Fix some issues introduced in batch skb dequeuing for input_pkt_queue.
The primary issue it that the queue head must be incremented only
after a packet has been processed, that is only after
__netif_receive_skb has been called. This is needed for the mechanism
to prevent OOO packet in RFS. Also when flushing the input_pkt_queue
and process_queue, the process queue should be done first to prevent
OOO packets.
Because the input_pkt_queue has been effectively split into two queues,
the calculation of the tail ptr is no longer correct. The correct value
would be head+input_pkt_queue->len+process_queue->len. To avoid
this calculation we added an explict input_queue_tail in softnet_data.
The tail value is simply incremented when queuing to input_pkt_queue.
Signed-off-by: Tom Herbert <therbert@google.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When GRO produces fraglist entries, and the resulting skb hits
an interface that is incapable of TSO but capable of FRAGLIST,
we end up producing a bogus packet with gso_size non-zero.
This was reported in the field with older versions of KVM that
did not set the TSO bits on tuntap.
This patch fixes that.
Reported-by: Igor Zhang <yugzhang@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add new netdev ops ndo_{set|get}_vf_port to allow setting of
port-profile on a netdev interface. Extends netlink socket RTM_SETLINK/
RTM_GETLINK with two new sub msgs called IFLA_VF_PORTS and IFLA_PORT_SELF
(added to end of IFLA_cmd list). These are both nested atrtibutes
using this layout:
[IFLA_NUM_VF]
[IFLA_VF_PORTS]
[IFLA_VF_PORT]
[IFLA_PORT_*], ...
[IFLA_VF_PORT]
[IFLA_PORT_*], ...
...
[IFLA_PORT_SELF]
[IFLA_PORT_*], ...
These attributes are design to be set and get symmetrically. VF_PORTS
is a list of VF_PORTs, one for each VF, when dealing with an SR-IOV
device. PORT_SELF is for the PF of the SR-IOV device, in case it wants
to also have a port-profile, or for the case where the VF==PF, like in
enic patch 2/2 of this patch set.
A port-profile is used to configure/enable the external switch virtual port
backing the netdev interface, not to configure the host-facing side of the
netdev. A port-profile is an identifier known to the switch. How port-
profiles are installed on the switch or how available port-profiles are
made know to the host is outside the scope of this patch.
There are two types of port-profiles specs in the netlink msg. The first spec
is for 802.1Qbg (pre-)standard, VDP protocol. The second spec is for devices
that run a similar protocol as VDP but in firmware, thus hiding the protocol
details. In either case, the specs have much in common and makes sense to
define the netlink msg as the union of the two specs. For example, both specs
have a notition of associating/deassociating a port-profile. And both specs
require some information from the hypervisor manager, such as client port
instance ID.
The general flow is the port-profile is applied to a host netdev interface
using RTM_SETLINK, the receiver of the RTM_SETLINK msg communicates with the
switch, and the switch virtual port backing the host netdev interface is
configured/enabled based on the settings defined by the port-profile. What
those settings comprise, and how those settings are managed is again
outside the scope of this patch, since this patch only deals with the
first step in the flow.
Signed-off-by: Scott Feldman <scofeldm@cisco.com>
Signed-off-by: Roopa Prabhu <roprabhu@cisco.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Also added an explicit break; to avoid
a fallthrough in net/ipv4/tcp_input.c
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use low order bit of skb->_skb_dst to tell dst is not refcounted.
Change _skb_dst to _skb_refdst to make sure all uses are catched.
skb_dst() returns the dst, regardless of noref bit set or not, but
with a lockdep check to make sure a noref dst is not given if current
user is not rcu protected.
New skb_dst_set_noref() helper to set an notrefcounted dst on a skb.
(with lockdep check)
skb_dst_drop() drops a reference only if skb dst was refcounted.
skb_dst_force() helper is used to force a refcount on dst, when skb
is queued and not anymore RCU protected.
Use skb_dst_force() in __sk_add_backlog(), __dev_xmit_skb() if
!IFF_XMIT_DST_RELEASE or skb enqueued on qdisc queue, in
sock_queue_rcv_skb(), in __nf_queue().
Use skb_dst_force() in dev_requeue_skb().
Note: dst_use_noref() still dirties dst, we might transform it
later to do one dirtying per jiffies.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If CONFIG_SMP=y, then we own a queue spinlock, we can avoid the atomic
test_and_set_bit() from napi_schedule_prep().
We now have same number of atomic ops per netif_rx() calls than with
pre-RPS kernel.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now we have a set of nested attributes:
IFLA_VFINFO_LIST (NESTED)
IFLA_VF_INFO (NESTED)
IFLA_VF_MAC
IFLA_VF_VLAN
IFLA_VF_TX_RATE
This allows a single set to operate on multiple attributes if desired.
Among other things, it means a dump can be replayed to set state.
The current interface has yet to be released, so this seems like
something to consider for 2.6.34.
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP-MD5 sessions have intermittent failures, when route cache is
invalidated. ip_queue_xmit() has to find a new route, calls
sk_setup_caps(sk, &rt->u.dst), destroying the
sk->sk_route_caps &= ~NETIF_F_GSO_MASK
that MD5 desperately try to make all over its way (from
tcp_transmit_skb() for example)
So we send few bad packets, and everything is fine when
tcp_transmit_skb() is called again for this socket.
Since ip_queue_xmit() is at a lower level than TCP-MD5, I chose to use a
socket field, sk_route_nocaps, containing bits to mask on sk_route_caps.
Reported-by: Bhaskar Dutta <bhaskie@gmail.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With RPS inclusion, skb timestamping is not consistent in RX path.
If netif_receive_skb() is used, its deferred after RPS dispatch.
If netif_rx() is used, its done before RPS dispatch.
This can give strange tcpdump timestamps results.
I think timestamping should be done as soon as possible in the receive
path, to get meaningful values (ie timestamps taken at the time packet
was delivered by NIC driver to our stack), even if NAPI already can
defer timestamping a bit (RPS can help to reduce the gap)
Tom Herbert prefer to sample timestamps after RPS dispatch. In case
sampling is expensive (HPET/acpi_pm on x86), this makes sense.
Let admins switch from one mode to another, using a new
sysctl, /proc/sys/net/core/netdev_tstamp_prequeue
Its default value (1), means timestamps are taken as soon as possible,
before backlog queueing, giving accurate timestamps.
Setting a 0 value permits to sample timestamps when processing backlog,
after RPS dispatch, to lower the load of the pre-RPS cpu.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now there's null check here and also again in the hook. Looking at bridge bits
which are simmilar, port structure is rcu_dereferenced right away in
handle_bridge and passed to hook. Looks nicer.
Signed-off-by: Jiri Pirko <jpirko@redhat.com>
Acked-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce ____napi_schedule() helper for callers in irq disabled
contexts. rps_trigger_softirq() becomes a leaf function.
Use container_of() in process_backlog() instead of accessing per_cpu
address.
Use a custom inlined version of __napi_complete() in process_backlog()
to avoid one locked instruction :
only current cpu owns and manipulates this napi,
and NAPI_STATE_SCHED is the only possible flag set on backlog.
we can use a plain write instead of clear_bit(),
and we dont need an smp_mb() memory barrier, since RPS is on,
backlog is protected by a spinlock.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In case of congestion, netif_rx() frees the skb, so we must assume
dev_forward_skb() also consume skb.
Bug introduced by commit 445409602c
(veth: move loopback logic to common location)
We must change dev_forward_skb() to always consume skb, and veth to not
double free it.
Bug report : http://marc.info/?l=linux-netdev&m=127310770900442&w=3
Reported-by: Martín Ferrari <martin.ferrari@gmail.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This whole patchset is for adding netpoll support to bridge and bonding
devices. I already tested it for bridge, bonding, bridge over bonding,
and bonding over bridge. It looks fine now.
To make bridge and bonding support netpoll, we need to adjust
some netpoll generic code. This patch does the following things:
1) introduce two new priv_flags for struct net_device:
IFF_IN_NETPOLL which identifies we are processing a netpoll;
IFF_DISABLE_NETPOLL is used to disable netpoll support for a device
at run-time;
2) introduce one new method for netdev_ops:
->ndo_netpoll_cleanup() is used to clean up netpoll when a device is
removed.
3) introduce netpoll_poll_dev() which takes a struct net_device * parameter;
export netpoll_send_skb() and netpoll_poll_dev() which will be used later;
4) hide a pointer to struct netpoll in struct netpoll_info, ditto.
5) introduce ->real_dev for struct netpoll.
6) introduce a new status NETDEV_BONDING_DESLAE, which is used to disable
netconsole before releasing a slave, to avoid deadlocks.
Cc: David Miller <davem@davemloft.net>
Cc: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: WANG Cong <amwang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With following patch I can reach maximum rate of my pktgen+udpsink
simulator :
- 'old' machine : dual quad core E5450 @3.00GHz
- 64 UDP rx flows (only differ by destination port)
- RPS enabled, NIC interrupts serviced on cpu0
- rps dispatched on 7 other cores. (~130.000 IPI per second)
- SLAB allocator (faster than SLUB in this workload)
- tg3 NIC
- 1.080.000 pps without a single drop at NIC level.
Idea is to add two prefetchw() calls in __alloc_skb(), one to prefetch
first sk_buff cache line, the second to prefetch the shinfo part.
Also using one memset() to initialize all skb_shared_info fields instead
of one by one to reduce number of instructions, using long word moves.
All skb_shared_info fields before 'dataref' are cleared in
__alloc_skb().
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 4b0b72f7dd ( net: speedup udp receive path )
introduced a bug in skb_free_datagram_locked().
We should not skb_orphan() skb if we dont have the guarantee we are the
last skb user, this might happen with MSG_PEEK concurrent users.
To keep socket locked for the smallest period of time, we split
consume_skb() logic, inlined in skb_free_datagram_locked()
Reported-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Per cpu variable softnet_data.total was shared between IRQ and SoftIRQ context
without any protection. And enqueue_to_backlog should update the netdev_rx_stat
of the target CPU.
This patch renames softnet_data.total to softnet_data.processed: the number of
packets processed in uppper levels(IP stacks).
softnet_stat data is moved into softnet_data.
Signed-off-by: Changli Gao <xiaosuo@gmail.com>
----
include/linux/netdevice.h | 17 +++++++----------
net/core/dev.c | 26 ++++++++++++--------------
net/sched/sch_generic.c | 2 +-
3 files changed, 20 insertions(+), 25 deletions(-)
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit 6be8ac2f ("[NET]: uninline skb_pull, de-bloats a lot")
we uninlined skb_pull.
But in some critical paths it makes sense to inline this thing
and it helps performance significantly.
Create an skb_pull_inline() so that we can do this in a way that
serves also as annotation.
Based upon a patch by Eric Dumazet.
Signed-off-by: David S. Miller <davem@davemloft.net>
sk_callback_lock rwlock actually protects sk->sk_sleep pointer, so we
need two atomic operations (and associated dirtying) per incoming
packet.
RCU conversion is pretty much needed :
1) Add a new structure, called "struct socket_wq" to hold all fields
that will need rcu_read_lock() protection (currently: a
wait_queue_head_t and a struct fasync_struct pointer).
[Future patch will add a list anchor for wakeup coalescing]
2) Attach one of such structure to each "struct socket" created in
sock_alloc_inode().
3) Respect RCU grace period when freeing a "struct socket_wq"
4) Change sk_sleep pointer in "struct sock" by sk_wq, pointer to "struct
socket_wq"
5) Change sk_sleep() function to use new sk->sk_wq instead of
sk->sk_sleep
6) Change sk_has_sleeper() to wq_has_sleeper() that must be used inside
a rcu_read_lock() section.
7) Change all sk_has_sleeper() callers to :
- Use rcu_read_lock() instead of read_lock(&sk->sk_callback_lock)
- Use wq_has_sleeper() to eventually wakeup tasks.
- Use rcu_read_unlock() instead of read_unlock(&sk->sk_callback_lock)
8) sock_wake_async() is modified to use rcu protection as well.
9) Exceptions :
macvtap, drivers/net/tun.c, af_unix use integrated "struct socket_wq"
instead of dynamically allocated ones. They dont need rcu freeing.
Some cleanups or followups are probably needed, (possible
sk_callback_lock conversion to a spinlock for example...).
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since commit 95766fff ([UDP]: Add memory accounting.),
each received packet needs one extra sock_lock()/sock_release() pair.
This added latency because of possible backlog handling. Then later,
ticket spinlocks added yet another latency source in case of DDOS.
This patch introduces lock_sock_bh() and unlock_sock_bh()
synchronization primitives, avoiding one atomic operation and backlog
processing.
skb_free_datagram_locked() uses them instead of full blown
lock_sock()/release_sock(). skb is orphaned inside locked section for
proper socket memory reclaim, and finally freed outside of it.
UDP receive path now take the socket spinlock only once.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now there's no need to use this fuction directly because it's handled by
register_pernet_device. So to make this simple and easy to understand,
make this static to do not tempt potentional users.
Signed-off-by: Jiri Pirko <jpirko@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current socket backlog limit is not enough to really stop DDOS attacks,
because user thread spend many time to process a full backlog each
round, and user might crazy spin on socket lock.
We should add backlog size and receive_queue size (aka rmem_alloc) to
pace writers, and let user run without being slow down too much.
Introduce a sk_rcvqueues_full() helper, to avoid taking socket lock in
stress situations.
Under huge stress from a multiqueue/RPS enabled NIC, a single flow udp
receiver can now process ~200.000 pps (instead of ~100 pps before the
patch) on a 8 core machine.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
batch skb dequeueing from softnet input_pkt_queue to reduce potential lock
contention when RPS is enabled.
Note: in the worst case, the number of packets in a softnet_data may
be double of netdev_max_backlog.
Signed-off-by: Changli Gao <xiaosuo@gmail.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
reimplement softnet_data.output_queue as a FIFO queue to keep the
fairness among the qdiscs rescheduled.
Signed-off-by: Changli Gao <xiaosuo@gmail.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
----
include/linux/netdevice.h | 1 +
net/core/dev.c | 22 ++++++++++++----------
2 files changed, 13 insertions(+), 10 deletions(-)
Signed-off-by: David S. Miller <davem@davemloft.net>
Decouple rtnetlink address families from real address families in socket.h to
be able to add rtnetlink interfaces to code that is not a real address family
without increasing AF_MAX/NPROTO.
This will be used to add support for multicast route dumping from all tables
as the proc interface can't be extended to support anything but the main table
without breaking compatibility.
This partialy undoes the patch to introduce independant families for routing
rules and converts ipmr routing rules to a new rtnetlink family. Similar to
that patch, values up to 127 are reserved for real address families, values
above that may be used arbitrarily.
Signed-off-by: Patrick McHardy <kaber@trash.net>
fib_rules_register() duplicates the template passed to it without modification,
mark the argument as const. Additionally the templates are only needed when
instantiating a new namespace, so mark them as __net_initdata, which means
they can be discarded when CONFIG_NET_NS=n.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Stay consistent with other functions and with comment also and name
pernet_operations parameter properly.
Signed-off-by: Jiri Pirko <jpirko@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
optimize rps_get_cpu().
don't initialize ports when we can get the ports. one memory access
for ports than two.
Signed-off-by: Changli Gao <xiaosuo@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add an SKF_AD_HATYPE field to the packet ancilliary data area, giving
access to skb->dev->type, as reported in the sll_hatype field.
When capturing packets on a PF_PACKET/SOCK_RAW socket bound to all
interfaces, there doesn't appear to be a way for the filter program to
actually find out the underlying hardware type the packet was captured
on. This patch adds such ability.
This patch also handles the case where skb->dev can be NULL, such as on
netlink sockets.
Signed-off-by: Paul Evans <leonerd@leonerd.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the original code, if rtnl_create_link() returned an ERR_PTR then that
would get passed to rtnl_configure_link() which dereferences it.
Signed-off-by: Dan Carpenter <error27@gmail.com>
Acked-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
This way GSO packets don't get handled differently.
With help from Eric Dumazet.
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
If some skb are queued to our backlog, we are delaying IPI sending at
the end of net_rx_action(), increasing latencies. This defeats the
queueing, since we want to quickly dispatch packets to the pool of
worker cpus, then eventually deeply process our packets.
It's better to send IPI before processing our packets in upper layers,
from process_backlog().
Change the _and_disable_irq suffix to _and_enable_irq(), since we enable
local irq in net_rps_action(), sorry for the confusion.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
At this point, skb->destructor is not the original one (stored in
DEV_GSO_CB(skb)->destructor)
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is no need to export skb_under_panic() and skb_over_panic() in
skbuff.c, since these methods are used only in skbuff.c ; this patch
removes these two exports. It also marks these functions as 'static'
and removeS the extern declarations of them from
include/linux/skbuff.h
Signed-off-by: Rami Rosen <ramirose@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Define a new function to return the waitqueue of a "struct sock".
static inline wait_queue_head_t *sk_sleep(struct sock *sk)
{
return sk->sk_sleep;
}
Change all read occurrences of sk_sleep by a call to this function.
Needed for a future RCU conversion. sk_sleep wont be a field directly
available.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since netdev_chain is guarded by rtnl_lock, ASSERT_RTNL should be
present here to make sure that all callers of call_netdevice_notifiers
does the locking properly.
Signed-off-by: Jiri Pirko <jpirko@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In case we compute a software skb->rxhash, we can generate a consistent
hash : Its value will be the same in both flow directions.
This helps some workloads, like conntracking, since the same state needs
to be accessed in both directions.
tbench + RFS + this patch gives better results than tbench with default
kernel configuration (no RPS, no RFS)
Also fixed some sparse warnings.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
struct softnet_data holds many queues, so consistent use "sd" name
instead of "queue" is better.
Adds a rps_ipi_queued() helper to cleanup enqueue_to_backlog()
Adds a _and_irq_disable suffix to net_rps_action() name, as David
suggested.
incr_input_queue_head() becomes input_queue_head_incr()
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
store_rps_map() & store_rps_dev_flow_table_cnt() are static.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net_rps_action() is a bit expensive on NR_CPUS=64..4096 kernels, even if
RPS is not active.
Tom Herbert used two bitmasks to hold information needed to send IPI,
but a single LIFO list seems more appropriate.
Move all RPS logic into net_rps_action() to cleanup net_rx_action() code
(remove two ifdefs)
Move rps_remote_softirq_cpus into softnet_data to share its first cache
line, filling an existing hole.
In a future patch, we could call net_rps_action() from process_backlog()
to make sure we send IPI before handling this cpu backlog.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Transmitted skb might be attached to a socket and a destructor, for
memory accounting purposes.
Traditionally, this destructor is called at tx completion time, when skb
is freed.
When tx completion is performed by another cpu than the sender, this
forces some cache lines to change ownership. XPS was an attempt to give
tx completion to initial cpu.
David idea is to call destructor right before giving skb to device (call
to ndo_start_xmit()). Because device queues are usually small, orphaning
skb before tx completion is not a big deal. Some drivers already do
this, we could do it in upper level.
There is one known exception to this early orphaning, called tx
timestamping. It needs to keep a reference to socket until device can
give a hardware or software timestamp.
This patch adds a skb_orphan_try() helper, to centralize all exceptions
to early orphaning in one spot, and use it in dev_hard_start_xmit().
"tbench 16" results on a Nehalem machine (2 X5570 @ 2.93GHz)
before: Throughput 4428.9 MB/sec 16 procs
after: Throughput 4448.14 MB/sec 16 procs
UDP should get even better results, its destructor being more complex,
since SOCK_USE_WRITE_QUEUE is not set (four atomic ops instead of one)
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
- There is no point to enforce a time limit in process_backlog(), since
other napi instances dont follow same rule. We can exit after only one
packet processed...
The normal quota of 64 packets per napi instance should be the norm, and
net_rx_action() already has its own time limit.
Note : /proc/net/core/dev_weight can be used to tune this 64 default
value.
- Use DEFINE_PER_CPU_ALIGNED for softnet_data definition.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements receive flow steering (RFS). RFS steers
received packets for layer 3 and 4 processing to the CPU where
the application for the corresponding flow is running. RFS is an
extension of Receive Packet Steering (RPS).
The basic idea of RFS is that when an application calls recvmsg
(or sendmsg) the application's running CPU is stored in a hash
table that is indexed by the connection's rxhash which is stored in
the socket structure. The rxhash is passed in skb's received on
the connection from netif_receive_skb. For each received packet,
the associated rxhash is used to look up the CPU in the hash table,
if a valid CPU is set then the packet is steered to that CPU using
the RPS mechanisms.
The convolution of the simple approach is that it would potentially
allow OOO packets. If threads are thrashing around CPUs or multiple
threads are trying to read from the same sockets, a quickly changing
CPU value in the hash table could cause rampant OOO packets--
we consider this a non-starter.
To avoid OOO packets, this solution implements two types of hash
tables: rps_sock_flow_table and rps_dev_flow_table.
rps_sock_table is a global hash table. Each entry is just a CPU
number and it is populated in recvmsg and sendmsg as described above.
This table contains the "desired" CPUs for flows.
rps_dev_flow_table is specific to each device queue. Each entry
contains a CPU and a tail queue counter. The CPU is the "current"
CPU for a matching flow. The tail queue counter holds the value
of a tail queue counter for the associated CPU's backlog queue at
the time of last enqueue for a flow matching the entry.
Each backlog queue has a queue head counter which is incremented
on dequeue, and so a queue tail counter is computed as queue head
count + queue length. When a packet is enqueued on a backlog queue,
the current value of the queue tail counter is saved in the hash
entry of the rps_dev_flow_table.
And now the trick: when selecting the CPU for RPS (get_rps_cpu)
the rps_sock_flow table and the rps_dev_flow table for the RX queue
are consulted. When the desired CPU for the flow (found in the
rps_sock_flow table) does not match the current CPU (found in the
rps_dev_flow table), the current CPU is changed to the desired CPU
if one of the following is true:
- The current CPU is unset (equal to RPS_NO_CPU)
- Current CPU is offline
- The current CPU's queue head counter >= queue tail counter in the
rps_dev_flow table. This checks if the queue tail has advanced
beyond the last packet that was enqueued using this table entry.
This guarantees that all packets queued using this entry have been
dequeued, thus preserving in order delivery.
Making each queue have its own rps_dev_flow table has two advantages:
1) the tail queue counters will be written on each receive, so
keeping the table local to interrupting CPU s good for locality. 2)
this allows lockless access to the table-- the CPU number and queue
tail counter need to be accessed together under mutual exclusion
from netif_receive_skb, we assume that this is only called from
device napi_poll which is non-reentrant.
This patch implements RFS for TCP and connected UDP sockets.
It should be usable for other flow oriented protocols.
There are two configuration parameters for RFS. The
"rps_flow_entries" kernel init parameter sets the number of
entries in the rps_sock_flow_table, the per rxqueue sysfs entry
"rps_flow_cnt" contains the number of entries in the rps_dev_flow
table for the rxqueue. Both are rounded to power of two.
The obvious benefit of RFS (over just RPS) is that it achieves
CPU locality between the receive processing for a flow and the
applications processing; this can result in increased performance
(higher pps, lower latency).
The benefits of RFS are dependent on cache hierarchy, application
load, and other factors. On simple benchmarks, we don't necessarily
see improvement and sometimes see degradation. However, for more
complex benchmarks and for applications where cache pressure is
much higher this technique seems to perform very well.
Below are some benchmark results which show the potential benfit of
this patch. The netperf test has 500 instances of netperf TCP_RR
test with 1 byte req. and resp. The RPC test is an request/response
test similar in structure to netperf RR test ith 100 threads on
each host, but does more work in userspace that netperf.
e1000e on 8 core Intel
No RFS or RPS 104K tps at 30% CPU
No RFS (best RPS config): 290K tps at 63% CPU
RFS 303K tps at 61% CPU
RPC test tps CPU% 50/90/99% usec latency Latency StdDev
No RFS/RPS 103K 48% 757/900/3185 4472.35
RPS only: 174K 73% 415/993/2468 491.66
RFS 223K 73% 379/651/1382 315.61
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When dev_pick_tx() caches tx queue_index on a socket, we must check
socket dst_entry matches skb one, or risk a crash later, as reported by
Denys Fedorysychenko, if old packets are in flight during a route
change, involving devices with different number of queues.
Bug introduced by commit a4ee3ce3
(net: Use sk_tx_queue_mapping for connected sockets)
Reported-by: Denys Fedorysychenko <nuclearcat@nuclearcat.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
