A single sendmsg or sendfile system call can contain multiple logical
messages that a BPF program may want to read and apply a verdict. But,
without an apply_bytes helper any verdict on the data applies to all
bytes in the sendmsg/sendfile. Alternatively, a BPF program may only
care to read the first N bytes of a msg. If the payload is large say
MB or even GB setting up and calling the BPF program repeatedly for
all bytes, even though the verdict is already known, creates
unnecessary overhead.
To allow BPF programs to control how many bytes a given verdict
applies to we implement a bpf_msg_apply_bytes() helper. When called
from within a BPF program this sets a counter, internal to the
BPF infrastructure, that applies the last verdict to the next N
bytes. If the N is smaller than the current data being processed
from a sendmsg/sendfile call, the first N bytes will be sent and
the BPF program will be re-run with start_data pointing to the N+1
byte. If N is larger than the current data being processed the
BPF verdict will be applied to multiple sendmsg/sendfile calls
until N bytes are consumed.
Note1 if a socket closes with apply_bytes counter non-zero this
is not a problem because data is not being buffered for N bytes
and is sent as its received.
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
This implements a BPF ULP layer to allow policy enforcement and
monitoring at the socket layer. In order to support this a new
program type BPF_PROG_TYPE_SK_MSG is used to run the policy at
the sendmsg/sendpage hook. To attach the policy to sockets a
sockmap is used with a new program attach type BPF_SK_MSG_VERDICT.
Similar to previous sockmap usages when a sock is added to a
sockmap, via a map update, if the map contains a BPF_SK_MSG_VERDICT
program type attached then the BPF ULP layer is created on the
socket and the attached BPF_PROG_TYPE_SK_MSG program is run for
every msg in sendmsg case and page/offset in sendpage case.
BPF_PROG_TYPE_SK_MSG Semantics/API:
BPF_PROG_TYPE_SK_MSG supports only two return codes SK_PASS and
SK_DROP. Returning SK_DROP free's the copied data in the sendmsg
case and in the sendpage case leaves the data untouched. Both cases
return -EACESS to the user. Returning SK_PASS will allow the msg to
be sent.
In the sendmsg case data is copied into kernel space buffers before
running the BPF program. The kernel space buffers are stored in a
scatterlist object where each element is a kernel memory buffer.
Some effort is made to coalesce data from the sendmsg call here.
For example a sendmsg call with many one byte iov entries will
likely be pushed into a single entry. The BPF program is run with
data pointers (start/end) pointing to the first sg element.
In the sendpage case data is not copied. We opt not to copy the
data by default here, because the BPF infrastructure does not
know what bytes will be needed nor when they will be needed. So
copying all bytes may be wasteful. Because of this the initial
start/end data pointers are (0,0). Meaning no data can be read or
written. This avoids reading data that may be modified by the
user. A new helper is added later in this series if reading and
writing the data is needed. The helper call will do a copy by
default so that the page is exclusively owned by the BPF call.
The verdict from the BPF_PROG_TYPE_SK_MSG applies to the entire msg
in the sendmsg() case and the entire page/offset in the sendpage case.
This avoids ambiguity on how to handle mixed return codes in the
sendmsg case. Again a helper is added later in the series if
a verdict needs to apply to multiple system calls and/or only
a subpart of the currently being processed message.
The helper msg_redirect_map() can be used to select the socket to
send the data on. This is used similar to existing redirect use
cases. This allows policy to redirect msgs.
Pseudo code simple example:
The basic logic to attach a program to a socket is as follows,
// load the programs
bpf_prog_load(SOCKMAP_TCP_MSG_PROG, BPF_PROG_TYPE_SK_MSG,
&obj, &msg_prog);
// lookup the sockmap
bpf_map_msg = bpf_object__find_map_by_name(obj, "my_sock_map");
// get fd for sockmap
map_fd_msg = bpf_map__fd(bpf_map_msg);
// attach program to sockmap
bpf_prog_attach(msg_prog, map_fd_msg, BPF_SK_MSG_VERDICT, 0);
Adding sockets to the map is done in the normal way,
// Add a socket 'fd' to sockmap at location 'i'
bpf_map_update_elem(map_fd_msg, &i, fd, BPF_ANY);
After the above any socket attached to "my_sock_map", in this case
'fd', will run the BPF msg verdict program (msg_prog) on every
sendmsg and sendpage system call.
For a complete example see BPF selftests or sockmap samples.
Implementation notes:
It seemed the simplest, to me at least, to use a refcnt to ensure
psock is not lost across the sendmsg copy into the sg, the bpf program
running on the data in sg_data, and the final pass to the TCP stack.
Some performance testing may show a better method to do this and avoid
the refcnt cost, but for now use the simpler method.
Another item that will come after basic support is in place is
supporting MSG_MORE flag. At the moment we call sendpages even if
the MSG_MORE flag is set. An enhancement would be to collect the
pages into a larger scatterlist and pass down the stack. Notice that
bpf_tcp_sendmsg() could support this with some additional state saved
across sendmsg calls. I built the code to support this without having
to do refactoring work. Other features TBD include ZEROCOPY and the
TCP_RECV_QUEUE/TCP_NO_QUEUE support. This will follow initial series
shortly.
Future work could improve size limits on the scatterlist rings used
here. Currently, we use MAX_SKB_FRAGS simply because this was being
used already in the TLS case. Future work could extend the kernel sk
APIs to tune this depending on workload. This is a trade-off
between memory usage and throughput performance.
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
The current implementation of sk_alloc_sg expects scatterlist to always
start at entry 0 and complete at entry MAX_SKB_FRAGS.
Future patches will want to support starting at arbitrary offset into
scatterlist so add an additional sg_start parameters and then default
to the current values in TLS code paths.
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
The TLS ULP module builds scatterlists from a sock using
page_frag_refill(). This is going to be useful for other ULPs
so move it into sock file for more general use.
In the process remove useless goto at end of while loop.
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Fix the SO_ZEROCOPY switch case on sock_setsockopt() avoiding the
ret values to be overwritten by the one set on the default case.
Fixes: 28190752c7 ("sock: permit SO_ZEROCOPY on PF_RDS socket")
Signed-off-by: Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The patch adds SLAB_ACCOUNT to flags of net_cachep cache,
which enables accounting of struct net memory to memcg kmem.
Since number of net_namespaces may be significant, user
want to know, how much there were consumed, and control.
Note, that we do not account net_generic to the same memcg,
where net was accounted, moreover, we don't do this at all (*).
We do not want the situation, when single memcg memory deficit
prevents us to register new pernet_operations.
(*)Even despite there is !current process accounting already
available in linux-next. See kmalloc_memcg() there for the details.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
All of the conflicts were cases of overlapping changes.
In net/core/devlink.c, we have to make care that the
resouce size_params have become a struct member rather
than a pointer to such an object.
Signed-off-by: David S. Miller <davem@davemloft.net>
The other dst_cache_{get,set}_ip{4,6} functions, and the doc comment for
dst_cache_set_ip6 use 'saddr' for their source address parameter. Rename
the parameter to increase consistency.
This fixes the following kernel-doc warnings:
./include/net/dst_cache.h:58: warning: Function parameter or member 'addr' not described in 'dst_cache_set_ip6'
./include/net/dst_cache.h:58: warning: Excess function parameter 'saddr' description in 'dst_cache_set_ip6'
Fixes: 911362c70d ("net: add dst_cache support")
Signed-off-by: Jonathan Neuschäfer <j.neuschaefer@gmx.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
LRO and RX-FCS offloads cannot be enabled at the same time since it is
not clear what should happen to the FCS of each coalesced packet.
The FCS is not really part of the TCP payload, hence cannot be merged
into one big packet. On the other hand, providing one big LRO packet
with one FCS contradicts the RX-FCS feature goal.
Use the fix features mechanism in order to prevent intersection of the
features and drop LRO in case RX-FCS is requested.
Enabling RX-FCS while LRO is enabled will result in:
$ ethtool -K ens6 rx-fcs on
Actual changes:
large-receive-offload: off [requested on]
rx-fcs: on
Signed-off-by: Gal Pressman <galp@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently GRE sequence number can only be used in native
tunnel mode. This patch adds sequence number support for
gre collect metadata mode. RFC2890 defines GRE sequence
number to be specific to the traffic flow identified by the
key. However, this patch does not implement per-key seqno.
The sequence number is shared in the same tunnel device.
That is, different tunnel keys using the same collect_md
tunnel share single sequence number.
Signed-off-by: William Tu <u9012063@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
They're very hard to use properly as they do not consider the
GSO_BY_FRAGS case. Code should use skb_gso_validate_network_len
and skb_gso_validate_mac_len as they do consider this case.
Make the seglen functions static, which stops people using them
outside of skbuff.c
Signed-off-by: Daniel Axtens <dja@axtens.net>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If you take a GSO skb, and split it into packets, will the network
length (L3 headers + L4 headers + payload) of those packets be small
enough to fit within a given MTU?
skb_gso_validate_mtu gives you the answer to that question. However,
we recently added to add a way to validate the MAC length of a split GSO
skb (L2+L3+L4+payload), and the names get confusing, so rename
skb_gso_validate_mtu to skb_gso_validate_network_len
Signed-off-by: Daniel Axtens <dja@axtens.net>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
__get_hash_from_flowi6 is still used for flowlabels, but the IPv4
variant and the wrappers to both are not used. Remove them.
Signed-off-by: David Ahern <dsahern@gmail.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Reviewed-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If ethtool_ops->get_fecparam returns an error, pass that error on to the
user, rather than ignoring it.
Fixes: 1a5f3da20b ("net: ethtool: add support for forward error correction modes")
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Setting an interface into a VRF fails with 'RTNETLINK answers: File
exists' if one of its VLAN interfaces is already in the same VRF.
As the VRF is an upper device of the VLAN interface, it is also showing
up as an upper device of the interface itself. The solution is to
restrict this check to devices other than master. As only one master
device can be linked to a device, the check in this case is that the
upper device (VRF) being linked to is not the same as the master device
instead of it not being any one of the upper devices.
The following example shows an interface ens12 (with a VLAN interface
ens12.10) being set into VRF green, which behaves as expected:
# ip link add link ens12 ens12.10 type vlan id 10
# ip link set dev ens12 master vrfgreen
# ip link show dev ens12
3: ens12: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel
master vrfgreen state UP mode DEFAULT group default qlen 1000
link/ether 52:54:00:4c:a0:45 brd ff:ff:ff:ff:ff:ff
But if the VLAN interface has previously been set into the same VRF,
then setting the interface into the VRF fails:
# ip link set dev ens12 nomaster
# ip link set dev ens12.10 master vrfgreen
# ip link show dev ens12.10
39: ens12.10@ens12: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500
qdisc noqueue master vrfgreen state UP mode DEFAULT group default
qlen 1000 link/ether 52:54:00:4c:a0:45 brd ff:ff:ff:ff:ff:ff
# ip link set dev ens12 master vrfgreen
RTNETLINK answers: File exists
The workaround is to move the VLAN interface back into the default VRF
beforehand, but it has to be shut first so as to avoid the risk of
traffic leaking from the VRF. This fix avoids needing this workaround.
Signed-off-by: Mike Manning <mmanning@att.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix trivial spelling mistake "greater then" -> "greater than".
Signed-off-by: Gal Pressman <galp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
uapi for ip_proto, sport and dport range match
in fib rules.
Signed-off-by: Roopa Prabhu <roopa@cumulusnetworks.com>
Acked-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current code uses global variables, adjusts them and passes pointer down
to devlink. With every other mlxsw_core instance, the previously passed
pointer values are rewritten. Fix this by de-globalize the variables and
also memcpy size_params during devlink resource registration.
Also, introduce a convenient size_param_init helper.
Fixes: ef3116e540 ("mlxsw: spectrum: Register KVD resources with devlink")
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix resource coverity errors.
Fixes: d9f9b9a4d0 ("devlink: Add support for resource abstraction")
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current implementation checks the combined size of the children with
the 'size' of the parent. The correct behavior is to check the combined
size vs the pending change and to compare vs the 'size_new'.
Fixes: d9f9b9a4d0 ("devlink: Add support for resource abstraction")
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Tested-by: Yuval Mintz <yuvalm@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
All kmem caches aren't reallocated once set up.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For ages iproute2 has used `struct rtmsg` as the ancillary header for
FIB rules and in the process set the protocol value to RTPROT_BOOT.
Until ca56209a66 ("net: Allow a rule to track originating protocol")
the kernel rules code ignored the protocol value sent from userspace
and always returned 0 in notifications. To avoid incompatibility with
existing iproute2, send the protocol as a new attribute.
Fixes: cac56209a6 ("net: Allow a rule to track originating protocol")
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
pfifo_fast got percpu stats lately, uncovering a bug I introduced last
year in linux-4.10.
I missed the fact that we have to clear our temporary storage
before calling __gnet_stats_copy_basic() in the case of percpu stats.
Without this fix, rate estimators (tc qd replace dev xxx root est 1sec
4sec pfifo_fast) are utterly broken.
Fixes: 1c0d32fde5 ("net_sched: gen_estimator: complete rewrite of rate estimators")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The only user of this variable is inside of an #ifdef, causing
a warning without CONFIG_INET:
net/core/filter.c: In function '____bpf_sock_ops_cb_flags_set':
net/core/filter.c:3382:6: error: unused variable 'val' [-Werror=unused-variable]
int val = argval & BPF_SOCK_OPS_ALL_CB_FLAGS;
This replaces the #ifdef with a nicer IS_ENABLED() check that
makes the code more readable and avoids the warning.
Fixes: b13d880721 ("bpf: Adds field bpf_sock_ops_cb_flags to tcp_sock")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Allow a rule that is being added/deleted/modified or
dumped to contain the originating protocol's id.
The protocol is handled just like a routes originating
protocol is. This is especially useful because there
is starting to be a plethora of different user space
programs adding rules.
Allow the vrf device to specify that the kernel is the originator
of the rule created for this device.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Oleksandr Natalenko reported performance issues with BBR without FQ
packet scheduler that were root caused to lack of SG and GSO/TSO on
his configuration.
In this mode, TCP internal pacing has to setup a high resolution timer
for each MSS sent.
We could implement in TCP a strategy similar to the one adopted
in commit fefa569a9d ("net_sched: sch_fq: account for schedule/timers drifts")
or decide to finally switch TCP stack to a GSO only mode.
This has many benefits :
1) Most TCP developments are done with TSO in mind.
2) Less high-resolution timers needs to be armed for TCP-pacing
3) GSO can benefit of xmit_more hint
4) Receiver GRO is more effective (as if TSO was used for real on sender)
-> Lower ACK traffic
5) Write queues have less overhead (one skb holds about 64KB of payload)
6) SACK coalescing just works.
7) rtx rb-tree contains less packets, SACK is cheaper.
This patch implements the minimum patch, but we can remove some legacy
code as follow ups.
Tested:
On 40Gbit link, one netperf -t TCP_STREAM
BBR+fq:
sg on: 26 Gbits/sec
sg off: 15.7 Gbits/sec (was 2.3 Gbit before patch)
BBR+pfifo_fast:
sg on: 24.2 Gbits/sec
sg off: 14.9 Gbits/sec (was 0.66 Gbit before patch !!! )
BBR+fq_codel:
sg on: 24.4 Gbits/sec
sg off: 15 Gbits/sec (was 0.66 Gbit before patch !!! )
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently the size validation is done via a cb, which is unneeded. The
size validation can be moved to core. The next patch will perform cleanup.
Signed-off-by: Arkadi Sharshevsky <arkadis@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When llist_add() returns false, cleanup_net() hasn't made its
llist_del_all(), while the work has already been scheduled
by the first queuer. So, we may skip queue_work() in this case.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This simplifies cleanup queueing and makes cleanup lists
to use llist primitives. Since llist has its own cmpxchg()
ordering, cleanup_list_lock is not more need.
Also, struct llist_node is smaller, than struct list_head,
so we save some bytes in struct net with this patch.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We take net_mutex, when there are !async pernet_operations
registered, and read locking of net_sem is not enough. But
we may get rid of taking the mutex, and just change the logic
to write lock net_sem in such cases. This obviously reduces
the number of lock operations, we do.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
allow the application to set SO_ZEROCOPY on the underlying sk
of a PF_RDS socket
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
RDS would like to use the helper functions for managing pinned pages
added by Commit a91dbff551 ("sock: ulimit on MSG_ZEROCOPY pages")
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
netif_set_real_num_tx_queues() can be called when netdev is up.
That usually happens when user requests change of number of
channels/rings with ethtool -L. The procedure for changing
the number of queues involves resetting the qdiscs and setting
dev->num_tx_queues to the new value. When the new value is
lower than the old one, extra care has to be taken to ensure
ordering of accesses to the number of queues vs qdisc reset.
Currently the queues are reset before new dev->num_tx_queues
is assigned, leaving a window of time where packets can be
enqueued onto the queues going down, leading to a likely
crash in the drivers, since most drivers don't check if TX
skbs are assigned to an active queue.
Fixes: e6484930d7 ("net: allocate tx queues in register_netdevice")
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
These pernet operations just create and destroy netlink
socket. The socket is pernet and else operations don't
touch it.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
These pernet operations consist of exit() and exit_batch() methods.
default_device_exit() moves not-local and virtual devices to init_net.
There is nothing exciting, because this may happen in any time
on a working system, and rtnl_lock() and synchronize_net() protect
us from all cases of external dereference.
The same for default_device_exit_batch(). Similar unregisteration
may happen in any time on a system. Here several lists (like todo_list),
which are accessed under rtnl_lock(). After rtnl_unlock() and
netdev_run_todo() all the devices are flushed.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
These pernet_operations register and destroy sysctl
directory, and it's not interesting for foreign
pernet_operations.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Both of them create and initialize lists, which are not touched
by another foreing pernet_operations.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are:
1)dev_proc_ops and dev_mc_net_ops, which create and destroy
pernet proc file and not interesting for another net namespaces;
2)netdev_net_ops, which creates pernet hashes, which are not
touched by another pernet_operations.
So, make them async.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch starts to convert pernet_subsys, registered
from subsys initcalls.
It seems safe to be executed in parallel with others,
as it's only creates/destoyes proc entry,
which nobody else is not interested in.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
rtnetlink_net_init() and rtnetlink_net_exit()
create and destroy netlink socket net::rtnl.
The socket is used to send rtnl notification via
rtnl_net_notifyid(). There is no a problem
to create and destroy it in parallel with other
pernet operations, as we link net in setup_net()
after the socket is created, and destroy
in cleanup_net() after net is unhashed from all
the lists and there is no RCU references on it.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net_defaults_ops introduce only net_defaults_init_net method,
and it acts on net::core::sysctl_somaxconn, which
is not interesting for the rest of pernet_subsys and
pernet_device lists. Then, make them async.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net_inuse_ops methods expose statistics in /proc.
No one from the rest of pernet_subsys or pernet_device
lists touch net::core::inuse.
So, it's safe to make net_inuse_ops async.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch starts to convert pernet_subsys, registered
from pure initcalls.
net_ns_ops::net_ns_net_init/net_ns_net_init, methods use only
ida_simple_* functions, which are not need a synchronization.
They are synchronized by idr subsystem.
So, net_ns_ops methods are able to be executed
in parallel with methods of other pernet operations.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds new pernet_operations::async flag to indicate operations,
which ->init(), ->exit() and ->exit_batch() methods are allowed
to be executed in parallel with the methods of any other pernet_operations.
When there are only asynchronous pernet_operations in the system,
net_mutex won't be taken for a net construction and destruction.
Also, remove BUG_ON(mutex_is_locked()) from net_assign_generic()
without replacing with the equivalent net_sem check, as there is
one more lockdep assert below.
v3: Add comment near net_mutex.
Suggested-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net_sem protects from pernet_list changing, while
ops_free_list() makes simple kfree(), and it can't
race with other pernet_operations callbacks.
So we may release net_mutex earlier then it was.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the mutex is mostly used to protect pernet operations
list. It orders setup_net() and cleanup_net() with parallel
{un,}register_pernet_operations() calls, so ->exit{,batch} methods
of the same pernet operations are executed for a dying net, as
were used to call ->init methods, even after the net namespace
is unlinked from net_namespace_list in cleanup_net().
But there are several problems with scalability. The first one
is that more than one net can't be created or destroyed
at the same moment on the node. For big machines with many cpus
running many containers it's very sensitive.
The second one is that it's need to synchronize_rcu() after net
is removed from net_namespace_list():
Destroy net_ns:
cleanup_net()
mutex_lock(&net_mutex)
list_del_rcu(&net->list)
synchronize_rcu() <--- Sleep there for ages
list_for_each_entry_reverse(ops, &pernet_list, list)
ops_exit_list(ops, &net_exit_list)
list_for_each_entry_reverse(ops, &pernet_list, list)
ops_free_list(ops, &net_exit_list)
mutex_unlock(&net_mutex)
This primitive is not fast, especially on the systems with many processors
and/or when preemptible RCU is enabled in config. So, all the time, while
cleanup_net() is waiting for RCU grace period, creation of new net namespaces
is not possible, the tasks, who makes it, are sleeping on the same mutex:
Create net_ns:
copy_net_ns()
mutex_lock_killable(&net_mutex) <--- Sleep there for ages
I observed 20-30 seconds hangs of "unshare -n" on ordinary 8-cpu laptop
with preemptible RCU enabled after CRIU tests round is finished.
The solution is to convert net_mutex to the rw_semaphore and add fine grain
locks to really small number of pernet_operations, what really need them.
Then, pernet_operations::init/::exit methods, modifying the net-related data,
will require down_read() locking only, while down_write() will be used
for changing pernet_list (i.e., when modules are being loaded and unloaded).
This gives signify performance increase, after all patch set is applied,
like you may see here:
%for i in {1..10000}; do unshare -n bash -c exit; done
*before*
real 1m40,377s
user 0m9,672s
sys 0m19,928s
*after*
real 0m17,007s
user 0m5,311s
sys 0m11,779
(5.8 times faster)
This patch starts replacing net_mutex to net_sem. It adds rw_semaphore,
describes the variables it protects, and makes to use, where appropriate.
net_mutex is still present, and next patches will kick it out step-by-step.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Line up destructors actions in the revers order
to constructors. Next patches will add more actions,
and this will be comfortable, if there is the such
order.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
