amount of outstanding data (pipe) below the ssthresh value selected by the congestion control algorithm and, 2) losses near the end of short flows where application runs out of data to send. As an example, with the existing rate halving implementation a single loss event can cause a connection carrying short Web transactions to go into the slow start mode after the recovery. This is because during recovery Linux pulls the congestion window down to packets_in_flight+1 on every ACK. A short Web response often runs out of new data to send and its pipe reduces to zero by the end of recovery when all its packets are drained from the network. Subsequent HTTP responses using the same connection will have to slow start to raise cwnd to ssthresh. PRR on the other hand aims for the cwnd to be as close as possible to ssthresh by the end of recovery. A description of PRR and a discussion of its performance can be found at the following links: - IETF Draft: http://tools.ietf.org/html/draft-mathis-tcpm-proportional-rate-reduction-01 - IETF Slides: http://www.ietf.org/proceedings/80/slides/tcpm-6.pdf http://tools.ietf.org/agenda/81/slides/tcpm-2.pdf - Paper to appear in Internet Measurements Conference (IMC) 2011: Improving TCP Loss Recovery Nandita Dukkipati, Matt Mathis, Yuchung Cheng Signed-off-by: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-08-22 04:21:57 +08:00
sent_pkts += tcp_skb_pcount(skb);
if (push_one)
break;
}
if (is_rwnd_limited)
tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
else
tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
if (likely(sent_pkts)) {
if (tcp_in_cwnd_reduction(sk))
tp->prr_out += sent_pkts;
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
/* Send one loss probe per tail loss episode. */
if (push_one != 2)
tcp: when scheduling TLP, time of RTO should account for current ACK Fix the TLP scheduling logic so that when scheduling a TLP probe, we ensure that the estimated time at which an RTO would fire accounts for the fact that ACKs indicating forward progress should push back RTO times. After the following fix: df92c8394e6e ("tcp: fix xmit timer to only be reset if data ACKed/SACKed") we had an unintentional behavior change in the following kind of scenario: suppose the RTT variance has been very low recently. Then suppose we send out a flight of N packets and our RTT is 100ms: t=0: send a flight of N packets t=100ms: receive an ACK for N-1 packets The response before df92c8394e6e that was: -> schedule a TLP for now + RTO_interval The response after df92c8394e6e is: -> schedule a TLP for t=0 + RTO_interval Since RTO_interval = srtt + RTT_variance, this means that we have scheduled a TLP timer at a point in the future that only accounts for RTT_variance. If the RTT_variance term is small, this means that the timer fires soon. Before df92c8394e6e this would not happen, because in that code, when we receive an ACK for a prefix of flight, we did: 1) Near the top of tcp_ack(), switch from TLP timer to RTO at write_queue_head->paket_tx_time + RTO_interval: if (icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) tcp_rearm_rto(sk); 2) In tcp_clean_rtx_queue(), update the RTO to now + RTO_interval: if (flag & FLAG_ACKED) { tcp_rearm_rto(sk); 3) In tcp_ack() after tcp_fastretrans_alert() switch from RTO to TLP at now + RTO_interval: if (icsk->icsk_pending == ICSK_TIME_RETRANS) tcp_schedule_loss_probe(sk); In df92c8394e6e we removed that 3-phase dance, and instead directly set the TLP timer once: we set the TLP timer in cases like this to write_queue_head->packet_tx_time + RTO_interval. So if the RTT variance is small, then this means that this is setting the TLP timer to fire quite soon. This means if the ACK for the tail of the flight takes longer than an RTT to arrive (often due to delayed ACKs), then the TLP timer fires too quickly. Fixes: df92c8394e6e ("tcp: fix xmit timer to only be reset if data ACKed/SACKed") Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-11-18 10:06:14 +08:00
tcp_schedule_loss_probe(sk, false);
tcp: Fix CWV being too strict on thin streams Application limited streams such as thin streams, that transmit small amounts of payload in relatively few packets per RTT, can be prevented from growing the CWND when in congestion avoidance. This leads to increased sojourn times for data segments in streams that often transmit time-dependent data. Currently, a connection is considered CWND limited only after having successfully transmitted at least one packet with new data, while at the same time failing to transmit some unsent data from the output queue because the CWND is full. Applications that produce small amounts of data may be left in a state where it is never considered to be CWND limited, because all unsent data is successfully transmitted each time an incoming ACK opens up for more data to be transmitted in the send window. Fix by always testing whether the CWND is fully used after successful packet transmissions, such that a connection is considered CWND limited whenever the CWND has been filled. This is the correct behavior as specified in RFC2861 (section 3.1). Cc: Andreas Petlund <apetlund@simula.no> Cc: Carsten Griwodz <griff@simula.no> Cc: Jonas Markussen <jonassm@ifi.uio.no> Cc: Kenneth Klette Jonassen <kennetkl@ifi.uio.no> Cc: Mads Johannessen <madsjoh@ifi.uio.no> Signed-off-by: Bendik Rønning Opstad <bro.devel+kernel@gmail.com> Acked-by: Eric Dumazet <edumazet@google.com> Tested-by: Eric Dumazet <edumazet@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Tested-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-09-24 00:49:53 +08:00
is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
tcp_cwnd_validate(sk, is_cwnd_limited);
return false;
}
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
return !tp->packets_out && !tcp_write_queue_empty(sk);
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
}
tcp: when scheduling TLP, time of RTO should account for current ACK Fix the TLP scheduling logic so that when scheduling a TLP probe, we ensure that the estimated time at which an RTO would fire accounts for the fact that ACKs indicating forward progress should push back RTO times. After the following fix: df92c8394e6e ("tcp: fix xmit timer to only be reset if data ACKed/SACKed") we had an unintentional behavior change in the following kind of scenario: suppose the RTT variance has been very low recently. Then suppose we send out a flight of N packets and our RTT is 100ms: t=0: send a flight of N packets t=100ms: receive an ACK for N-1 packets The response before df92c8394e6e that was: -> schedule a TLP for now + RTO_interval The response after df92c8394e6e is: -> schedule a TLP for t=0 + RTO_interval Since RTO_interval = srtt + RTT_variance, this means that we have scheduled a TLP timer at a point in the future that only accounts for RTT_variance. If the RTT_variance term is small, this means that the timer fires soon. Before df92c8394e6e this would not happen, because in that code, when we receive an ACK for a prefix of flight, we did: 1) Near the top of tcp_ack(), switch from TLP timer to RTO at write_queue_head->paket_tx_time + RTO_interval: if (icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) tcp_rearm_rto(sk); 2) In tcp_clean_rtx_queue(), update the RTO to now + RTO_interval: if (flag & FLAG_ACKED) { tcp_rearm_rto(sk); 3) In tcp_ack() after tcp_fastretrans_alert() switch from RTO to TLP at now + RTO_interval: if (icsk->icsk_pending == ICSK_TIME_RETRANS) tcp_schedule_loss_probe(sk); In df92c8394e6e we removed that 3-phase dance, and instead directly set the TLP timer once: we set the TLP timer in cases like this to write_queue_head->packet_tx_time + RTO_interval. So if the RTT variance is small, then this means that this is setting the TLP timer to fire quite soon. This means if the ACK for the tail of the flight takes longer than an RTT to arrive (often due to delayed ACKs), then the TLP timer fires too quickly. Fixes: df92c8394e6e ("tcp: fix xmit timer to only be reset if data ACKed/SACKed") Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-11-18 10:06:14 +08:00
bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
{
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
u32 timeout, rto_delta_us;
int early_retrans;
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
/* Don't do any loss probe on a Fast Open connection before 3WHS
* finishes.
*/
if (rcu_access_pointer(tp->fastopen_rsk))
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
return false;
early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
/* Schedule a loss probe in 2*RTT for SACK capable connections
tcp: allow TLP in ECN CWR This patch enables tail loss probe in cwnd reduction (CWR) state to detect potential losses. Prior to this patch, since the sender uses PRR to determine the cwnd in CWR state, the combination of CWR+PRR plus tcp_tso_should_defer() could cause unnecessary stalls upon losses: PRR makes cwnd so gentle that tcp_tso_should_defer() defers sending wait for more ACKs. The ACKs may not come due to packet losses. Disallowing TLP when there is unused cwnd had the primary effect of disallowing TLP when there is TSO deferral, Nagle deferral, or we hit the rwin limit. Because basically every application write() or incoming ACK will cause us to run tcp_write_xmit() to see if we can send more, and then if we sent something we call tcp_schedule_loss_probe() to see if we should schedule a TLP. At that point, there are a few common reasons why some cwnd budget could still be unused: (a) rwin limit (b) nagle check (c) TSO deferral (d) TSQ For (d), after the next packet tx completion the TSQ mechanism will allow us to send more packets, so we don't really need a TLP (in practice it shouldn't matter whether we schedule one or not). But for (a), (b), (c) the sender won't send any more packets until it gets another ACK. But if the whole flight was lost, or all the ACKs were lost, then we won't get any more ACKs, and ideally we should schedule and send a TLP to get more feedback. In particular for a long time we have wanted some kind of timer for TSO deferral, and at least this would give us some kind of timer Reported-by: Steve Ibanez <sibanez@stanford.edu> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Reviewed-by: Nandita Dukkipati <nanditad@google.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-12-12 07:42:53 +08:00
* not in loss recovery, that are either limited by cwnd or application.
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
*/
if ((early_retrans != 3 && early_retrans != 4) ||
!tp->packets_out || !tcp_is_sack(tp) ||
tcp: allow TLP in ECN CWR This patch enables tail loss probe in cwnd reduction (CWR) state to detect potential losses. Prior to this patch, since the sender uses PRR to determine the cwnd in CWR state, the combination of CWR+PRR plus tcp_tso_should_defer() could cause unnecessary stalls upon losses: PRR makes cwnd so gentle that tcp_tso_should_defer() defers sending wait for more ACKs. The ACKs may not come due to packet losses. Disallowing TLP when there is unused cwnd had the primary effect of disallowing TLP when there is TSO deferral, Nagle deferral, or we hit the rwin limit. Because basically every application write() or incoming ACK will cause us to run tcp_write_xmit() to see if we can send more, and then if we sent something we call tcp_schedule_loss_probe() to see if we should schedule a TLP. At that point, there are a few common reasons why some cwnd budget could still be unused: (a) rwin limit (b) nagle check (c) TSO deferral (d) TSQ For (d), after the next packet tx completion the TSQ mechanism will allow us to send more packets, so we don't really need a TLP (in practice it shouldn't matter whether we schedule one or not). But for (a), (b), (c) the sender won't send any more packets until it gets another ACK. But if the whole flight was lost, or all the ACKs were lost, then we won't get any more ACKs, and ideally we should schedule and send a TLP to get more feedback. In particular for a long time we have wanted some kind of timer for TSO deferral, and at least this would give us some kind of timer Reported-by: Steve Ibanez <sibanez@stanford.edu> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Reviewed-by: Nandita Dukkipati <nanditad@google.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-12-12 07:42:53 +08:00
(icsk->icsk_ca_state != TCP_CA_Open &&
icsk->icsk_ca_state != TCP_CA_CWR))
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
return false;
/* Probe timeout is 2*rtt. Add minimum RTO to account
* for delayed ack when there's one outstanding packet. If no RTT
* sample is available then probe after TCP_TIMEOUT_INIT.
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
*/
if (tp->srtt_us) {
timeout = usecs_to_jiffies(tp->srtt_us >> 2);
if (tp->packets_out == 1)
timeout += TCP_RTO_MIN;
else
timeout += TCP_TIMEOUT_MIN;
} else {
timeout = TCP_TIMEOUT_INIT;
}
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
/* If the RTO formula yields an earlier time, then use that time. */
tcp: when scheduling TLP, time of RTO should account for current ACK Fix the TLP scheduling logic so that when scheduling a TLP probe, we ensure that the estimated time at which an RTO would fire accounts for the fact that ACKs indicating forward progress should push back RTO times. After the following fix: df92c8394e6e ("tcp: fix xmit timer to only be reset if data ACKed/SACKed") we had an unintentional behavior change in the following kind of scenario: suppose the RTT variance has been very low recently. Then suppose we send out a flight of N packets and our RTT is 100ms: t=0: send a flight of N packets t=100ms: receive an ACK for N-1 packets The response before df92c8394e6e that was: -> schedule a TLP for now + RTO_interval The response after df92c8394e6e is: -> schedule a TLP for t=0 + RTO_interval Since RTO_interval = srtt + RTT_variance, this means that we have scheduled a TLP timer at a point in the future that only accounts for RTT_variance. If the RTT_variance term is small, this means that the timer fires soon. Before df92c8394e6e this would not happen, because in that code, when we receive an ACK for a prefix of flight, we did: 1) Near the top of tcp_ack(), switch from TLP timer to RTO at write_queue_head->paket_tx_time + RTO_interval: if (icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) tcp_rearm_rto(sk); 2) In tcp_clean_rtx_queue(), update the RTO to now + RTO_interval: if (flag & FLAG_ACKED) { tcp_rearm_rto(sk); 3) In tcp_ack() after tcp_fastretrans_alert() switch from RTO to TLP at now + RTO_interval: if (icsk->icsk_pending == ICSK_TIME_RETRANS) tcp_schedule_loss_probe(sk); In df92c8394e6e we removed that 3-phase dance, and instead directly set the TLP timer once: we set the TLP timer in cases like this to write_queue_head->packet_tx_time + RTO_interval. So if the RTT variance is small, then this means that this is setting the TLP timer to fire quite soon. This means if the ACK for the tail of the flight takes longer than an RTT to arrive (often due to delayed ACKs), then the TLP timer fires too quickly. Fixes: df92c8394e6e ("tcp: fix xmit timer to only be reset if data ACKed/SACKed") Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-11-18 10:06:14 +08:00
rto_delta_us = advancing_rto ?
jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
tcp_rto_delta_us(sk); /* How far in future is RTO? */
if (rto_delta_us > 0)
timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
TCP_RTO_MAX, NULL);
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
return true;
}
/* Thanks to skb fast clones, we can detect if a prior transmit of
* a packet is still in a qdisc or driver queue.
* In this case, there is very little point doing a retransmit !
*/
static bool skb_still_in_host_queue(const struct sock *sk,
const struct sk_buff *skb)
{
if (unlikely(skb_fclone_busy(sk, skb))) {
NET_INC_STATS(sock_net(sk),
LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
return true;
}
return false;
}
/* When probe timeout (PTO) fires, try send a new segment if possible, else
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
* retransmit the last segment.
*/
void tcp_send_loss_probe(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
struct sk_buff *skb;
int pcount;
int mss = tcp_current_mss(sk);
skb = tcp_send_head(sk);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
pcount = tp->packets_out;
tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
if (tp->packets_out > pcount)
goto probe_sent;
goto rearm_timer;
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
}
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
skb = skb_rb_last(&sk->tcp_rtx_queue);
if (unlikely(!skb)) {
WARN_ONCE(tp->packets_out,
"invalid inflight: %u state %u cwnd %u mss %d\n",
tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
inet_csk(sk)->icsk_pending = 0;
return;
}
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
/* At most one outstanding TLP retransmission. */
if (tp->tlp_high_seq)
goto rearm_timer;
if (skb_still_in_host_queue(sk, skb))
goto rearm_timer;
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
pcount = tcp_skb_pcount(skb);
if (WARN_ON(!pcount))
goto rearm_timer;
if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
(pcount - 1) * mss, mss,
GFP_ATOMIC)))
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
goto rearm_timer;
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
skb = skb_rb_next(skb);
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
}
if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
goto rearm_timer;
if (__tcp_retransmit_skb(sk, skb, 1))
goto rearm_timer;
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
/* Record snd_nxt for loss detection. */
tp->tlp_high_seq = tp->snd_nxt;
probe_sent:
NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
/* Reset s.t. tcp_rearm_rto will restart timer from now */
inet_csk(sk)->icsk_pending = 0;
tcp: Tail loss probe (TLP) This patch series implement the Tail loss probe (TLP) algorithm described in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The first patch implements the basic algorithm. TLP's goal is to reduce tail latency of short transactions. It achieves this by converting retransmission timeouts (RTOs) occuring due to tail losses (losses at end of transactions) into fast recovery. TLP transmits one packet in two round-trips when a connection is in Open state and isn't receiving any ACKs. The transmitted packet, aka loss probe, can be either new or a retransmission. When there is tail loss, the ACK from a loss probe triggers FACK/early-retransmit based fast recovery, thus avoiding a costly RTO. In the absence of loss, there is no change in the connection state. PTO stands for probe timeout. It is a timer event indicating that an ACK is overdue and triggers a loss probe packet. The PTO value is set to max(2*SRTT, 10ms) and is adjusted to account for delayed ACK timer when there is only one oustanding packet. TLP Algorithm On transmission of new data in Open state: -> packets_out > 1: schedule PTO in max(2*SRTT, 10ms). -> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms) -> PTO = min(PTO, RTO) Conditions for scheduling PTO: -> Connection is in Open state. -> Connection is either cwnd limited or no new data to send. -> Number of probes per tail loss episode is limited to one. -> Connection is SACK enabled. When PTO fires: new_segment_exists: -> transmit new segment. -> packets_out++. cwnd remains same. no_new_packet: -> retransmit the last segment. Its ACK triggers FACK or early retransmit based recovery. ACK path: -> rearm RTO at start of ACK processing. -> reschedule PTO if need be. In addition, the patch includes a small variation to the Early Retransmit (ER) algorithm, such that ER and TLP together can in principle recover any N-degree of tail loss through fast recovery. TLP is controlled by the same sysctl as ER, tcp_early_retrans sysctl. tcp_early_retrans==0; disables TLP and ER. ==1; enables RFC5827 ER. ==2; delayed ER. ==3; TLP and delayed ER. [DEFAULT] ==4; TLP only. The TLP patch series have been extensively tested on Google Web servers. It is most effective for short Web trasactions, where it reduced RTOs by 15% and improved HTTP response time (average by 6%, 99th percentile by 10%). The transmitted probes account for <0.5% of the overall transmissions. Signed-off-by: Nandita Dukkipati <nanditad@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-11 18:00:43 +08:00
rearm_timer:
tcp_rearm_rto(sk);
}
/* Push out any pending frames which were held back due to
* TCP_CORK or attempt at coalescing tiny packets.
* The socket must be locked by the caller.
*/
[TCP]: Sed magic converts func(sk, tp, ...) -> func(sk, ...) This is (mostly) automated change using magic: sed -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e '/struct sock \*sk/ N' -e 's|struct sock \*sk,[\n\t ]*struct tcp_sock \*tp\([^{]*\n{\n\)| struct sock \*sk\1\tstruct tcp_sock *tp = tcp_sk(sk);\n|g' -e 's|struct sock \*sk, struct tcp_sock \*tp| struct sock \*sk|g' -e 's|sk, tp\([^-]\)|sk\1|g' Fixed four unused variable (tp) warnings that were introduced. In addition, manually added newlines after local variables and tweaked function arguments positioning. $ gcc --version gcc (GCC) 4.1.1 20060525 (Red Hat 4.1.1-1) ... $ codiff -fV built-in.o.old built-in.o.new net/ipv4/route.c: rt_cache_flush | +14 1 function changed, 14 bytes added net/ipv4/tcp.c: tcp_setsockopt | -5 tcp_sendpage | -25 tcp_sendmsg | -16 3 functions changed, 46 bytes removed net/ipv4/tcp_input.c: tcp_try_undo_recovery | +3 tcp_try_undo_dsack | +2 tcp_mark_head_lost | -12 tcp_ack | -15 tcp_event_data_recv | -32 tcp_rcv_state_process | -10 tcp_rcv_established | +1 7 functions changed, 6 bytes added, 69 bytes removed, diff: -63 net/ipv4/tcp_output.c: update_send_head | -9 tcp_transmit_skb | +19 tcp_cwnd_validate | +1 tcp_write_wakeup | -17 __tcp_push_pending_frames | -25 tcp_push_one | -8 tcp_send_fin | -4 7 functions changed, 20 bytes added, 63 bytes removed, diff: -43 built-in.o.new: 18 functions changed, 40 bytes added, 178 bytes removed, diff: -138 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-04-21 13:18:02 +08:00
void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
int nonagle)
{
/* If we are closed, the bytes will have to remain here.
* In time closedown will finish, we empty the write queue and
* all will be happy.
*/
if (unlikely(sk->sk_state == TCP_CLOSE))
return;
if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
sk_gfp_mask(sk, GFP_ATOMIC)))
tcp_check_probe_timer(sk);
}
/* Send _single_ skb sitting at the send head. This function requires
* true push pending frames to setup probe timer etc.
*/
void tcp_push_one(struct sock *sk, unsigned int mss_now)
{
struct sk_buff *skb = tcp_send_head(sk);
BUG_ON(!skb || skb->len < mss_now);
tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
}
/* This function returns the amount that we can raise the
* usable window based on the following constraints
*
* 1. The window can never be shrunk once it is offered (RFC 793)
* 2. We limit memory per socket
*
* RFC 1122:
* "the suggested [SWS] avoidance algorithm for the receiver is to keep
* RECV.NEXT + RCV.WIN fixed until:
* RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
*
* i.e. don't raise the right edge of the window until you can raise
* it at least MSS bytes.
*
* Unfortunately, the recommended algorithm breaks header prediction,
* since header prediction assumes th->window stays fixed.
*
* Strictly speaking, keeping th->window fixed violates the receiver
* side SWS prevention criteria. The problem is that under this rule
* a stream of single byte packets will cause the right side of the
* window to always advance by a single byte.
*
* Of course, if the sender implements sender side SWS prevention
* then this will not be a problem.
*
* BSD seems to make the following compromise:
*
* If the free space is less than the 1/4 of the maximum
* space available and the free space is less than 1/2 mss,
* then set the window to 0.
* [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
* Otherwise, just prevent the window from shrinking
* and from being larger than the largest representable value.
*
* This prevents incremental opening of the window in the regime
* where TCP is limited by the speed of the reader side taking
* data out of the TCP receive queue. It does nothing about
* those cases where the window is constrained on the sender side
* because the pipeline is full.
*
* BSD also seems to "accidentally" limit itself to windows that are a
* multiple of MSS, at least until the free space gets quite small.
* This would appear to be a side effect of the mbuf implementation.
* Combining these two algorithms results in the observed behavior
* of having a fixed window size at almost all times.
*
* Below we obtain similar behavior by forcing the offered window to
* a multiple of the mss when it is feasible to do so.
*
* Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
* Regular options like TIMESTAMP are taken into account.
*/
u32 __tcp_select_window(struct sock *sk)
{
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
/* MSS for the peer's data. Previous versions used mss_clamp
* here. I don't know if the value based on our guesses
* of peer's MSS is better for the performance. It's more correct
* but may be worse for the performance because of rcv_mss
* fluctuations. --SAW 1998/11/1
*/
int mss = icsk->icsk_ack.rcv_mss;
int free_space = tcp_space(sk);
tcp: use zero-window when free_space is low Currently the kernel tries to announce a zero window when free_space is below the current receiver mss estimate. When a sender is transmitting small packets and reader consumes data slowly (or not at all), receiver might be unable to shrink the receive win because a) we cannot withdraw already-commited receive window, and, b) we have to round the current rwin up to a multiple of the wscale factor, else we would shrink the current window. This causes the receive buffer to fill up until the rmem limit is hit. When this happens, we start dropping packets. Moreover, tcp_clamp_window may continue to grow sk_rcvbuf towards rmem[2] even if socket is not being read from. As we cannot avoid the "current_win is rounded up to multiple of mss" issue [we would violate a) above] at least try to prevent the receive buf growth towards tcp_rmem[2] limit by attempting to move to zero-window announcement when free_space becomes less than 1/16 of the current allowed receive buffer maximum. If tcp_rmem[2] is large, this will increase our chances to get a zero-window announcement out in time. Reproducer: On server: $ nc -l -p 12345 <suspend it: CTRL-Z> Client: #!/usr/bin/env python import socket import time sock = socket.socket() sock.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1) sock.connect(("192.168.4.1", 12345)); while True: sock.send('A' * 23) time.sleep(0.005) socket buffer on server-side will grow until tcp_rmem[2] is hit, at which point the client rexmits data until -EDTIMEOUT: tcp_data_queue invokes tcp_try_rmem_schedule which will call tcp_prune_queue which calls tcp_clamp_window(). And that function will grow sk->sk_rcvbuf up until it eventually hits tcp_rmem[2]. Thanks to Eric Dumazet for running regression tests. Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Tested-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-02-19 19:51:10 +08:00
int allowed_space = tcp_full_space(sk);
int full_space = min_t(int, tp->window_clamp, allowed_space);
int window;
if (unlikely(mss > full_space)) {
mss = full_space;
if (mss <= 0)
return 0;
}
if (free_space < (full_space >> 1)) {
icsk->icsk_ack.quick = 0;
if (tcp_under_memory_pressure(sk))
tp->rcv_ssthresh = min(tp->rcv_ssthresh,
4U * tp->advmss);
tcp: use zero-window when free_space is low Currently the kernel tries to announce a zero window when free_space is below the current receiver mss estimate. When a sender is transmitting small packets and reader consumes data slowly (or not at all), receiver might be unable to shrink the receive win because a) we cannot withdraw already-commited receive window, and, b) we have to round the current rwin up to a multiple of the wscale factor, else we would shrink the current window. This causes the receive buffer to fill up until the rmem limit is hit. When this happens, we start dropping packets. Moreover, tcp_clamp_window may continue to grow sk_rcvbuf towards rmem[2] even if socket is not being read from. As we cannot avoid the "current_win is rounded up to multiple of mss" issue [we would violate a) above] at least try to prevent the receive buf growth towards tcp_rmem[2] limit by attempting to move to zero-window announcement when free_space becomes less than 1/16 of the current allowed receive buffer maximum. If tcp_rmem[2] is large, this will increase our chances to get a zero-window announcement out in time. Reproducer: On server: $ nc -l -p 12345 <suspend it: CTRL-Z> Client: #!/usr/bin/env python import socket import time sock = socket.socket() sock.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1) sock.connect(("192.168.4.1", 12345)); while True: sock.send('A' * 23) time.sleep(0.005) socket buffer on server-side will grow until tcp_rmem[2] is hit, at which point the client rexmits data until -EDTIMEOUT: tcp_data_queue invokes tcp_try_rmem_schedule which will call tcp_prune_queue which calls tcp_clamp_window(). And that function will grow sk->sk_rcvbuf up until it eventually hits tcp_rmem[2]. Thanks to Eric Dumazet for running regression tests. Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Tested-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-02-19 19:51:10 +08:00
/* free_space might become our new window, make sure we don't
* increase it due to wscale.
*/
free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
/* if free space is less than mss estimate, or is below 1/16th
* of the maximum allowed, try to move to zero-window, else
* tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
* new incoming data is dropped due to memory limits.
* With large window, mss test triggers way too late in order
* to announce zero window in time before rmem limit kicks in.
*/
if (free_space < (allowed_space >> 4) || free_space < mss)
return 0;
}
if (free_space > tp->rcv_ssthresh)
free_space = tp->rcv_ssthresh;
/* Don't do rounding if we are using window scaling, since the
* scaled window will not line up with the MSS boundary anyway.
*/
if (tp->rx_opt.rcv_wscale) {
window = free_space;
/* Advertise enough space so that it won't get scaled away.
* Import case: prevent zero window announcement if
* 1<<rcv_wscale > mss.
*/
window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
} else {
window = tp->rcv_wnd;
/* Get the largest window that is a nice multiple of mss.
* Window clamp already applied above.
* If our current window offering is within 1 mss of the
* free space we just keep it. This prevents the divide
* and multiply from happening most of the time.
* We also don't do any window rounding when the free space
* is too small.
*/
if (window <= free_space - mss || window > free_space)
window = rounddown(free_space, mss);
else if (mss == full_space &&
free_space > window + (full_space >> 1))
window = free_space;
}
return window;
}
tcp: Merge tx_flags and tskey in tcp_shifted_skb After receiving sacks, tcp_shifted_skb() will collapse skbs if possible. tx_flags and tskey also have to be merged. This patch reuses the tcp_skb_collapse_tstamp() to handle them. BPF Output Before: ~~~~~ <no-output-due-to-missing-tstamp-event> BPF Output After: ~~~~~ <...>-2024 [007] d.s. 88.644374: : ee_data:14599 Packetdrill Script: ~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 write(4, ..., 1460) = 1460 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 13140) = 13140 0.200 > P. 1:1461(1460) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:14601(5840) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:14601,nop,nop> 0.300 > P. 1:1461(1460) ack 1 0.400 < . 1:1(0) ack 14601 win 257 0.400 close(4) = 0 0.400 > F. 14601:14601(0) ack 1 0.500 < F. 1:1(0) ack 14602 win 257 0.500 > . 14602:14602(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Tested-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-20 13:39:29 +08:00
void tcp_skb_collapse_tstamp(struct sk_buff *skb,
const struct sk_buff *next_skb)
tcp: Merge tx_flags and tskey in tcp_collapse_retrans If two skbs are merged/collapsed during retransmission, the current logic does not merge the tx_flags and tskey. The end result is the SCM_TSTAMP_ACK timestamp could be missing for a packet. The patch: 1. Merge the tx_flags 2. Overwrite the prev_skb's tskey with the next_skb's tskey BPF Output Before: ~~~~~~ <no-output-due-to-missing-tstamp-event> BPF Output After: ~~~~~~ packetdrill-2092 [001] d.s. 453.998486: : ee_data:1459 Packetdrill Script: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0 0.200 write(4, ..., 11680) = 11680 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 > P. 1:731(730) ack 1 0.200 > P. 731:1461(730) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:13141(4380) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:2921,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:4381,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:5841,nop,nop> 0.300 > P. 1:1461(1460) ack 1 0.400 < . 1:1(0) ack 13141 win 257 0.400 close(4) = 0 0.400 > F. 13141:13141(0) ack 1 0.500 < F. 1:1(0) ack 13142 win 257 0.500 > . 13142:13142(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Tested-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-20 13:39:28 +08:00
{
if (unlikely(tcp_has_tx_tstamp(next_skb))) {
const struct skb_shared_info *next_shinfo =
skb_shinfo(next_skb);
tcp: Merge tx_flags and tskey in tcp_collapse_retrans If two skbs are merged/collapsed during retransmission, the current logic does not merge the tx_flags and tskey. The end result is the SCM_TSTAMP_ACK timestamp could be missing for a packet. The patch: 1. Merge the tx_flags 2. Overwrite the prev_skb's tskey with the next_skb's tskey BPF Output Before: ~~~~~~ <no-output-due-to-missing-tstamp-event> BPF Output After: ~~~~~~ packetdrill-2092 [001] d.s. 453.998486: : ee_data:1459 Packetdrill Script: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0 0.200 write(4, ..., 11680) = 11680 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 > P. 1:731(730) ack 1 0.200 > P. 731:1461(730) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:13141(4380) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:2921,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:4381,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:5841,nop,nop> 0.300 > P. 1:1461(1460) ack 1 0.400 < . 1:1(0) ack 13141 win 257 0.400 close(4) = 0 0.400 > F. 13141:13141(0) ack 1 0.500 < F. 1:1(0) ack 13142 win 257 0.500 > . 13142:13142(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Tested-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-20 13:39:28 +08:00
struct skb_shared_info *shinfo = skb_shinfo(skb);
shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
tcp: Merge tx_flags and tskey in tcp_collapse_retrans If two skbs are merged/collapsed during retransmission, the current logic does not merge the tx_flags and tskey. The end result is the SCM_TSTAMP_ACK timestamp could be missing for a packet. The patch: 1. Merge the tx_flags 2. Overwrite the prev_skb's tskey with the next_skb's tskey BPF Output Before: ~~~~~~ <no-output-due-to-missing-tstamp-event> BPF Output After: ~~~~~~ packetdrill-2092 [001] d.s. 453.998486: : ee_data:1459 Packetdrill Script: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0 0.200 write(4, ..., 11680) = 11680 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 > P. 1:731(730) ack 1 0.200 > P. 731:1461(730) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:13141(4380) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:2921,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:4381,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:5841,nop,nop> 0.300 > P. 1:1461(1460) ack 1 0.400 < . 1:1(0) ack 13141 win 257 0.400 close(4) = 0 0.400 > F. 13141:13141(0) ack 1 0.500 < F. 1:1(0) ack 13142 win 257 0.500 > . 13142:13142(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Tested-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-20 13:39:28 +08:00
shinfo->tskey = next_shinfo->tskey;
tcp: Merge txstamp_ack in tcp_skb_collapse_tstamp When collapsing skbs, txstamp_ack also needs to be merged. Retrans Collapse Test: ~~~~~~ 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0 0.200 write(4, ..., 11680) = 11680 0.200 > P. 1:731(730) ack 1 0.200 > P. 731:1461(730) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:13141(4380) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:2921,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:4381,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:5841,nop,nop> 0.300 > P. 1:1461(1460) ack 1 0.400 < . 1:1(0) ack 13141 win 257 BPF Output Before: ~~~~~ <No output due to missing SCM_TSTAMP_ACK timestamp> BPF Output After: ~~~~~ <...>-2027 [007] d.s. 79.765921: : ee_data:1459 Sacks Collapse Test: ~~~~~ 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 write(4, ..., 1460) = 1460 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 13140) = 13140 +0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0 0.200 > P. 1:1461(1460) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:14601(5840) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:14601,nop,nop> 0.300 > P. 1:1461(1460) ack 1 0.400 < . 1:1(0) ack 14601 win 257 BPF Output Before: ~~~~~ <No output due to missing SCM_TSTAMP_ACK timestamp> BPF Output After: ~~~~~ <...>-2049 [007] d.s. 89.185538: : ee_data:14599 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Tested-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-20 13:50:48 +08:00
TCP_SKB_CB(skb)->txstamp_ack |=
TCP_SKB_CB(next_skb)->txstamp_ack;
tcp: Merge tx_flags and tskey in tcp_collapse_retrans If two skbs are merged/collapsed during retransmission, the current logic does not merge the tx_flags and tskey. The end result is the SCM_TSTAMP_ACK timestamp could be missing for a packet. The patch: 1. Merge the tx_flags 2. Overwrite the prev_skb's tskey with the next_skb's tskey BPF Output Before: ~~~~~~ <no-output-due-to-missing-tstamp-event> BPF Output After: ~~~~~~ packetdrill-2092 [001] d.s. 453.998486: : ee_data:1459 Packetdrill Script: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0 0.200 write(4, ..., 11680) = 11680 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 > P. 1:731(730) ack 1 0.200 > P. 731:1461(730) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:13141(4380) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:2921,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:4381,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:5841,nop,nop> 0.300 > P. 1:1461(1460) ack 1 0.400 < . 1:1(0) ack 13141 win 257 0.400 close(4) = 0 0.400 > F. 13141:13141(0) ack 1 0.500 < F. 1:1(0) ack 13142 win 257 0.500 > . 13142:13142(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Tested-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-20 13:39:28 +08:00
}
}
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
/* Collapses two adjacent SKB's during retransmission. */
tcp: enhance tcp_collapse_retrans() with skb_shift() In commit 2331ccc5b323 ("tcp: enhance tcp collapsing"), we made a first step allowing copying right skb to left skb head. Since all skbs in socket write queue are headless (but possibly the very first one), this strategy often does not work. This patch extends tcp_collapse_retrans() to perform frag shifting, thanks to skb_shift() helper. This helper needs to not BUG on non headless skbs, as callers are ok with that. Tested: Following packetdrill test now passes : 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 8> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8> +.100 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 +0 write(4, ..., 200) = 200 +0 > P. 1:201(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 201:401(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 401:601(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 601:801(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 801:1001(200) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1001:1101(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1101:1201(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1201:1301(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1301:1401(100) ack 1 +.099 < . 1:1(0) ack 201 win 257 +.001 < . 1:1(0) ack 201 win 257 <nop,nop,sack 1001:1401> +0 > P. 201:1001(800) ack 1 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-16 04:51:50 +08:00
static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
{
struct tcp_sock *tp = tcp_sk(sk);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
struct sk_buff *next_skb = skb_rb_next(skb);
int next_skb_size;
next_skb_size = next_skb->len;
BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
tcp: enhance tcp_collapse_retrans() with skb_shift() In commit 2331ccc5b323 ("tcp: enhance tcp collapsing"), we made a first step allowing copying right skb to left skb head. Since all skbs in socket write queue are headless (but possibly the very first one), this strategy often does not work. This patch extends tcp_collapse_retrans() to perform frag shifting, thanks to skb_shift() helper. This helper needs to not BUG on non headless skbs, as callers are ok with that. Tested: Following packetdrill test now passes : 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 8> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8> +.100 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 +0 write(4, ..., 200) = 200 +0 > P. 1:201(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 201:401(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 401:601(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 601:801(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 801:1001(200) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1001:1101(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1101:1201(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1201:1301(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1301:1401(100) ack 1 +.099 < . 1:1(0) ack 201 win 257 +.001 < . 1:1(0) ack 201 win 257 <nop,nop,sack 1001:1401> +0 > P. 201:1001(800) ack 1 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-16 04:51:50 +08:00
if (next_skb_size) {
if (next_skb_size <= skb_availroom(skb))
skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
next_skb_size);
else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
tcp: enhance tcp_collapse_retrans() with skb_shift() In commit 2331ccc5b323 ("tcp: enhance tcp collapsing"), we made a first step allowing copying right skb to left skb head. Since all skbs in socket write queue are headless (but possibly the very first one), this strategy often does not work. This patch extends tcp_collapse_retrans() to perform frag shifting, thanks to skb_shift() helper. This helper needs to not BUG on non headless skbs, as callers are ok with that. Tested: Following packetdrill test now passes : 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 8> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8> +.100 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 +0 write(4, ..., 200) = 200 +0 > P. 1:201(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 201:401(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 401:601(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 601:801(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 801:1001(200) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1001:1101(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1101:1201(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1201:1301(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1301:1401(100) ack 1 +.099 < . 1:1(0) ack 201 win 257 +.001 < . 1:1(0) ack 201 win 257 <nop,nop,sack 1001:1401> +0 > P. 201:1001(800) ack 1 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-16 04:51:50 +08:00
return false;
}
tcp_highest_sack_replace(sk, next_skb, skb);
/* Update sequence range on original skb. */
TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
/* Merge over control information. This moves PSH/FIN etc. over */
TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
/* All done, get rid of second SKB and account for it so
* packet counting does not break.
*/
TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
tcp: Handle eor bit when coalescing skb This patch: 1. Prevent next_skb from coalescing to the prev_skb if TCP_SKB_CB(prev_skb)->eor is set 2. Update the TCP_SKB_CB(prev_skb)->eor if coalescing is allowed Packetdrill script for testing: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 sendto(4, ..., 730, MSG_EOR, ..., ...) = 730 0.200 sendto(4, ..., 730, MSG_EOR, ..., ...) = 730 0.200 write(4, ..., 11680) = 11680 0.200 > P. 1:731(730) ack 1 0.200 > P. 731:1461(730) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:13141(4380) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:13141,nop,nop> 0.300 > P. 1:731(730) ack 1 0.300 > P. 731:1461(730) ack 1 0.400 < . 1:1(0) ack 13141 win 257 0.400 close(4) = 0 0.400 > F. 13141:13141(0) ack 1 0.500 < F. 1:1(0) ack 13142 win 257 0.500 > . 13142:13142(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-26 05:44:49 +08:00
TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
/* changed transmit queue under us so clear hints */
tcp_clear_retrans_hints_partial(tp);
if (next_skb == tp->retransmit_skb_hint)
tp->retransmit_skb_hint = skb;
tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
tcp: Merge tx_flags and tskey in tcp_collapse_retrans If two skbs are merged/collapsed during retransmission, the current logic does not merge the tx_flags and tskey. The end result is the SCM_TSTAMP_ACK timestamp could be missing for a packet. The patch: 1. Merge the tx_flags 2. Overwrite the prev_skb's tskey with the next_skb's tskey BPF Output Before: ~~~~~~ <no-output-due-to-missing-tstamp-event> BPF Output After: ~~~~~~ packetdrill-2092 [001] d.s. 453.998486: : ee_data:1459 Packetdrill Script: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 write(4, ..., 730) = 730 +0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0 0.200 write(4, ..., 11680) = 11680 +0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0 0.200 > P. 1:731(730) ack 1 0.200 > P. 731:1461(730) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:13141(4380) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:2921,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:4381,nop,nop> 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:5841,nop,nop> 0.300 > P. 1:1461(1460) ack 1 0.400 < . 1:1(0) ack 13141 win 257 0.400 close(4) = 0 0.400 > F. 13141:13141(0) ack 1 0.500 < F. 1:1(0) ack 13142 win 257 0.500 > . 13142:13142(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Tested-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-20 13:39:28 +08:00
tcp_skb_collapse_tstamp(skb, next_skb);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
tcp_rtx_queue_unlink_and_free(next_skb, sk);
tcp: enhance tcp_collapse_retrans() with skb_shift() In commit 2331ccc5b323 ("tcp: enhance tcp collapsing"), we made a first step allowing copying right skb to left skb head. Since all skbs in socket write queue are headless (but possibly the very first one), this strategy often does not work. This patch extends tcp_collapse_retrans() to perform frag shifting, thanks to skb_shift() helper. This helper needs to not BUG on non headless skbs, as callers are ok with that. Tested: Following packetdrill test now passes : 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 8> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8> +.100 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 +0 write(4, ..., 200) = 200 +0 > P. 1:201(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 201:401(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 401:601(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 601:801(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 801:1001(200) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1001:1101(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1101:1201(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1201:1301(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1301:1401(100) ack 1 +.099 < . 1:1(0) ack 201 win 257 +.001 < . 1:1(0) ack 201 win 257 <nop,nop,sack 1001:1401> +0 > P. 201:1001(800) ack 1 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-16 04:51:50 +08:00
return true;
}
/* Check if coalescing SKBs is legal. */
static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
{
if (tcp_skb_pcount(skb) > 1)
return false;
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
if (skb_cloned(skb))
return false;
tcp: enhance tcp collapsing As Ilya Lesokhin suggested, we can collapse two skbs at retransmit time even if the skb at the right has fragments. We simply have to use more generic skb_copy_bits() instead of skb_copy_from_linear_data() in tcp_collapse_retrans() Also need to guard this skb_copy_bits() in case there is nothing to copy, otherwise skb_put() could panic if left skb has frags. Tested: Used following packetdrill test // Establish a connection. 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 8> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8> +.100 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 +0 write(4, ..., 200) = 200 +0 > P. 1:201(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 201:401(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 401:601(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 601:801(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 801:1001(200) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1001:1101(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1101:1201(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1201:1301(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1301:1401(100) ack 1 +.100 < . 1:1(0) ack 1 win 257 <nop,nop,sack 1001:1401> // Check that TCP collapse works : +0 > P. 1:1001(1000) ack 1 Reported-by: Ilya Lesokhin <ilyal@mellanox.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-02 01:53:42 +08:00
/* Some heuristics for collapsing over SACK'd could be invented */
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
return false;
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
return true;
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
}
/* Collapse packets in the retransmit queue to make to create
* less packets on the wire. This is only done on retransmission.
*/
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
int space)
{
struct tcp_sock *tp = tcp_sk(sk);
struct sk_buff *skb = to, *tmp;
bool first = true;
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
return;
if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
return;
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
skb_rbtree_walk_from_safe(skb, tmp) {
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
if (!tcp_can_collapse(sk, skb))
break;
tcp: Handle eor bit when coalescing skb This patch: 1. Prevent next_skb from coalescing to the prev_skb if TCP_SKB_CB(prev_skb)->eor is set 2. Update the TCP_SKB_CB(prev_skb)->eor if coalescing is allowed Packetdrill script for testing: ~~~~~~ +0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10` +0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1` +0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7> 0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7> 0.200 < . 1:1(0) ack 1 win 257 0.200 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 0.200 sendto(4, ..., 730, MSG_EOR, ..., ...) = 730 0.200 sendto(4, ..., 730, MSG_EOR, ..., ...) = 730 0.200 write(4, ..., 11680) = 11680 0.200 > P. 1:731(730) ack 1 0.200 > P. 731:1461(730) ack 1 0.200 > . 1461:8761(7300) ack 1 0.200 > P. 8761:13141(4380) ack 1 0.300 < . 1:1(0) ack 1 win 257 <sack 1461:13141,nop,nop> 0.300 > P. 1:731(730) ack 1 0.300 > P. 731:1461(730) ack 1 0.400 < . 1:1(0) ack 13141 win 257 0.400 close(4) = 0 0.400 > F. 13141:13141(0) ack 1 0.500 < F. 1:1(0) ack 13142 win 257 0.500 > . 13142:13142(0) ack 2 Signed-off-by: Martin KaFai Lau <kafai@fb.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Soheil Hassas Yeganeh <soheil@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-04-26 05:44:49 +08:00
if (!tcp_skb_can_collapse_to(to))
break;
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
space -= skb->len;
if (first) {
first = false;
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
continue;
}
if (space < 0)
break;
if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
break;
tcp: enhance tcp_collapse_retrans() with skb_shift() In commit 2331ccc5b323 ("tcp: enhance tcp collapsing"), we made a first step allowing copying right skb to left skb head. Since all skbs in socket write queue are headless (but possibly the very first one), this strategy often does not work. This patch extends tcp_collapse_retrans() to perform frag shifting, thanks to skb_shift() helper. This helper needs to not BUG on non headless skbs, as callers are ok with that. Tested: Following packetdrill test now passes : 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 8> +0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8> +.100 < . 1:1(0) ack 1 win 257 +0 accept(3, ..., ...) = 4 +0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0 +0 write(4, ..., 200) = 200 +0 > P. 1:201(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 201:401(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 401:601(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 601:801(200) ack 1 +.001 write(4, ..., 200) = 200 +0 > P. 801:1001(200) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1001:1101(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1101:1201(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1201:1301(100) ack 1 +.001 write(4, ..., 100) = 100 +0 > P. 1301:1401(100) ack 1 +.099 < . 1:1(0) ack 201 win 257 +.001 < . 1:1(0) ack 201 win 257 <nop,nop,sack 1001:1401> +0 > P. 201:1001(800) ack 1 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-16 04:51:50 +08:00
if (!tcp_collapse_retrans(sk, to))
break;
tcp: collapse more than two on retransmission I always had thought that collapsing up to two at a time was intentional decision to avoid excessive processing if 1 byte sized skbs are to be combined for a full mtu, and consecutive retransmissions would make the size of the retransmittee double each round anyway, but some recent discussion made me to understand that was not the case. Thus make collapse work more and wait less. It would be possible to take advantage of the shifting machinery (added in the later patch) in the case of paged data but that can be implemented on top of this change. tcp_skb_is_last check is now provided by the loop. I tested a bit (ss-after-idle-off, fill 4096x4096B xfer, 10s sleep + 4096 x 1byte writes while dropping them for some a while with netem): . 16774097:16775545(1448) ack 1 win 46 . 16775545:16776993(1448) ack 1 win 46 . ack 16759617 win 2399 P 16776993:16777217(224) ack 1 win 46 . ack 16762513 win 2399 . ack 16765409 win 2399 . ack 16768305 win 2399 . ack 16771201 win 2399 . ack 16774097 win 2399 . ack 16776993 win 2399 . ack 16777217 win 2399 P 16777217:16777257(40) ack 1 win 46 . ack 16777257 win 2399 P 16777257:16778705(1448) ack 1 win 46 P 16778705:16780153(1448) ack 1 win 46 FP 16780153:16781313(1160) ack 1 win 46 . ack 16778705 win 2399 . ack 16780153 win 2399 F 1:1(0) ack 16781314 win 2399 While without drop-all period I get this: . 16773585:16775033(1448) ack 1 win 46 . ack 16764897 win 9367 . ack 16767793 win 9367 . ack 16770689 win 9367 . ack 16773585 win 9367 . 16775033:16776481(1448) ack 1 win 46 P 16776481:16777217(736) ack 1 win 46 . ack 16776481 win 9367 . ack 16777217 win 9367 P 16777217:16777218(1) ack 1 win 46 P 16777218:16777219(1) ack 1 win 46 P 16777219:16777220(1) ack 1 win 46 ... P 16777247:16777248(1) ack 1 win 46 . ack 16777218 win 9367 . ack 16777219 win 9367 ... . ack 16777233 win 9367 . ack 16777248 win 9367 P 16777248:16778696(1448) ack 1 win 46 P 16778696:16780144(1448) ack 1 win 46 FP 16780144:16781313(1169) ack 1 win 46 . ack 16780144 win 9367 F 1:1(0) ack 16781314 win 9367 The window seems to be 30-40 segments, which were successfully combined into: P 16777217:16777257(40) ack 1 win 46 Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-25 13:03:43 +08:00
}
}
/* This retransmits one SKB. Policy decisions and retransmit queue
* state updates are done by the caller. Returns non-zero if an
* error occurred which prevented the send.
*/
int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
{
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
unsigned int cur_mss;
int diff, len, err;
/* Inconclusive MTU probe */
if (icsk->icsk_mtup.probe_size)
icsk->icsk_mtup.probe_size = 0;
/* Do not sent more than we queued. 1/4 is reserved for possible
* copying overhead: fragmentation, tunneling, mangling etc.
*/
if (refcount_read(&sk->sk_wmem_alloc) >
min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
sk->sk_sndbuf))
return -EAGAIN;
if (skb_still_in_host_queue(sk, skb))
return -EBUSY;
if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
tcp: purge write queue in tcp_connect_init() syzkaller found a reliable way to crash the host, hitting a BUG() in __tcp_retransmit_skb() Malicous MSG_FASTOPEN is the root cause. We need to purge write queue in tcp_connect_init() at the point we init snd_una/write_seq. This patch also replaces the BUG() by a less intrusive WARN_ON_ONCE() kernel BUG at net/ipv4/tcp_output.c:2837! invalid opcode: 0000 [#1] SMP KASAN Dumping ftrace buffer: (ftrace buffer empty) Modules linked in: CPU: 0 PID: 5276 Comm: syz-executor0 Not tainted 4.17.0-rc3+ #51 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 RIP: 0010:__tcp_retransmit_skb+0x2992/0x2eb0 net/ipv4/tcp_output.c:2837 RSP: 0000:ffff8801dae06ff8 EFLAGS: 00010206 RAX: ffff8801b9fe61c0 RBX: 00000000ffc18a16 RCX: ffffffff864e1a49 RDX: 0000000000000100 RSI: ffffffff864e2e12 RDI: 0000000000000005 RBP: ffff8801dae073a0 R08: ffff8801b9fe61c0 R09: ffffed0039c40dd2 R10: ffffed0039c40dd2 R11: ffff8801ce206e93 R12: 00000000421eeaad R13: ffff8801ce206d4e R14: ffff8801ce206cc0 R15: ffff8801cd4f4a80 FS: 0000000000000000(0000) GS:ffff8801dae00000(0063) knlGS:00000000096bc900 CS: 0010 DS: 002b ES: 002b CR0: 0000000080050033 CR2: 0000000020000000 CR3: 00000001c47b6000 CR4: 00000000001406f0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <IRQ> tcp_retransmit_skb+0x2e/0x250 net/ipv4/tcp_output.c:2923 tcp_retransmit_timer+0xc50/0x3060 net/ipv4/tcp_timer.c:488 tcp_write_timer_handler+0x339/0x960 net/ipv4/tcp_timer.c:573 tcp_write_timer+0x111/0x1d0 net/ipv4/tcp_timer.c:593 call_timer_fn+0x230/0x940 kernel/time/timer.c:1326 expire_timers kernel/time/timer.c:1363 [inline] __run_timers+0x79e/0xc50 kernel/time/timer.c:1666 run_timer_softirq+0x4c/0x70 kernel/time/timer.c:1692 __do_softirq+0x2e0/0xaf5 kernel/softirq.c:285 invoke_softirq kernel/softirq.c:365 [inline] irq_exit+0x1d1/0x200 kernel/softirq.c:405 exiting_irq arch/x86/include/asm/apic.h:525 [inline] smp_apic_timer_interrupt+0x17e/0x710 arch/x86/kernel/apic/apic.c:1052 apic_timer_interrupt+0xf/0x20 arch/x86/entry/entry_64.S:863 Fixes: cf60af03ca4e ("net-tcp: Fast Open client - sendmsg(MSG_FASTOPEN)") Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Reported-by: syzbot <syzkaller@googlegroups.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-05-15 12:14:26 +08:00
if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
WARN_ON_ONCE(1);
return -EINVAL;
}
if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
return -ENOMEM;
}
if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
return -EHOSTUNREACH; /* Routing failure or similar. */
cur_mss = tcp_current_mss(sk);
/* If receiver has shrunk his window, and skb is out of
* new window, do not retransmit it. The exception is the
* case, when window is shrunk to zero. In this case
* our retransmit serves as a zero window probe.
*/
if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
TCP_SKB_CB(skb)->seq != tp->snd_una)
return -EAGAIN;
len = cur_mss * segs;
if (skb->len > len) {
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
cur_mss, GFP_ATOMIC))
return -ENOMEM; /* We'll try again later. */
} else {
if (skb_unclone(skb, GFP_ATOMIC))
return -ENOMEM;
diff = tcp_skb_pcount(skb);
tcp_set_skb_tso_segs(skb, cur_mss);
diff -= tcp_skb_pcount(skb);
if (diff)
tcp_adjust_pcount(sk, skb, diff);
if (skb->len < cur_mss)
tcp_retrans_try_collapse(sk, skb, cur_mss);
}
tcp: add rfc3168, section 6.1.1.1. fallback This work as a follow-up of commit f7b3bec6f516 ("net: allow setting ecn via routing table") and adds RFC3168 section 6.1.1.1. fallback for outgoing ECN connections. In other words, this work adds a retry with a non-ECN setup SYN packet, as suggested from the RFC on the first timeout: [...] A host that receives no reply to an ECN-setup SYN within the normal SYN retransmission timeout interval MAY resend the SYN and any subsequent SYN retransmissions with CWR and ECE cleared. [...] Schematic client-side view when assuming the server is in tcp_ecn=2 mode, that is, Linux default since 2009 via commit 255cac91c3c9 ("tcp: extend ECN sysctl to allow server-side only ECN"): 1) Normal ECN-capable path: SYN ECE CWR -----> <----- SYN ACK ECE ACK -----> 2) Path with broken middlebox, when client has fallback: SYN ECE CWR ----X crappy middlebox drops packet (timeout, rtx) SYN -----> <----- SYN ACK ACK -----> In case we would not have the fallback implemented, the middlebox drop point would basically end up as: SYN ECE CWR ----X crappy middlebox drops packet (timeout, rtx) SYN ECE CWR ----X crappy middlebox drops packet (timeout, rtx) SYN ECE CWR ----X crappy middlebox drops packet (timeout, rtx) In any case, it's rather a smaller percentage of sites where there would occur such additional setup latency: it was found in end of 2014 that ~56% of IPv4 and 65% of IPv6 servers of Alexa 1 million list would negotiate ECN (aka tcp_ecn=2 default), 0.42% of these webservers will fail to connect when trying to negotiate with ECN (tcp_ecn=1) due to timeouts, which the fallback would mitigate with a slight latency trade-off. Recent related paper on this topic: Brian Trammell, Mirja Kühlewind, Damiano Boppart, Iain Learmonth, Gorry Fairhurst, and Richard Scheffenegger: "Enabling Internet-Wide Deployment of Explicit Congestion Notification." Proc. PAM 2015, New York. http://ecn.ethz.ch/ecn-pam15.pdf Thus, when net.ipv4.tcp_ecn=1 is being set, the patch will perform RFC3168, section 6.1.1.1. fallback on timeout. For users explicitly not wanting this which can be in DC use case, we add a net.ipv4.tcp_ecn_fallback knob that allows for disabling the fallback. tp->ecn_flags are not being cleared in tcp_ecn_clear_syn() on output, but rather we let tcp_ecn_rcv_synack() take that over on input path in case a SYN ACK ECE was delayed. Thus a spurious SYN retransmission will not prevent ECN being negotiated eventually in that case. Reference: https://www.ietf.org/proceedings/92/slides/slides-92-iccrg-1.pdf Reference: https://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: Mirja Kühlewind <mirja.kuehlewind@tik.ee.ethz.ch> Signed-off-by: Brian Trammell <trammell@tik.ee.ethz.ch> Cc: Eric Dumazet <edumazet@google.com> Cc: Dave That <dave.taht@gmail.com> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-20 03:04:22 +08:00
/* RFC3168, section 6.1.1.1. ECN fallback */
if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
tcp_ecn_clear_syn(sk, skb);
/* Update global and local TCP statistics. */
segs = tcp_skb_pcount(skb);
TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
tp->total_retrans += segs;
tp->bytes_retrans += skb->len;
/* make sure skb->data is aligned on arches that require it
* and check if ack-trimming & collapsing extended the headroom
* beyond what csum_start can cover.
*/
if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
skb_headroom(skb) >= 0xFFFF)) {
struct sk_buff *nskb;
tcp: new list for sent but unacked skbs for RACK recovery This patch adds a new queue (list) that tracks the sent but not yet acked or SACKed skbs for a TCP connection. The list is chronologically ordered by skb->skb_mstamp (the head is the oldest sent skb). This list will be used to optimize TCP Rack recovery, which checks an skb's timestamp to judge if it has been lost and needs to be retransmitted. Since TCP write queue is ordered by sequence instead of sent time, RACK has to scan over the write queue to catch all eligible packets to detect lost retransmission, and iterates through SACKed skbs repeatedly. Special cares for rare events: 1. TCP repair fakes skb transmission so the send queue needs adjusted 2. SACK reneging would require re-inserting SACKed skbs into the send queue. For now I believe it's not worth the complexity to make RACK work perfectly on SACK reneging, so we do nothing here. 3. Fast Open: currently for non-TFO, send-queue correctly queues the pure SYN packet. For TFO which queues a pure SYN and then a data packet, send-queue only queues the data packet but not the pure SYN due to the structure of TFO code. This is okay because the SYN receiver would never respond with a SACK on a missing SYN (i.e. SYN is never fast-retransmitted by SACK/RACK). In order to not grow sk_buff, we use an union for the new list and _skb_refdst/destructor fields. This is a bit complicated because we need to make sure _skb_refdst and destructor are properly zeroed before skb is cloned/copied at transmit, and before being freed. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-05 03:59:58 +08:00
tcp_skb_tsorted_save(skb) {
nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
-ENOBUFS;
} tcp_skb_tsorted_restore(skb);
if (!err) {
tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
tcp_rate_skb_sent(sk, skb);
}
} else {
err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
}
/* To avoid taking spuriously low RTT samples based on a timestamp
* for a transmit that never happened, always mark EVER_RETRANS
*/
TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
TCP_SKB_CB(skb)->seq, segs, err);
if (likely(!err)) {
trace_tcp_retransmit_skb(sk, skb);
} else if (err != -EBUSY) {
NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
}
return err;
}
int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
{
struct tcp_sock *tp = tcp_sk(sk);
int err = __tcp_retransmit_skb(sk, skb, segs);
if (err == 0) {
#if FASTRETRANS_DEBUG > 0
if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
net_dbg_ratelimited("retrans_out leaked\n");
}
#endif
TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
tp->retrans_out += tcp_skb_pcount(skb);
}
/* Save stamp of the first (attempted) retransmit. */
if (!tp->retrans_stamp)
tp->retrans_stamp = tcp_skb_timestamp(skb);
if (tp->undo_retrans < 0)
tp->undo_retrans = 0;
tp->undo_retrans += tcp_skb_pcount(skb);
return err;
}
/* This gets called after a retransmit timeout, and the initially
* retransmitted data is acknowledged. It tries to continue
* resending the rest of the retransmit queue, until either
* we've sent it all or the congestion window limit is reached.
*/
void tcp_xmit_retransmit_queue(struct sock *sk)
{
const struct inet_connection_sock *icsk = inet_csk(sk);
struct sk_buff *skb, *rtx_head, *hole = NULL;
struct tcp_sock *tp = tcp_sk(sk);
u32 max_segs;
int mib_idx;
if (!tp->packets_out)
return;
rtx_head = tcp_rtx_queue_head(sk);
skb = tp->retransmit_skb_hint ?: rtx_head;
max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
skb_rbtree_walk_from(skb) {
__u8 sacked;
int segs;
tcp: internal implementation for pacing BBR congestion control depends on pacing, and pacing is currently handled by sch_fq packet scheduler for performance reasons, and also because implemening pacing with FQ was convenient to truly avoid bursts. However there are many cases where this packet scheduler constraint is not practical. - Many linux hosts are not focusing on handling thousands of TCP flows in the most efficient way. - Some routers use fq_codel or other AQM, but still would like to use BBR for the few TCP flows they initiate/terminate. This patch implements an automatic fallback to internal pacing. Pacing is requested either by BBR or use of SO_MAX_PACING_RATE option. If sch_fq happens to be in the egress path, pacing is delegated to the qdisc, otherwise pacing is done by TCP itself. One advantage of pacing from TCP stack is to get more precise rtt estimations, and less work done from TX completion, since TCP Small queue limits are not generally hit. Setups with single TX queue but many cpus might even benefit from this. Note that unlike sch_fq, we do not take into account header sizes. Taking care of these headers would add additional complexity for no practical differences in behavior. Some performance numbers using 800 TCP_STREAM flows rate limited to ~48 Mbit per second on 40Gbit NIC. If MQ+pfifo_fast is used on the NIC : $ sar -n DEV 1 5 | grep eth 14:48:44 eth0 725743.00 2932134.00 46776.76 4335184.68 0.00 0.00 1.00 14:48:45 eth0 725349.00 2932112.00 46751.86 4335158.90 0.00 0.00 0.00 14:48:46 eth0 725101.00 2931153.00 46735.07 4333748.63 0.00 0.00 0.00 14:48:47 eth0 725099.00 2931161.00 46735.11 4333760.44 0.00 0.00 1.00 14:48:48 eth0 725160.00 2931731.00 46738.88 4334606.07 0.00 0.00 0.00 Average: eth0 725290.40 2931658.20 46747.54 4334491.74 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 4 0 0 259825920 45644 2708324 0 0 21 2 247 98 0 0 100 0 0 4 0 0 259823744 45644 2708356 0 0 0 0 2400825 159843 0 19 81 0 0 0 0 0 259824208 45644 2708072 0 0 0 0 2407351 159929 0 19 81 0 0 1 0 0 259824592 45644 2708128 0 0 0 0 2405183 160386 0 19 80 0 0 1 0 0 259824272 45644 2707868 0 0 0 32 2396361 158037 0 19 81 0 0 Now use MQ+FQ : lpaa23:~# echo fq >/proc/sys/net/core/default_qdisc lpaa23:~# tc qdisc replace dev eth0 root mq $ sar -n DEV 1 5 | grep eth 14:49:57 eth0 678614.00 2727930.00 43739.13 4033279.14 0.00 0.00 0.00 14:49:58 eth0 677620.00 2723971.00 43674.69 4027429.62 0.00 0.00 1.00 14:49:59 eth0 676396.00 2719050.00 43596.83 4020125.02 0.00 0.00 0.00 14:50:00 eth0 675197.00 2714173.00 43518.62 4012938.90 0.00 0.00 1.00 14:50:01 eth0 676388.00 2719063.00 43595.47 4020171.64 0.00 0.00 0.00 Average: eth0 676843.00 2720837.40 43624.95 4022788.86 0.00 0.00 0.40 $ vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 2 0 0 259832240 46008 2710912 0 0 21 2 223 192 0 1 99 0 0 1 0 0 259832896 46008 2710744 0 0 0 0 1702206 198078 0 17 82 0 0 0 0 0 259830272 46008 2710596 0 0 0 0 1696340 197756 1 17 83 0 0 4 0 0 259829168 46024 2710584 0 0 16 0 1688472 197158 1 17 82 0 0 3 0 0 259830224 46024 2710408 0 0 0 0 1692450 197212 0 18 82 0 0 As expected, number of interrupts per second is very different. Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Soheil Hassas Yeganeh <soheil@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Van Jacobson <vanj@google.com> Cc: Jerry Chu <hkchu@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-16 19:24:36 +08:00
if (tcp_pacing_check(sk))
break;
/* we could do better than to assign each time */
if (!hole)
tp->retransmit_skb_hint = skb;
segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
if (segs <= 0)
return;
sacked = TCP_SKB_CB(skb)->sacked;
/* In case tcp_shift_skb_data() have aggregated large skbs,
* we need to make sure not sending too bigs TSO packets
*/
segs = min_t(int, segs, max_segs);
if (tp->retrans_out >= tp->lost_out) {
break;
} else if (!(sacked & TCPCB_LOST)) {
if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
hole = skb;
continue;
} else {
if (icsk->icsk_ca_state != TCP_CA_Loss)
mib_idx = LINUX_MIB_TCPFASTRETRANS;
else
mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
}
if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
continue;
if (tcp_small_queue_check(sk, skb, 1))
return;
if (tcp_retransmit_skb(sk, skb, segs))
return;
NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
if (tcp_in_cwnd_reduction(sk))
Proportional Rate Reduction for TCP. This patch implements Proportional Rate Reduction (PRR) for TCP. PRR is an algorithm that determines TCP's sending rate in fast recovery. PRR avoids excessive window reductions and aims for the actual congestion window size at the end of recovery to be as close as possible to the window determined by the congestion control algorithm. PRR also improves accuracy of the amount of data sent during loss recovery. The patch implements the recommended flavor of PRR called PRR-SSRB (Proportional rate reduction with slow start reduction bound) and replaces the existing rate halving algorithm. PRR improves upon the existing Linux fast recovery under a number of conditions including: 1) burst losses where the losses implicitly reduce the amount of outstanding data (pipe) below the ssthresh value selected by the congestion control algorithm and, 2) losses near the end of short flows where application runs out of data to send. As an example, with the existing rate halving implementation a single loss event can cause a connection carrying short Web transactions to go into the slow start mode after the recovery. This is because during recovery Linux pulls the congestion window down to packets_in_flight+1 on every ACK. A short Web response often runs out of new data to send and its pipe reduces to zero by the end of recovery when all its packets are drained from the network. Subsequent HTTP responses using the same connection will have to slow start to raise cwnd to ssthresh. PRR on the other hand aims for the cwnd to be as close as possible to ssthresh by the end of recovery. A description of PRR and a discussion of its performance can be found at the following links: - IETF Draft: http://tools.ietf.org/html/draft-mathis-tcpm-proportional-rate-reduction-01 - IETF Slides: http://www.ietf.org/proceedings/80/slides/tcpm-6.pdf http://tools.ietf.org/agenda/81/slides/tcpm-2.pdf - Paper to appear in Internet Measurements Conference (IMC) 2011: Improving TCP Loss Recovery Nandita Dukkipati, Matt Mathis, Yuchung Cheng Signed-off-by: Nandita Dukkipati <nanditad@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-08-22 04:21:57 +08:00
tp->prr_out += tcp_skb_pcount(skb);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (skb == rtx_head &&
tcp: add reordering timer in RACK loss detection This patch makes RACK install a reordering timer when it suspects some packets might be lost, but wants to delay the decision a little bit to accomodate reordering. It does not create a new timer but instead repurposes the existing RTO timer, because both are meant to retransmit packets. Specifically it arms a timer ICSK_TIME_REO_TIMEOUT when the RACK timing check fails. The wait time is set to RACK.RTT + RACK.reo_wnd - (NOW - Packet.xmit_time) + fudge This translates to expecting a packet (Packet) should take (RACK.RTT + RACK.reo_wnd + fudge) to deliver after it was sent. When there are multiple packets that need a timer, we use one timer with the maximum timeout. Therefore the timer conservatively uses the maximum window to expire N packets by one timeout, instead of N timeouts to expire N packets sent at different times. The fudge factor is 2 jiffies to ensure when the timer fires, all the suspected packets would exceed the deadline and be marked lost by tcp_rack_detect_loss(). It has to be at least 1 jiffy because the clock may tick between calling icsk_reset_xmit_timer(timeout) and actually hang the timer. The next jiffy is to lower-bound the timeout to 2 jiffies when reo_wnd is < 1ms. When the reordering timer fires (tcp_rack_reo_timeout): If we aren't in Recovery we'll enter fast recovery and force fast retransmit. This is very similar to the early retransmit (RFC5827) except RACK is not constrained to only enter recovery for small outstanding flights. Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-01-13 14:11:33 +08:00
icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
inet_csk(sk)->icsk_rto,
TCP_RTO_MAX,
skb);
}
}
/* We allow to exceed memory limits for FIN packets to expedite
* connection tear down and (memory) recovery.
* Otherwise tcp_send_fin() could be tempted to either delay FIN
* or even be forced to close flow without any FIN.
* In general, we want to allow one skb per socket to avoid hangs
* with edge trigger epoll()
*/
void sk_forced_mem_schedule(struct sock *sk, int size)
{
net: tcp_memcontrol: sanitize tcp memory accounting callbacks There won't be a tcp control soft limit, so integrating the memcg code into the global skmem limiting scheme complicates things unnecessarily. Replace this with simple and clear charge and uncharge calls--hidden behind a jump label--to account skb memory. Note that this is not purely aesthetic: as a result of shoehorning the per-memcg code into the same memory accounting functions that handle the global level, the old code would compare the per-memcg consumption against the smaller of the per-memcg limit and the global limit. This allowed the total consumption of multiple sockets to exceed the global limit, as long as the individual sockets stayed within bounds. After this change, the code will always compare the per-memcg consumption to the per-memcg limit, and the global consumption to the global limit, and thus close this loophole. Without a soft limit, the per-memcg memory pressure state in sockets is generally questionable. However, we did it until now, so we continue to enter it when the hard limit is hit, and packets are dropped, to let other sockets in the cgroup know that they shouldn't grow their transmit windows, either. However, keep it simple in the new callback model and leave memory pressure lazily when the next packet is accepted (as opposed to doing it synchroneously when packets are processed). When packets are dropped, network performance will already be in the toilet, so that should be a reasonable trade-off. As described above, consumption is now checked on the per-memcg level and the global level separately. Likewise, memory pressure states are maintained on both the per-memcg level and the global level, and a socket is considered under pressure when either level asserts as much. Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Reviewed-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: David S. Miller <davem@davemloft.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-15 07:21:14 +08:00
int amt;
if (size <= sk->sk_forward_alloc)
return;
amt = sk_mem_pages(size);
sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
net: tcp_memcontrol: sanitize tcp memory accounting callbacks There won't be a tcp control soft limit, so integrating the memcg code into the global skmem limiting scheme complicates things unnecessarily. Replace this with simple and clear charge and uncharge calls--hidden behind a jump label--to account skb memory. Note that this is not purely aesthetic: as a result of shoehorning the per-memcg code into the same memory accounting functions that handle the global level, the old code would compare the per-memcg consumption against the smaller of the per-memcg limit and the global limit. This allowed the total consumption of multiple sockets to exceed the global limit, as long as the individual sockets stayed within bounds. After this change, the code will always compare the per-memcg consumption to the per-memcg limit, and the global consumption to the global limit, and thus close this loophole. Without a soft limit, the per-memcg memory pressure state in sockets is generally questionable. However, we did it until now, so we continue to enter it when the hard limit is hit, and packets are dropped, to let other sockets in the cgroup know that they shouldn't grow their transmit windows, either. However, keep it simple in the new callback model and leave memory pressure lazily when the next packet is accepted (as opposed to doing it synchroneously when packets are processed). When packets are dropped, network performance will already be in the toilet, so that should be a reasonable trade-off. As described above, consumption is now checked on the per-memcg level and the global level separately. Likewise, memory pressure states are maintained on both the per-memcg level and the global level, and a socket is considered under pressure when either level asserts as much. Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Reviewed-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: David S. Miller <davem@davemloft.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-15 07:21:14 +08:00
sk_memory_allocated_add(sk, amt);
if (mem_cgroup_sockets_enabled && sk->sk_memcg)
mem_cgroup_charge_skmem(sk->sk_memcg, amt);
}
/* Send a FIN. The caller locks the socket for us.
* We should try to send a FIN packet really hard, but eventually give up.
*/
void tcp_send_fin(struct sock *sk)
{
struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
struct tcp_sock *tp = tcp_sk(sk);
/* Optimization, tack on the FIN if we have one skb in write queue and
* this skb was not yet sent, or we are under memory pressure.
* Note: in the latter case, FIN packet will be sent after a timeout,
* as TCP stack thinks it has already been transmitted.
*/
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (!tskb && tcp_under_memory_pressure(sk))
tskb = skb_rb_last(&sk->tcp_rtx_queue);
if (tskb) {
TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
TCP_SKB_CB(tskb)->end_seq++;
tp->write_seq++;
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (tcp_write_queue_empty(sk)) {
/* This means tskb was already sent.
* Pretend we included the FIN on previous transmit.
* We need to set tp->snd_nxt to the value it would have
* if FIN had been sent. This is because retransmit path
* does not change tp->snd_nxt.
*/
WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
return;
}
} else {
skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
if (unlikely(!skb))
return;
tcp: new list for sent but unacked skbs for RACK recovery This patch adds a new queue (list) that tracks the sent but not yet acked or SACKed skbs for a TCP connection. The list is chronologically ordered by skb->skb_mstamp (the head is the oldest sent skb). This list will be used to optimize TCP Rack recovery, which checks an skb's timestamp to judge if it has been lost and needs to be retransmitted. Since TCP write queue is ordered by sequence instead of sent time, RACK has to scan over the write queue to catch all eligible packets to detect lost retransmission, and iterates through SACKed skbs repeatedly. Special cares for rare events: 1. TCP repair fakes skb transmission so the send queue needs adjusted 2. SACK reneging would require re-inserting SACKed skbs into the send queue. For now I believe it's not worth the complexity to make RACK work perfectly on SACK reneging, so we do nothing here. 3. Fast Open: currently for non-TFO, send-queue correctly queues the pure SYN packet. For TFO which queues a pure SYN and then a data packet, send-queue only queues the data packet but not the pure SYN due to the structure of TFO code. This is okay because the SYN receiver would never respond with a SACK on a missing SYN (i.e. SYN is never fast-retransmitted by SACK/RACK). In order to not grow sk_buff, we use an union for the new list and _skb_refdst/destructor fields. This is a bit complicated because we need to make sure _skb_refdst and destructor are properly zeroed before skb is cloned/copied at transmit, and before being freed. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-05 03:59:58 +08:00
INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
skb_reserve(skb, MAX_TCP_HEADER);
sk_forced_mem_schedule(sk, skb->truesize);
/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
tcp_init_nondata_skb(skb, tp->write_seq,
TCPHDR_ACK | TCPHDR_FIN);
tcp_queue_skb(sk, skb);
}
__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
}
/* We get here when a process closes a file descriptor (either due to
* an explicit close() or as a byproduct of exit()'ing) and there
* was unread data in the receive queue. This behavior is recommended
* by RFC 2525, section 2.17. -DaveM
*/
void tcp_send_active_reset(struct sock *sk, gfp_t priority)
{
struct sk_buff *skb;
TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
/* NOTE: No TCP options attached and we never retransmit this. */
skb = alloc_skb(MAX_TCP_HEADER, priority);
if (!skb) {
NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
return;
}
/* Reserve space for headers and prepare control bits. */
skb_reserve(skb, MAX_TCP_HEADER);
tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
TCPHDR_ACK | TCPHDR_RST);
tcp_mstamp_refresh(tcp_sk(sk));
/* Send it off. */
if (tcp_transmit_skb(sk, skb, 0, priority))
NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
* skb here is different to the troublesome skb, so use NULL
*/
trace_tcp_send_reset(sk, NULL);
}
/* Send a crossed SYN-ACK during socket establishment.
* WARNING: This routine must only be called when we have already sent
* a SYN packet that crossed the incoming SYN that caused this routine
* to get called. If this assumption fails then the initial rcv_wnd
* and rcv_wscale values will not be correct.
*/
int tcp_send_synack(struct sock *sk)
{
struct sk_buff *skb;
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
skb = tcp_rtx_queue_head(sk);
if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
pr_err("%s: wrong queue state\n", __func__);
return -EFAULT;
}
if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
if (skb_cloned(skb)) {
tcp: new list for sent but unacked skbs for RACK recovery This patch adds a new queue (list) that tracks the sent but not yet acked or SACKed skbs for a TCP connection. The list is chronologically ordered by skb->skb_mstamp (the head is the oldest sent skb). This list will be used to optimize TCP Rack recovery, which checks an skb's timestamp to judge if it has been lost and needs to be retransmitted. Since TCP write queue is ordered by sequence instead of sent time, RACK has to scan over the write queue to catch all eligible packets to detect lost retransmission, and iterates through SACKed skbs repeatedly. Special cares for rare events: 1. TCP repair fakes skb transmission so the send queue needs adjusted 2. SACK reneging would require re-inserting SACKed skbs into the send queue. For now I believe it's not worth the complexity to make RACK work perfectly on SACK reneging, so we do nothing here. 3. Fast Open: currently for non-TFO, send-queue correctly queues the pure SYN packet. For TFO which queues a pure SYN and then a data packet, send-queue only queues the data packet but not the pure SYN due to the structure of TFO code. This is okay because the SYN receiver would never respond with a SACK on a missing SYN (i.e. SYN is never fast-retransmitted by SACK/RACK). In order to not grow sk_buff, we use an union for the new list and _skb_refdst/destructor fields. This is a bit complicated because we need to make sure _skb_refdst and destructor are properly zeroed before skb is cloned/copied at transmit, and before being freed. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-05 03:59:58 +08:00
struct sk_buff *nskb;
tcp_skb_tsorted_save(skb) {
nskb = skb_copy(skb, GFP_ATOMIC);
} tcp_skb_tsorted_restore(skb);
if (!nskb)
return -ENOMEM;
tcp: new list for sent but unacked skbs for RACK recovery This patch adds a new queue (list) that tracks the sent but not yet acked or SACKed skbs for a TCP connection. The list is chronologically ordered by skb->skb_mstamp (the head is the oldest sent skb). This list will be used to optimize TCP Rack recovery, which checks an skb's timestamp to judge if it has been lost and needs to be retransmitted. Since TCP write queue is ordered by sequence instead of sent time, RACK has to scan over the write queue to catch all eligible packets to detect lost retransmission, and iterates through SACKed skbs repeatedly. Special cares for rare events: 1. TCP repair fakes skb transmission so the send queue needs adjusted 2. SACK reneging would require re-inserting SACKed skbs into the send queue. For now I believe it's not worth the complexity to make RACK work perfectly on SACK reneging, so we do nothing here. 3. Fast Open: currently for non-TFO, send-queue correctly queues the pure SYN packet. For TFO which queues a pure SYN and then a data packet, send-queue only queues the data packet but not the pure SYN due to the structure of TFO code. This is okay because the SYN receiver would never respond with a SACK on a missing SYN (i.e. SYN is never fast-retransmitted by SACK/RACK). In order to not grow sk_buff, we use an union for the new list and _skb_refdst/destructor fields. This is a bit complicated because we need to make sure _skb_refdst and destructor are properly zeroed before skb is cloned/copied at transmit, and before being freed. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-05 03:59:58 +08:00
INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
tcp_rtx_queue_unlink_and_free(skb, sk);
__skb_header_release(nskb);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
sk_wmem_queued_add(sk, nskb->truesize);
[NET] CORE: Introducing new memory accounting interface. This patch introduces new memory accounting functions for each network protocol. Most of them are renamed from memory accounting functions for stream protocols. At the same time, some stream memory accounting functions are removed since other functions do same thing. Renaming: sk_stream_free_skb() -> sk_wmem_free_skb() __sk_stream_mem_reclaim() -> __sk_mem_reclaim() sk_stream_mem_reclaim() -> sk_mem_reclaim() sk_stream_mem_schedule -> __sk_mem_schedule() sk_stream_pages() -> sk_mem_pages() sk_stream_rmem_schedule() -> sk_rmem_schedule() sk_stream_wmem_schedule() -> sk_wmem_schedule() sk_charge_skb() -> sk_mem_charge() Removeing sk_stream_rfree(): consolidates into sock_rfree() sk_stream_set_owner_r(): consolidates into skb_set_owner_r() sk_stream_mem_schedule() The following functions are added. sk_has_account(): check if the protocol supports accounting sk_mem_uncharge(): do the opposite of sk_mem_charge() In addition, to achieve consolidation, updating sk_wmem_queued is removed from sk_mem_charge(). Next, to consolidate memory accounting functions, this patch adds memory accounting calls to network core functions. Moreover, present memory accounting call is renamed to new accounting call. Finally we replace present memory accounting calls with new interface in TCP and SCTP. Signed-off-by: Takahiro Yasui <tyasui@redhat.com> Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-12-31 16:11:19 +08:00
sk_mem_charge(sk, nskb->truesize);
skb = nskb;
}
TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
tcp_ecn_send_synack(sk, skb);
}
return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
}
/**
* tcp_make_synack - Prepare a SYN-ACK.
* sk: listener socket
* dst: dst entry attached to the SYNACK
* req: request_sock pointer
*
* Allocate one skb and build a SYNACK packet.
* @dst is consumed : Caller should not use it again.
*/
struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
struct request_sock *req,
struct tcp_fastopen_cookie *foc,
enum tcp_synack_type synack_type)
{
struct inet_request_sock *ireq = inet_rsk(req);
const struct tcp_sock *tp = tcp_sk(sk);
struct tcp_md5sig_key *md5 = NULL;
struct tcp_out_options opts;
struct sk_buff *skb;
int tcp_header_size;
struct tcphdr *th;
int mss;
tcp: add optional per socket transmit delay Adding delays to TCP flows is crucial for studying behavior of TCP stacks, including congestion control modules. Linux offers netem module, but it has unpractical constraints : - Need root access to change qdisc - Hard to setup on egress if combined with non trivial qdisc like FQ - Single delay for all flows. EDT (Earliest Departure Time) adoption in TCP stack allows us to enable a per socket delay at a very small cost. Networking tools can now establish thousands of flows, each of them with a different delay, simulating real world conditions. This requires FQ packet scheduler or a EDT-enabled NIC. This patchs adds TCP_TX_DELAY socket option, to set a delay in usec units. unsigned int tx_delay = 10000; /* 10 msec */ setsockopt(fd, SOL_TCP, TCP_TX_DELAY, &tx_delay, sizeof(tx_delay)); Note that FQ packet scheduler limits might need some tweaking : man tc-fq PARAMETERS limit Hard limit on the real queue size. When this limit is reached, new packets are dropped. If the value is lowered, packets are dropped so that the new limit is met. Default is 10000 packets. flow_limit Hard limit on the maximum number of packets queued per flow. Default value is 100. Use of TCP_TX_DELAY option will increase number of skbs in FQ qdisc, so packets would be dropped if any of the previous limit is hit. Use of a jump label makes this support runtime-free, for hosts never using the option. Also note that TSQ (TCP Small Queues) limits are slightly changed with this patch : we need to account that skbs artificially delayed wont stop us providind more skbs to feed the pipe (netem uses skb_orphan_partial() for this purpose, but FQ can not use this trick) Because of that, using big delays might very well trigger old bugs in TSO auto defer logic and/or sndbuf limited detection. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-13 02:57:25 +08:00
u64 now;
skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
if (unlikely(!skb)) {
dst_release(dst);
return NULL;
}
/* Reserve space for headers. */
skb_reserve(skb, MAX_TCP_HEADER);
switch (synack_type) {
case TCP_SYNACK_NORMAL:
skb_set_owner_w(skb, req_to_sk(req));
break;
case TCP_SYNACK_COOKIE:
/* Under synflood, we do not attach skb to a socket,
* to avoid false sharing.
*/
break;
case TCP_SYNACK_FASTOPEN:
/* sk is a const pointer, because we want to express multiple
* cpu might call us concurrently.
* sk->sk_wmem_alloc in an atomic, we can promote to rw.
*/
skb_set_owner_w(skb, (struct sock *)sk);
break;
}
skb_dst_set(skb, dst);
mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
memset(&opts, 0, sizeof(opts));
tcp: add optional per socket transmit delay Adding delays to TCP flows is crucial for studying behavior of TCP stacks, including congestion control modules. Linux offers netem module, but it has unpractical constraints : - Need root access to change qdisc - Hard to setup on egress if combined with non trivial qdisc like FQ - Single delay for all flows. EDT (Earliest Departure Time) adoption in TCP stack allows us to enable a per socket delay at a very small cost. Networking tools can now establish thousands of flows, each of them with a different delay, simulating real world conditions. This requires FQ packet scheduler or a EDT-enabled NIC. This patchs adds TCP_TX_DELAY socket option, to set a delay in usec units. unsigned int tx_delay = 10000; /* 10 msec */ setsockopt(fd, SOL_TCP, TCP_TX_DELAY, &tx_delay, sizeof(tx_delay)); Note that FQ packet scheduler limits might need some tweaking : man tc-fq PARAMETERS limit Hard limit on the real queue size. When this limit is reached, new packets are dropped. If the value is lowered, packets are dropped so that the new limit is met. Default is 10000 packets. flow_limit Hard limit on the maximum number of packets queued per flow. Default value is 100. Use of TCP_TX_DELAY option will increase number of skbs in FQ qdisc, so packets would be dropped if any of the previous limit is hit. Use of a jump label makes this support runtime-free, for hosts never using the option. Also note that TSQ (TCP Small Queues) limits are slightly changed with this patch : we need to account that skbs artificially delayed wont stop us providind more skbs to feed the pipe (netem uses skb_orphan_partial() for this purpose, but FQ can not use this trick) Because of that, using big delays might very well trigger old bugs in TSO auto defer logic and/or sndbuf limited detection. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-13 02:57:25 +08:00
now = tcp_clock_ns();
#ifdef CONFIG_SYN_COOKIES
if (unlikely(req->cookie_ts))
skb->skb_mstamp_ns = cookie_init_timestamp(req);
else
#endif
{
tcp: add optional per socket transmit delay Adding delays to TCP flows is crucial for studying behavior of TCP stacks, including congestion control modules. Linux offers netem module, but it has unpractical constraints : - Need root access to change qdisc - Hard to setup on egress if combined with non trivial qdisc like FQ - Single delay for all flows. EDT (Earliest Departure Time) adoption in TCP stack allows us to enable a per socket delay at a very small cost. Networking tools can now establish thousands of flows, each of them with a different delay, simulating real world conditions. This requires FQ packet scheduler or a EDT-enabled NIC. This patchs adds TCP_TX_DELAY socket option, to set a delay in usec units. unsigned int tx_delay = 10000; /* 10 msec */ setsockopt(fd, SOL_TCP, TCP_TX_DELAY, &tx_delay, sizeof(tx_delay)); Note that FQ packet scheduler limits might need some tweaking : man tc-fq PARAMETERS limit Hard limit on the real queue size. When this limit is reached, new packets are dropped. If the value is lowered, packets are dropped so that the new limit is met. Default is 10000 packets. flow_limit Hard limit on the maximum number of packets queued per flow. Default value is 100. Use of TCP_TX_DELAY option will increase number of skbs in FQ qdisc, so packets would be dropped if any of the previous limit is hit. Use of a jump label makes this support runtime-free, for hosts never using the option. Also note that TSQ (TCP Small Queues) limits are slightly changed with this patch : we need to account that skbs artificially delayed wont stop us providind more skbs to feed the pipe (netem uses skb_orphan_partial() for this purpose, but FQ can not use this trick) Because of that, using big delays might very well trigger old bugs in TSO auto defer logic and/or sndbuf limited detection. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-13 02:57:25 +08:00
skb->skb_mstamp_ns = now;
if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
}
#ifdef CONFIG_TCP_MD5SIG
rcu_read_lock();
md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
#endif
skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
foc) + sizeof(*th);
skb_push(skb, tcp_header_size);
skb_reset_transport_header(skb);
th = (struct tcphdr *)skb->data;
memset(th, 0, sizeof(struct tcphdr));
th->syn = 1;
th->ack = 1;
tcp_ecn_make_synack(req, th);
th->source = htons(ireq->ir_num);
th->dest = ireq->ir_rmt_port;
skb->mark = ireq->ir_mark;
skb->ip_summed = CHECKSUM_PARTIAL;
th->seq = htonl(tcp_rsk(req)->snt_isn);
/* XXX data is queued and acked as is. No buffer/window check */
th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
th->window = htons(min(req->rsk_rcv_wnd, 65535U));
tcp_options_write((__be32 *)(th + 1), NULL, &opts);
th->doff = (tcp_header_size >> 2);
__TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
#ifdef CONFIG_TCP_MD5SIG
/* Okay, we have all we need - do the md5 hash if needed */
if (md5)
tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
md5, req_to_sk(req), skb);
rcu_read_unlock();
#endif
tcp: add optional per socket transmit delay Adding delays to TCP flows is crucial for studying behavior of TCP stacks, including congestion control modules. Linux offers netem module, but it has unpractical constraints : - Need root access to change qdisc - Hard to setup on egress if combined with non trivial qdisc like FQ - Single delay for all flows. EDT (Earliest Departure Time) adoption in TCP stack allows us to enable a per socket delay at a very small cost. Networking tools can now establish thousands of flows, each of them with a different delay, simulating real world conditions. This requires FQ packet scheduler or a EDT-enabled NIC. This patchs adds TCP_TX_DELAY socket option, to set a delay in usec units. unsigned int tx_delay = 10000; /* 10 msec */ setsockopt(fd, SOL_TCP, TCP_TX_DELAY, &tx_delay, sizeof(tx_delay)); Note that FQ packet scheduler limits might need some tweaking : man tc-fq PARAMETERS limit Hard limit on the real queue size. When this limit is reached, new packets are dropped. If the value is lowered, packets are dropped so that the new limit is met. Default is 10000 packets. flow_limit Hard limit on the maximum number of packets queued per flow. Default value is 100. Use of TCP_TX_DELAY option will increase number of skbs in FQ qdisc, so packets would be dropped if any of the previous limit is hit. Use of a jump label makes this support runtime-free, for hosts never using the option. Also note that TSQ (TCP Small Queues) limits are slightly changed with this patch : we need to account that skbs artificially delayed wont stop us providind more skbs to feed the pipe (netem uses skb_orphan_partial() for this purpose, but FQ can not use this trick) Because of that, using big delays might very well trigger old bugs in TSO auto defer logic and/or sndbuf limited detection. Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-13 02:57:25 +08:00
skb->skb_mstamp_ns = now;
tcp_add_tx_delay(skb, tp);
return skb;
}
EXPORT_SYMBOL(tcp_make_synack);
net: tcp: add per route congestion control This work adds the possibility to define a per route/destination congestion control algorithm. Generally, this opens up the possibility for a machine with different links to enforce specific congestion control algorithms with optimal strategies for each of them based on their network characteristics, even transparently for a single application listening on all links. For our specific use case, this additionally facilitates deployment of DCTCP, for example, applications can easily serve internal traffic/dsts in DCTCP and external one with CUBIC. Other scenarios would also allow for utilizing e.g. long living, low priority background flows for certain destinations/routes while still being able for normal traffic to utilize the default congestion control algorithm. We also thought about a per netns setting (where different defaults are possible), but given its actually a link specific property, we argue that a per route/destination setting is the most natural and flexible. The administrator can utilize this through ip-route(8) by appending "congctl [lock] <name>", where <name> denotes the name of a congestion control algorithm and the optional lock parameter allows to enforce the given algorithm so that applications in user space would not be allowed to overwrite that algorithm for that destination. The dst metric lookups are being done when a dst entry is already available in order to avoid a costly lookup and still before the algorithms are being initialized, thus overhead is very low when the feature is not being used. While the client side would need to drop the current reference on the module, on server side this can actually even be avoided as we just got a flat-copied socket clone. Joint work with Florian Westphal. Suggested-by: Hannes Frederic Sowa <hannes@stressinduktion.org> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-01-06 06:57:48 +08:00
static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
{
struct inet_connection_sock *icsk = inet_csk(sk);
const struct tcp_congestion_ops *ca;
u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
if (ca_key == TCP_CA_UNSPEC)
return;
rcu_read_lock();
ca = tcp_ca_find_key(ca_key);
if (likely(ca && try_module_get(ca->owner))) {
module_put(icsk->icsk_ca_ops->owner);
icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
icsk->icsk_ca_ops = ca;
}
rcu_read_unlock();
}
/* Do all connect socket setups that can be done AF independent. */
static void tcp_connect_init(struct sock *sk)
{
const struct dst_entry *dst = __sk_dst_get(sk);
struct tcp_sock *tp = tcp_sk(sk);
__u8 rcv_wscale;
u32 rcv_wnd;
/* We'll fix this up when we get a response from the other end.
* See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
*/
tp->tcp_header_len = sizeof(struct tcphdr);
if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
#ifdef CONFIG_TCP_MD5SIG
if (tp->af_specific->md5_lookup(sk, sk))
tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
#endif
/* If user gave his TCP_MAXSEG, record it to clamp */
if (tp->rx_opt.user_mss)
tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
tp->max_window = 0;
tcp_mtup_init(sk);
tcp_sync_mss(sk, dst_mtu(dst));
net: tcp: add per route congestion control This work adds the possibility to define a per route/destination congestion control algorithm. Generally, this opens up the possibility for a machine with different links to enforce specific congestion control algorithms with optimal strategies for each of them based on their network characteristics, even transparently for a single application listening on all links. For our specific use case, this additionally facilitates deployment of DCTCP, for example, applications can easily serve internal traffic/dsts in DCTCP and external one with CUBIC. Other scenarios would also allow for utilizing e.g. long living, low priority background flows for certain destinations/routes while still being able for normal traffic to utilize the default congestion control algorithm. We also thought about a per netns setting (where different defaults are possible), but given its actually a link specific property, we argue that a per route/destination setting is the most natural and flexible. The administrator can utilize this through ip-route(8) by appending "congctl [lock] <name>", where <name> denotes the name of a congestion control algorithm and the optional lock parameter allows to enforce the given algorithm so that applications in user space would not be allowed to overwrite that algorithm for that destination. The dst metric lookups are being done when a dst entry is already available in order to avoid a costly lookup and still before the algorithms are being initialized, thus overhead is very low when the feature is not being used. While the client side would need to drop the current reference on the module, on server side this can actually even be avoided as we just got a flat-copied socket clone. Joint work with Florian Westphal. Suggested-by: Hannes Frederic Sowa <hannes@stressinduktion.org> Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: Daniel Borkmann <dborkman@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-01-06 06:57:48 +08:00
tcp_ca_dst_init(sk, dst);
if (!tp->window_clamp)
tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
tcp_initialize_rcv_mss(sk);
tcp: allow effective reduction of TCP's rcv-buffer via setsockopt Via setsockopt it is possible to reduce the socket RX buffer (SO_RCVBUF). TCP method to select the initial window and window scaling option in tcp_select_initial_window() currently misbehaves and do not consider a reduced RX socket buffer via setsockopt. Even though the server's RX buffer is reduced via setsockopt() to 256 byte (Initial Window 384 byte => 256 * 2 - (256 * 2 / 4)) the window scale option is still 7: 192.168.1.38.40676 > 78.47.222.210.5001: Flags [S], seq 2577214362, win 5840, options [mss 1460,sackOK,TS val 338417 ecr 0,nop,wscale 0], length 0 78.47.222.210.5001 > 192.168.1.38.40676: Flags [S.], seq 1570631029, ack 2577214363, win 384, options [mss 1452,sackOK,TS val 2435248895 ecr 338417,nop,wscale 7], length 0 192.168.1.38.40676 > 78.47.222.210.5001: Flags [.], ack 1, win 5840, options [nop,nop,TS val 338421 ecr 2435248895], length 0 Within tcp_select_initial_window() the original space argument - a representation of the rx buffer size - is expanded during tcp_select_initial_window(). Only sysctl_tcp_rmem[2], sysctl_rmem_max and window_clamp are considered to calculate the initial window. This patch adjust the window_clamp argument if the user explicitly reduce the receive buffer. Signed-off-by: Hagen Paul Pfeifer <hagen@jauu.net> Cc: David S. Miller <davem@davemloft.net> Cc: Patrick McHardy <kaber@trash.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-08-19 14:33:05 +08:00
/* limit the window selection if the user enforce a smaller rx buffer */
if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
(tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
tp->window_clamp = tcp_full_space(sk);
rcv_wnd = tcp_rwnd_init_bpf(sk);
if (rcv_wnd == 0)
rcv_wnd = dst_metric(dst, RTAX_INITRWND);
tcp_select_initial_window(sk, tcp_full_space(sk),
tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
&tp->rcv_wnd,
&tp->window_clamp,
sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
&rcv_wscale,
rcv_wnd);
tp->rx_opt.rcv_wscale = rcv_wscale;
tp->rcv_ssthresh = tp->rcv_wnd;
sk->sk_err = 0;
sock_reset_flag(sk, SOCK_DONE);
tp->snd_wnd = 0;
tcp_init_wl(tp, 0);
tcp: purge write queue in tcp_connect_init() syzkaller found a reliable way to crash the host, hitting a BUG() in __tcp_retransmit_skb() Malicous MSG_FASTOPEN is the root cause. We need to purge write queue in tcp_connect_init() at the point we init snd_una/write_seq. This patch also replaces the BUG() by a less intrusive WARN_ON_ONCE() kernel BUG at net/ipv4/tcp_output.c:2837! invalid opcode: 0000 [#1] SMP KASAN Dumping ftrace buffer: (ftrace buffer empty) Modules linked in: CPU: 0 PID: 5276 Comm: syz-executor0 Not tainted 4.17.0-rc3+ #51 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 RIP: 0010:__tcp_retransmit_skb+0x2992/0x2eb0 net/ipv4/tcp_output.c:2837 RSP: 0000:ffff8801dae06ff8 EFLAGS: 00010206 RAX: ffff8801b9fe61c0 RBX: 00000000ffc18a16 RCX: ffffffff864e1a49 RDX: 0000000000000100 RSI: ffffffff864e2e12 RDI: 0000000000000005 RBP: ffff8801dae073a0 R08: ffff8801b9fe61c0 R09: ffffed0039c40dd2 R10: ffffed0039c40dd2 R11: ffff8801ce206e93 R12: 00000000421eeaad R13: ffff8801ce206d4e R14: ffff8801ce206cc0 R15: ffff8801cd4f4a80 FS: 0000000000000000(0000) GS:ffff8801dae00000(0063) knlGS:00000000096bc900 CS: 0010 DS: 002b ES: 002b CR0: 0000000080050033 CR2: 0000000020000000 CR3: 00000001c47b6000 CR4: 00000000001406f0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <IRQ> tcp_retransmit_skb+0x2e/0x250 net/ipv4/tcp_output.c:2923 tcp_retransmit_timer+0xc50/0x3060 net/ipv4/tcp_timer.c:488 tcp_write_timer_handler+0x339/0x960 net/ipv4/tcp_timer.c:573 tcp_write_timer+0x111/0x1d0 net/ipv4/tcp_timer.c:593 call_timer_fn+0x230/0x940 kernel/time/timer.c:1326 expire_timers kernel/time/timer.c:1363 [inline] __run_timers+0x79e/0xc50 kernel/time/timer.c:1666 run_timer_softirq+0x4c/0x70 kernel/time/timer.c:1692 __do_softirq+0x2e0/0xaf5 kernel/softirq.c:285 invoke_softirq kernel/softirq.c:365 [inline] irq_exit+0x1d1/0x200 kernel/softirq.c:405 exiting_irq arch/x86/include/asm/apic.h:525 [inline] smp_apic_timer_interrupt+0x17e/0x710 arch/x86/kernel/apic/apic.c:1052 apic_timer_interrupt+0xf/0x20 arch/x86/entry/entry_64.S:863 Fixes: cf60af03ca4e ("net-tcp: Fast Open client - sendmsg(MSG_FASTOPEN)") Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Reported-by: syzbot <syzkaller@googlegroups.com> Acked-by: Neal Cardwell <ncardwell@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-05-15 12:14:26 +08:00
tcp_write_queue_purge(sk);
tp->snd_una = tp->write_seq;
tp->snd_sml = tp->write_seq;
tp->snd_up = tp->write_seq;
WRITE_ONCE(tp->snd_nxt, tp->write_seq);
if (likely(!tp->repair))
tp->rcv_nxt = 0;
else
tp->rcv_tstamp = tcp_jiffies32;
tp->rcv_wup = tp->rcv_nxt;
WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
inet_csk(sk)->icsk_retransmits = 0;
tcp_clear_retrans(tp);
}
static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
{
struct tcp_sock *tp = tcp_sk(sk);
struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
tcb->end_seq += skb->len;
__skb_header_release(skb);
sk_wmem_queued_add(sk, skb->truesize);
sk_mem_charge(sk, skb->truesize);
WRITE_ONCE(tp->write_seq, tcb->end_seq);
tp->packets_out += tcp_skb_pcount(skb);
}
/* Build and send a SYN with data and (cached) Fast Open cookie. However,
* queue a data-only packet after the regular SYN, such that regular SYNs
* are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
* only the SYN sequence, the data are retransmitted in the first ACK.
* If cookie is not cached or other error occurs, falls back to send a
* regular SYN with Fast Open cookie request option.
*/
static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
{
struct tcp_sock *tp = tcp_sk(sk);
struct tcp_fastopen_request *fo = tp->fastopen_req;
int space, err = 0;
struct sk_buff *syn_data;
tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
goto fallback;
/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
* user-MSS. Reserve maximum option space for middleboxes that add
* private TCP options. The cost is reduced data space in SYN :(
*/
tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
MAX_TCP_OPTION_SPACE;
space = min_t(size_t, space, fo->size);
/* limit to order-0 allocations */
space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
if (!syn_data)
goto fallback;
syn_data->ip_summed = CHECKSUM_PARTIAL;
memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
if (space) {
int copied = copy_from_iter(skb_put(syn_data, space), space,
&fo->data->msg_iter);
if (unlikely(!copied)) {
tcp_skb_tsorted_anchor_cleanup(syn_data);
kfree_skb(syn_data);
goto fallback;
}
if (copied != space) {
skb_trim(syn_data, copied);
space = copied;
}
skb_zcopy_set(syn_data, fo->uarg, NULL);
}
/* No more data pending in inet_wait_for_connect() */
if (space == fo->size)
fo->data = NULL;
fo->copied = space;
tcp_connect_queue_skb(sk, syn_data);
if (syn_data->len)
tcp_chrono_start(sk, TCP_CHRONO_BUSY);
err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
/* Now full SYN+DATA was cloned and sent (or not),
* remove the SYN from the original skb (syn_data)
* we keep in write queue in case of a retransmit, as we
* also have the SYN packet (with no data) in the same queue.
*/
TCP_SKB_CB(syn_data)->seq++;
TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
if (!err) {
tp->syn_data = (fo->copied > 0);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
goto done;
}
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
/* data was not sent, put it in write_queue */
__skb_queue_tail(&sk->sk_write_queue, syn_data);
tcp: fastopen: fix on syn-data transmit failure Our recent change exposed a bug in TCP Fastopen Client that syzkaller found right away [1] When we prepare skb with SYN+DATA, we attempt to transmit it, and we update socket state as if the transmit was a success. In socket RTX queue we have two skbs, one with the SYN alone, and a second one containing the DATA. When (malicious) ACK comes in, we now complain that second one had no skb_mstamp. The proper fix is to make sure that if the transmit failed, we do not pretend we sent the DATA skb, and make it our send_head. When 3WHS completes, we can now send the DATA right away, without having to wait for a timeout. [1] WARNING: CPU: 0 PID: 100189 at net/ipv4/tcp_input.c:3117 tcp_clean_rtx_queue+0x2057/0x2ab0 net/ipv4/tcp_input.c:3117() WARN_ON_ONCE(last_ackt == 0); Modules linked in: CPU: 0 PID: 100189 Comm: syz-executor1 Not tainted Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 0000000000000000 ffff8800b35cb1d8 ffffffff81cad00d 0000000000000000 ffffffff828a4347 ffff88009f86c080 ffffffff8316eb20 0000000000000d7f ffff8800b35cb220 ffffffff812c33c2 ffff8800baad2440 00000009d46575c0 Call Trace: [<ffffffff81cad00d>] __dump_stack [<ffffffff81cad00d>] dump_stack+0xc1/0x124 [<ffffffff812c33c2>] warn_slowpath_common+0xe2/0x150 [<ffffffff812c361e>] warn_slowpath_null+0x2e/0x40 [<ffffffff828a4347>] tcp_clean_rtx_queue+0x2057/0x2ab0 n [<ffffffff828ae6fd>] tcp_ack+0x151d/0x3930 [<ffffffff828baa09>] tcp_rcv_state_process+0x1c69/0x4fd0 [<ffffffff828efb7f>] tcp_v4_do_rcv+0x54f/0x7c0 [<ffffffff8258aacb>] sk_backlog_rcv [<ffffffff8258aacb>] __release_sock+0x12b/0x3a0 [<ffffffff8258ad9e>] release_sock+0x5e/0x1c0 [<ffffffff8294a785>] inet_wait_for_connect [<ffffffff8294a785>] __inet_stream_connect+0x545/0xc50 [<ffffffff82886f08>] tcp_sendmsg_fastopen [<ffffffff82886f08>] tcp_sendmsg+0x2298/0x35a0 [<ffffffff82952515>] inet_sendmsg+0xe5/0x520 [<ffffffff8257152f>] sock_sendmsg_nosec [<ffffffff8257152f>] sock_sendmsg+0xcf/0x110 Fixes: 8c72c65b426b ("tcp: update skb->skb_mstamp more carefully") Fixes: 783237e8daf1 ("net-tcp: Fast Open client - sending SYN-data") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-20 01:05:57 +08:00
tp->packets_out -= tcp_skb_pcount(syn_data);
fallback:
/* Send a regular SYN with Fast Open cookie request option */
if (fo->cookie.len > 0)
fo->cookie.len = 0;
err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
if (err)
tp->syn_fastopen = 0;
done:
fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
return err;
}
/* Build a SYN and send it off. */
int tcp_connect(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
struct sk_buff *buff;
int err;
tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
return -EHOSTUNREACH; /* Routing failure or similar. */
tcp_connect_init(sk);
if (unlikely(tp->repair)) {
tcp_finish_connect(sk, NULL);
return 0;
}
buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
if (unlikely(!buff))
return -ENOBUFS;
tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
tcp_mstamp_refresh(tp);
tp->retrans_stamp = tcp_time_stamp(tp);
tcp_connect_queue_skb(sk, buff);
tcp_ecn_send_syn(sk, buff);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
/* Send off SYN; include data in Fast Open. */
err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
if (err == -ECONNREFUSED)
return err;
/* We change tp->snd_nxt after the tcp_transmit_skb() call
* in order to make this packet get counted in tcpOutSegs.
*/
WRITE_ONCE(tp->snd_nxt, tp->write_seq);
tp->pushed_seq = tp->write_seq;
tcp: fastopen: fix on syn-data transmit failure Our recent change exposed a bug in TCP Fastopen Client that syzkaller found right away [1] When we prepare skb with SYN+DATA, we attempt to transmit it, and we update socket state as if the transmit was a success. In socket RTX queue we have two skbs, one with the SYN alone, and a second one containing the DATA. When (malicious) ACK comes in, we now complain that second one had no skb_mstamp. The proper fix is to make sure that if the transmit failed, we do not pretend we sent the DATA skb, and make it our send_head. When 3WHS completes, we can now send the DATA right away, without having to wait for a timeout. [1] WARNING: CPU: 0 PID: 100189 at net/ipv4/tcp_input.c:3117 tcp_clean_rtx_queue+0x2057/0x2ab0 net/ipv4/tcp_input.c:3117() WARN_ON_ONCE(last_ackt == 0); Modules linked in: CPU: 0 PID: 100189 Comm: syz-executor1 Not tainted Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 0000000000000000 ffff8800b35cb1d8 ffffffff81cad00d 0000000000000000 ffffffff828a4347 ffff88009f86c080 ffffffff8316eb20 0000000000000d7f ffff8800b35cb220 ffffffff812c33c2 ffff8800baad2440 00000009d46575c0 Call Trace: [<ffffffff81cad00d>] __dump_stack [<ffffffff81cad00d>] dump_stack+0xc1/0x124 [<ffffffff812c33c2>] warn_slowpath_common+0xe2/0x150 [<ffffffff812c361e>] warn_slowpath_null+0x2e/0x40 [<ffffffff828a4347>] tcp_clean_rtx_queue+0x2057/0x2ab0 n [<ffffffff828ae6fd>] tcp_ack+0x151d/0x3930 [<ffffffff828baa09>] tcp_rcv_state_process+0x1c69/0x4fd0 [<ffffffff828efb7f>] tcp_v4_do_rcv+0x54f/0x7c0 [<ffffffff8258aacb>] sk_backlog_rcv [<ffffffff8258aacb>] __release_sock+0x12b/0x3a0 [<ffffffff8258ad9e>] release_sock+0x5e/0x1c0 [<ffffffff8294a785>] inet_wait_for_connect [<ffffffff8294a785>] __inet_stream_connect+0x545/0xc50 [<ffffffff82886f08>] tcp_sendmsg_fastopen [<ffffffff82886f08>] tcp_sendmsg+0x2298/0x35a0 [<ffffffff82952515>] inet_sendmsg+0xe5/0x520 [<ffffffff8257152f>] sock_sendmsg_nosec [<ffffffff8257152f>] sock_sendmsg+0xcf/0x110 Fixes: 8c72c65b426b ("tcp: update skb->skb_mstamp more carefully") Fixes: 783237e8daf1 ("net-tcp: Fast Open client - sending SYN-data") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-20 01:05:57 +08:00
buff = tcp_send_head(sk);
if (unlikely(buff)) {
WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
tcp: fastopen: fix on syn-data transmit failure Our recent change exposed a bug in TCP Fastopen Client that syzkaller found right away [1] When we prepare skb with SYN+DATA, we attempt to transmit it, and we update socket state as if the transmit was a success. In socket RTX queue we have two skbs, one with the SYN alone, and a second one containing the DATA. When (malicious) ACK comes in, we now complain that second one had no skb_mstamp. The proper fix is to make sure that if the transmit failed, we do not pretend we sent the DATA skb, and make it our send_head. When 3WHS completes, we can now send the DATA right away, without having to wait for a timeout. [1] WARNING: CPU: 0 PID: 100189 at net/ipv4/tcp_input.c:3117 tcp_clean_rtx_queue+0x2057/0x2ab0 net/ipv4/tcp_input.c:3117() WARN_ON_ONCE(last_ackt == 0); Modules linked in: CPU: 0 PID: 100189 Comm: syz-executor1 Not tainted Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 0000000000000000 ffff8800b35cb1d8 ffffffff81cad00d 0000000000000000 ffffffff828a4347 ffff88009f86c080 ffffffff8316eb20 0000000000000d7f ffff8800b35cb220 ffffffff812c33c2 ffff8800baad2440 00000009d46575c0 Call Trace: [<ffffffff81cad00d>] __dump_stack [<ffffffff81cad00d>] dump_stack+0xc1/0x124 [<ffffffff812c33c2>] warn_slowpath_common+0xe2/0x150 [<ffffffff812c361e>] warn_slowpath_null+0x2e/0x40 [<ffffffff828a4347>] tcp_clean_rtx_queue+0x2057/0x2ab0 n [<ffffffff828ae6fd>] tcp_ack+0x151d/0x3930 [<ffffffff828baa09>] tcp_rcv_state_process+0x1c69/0x4fd0 [<ffffffff828efb7f>] tcp_v4_do_rcv+0x54f/0x7c0 [<ffffffff8258aacb>] sk_backlog_rcv [<ffffffff8258aacb>] __release_sock+0x12b/0x3a0 [<ffffffff8258ad9e>] release_sock+0x5e/0x1c0 [<ffffffff8294a785>] inet_wait_for_connect [<ffffffff8294a785>] __inet_stream_connect+0x545/0xc50 [<ffffffff82886f08>] tcp_sendmsg_fastopen [<ffffffff82886f08>] tcp_sendmsg+0x2298/0x35a0 [<ffffffff82952515>] inet_sendmsg+0xe5/0x520 [<ffffffff8257152f>] sock_sendmsg_nosec [<ffffffff8257152f>] sock_sendmsg+0xcf/0x110 Fixes: 8c72c65b426b ("tcp: update skb->skb_mstamp more carefully") Fixes: 783237e8daf1 ("net-tcp: Fast Open client - sending SYN-data") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Cc: Neal Cardwell <ncardwell@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-20 01:05:57 +08:00
tp->pushed_seq = TCP_SKB_CB(buff)->seq;
}
TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
/* Timer for repeating the SYN until an answer. */
inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
return 0;
}
EXPORT_SYMBOL(tcp_connect);
/* Send out a delayed ack, the caller does the policy checking
* to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
* for details.
*/
void tcp_send_delayed_ack(struct sock *sk)
{
struct inet_connection_sock *icsk = inet_csk(sk);
int ato = icsk->icsk_ack.ato;
unsigned long timeout;
if (ato > TCP_DELACK_MIN) {
const struct tcp_sock *tp = tcp_sk(sk);
int max_ato = HZ / 2;
if (inet_csk_in_pingpong_mode(sk) ||
(icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
max_ato = TCP_DELACK_MAX;
/* Slow path, intersegment interval is "high". */
/* If some rtt estimate is known, use it to bound delayed ack.
* Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
* directly.
*/
tcp: switch rtt estimations to usec resolution Upcoming congestion controls for TCP require usec resolution for RTT estimations. Millisecond resolution is simply not enough these days. FQ/pacing in DC environments also require this change for finer control and removal of bimodal behavior due to the current hack in tcp_update_pacing_rate() for 'small rtt' TCP_CONG_RTT_STAMP is no longer needed. As Julian Anastasov pointed out, we need to keep user compatibility : tcp_metrics used to export RTT and RTTVAR in msec resolution, so we added RTT_US and RTTVAR_US. An iproute2 patch is needed to use the new attributes if provided by the kernel. In this example ss command displays a srtt of 32 usecs (10Gbit link) lpk51:~# ./ss -i dst lpk52 Netid State Recv-Q Send-Q Local Address:Port Peer Address:Port tcp ESTAB 0 1 10.246.11.51:42959 10.246.11.52:64614 cubic wscale:6,6 rto:201 rtt:0.032/0.001 ato:40 mss:1448 cwnd:10 send 3620.0Mbps pacing_rate 7240.0Mbps unacked:1 rcv_rtt:993 rcv_space:29559 Updated iproute2 ip command displays : lpk51:~# ./ip tcp_metrics | grep 10.246.11.52 10.246.11.52 age 561.914sec cwnd 10 rtt 274us rttvar 213us source 10.246.11.51 Old binary displays : lpk51:~# ip tcp_metrics | grep 10.246.11.52 10.246.11.52 age 561.914sec cwnd 10 rtt 250us rttvar 125us source 10.246.11.51 With help from Julian Anastasov, Stephen Hemminger and Yuchung Cheng Signed-off-by: Eric Dumazet <edumazet@google.com> Acked-by: Neal Cardwell <ncardwell@google.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Yuchung Cheng <ycheng@google.com> Cc: Larry Brakmo <brakmo@google.com> Cc: Julian Anastasov <ja@ssi.bg> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-02-27 06:02:48 +08:00
if (tp->srtt_us) {
int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
TCP_DELACK_MIN);
if (rtt < max_ato)
max_ato = rtt;
}
ato = min(ato, max_ato);
}
/* Stay within the limit we were given */
timeout = jiffies + ato;
/* Use new timeout only if there wasn't a older one earlier. */
if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
/* If delack timer was blocked or is about to expire,
* send ACK now.
*/
if (icsk->icsk_ack.blocked ||
time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
tcp_send_ack(sk);
return;
}
if (!time_before(timeout, icsk->icsk_ack.timeout))
timeout = icsk->icsk_ack.timeout;
}
icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
icsk->icsk_ack.timeout = timeout;
sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
}
/* This routine sends an ack and also updates the window. */
void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
{
struct sk_buff *buff;
/* If we have been reset, we may not send again. */
if (sk->sk_state == TCP_CLOSE)
return;
/* We are not putting this on the write queue, so
* tcp_transmit_skb() will set the ownership to this
* sock.
*/
buff = alloc_skb(MAX_TCP_HEADER,
sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
if (unlikely(!buff)) {
inet_csk_schedule_ack(sk);
inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
TCP_DELACK_MAX, TCP_RTO_MAX);
return;
}
/* Reserve space for headers and prepare control bits. */
skb_reserve(buff, MAX_TCP_HEADER);
tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
/* We do not want pure acks influencing TCP Small Queues or fq/pacing
* too much.
* SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
*/
skb_set_tcp_pure_ack(buff);
/* Send it off, this clears delayed acks for us. */
__tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
}
EXPORT_SYMBOL_GPL(__tcp_send_ack);
void tcp_send_ack(struct sock *sk)
{
__tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
}
/* This routine sends a packet with an out of date sequence
* number. It assumes the other end will try to ack it.
*
* Question: what should we make while urgent mode?
* 4.4BSD forces sending single byte of data. We cannot send
* out of window data, because we have SND.NXT==SND.MAX...
*
* Current solution: to send TWO zero-length segments in urgent mode:
* one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
* out-of-date with SND.UNA-1 to probe window.
*/
static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
{
struct tcp_sock *tp = tcp_sk(sk);
struct sk_buff *skb;
/* We don't queue it, tcp_transmit_skb() sets ownership. */
skb = alloc_skb(MAX_TCP_HEADER,
sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
if (!skb)
return -1;
/* Reserve space for headers and set control bits. */
skb_reserve(skb, MAX_TCP_HEADER);
/* Use a previous sequence. This should cause the other
* end to send an ack. Don't queue or clone SKB, just
* send it.
*/
tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
tcp: remove improper preemption check in tcp_xmit_probe_skb() Commit e520af48c7e5a introduced the following bug when setting the TCP_REPAIR sockoption: [ 2860.657036] BUG: using __this_cpu_add() in preemptible [00000000] code: daemon/12164 [ 2860.657045] caller is __this_cpu_preempt_check+0x13/0x20 [ 2860.657049] CPU: 1 PID: 12164 Comm: daemon Not tainted 4.2.3 #1 [ 2860.657051] Hardware name: Dell Inc. PowerEdge R210 II/0JP7TR, BIOS 2.0.5 03/13/2012 [ 2860.657054] ffffffff81c7f071 ffff880231e9fdf8 ffffffff8185d765 0000000000000002 [ 2860.657058] 0000000000000001 ffff880231e9fe28 ffffffff8146ed91 ffff880231e9fe18 [ 2860.657062] ffffffff81cd1a5d ffff88023534f200 ffff8800b9811000 ffff880231e9fe38 [ 2860.657065] Call Trace: [ 2860.657072] [<ffffffff8185d765>] dump_stack+0x4f/0x7b [ 2860.657075] [<ffffffff8146ed91>] check_preemption_disabled+0xe1/0xf0 [ 2860.657078] [<ffffffff8146edd3>] __this_cpu_preempt_check+0x13/0x20 [ 2860.657082] [<ffffffff817e0bc7>] tcp_xmit_probe_skb+0xc7/0x100 [ 2860.657085] [<ffffffff817e1e2d>] tcp_send_window_probe+0x2d/0x30 [ 2860.657089] [<ffffffff817d1d8c>] do_tcp_setsockopt.isra.29+0x74c/0x830 [ 2860.657093] [<ffffffff817d1e9c>] tcp_setsockopt+0x2c/0x30 [ 2860.657097] [<ffffffff81767b74>] sock_common_setsockopt+0x14/0x20 [ 2860.657100] [<ffffffff817669e1>] SyS_setsockopt+0x71/0xc0 [ 2860.657104] [<ffffffff81865172>] entry_SYSCALL_64_fastpath+0x16/0x75 Since tcp_xmit_probe_skb() can be called from process context, use NET_INC_STATS() instead of NET_INC_STATS_BH(). Fixes: e520af48c7e5 ("tcp: add TCPWinProbe and TCPKeepAlive SNMP counters") Signed-off-by: Renato Westphal <renatow@taghos.com.br> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-10-20 04:51:34 +08:00
NET_INC_STATS(sock_net(sk), mib);
return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
}
/* Called from setsockopt( ... TCP_REPAIR ) */
void tcp_send_window_probe(struct sock *sk)
{
if (sk->sk_state == TCP_ESTABLISHED) {
tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
tcp_mstamp_refresh(tcp_sk(sk));
tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
}
}
/* Initiate keepalive or window probe from timer. */
int tcp_write_wakeup(struct sock *sk, int mib)
{
struct tcp_sock *tp = tcp_sk(sk);
struct sk_buff *skb;
if (sk->sk_state == TCP_CLOSE)
return -1;
skb = tcp_send_head(sk);
if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
int err;
unsigned int mss = tcp_current_mss(sk);
unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
/* We are probing the opening of a window
* but the window size is != 0
* must have been a result SWS avoidance ( sender )
*/
if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
skb->len > mss) {
seg_size = min(seg_size, mss);
TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
skb, seg_size, mss, GFP_ATOMIC))
return -1;
} else if (!tcp_skb_pcount(skb))
tcp_set_skb_tso_segs(skb, mss);
TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
if (!err)
tcp_event_new_data_sent(sk, skb);
return err;
} else {
if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
tcp_xmit_probe_skb(sk, 1, mib);
return tcp_xmit_probe_skb(sk, 0, mib);
}
}
/* A window probe timeout has occurred. If window is not closed send
* a partial packet else a zero probe.
*/
void tcp_send_probe0(struct sock *sk)
{
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
struct net *net = sock_net(sk);
unsigned long timeout;
int err;
err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
tcp: implement rb-tree based retransmit queue Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-10-06 13:21:27 +08:00
if (tp->packets_out || tcp_write_queue_empty(sk)) {
/* Cancel probe timer, if it is not required. */
icsk->icsk_probes_out = 0;
icsk->icsk_backoff = 0;
return;
}
icsk->icsk_probes_out++;
if (err <= 0) {
if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
icsk->icsk_backoff++;
timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
} else {
/* If packet was not sent due to local congestion,
* Let senders fight for local resources conservatively.
*/
timeout = TCP_RESOURCE_PROBE_INTERVAL;
}
tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL);
}
int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
{
const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
struct flowi fl;
int res;
tcp_rsk(req)->txhash = net_tx_rndhash();
res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
if (!res) {
__TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
if (unlikely(tcp_passive_fastopen(sk)))
tcp_sk(sk)->total_retrans++;
trace_tcp_retransmit_synack(sk, req);
}
return res;
}
EXPORT_SYMBOL(tcp_rtx_synack);