The more sockets we have in the hash table, the longer we spend looking
up the socket. While running a number of small workloads on the same
host, they penalise each other and cause performance degradation.
The root cause might be a single workload that consumes much more
resources than the others. It often happens on a cloud service where
different workloads share the same computing resource.
On EC2 c5.24xlarge instance (196 GiB memory and 524288 (1Mi / 2) ehash
entries), after running iperf3 in different netns, creating 24Mi sockets
without data transfer in the root netns causes about 10% performance
regression for the iperf3's connection.
thash_entries sockets length Gbps
524288 1 1 50.7
24Mi 48 45.1
It is basically related to the length of the list of each hash bucket.
For testing purposes to see how performance drops along the length,
I set 131072 (1Mi / 8) to thash_entries, and here's the result.
thash_entries sockets length Gbps
131072 1 1 50.7
1Mi 8 49.9
2Mi 16 48.9
4Mi 32 47.3
8Mi 64 44.6
16Mi 128 40.6
24Mi 192 36.3
32Mi 256 32.5
40Mi 320 27.0
48Mi 384 25.0
To resolve the socket lookup degradation, we introduce an optional
per-netns hash table for TCP, but it's just ehash, and we still share
the global bhash, bhash2 and lhash2.
With a smaller ehash, we can look up non-listener sockets faster and
isolate such noisy neighbours. In addition, we can reduce lock contention.
We can control the ehash size by a new sysctl knob. However, depending
on workloads, it will require very sensitive tuning, so we disable the
feature by default (net.ipv4.tcp_child_ehash_entries == 0). Moreover,
we can fall back to using the global ehash in case we fail to allocate
enough memory for a new ehash. The maximum size is 16Mi, which is large
enough that even if we have 48Mi sockets, the average list length is 3,
and regression would be less than 1%.
We can check the current ehash size by another read-only sysctl knob,
net.ipv4.tcp_ehash_entries. A negative value means the netns shares
the global ehash (per-netns ehash is disabled or failed to allocate
memory).
# dmesg | cut -d ' ' -f 5- | grep "established hash"
TCP established hash table entries: 524288 (order: 10, 4194304 bytes, vmalloc hugepage)
# sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = 524288 # can be changed by thash_entries
# sysctl net.ipv4.tcp_child_ehash_entries
net.ipv4.tcp_child_ehash_entries = 0 # disabled by default
# ip netns add test1
# ip netns exec test1 sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = -524288 # share the global ehash
# sysctl -w net.ipv4.tcp_child_ehash_entries=100
net.ipv4.tcp_child_ehash_entries = 100
# ip netns add test2
# ip netns exec test2 sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = 128 # own a per-netns ehash with 2^n buckets
When more than two processes in the same netns create per-netns ehash
concurrently with different sizes, we need to guarantee the size in
one of the following ways:
1) Share the global ehash and create per-netns ehash
First, unshare() with tcp_child_ehash_entries==0. It creates dedicated
netns sysctl knobs where we can safely change tcp_child_ehash_entries
and clone()/unshare() to create a per-netns ehash.
2) Control write on sysctl by BPF
We can use BPF_PROG_TYPE_CGROUP_SYSCTL to allow/deny read/write on
sysctl knobs.
Note that the global ehash allocated at the boot time is spread over
available NUMA nodes, but inet_pernet_hashinfo_alloc() will allocate
pages for each per-netns ehash depending on the current process's NUMA
policy. By default, the allocation is done in the local node only, so
the per-netns hash table could fully reside on a random node. Thus,
depending on the NUMA policy the netns is created with and the CPU the
current thread is running on, we could see some performance differences
for highly optimised networking applications.
Note also that the default values of two sysctl knobs depend on the ehash
size and should be tuned carefully:
tcp_max_tw_buckets : tcp_child_ehash_entries / 2
tcp_max_syn_backlog : max(128, tcp_child_ehash_entries / 128)
As a bonus, we can dismantle netns faster. Currently, while destroying
netns, we call inet_twsk_purge(), which walks through the global ehash.
It can be potentially big because it can have many sockets other than
TIME_WAIT in all netns. Splitting ehash changes that situation, where
it's only necessary for inet_twsk_purge() to clean up TIME_WAIT sockets
in each netns.
With regard to this, we do not free the per-netns ehash in inet_twsk_kill()
to avoid UAF while iterating the per-netns ehash in inet_twsk_purge().
Instead, we do it in tcp_sk_exit_batch() after calling tcp_twsk_purge() to
keep it protocol-family-independent.
In the future, we could optimise ehash lookup/iteration further by removing
netns comparison for the per-netns ehash.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While destroying netns, we call inet_twsk_purge() in tcp_sk_exit_batch()
and tcpv6_net_exit_batch() for AF_INET and AF_INET6. These commands
trigger the kernel to walk through the potentially big ehash twice even
though the netns has no TIME_WAIT sockets.
# ip netns add test
# ip netns del test
or
# unshare -n /bin/true >/dev/null
When tw_refcount is 1, we need not call inet_twsk_purge() at least
for the net. We can save such unneeded iterations if all netns in
net_exit_list have no TIME_WAIT sockets. This change eliminates
the tax by the additional unshare() described in the next patch to
guarantee the per-netns ehash size.
Tested:
# mount -t debugfs none /sys/kernel/debug/
# echo cleanup_net > /sys/kernel/debug/tracing/set_ftrace_filter
# echo inet_twsk_purge >> /sys/kernel/debug/tracing/set_ftrace_filter
# echo function > /sys/kernel/debug/tracing/current_tracer
# cat ./add_del_unshare.sh
for i in `seq 1 40`
do
(for j in `seq 1 100` ; do unshare -n /bin/true >/dev/null ; done) &
done
wait;
# ./add_del_unshare.sh
Before the patch:
# cat /sys/kernel/debug/tracing/trace_pipe
kworker/u128:0-8 [031] ...1. 174.162765: cleanup_net <-process_one_work
kworker/u128:0-8 [031] ...1. 174.240796: inet_twsk_purge <-cleanup_net
kworker/u128:0-8 [032] ...1. 174.244759: inet_twsk_purge <-tcp_sk_exit_batch
kworker/u128:0-8 [034] ...1. 174.290861: cleanup_net <-process_one_work
kworker/u128:0-8 [039] ...1. 175.245027: inet_twsk_purge <-cleanup_net
kworker/u128:0-8 [046] ...1. 175.290541: inet_twsk_purge <-tcp_sk_exit_batch
kworker/u128:0-8 [037] ...1. 175.321046: cleanup_net <-process_one_work
kworker/u128:0-8 [024] ...1. 175.941633: inet_twsk_purge <-cleanup_net
kworker/u128:0-8 [025] ...1. 176.242539: inet_twsk_purge <-tcp_sk_exit_batch
After:
# cat /sys/kernel/debug/tracing/trace_pipe
kworker/u128:0-8 [038] ...1. 428.116174: cleanup_net <-process_one_work
kworker/u128:0-8 [038] ...1. 428.262532: cleanup_net <-process_one_work
kworker/u128:0-8 [030] ...1. 429.292645: cleanup_net <-process_one_work
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We will soon introduce an optional per-netns ehash.
This means we cannot use tcp_hashinfo directly in most places.
Instead, access it via net->ipv4.tcp_death_row.hashinfo.
The access will be valid only while initialising tcp_hashinfo
itself and creating/destroying each netns.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We will soon introduce an optional per-netns ehash.
This means we cannot use the global sk->sk_prot->h.hashinfo
to fetch a TCP hashinfo.
Instead, set NULL to sk->sk_prot->h.hashinfo for TCP and get
a proper hashinfo from net->ipv4.tcp_death_row.hashinfo.
Note that we need not use sk->sk_prot->h.hashinfo if DCCP is
disabled.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We will soon introduce an optional per-netns ehash and access hash
tables via net->ipv4.tcp_death_row->hashinfo instead of &tcp_hashinfo
in most places.
It could harm the fast path because dereferences of two fields in net
and tcp_death_row might incur two extra cache line misses. To save one
dereference, let's place tcp_death_row back in netns_ipv4 and fetch
hashinfo via net->ipv4.tcp_death_row"."hashinfo.
Note tcp_death_row was initially placed in netns_ipv4, and commit
fbb8295248 ("tcp: allocate tcp_death_row outside of struct netns_ipv4")
changed it to a pointer so that we can fire TIME_WAIT timers after freeing
net. However, we don't do so after commit 04c494e68a ("Revert "tcp/dccp:
get rid of inet_twsk_purge()""), so we need not define tcp_death_row as a
pointer.
Also, we move refcount_dec_and_test(&tw_refcount) from tcp_sk_exit() to
tcp_sk_exit_batch() as a debug check.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch adds no functional change and cleans up some functions
that the following patches touch around so that we make them tidy
and easy to review/revert. The changes are
- Keep reverse christmas tree order
- Remove unnecessary init of port in inet_csk_find_open_port()
- Use req_to_sk() once in reqsk_queue_unlink()
- Use sock_net(sk) once in tcp_time_wait() and tcp_v[46]_connect()
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Because per host rate limiting has been proven problematic (side channel
attacks can be based on it), per host rate limiting of challenge acks ideally
should be per netns and turned off by default.
This is a long due followup of following commits:
083ae30828 ("tcp: enable per-socket rate limiting of all 'challenge acks'")
f2b2c582e8 ("tcp: mitigate ACK loops for connections as tcp_sock")
75ff39ccc1 ("tcp: make challenge acks less predictable")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jason Baron <jbaron@akamai.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The current bind hashtable (bhash) is hashed by port only.
In the socket bind path, we have to check for bind conflicts by
traversing the specified port's inet_bind_bucket while holding the
hashbucket's spinlock (see inet_csk_get_port() and
inet_csk_bind_conflict()). In instances where there are tons of
sockets hashed to the same port at different addresses, the bind
conflict check is time-intensive and can cause softirq cpu lockups,
as well as stops new tcp connections since __inet_inherit_port()
also contests for the spinlock.
This patch adds a second bind table, bhash2, that hashes by
port and sk->sk_rcv_saddr (ipv4) and sk->sk_v6_rcv_saddr (ipv6).
Searching the bhash2 table leads to significantly faster conflict
resolution and less time holding the hashbucket spinlock.
Please note a few things:
* There can be the case where the a socket's address changes after it
has been bound. There are two cases where this happens:
1) The case where there is a bind() call on INADDR_ANY (ipv4) or
IPV6_ADDR_ANY (ipv6) and then a connect() call. The kernel will
assign the socket an address when it handles the connect()
2) In inet_sk_reselect_saddr(), which is called when rebuilding the
sk header and a few pre-conditions are met (eg rerouting fails).
In these two cases, we need to update the bhash2 table by removing the
entry for the old address, and add a new entry reflecting the updated
address.
* The bhash2 table must have its own lock, even though concurrent
accesses on the same port are protected by the bhash lock. Bhash2 must
have its own lock to protect against cases where sockets on different
ports hash to different bhash hashbuckets but to the same bhash2
hashbucket.
This brings up a few stipulations:
1) When acquiring both the bhash and the bhash2 lock, the bhash2 lock
will always be acquired after the bhash lock and released before the
bhash lock is released.
2) There are no nested bhash2 hashbucket locks. A bhash2 lock is always
acquired+released before another bhash2 lock is acquired+released.
* The bhash table cannot be superseded by the bhash2 table because for
bind requests on INADDR_ANY (ipv4) or IPV6_ADDR_ANY (ipv6), every socket
bound to that port must be checked for a potential conflict. The bhash
table is the only source of port->socket associations.
Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While reading sysctl_tcp_reflect_tos, it can be changed concurrently.
Thus, we need to add READ_ONCE() to its readers.
Fixes: ac8f1710c1 ("tcp: reflect tos value received in SYN to the socket")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Wei Wang <weiwan@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While reading sysctl_tcp_tw_reuse, it can be changed concurrently.
Thus, we need to add READ_ONCE() to its reader.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If we set XFRM security policy by calling setsockopt with option
IPV6_XFRM_POLICY, the policy will be stored in 'sock_policy' in 'sock'
struct. However tcp_v6_send_response doesn't look up dst_entry with the
actual socket but looks up with tcp control socket. This may cause a
problem that a RST packet is sent without ESP encryption & peer's TCP
socket can't receive it.
This patch will make the function look up dest_entry with actual socket,
if the socket has XFRM policy(sock_policy), so that the TCP response
packet via this function can be encrypted, & aligned on the encrypted
TCP socket.
Tested: We encountered this problem when a TCP socket which is encrypted
in ESP transport mode encryption, receives challenge ACK at SYN_SENT
state. After receiving challenge ACK, TCP needs to send RST to
establish the socket at next SYN try. But the RST was not encrypted &
peer TCP socket still remains on ESTABLISHED state.
So we verified this with test step as below.
[Test step]
1. Making a TCP state mismatch between client(IDLE) & server(ESTABLISHED).
2. Client tries a new connection on the same TCP ports(src & dst).
3. Server will return challenge ACK instead of SYN,ACK.
4. Client will send RST to server to clear the SOCKET.
5. Client will retransmit SYN to server on the same TCP ports.
[Expected result]
The TCP connection should be established.
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Steffen Klassert <steffen.klassert@secunet.com>
Cc: Sehee Lee <seheele@google.com>
Signed-off-by: Sewook Seo <sewookseo@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/microchip/sparx5/sparx5_switchdev.c
9c5de246c1 ("net: sparx5: mdb add/del handle non-sparx5 devices")
fbb89d02e3 ("net: sparx5: Allow mdb entries to both CPU and ports")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When the third packet of 3WHS connection establishment
contains payload, it is added into socket receive queue
without the XFRM check and the drop of connection tracking
context.
This means that if the data is left unread in the socket
receive queue, conntrack module can not be unloaded.
As most applications usually reads the incoming data
immediately after accept(), bug has been hiding for
quite a long time.
Commit 68822bdf76 ("net: generalize skb freeing
deferral to per-cpu lists") exposed this bug because
even if the application reads this data, the skb
with nfct state could stay in a per-cpu cache for
an arbitrary time, if said cpu no longer process RX softirqs.
Many thanks to Ilya Maximets for reporting this issue,
and for testing various patches:
https://lore.kernel.org/netdev/20220619003919.394622-1-i.maximets@ovn.org/
Note that I also added a missing xfrm4_policy_check() call,
although this is probably not a big issue, as the SYN
packet should have been dropped earlier.
Fixes: b59c270104 ("[NETFILTER]: Keep conntrack reference until IPsec policy checks are done")
Reported-by: Ilya Maximets <i.maximets@ovn.org>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Florian Westphal <fw@strlen.de>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>
Cc: Steffen Klassert <steffen.klassert@secunet.com>
Tested-by: Ilya Maximets <i.maximets@ovn.org>
Reviewed-by: Ilya Maximets <i.maximets@ovn.org>
Link: https://lore.kernel.org/r/20220623050436.1290307-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Each protocol having a ->memory_allocated pointer gets a corresponding
per-cpu reserve, that following patches will use.
Instead of having reserved bytes per socket,
we want to have per-cpu reserves.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Fix the following build warning when CONFIG_IPV6 is not set:
In function ‘fortify_memcpy_chk’,
inlined from ‘tcp_md5_do_add’ at net/ipv4/tcp_ipv4.c:1210:2:
./include/linux/fortify-string.h:328:4: error: call to ‘__write_overflow_field’ declared with attribute warning: detected write beyond size of field (1st parameter); maybe use struct_group()? [-Werror=attribute-warning]
328 | __write_overflow_field(p_size_field, size);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Suggested-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: huhai <huhai@kylinos.cn>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/r/20220526101213.2392980-1-zhanggenjian@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The 'drop_reason' that passed to kfree_skb_reason() in tcp_v4_rcv()
and tcp_v6_rcv() can be SKB_NOT_DROPPED_YET(0), as it is used as the
return value of tcp_inbound_md5_hash().
And it can panic the kernel with NULL pointer in
net_dm_packet_report_size() if the reason is 0, as drop_reasons[0]
is NULL.
Fixes: 1330b6ef33 ("skb: make drop reason booleanable")
Reviewed-by: Jiang Biao <benbjiang@tencent.com>
Reviewed-by: Hao Peng <flyingpeng@tencent.com>
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts commits:
0dad4087a8 ("tcp/dccp: get rid of inet_twsk_purge()")
d507204d3c ("tcp/dccp: add tw->tw_bslot")
As Leonard pointed out, a newly allocated netns can happen
to reuse a freed 'struct net'.
While TCP TW timers were covered by my patches, other things were not:
1) Lookups in rx path (INET_MATCH() and INET6_MATCH()), as they look
at 4-tuple plus the 'struct net' pointer.
2) /proc/net/tcp[6] and inet_diag, same reason.
3) hashinfo->bhash[], same reason.
Fixing all this seems risky, lets instead revert.
In the future, we might have a per netns tcp hash table, or
a per netns list of timewait sockets...
Fixes: 0dad4087a8 ("tcp/dccp: get rid of inet_twsk_purge()")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Leonard Crestez <cdleonard@gmail.com>
Tested-by: Leonard Crestez <cdleonard@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The listen sk is currently stored in two hash tables,
listening_hash (hashed by port) and lhash2 (hashed by port and address).
After commit 0ee58dad5b ("net: tcp6: prefer listeners bound to an address")
and commit d9fbc7f643 ("net: tcp: prefer listeners bound to an address"),
the TCP-SYN lookup fast path does not use listening_hash.
The commit 05c0b35709 ("tcp: seq_file: Replace listening_hash with lhash2")
also moved the seq_file (/proc/net/tcp) iteration usage from
listening_hash to lhash2.
There are still a few listening_hash usages left.
One of them is inet_reuseport_add_sock() which uses the listening_hash
to search a listen sk during the listen() system call. This turns
out to be very slow on use cases that listen on many different
VIPs at a popular port (e.g. 443). [ On top of the slowness in
adding to the tail in the IPv6 case ]. The latter patch has a
selftest to demonstrate this case.
This patch takes this chance to move all remaining listening_hash
usages to lhash2 and then retire listening_hash.
Since most changes need to be done together, it is hard to cut
the listening_hash to lhash2 switch into small patches. The
changes in this patch is highlighted here for the review
purpose.
1. Because of the listening_hash removal, lhash2 can use the
sk->sk_nulls_node instead of the icsk->icsk_listen_portaddr_node.
This will also keep the sk_unhashed() check to work as is
after stop adding sk to listening_hash.
The union is removed from inet_listen_hashbucket because
only nulls_head is needed.
2. icsk->icsk_listen_portaddr_node and its helpers are removed.
3. The current lhash2 users needs to iterate with sk_nulls_node
instead of icsk_listen_portaddr_node.
One case is in the inet[6]_lhash2_lookup().
Another case is the seq_file iterator in tcp_ipv4.c.
One thing to note is sk_nulls_next() is needed
because the old inet_lhash2_for_each_icsk_continue()
does a "next" first before iterating.
4. Move the remaining listening_hash usage to lhash2
inet_reuseport_add_sock() which this series is
trying to improve.
inet_diag.c and mptcp_diag.c are the final two
remaining use cases and is moved to lhash2 now also.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Logic added in commit f35f821935 ("tcp: defer skb freeing after socket
lock is released") helped bulk TCP flows to move the cost of skbs
frees outside of critical section where socket lock was held.
But for RPC traffic, or hosts with RFS enabled, the solution is far from
being ideal.
For RPC traffic, recvmsg() has to return to user space right after
skb payload has been consumed, meaning that BH handler has no chance
to pick the skb before recvmsg() thread. This issue is more visible
with BIG TCP, as more RPC fit one skb.
For RFS, even if BH handler picks the skbs, they are still picked
from the cpu on which user thread is running.
Ideally, it is better to free the skbs (and associated page frags)
on the cpu that originally allocated them.
This patch removes the per socket anchor (sk->defer_list) and
instead uses a per-cpu list, which will hold more skbs per round.
This new per-cpu list is drained at the end of net_action_rx(),
after incoming packets have been processed, to lower latencies.
In normal conditions, skbs are added to the per-cpu list with
no further action. In the (unlikely) cases where the cpu does not
run net_action_rx() handler fast enough, we use an IPI to raise
NET_RX_SOFTIRQ on the remote cpu.
Also, we do not bother draining the per-cpu list from dev_cpu_dead()
This is because skbs in this list have no requirement on how fast
they should be freed.
Note that we can add in the future a small per-cpu cache
if we see any contention on sd->defer_lock.
Tested on a pair of hosts with 100Gbit NIC, RFS enabled,
and /proc/sys/net/ipv4/tcp_rmem[2] tuned to 16MB to work around
page recycling strategy used by NIC driver (its page pool capacity
being too small compared to number of skbs/pages held in sockets
receive queues)
Note that this tuning was only done to demonstrate worse
conditions for skb freeing for this particular test.
These conditions can happen in more general production workload.
10 runs of one TCP_STREAM flow
Before:
Average throughput: 49685 Mbit.
Kernel profiles on cpu running user thread recvmsg() show high cost for
skb freeing related functions (*)
57.81% [kernel] [k] copy_user_enhanced_fast_string
(*) 12.87% [kernel] [k] skb_release_data
(*) 4.25% [kernel] [k] __free_one_page
(*) 3.57% [kernel] [k] __list_del_entry_valid
1.85% [kernel] [k] __netif_receive_skb_core
1.60% [kernel] [k] __skb_datagram_iter
(*) 1.59% [kernel] [k] free_unref_page_commit
(*) 1.16% [kernel] [k] __slab_free
1.16% [kernel] [k] _copy_to_iter
(*) 1.01% [kernel] [k] kfree
(*) 0.88% [kernel] [k] free_unref_page
0.57% [kernel] [k] ip6_rcv_core
0.55% [kernel] [k] ip6t_do_table
0.54% [kernel] [k] flush_smp_call_function_queue
(*) 0.54% [kernel] [k] free_pcppages_bulk
0.51% [kernel] [k] llist_reverse_order
0.38% [kernel] [k] process_backlog
(*) 0.38% [kernel] [k] free_pcp_prepare
0.37% [kernel] [k] tcp_recvmsg_locked
(*) 0.37% [kernel] [k] __list_add_valid
0.34% [kernel] [k] sock_rfree
0.34% [kernel] [k] _raw_spin_lock_irq
(*) 0.33% [kernel] [k] __page_cache_release
0.33% [kernel] [k] tcp_v6_rcv
(*) 0.33% [kernel] [k] __put_page
(*) 0.29% [kernel] [k] __mod_zone_page_state
0.27% [kernel] [k] _raw_spin_lock
After patch:
Average throughput: 73076 Mbit.
Kernel profiles on cpu running user thread recvmsg() looks better:
81.35% [kernel] [k] copy_user_enhanced_fast_string
1.95% [kernel] [k] _copy_to_iter
1.95% [kernel] [k] __skb_datagram_iter
1.27% [kernel] [k] __netif_receive_skb_core
1.03% [kernel] [k] ip6t_do_table
0.60% [kernel] [k] sock_rfree
0.50% [kernel] [k] tcp_v6_rcv
0.47% [kernel] [k] ip6_rcv_core
0.45% [kernel] [k] read_tsc
0.44% [kernel] [k] _raw_spin_lock_irqsave
0.37% [kernel] [k] _raw_spin_lock
0.37% [kernel] [k] native_irq_return_iret
0.33% [kernel] [k] __inet6_lookup_established
0.31% [kernel] [k] ip6_protocol_deliver_rcu
0.29% [kernel] [k] tcp_rcv_established
0.29% [kernel] [k] llist_reverse_order
v2: kdoc issue (kernel bots)
do not defer if (alloc_cpu == smp_processor_id()) (Paolo)
replace the sk_buff_head with a single-linked list (Jakub)
add a READ_ONCE()/WRITE_ONCE() for the lockless read of sd->defer_list
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20220422201237.416238-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Now that ip_rt_fix_tos() doesn't reset ->flowi4_scope unconditionally,
we don't have to rely on the RTO_ONLINK bit to properly set the scope
of a flowi4 structure. We can just set ->flowi4_scope explicitly and
avoid using RTO_ONLINK in ->flowi4_tos.
This patch converts callers of ip_route_connect(). Instead of setting
the tos parameter with RT_CONN_FLAGS(sk), as all callers do, we can:
1- Drop the tos parameter from ip_route_connect(): its value was
entirely based on sk, which is also passed as parameter.
2- Set ->flowi4_scope depending on the SOCK_LOCALROUTE socket option
instead of always initialising it with RT_SCOPE_UNIVERSE (let's
define ip_sock_rt_scope() for this purpose).
3- Avoid overloading ->flowi4_tos with RTO_ONLINK: since the scope is
now properly initialised, we don't need to tell ip_rt_fix_tos() to
adjust ->flowi4_scope for us. So let's define ip_sock_rt_tos(),
which is the same as RT_CONN_FLAGS() but without the RTO_ONLINK
bit overload.
Note:
In the original ip_route_connect() code, __ip_route_output_key()
might clear the RTO_ONLINK bit of fl4->flowi4_tos (because of
ip_rt_fix_tos()). Therefore flowi4_update_output() had to reuse the
original tos variable. Now that we don't set RTO_ONLINK any more,
this is not a problem and we can use fl4->flowi4_tos in
flowi4_update_output().
Signed-off-by: Guillaume Nault <gnault@redhat.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
We had various bugs over the years with code
breaking the assumption that tp->snd_cwnd is greater
than zero.
Lately, syzbot reported the WARN_ON_ONCE(!tp->prior_cwnd) added
in commit 8b8a321ff7 ("tcp: fix zero cwnd in tcp_cwnd_reduction")
can trigger, and without a repro we would have to spend
considerable time finding the bug.
Instead of complaining too late, we want to catch where
and when tp->snd_cwnd is set to an illegal value.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Suggested-by: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Link: https://lore.kernel.org/r/20220405233538.947344-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Back when tcp_tso_autosize() and TCP pacing were introduced,
our focus was really to reduce burst sizes for long distance
flows.
The simple heuristic of using sk_pacing_rate/1024 has worked
well, but can lead to too small packets for hosts in the same
rack/cluster, when thousands of flows compete for the bottleneck.
Neal Cardwell had the idea of making the TSO burst size
a function of both sk_pacing_rate and tcp_min_rtt()
Indeed, for local flows, sending bigger bursts is better
to reduce cpu costs, as occasional losses can be repaired
quite fast.
This patch is based on Neal Cardwell implementation
done more than two years ago.
bbr is adjusting max_pacing_rate based on measured bandwidth,
while cubic would over estimate max_pacing_rate.
/proc/sys/net/ipv4/tcp_tso_rtt_log can be used to tune or disable
this new feature, in logarithmic steps.
Tested:
100Gbit NIC, two hosts in the same rack, 4K MTU.
600 flows rate-limited to 20000000 bytes per second.
Before patch: (TSO sizes would be limited to 20000000/1024/4096 -> 4 segments per TSO)
~# echo 0 >/proc/sys/net/ipv4/tcp_tso_rtt_log
~# nstat -n;perf stat ./super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000;nstat|egrep "TcpInSegs|TcpOutSegs|TcpRetransSegs|Delivered"
96005
Performance counter stats for './super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000':
65,945.29 msec task-clock # 2.845 CPUs utilized
1,314,632 context-switches # 19935.279 M/sec
5,292 cpu-migrations # 80.249 M/sec
940,641 page-faults # 14264.023 M/sec
201,117,030,926 cycles # 3049769.216 GHz (83.45%)
17,699,435,405 stalled-cycles-frontend # 8.80% frontend cycles idle (83.48%)
136,584,015,071 stalled-cycles-backend # 67.91% backend cycles idle (83.44%)
53,809,530,436 instructions # 0.27 insn per cycle
# 2.54 stalled cycles per insn (83.36%)
9,062,315,523 branches # 137422329.563 M/sec (83.22%)
153,008,621 branch-misses # 1.69% of all branches (83.32%)
23.182970846 seconds time elapsed
TcpInSegs 15648792 0.0
TcpOutSegs 58659110 0.0 # Average of 3.7 4K segments per TSO packet
TcpExtTCPDelivered 58654791 0.0
TcpExtTCPDeliveredCE 19 0.0
After patch:
~# echo 9 >/proc/sys/net/ipv4/tcp_tso_rtt_log
~# nstat -n;perf stat ./super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000;nstat|egrep "TcpInSegs|TcpOutSegs|TcpRetransSegs|Delivered"
96046
Performance counter stats for './super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000':
48,982.58 msec task-clock # 2.104 CPUs utilized
186,014 context-switches # 3797.599 M/sec
3,109 cpu-migrations # 63.472 M/sec
941,180 page-faults # 19214.814 M/sec
153,459,763,868 cycles # 3132982.807 GHz (83.56%)
12,069,861,356 stalled-cycles-frontend # 7.87% frontend cycles idle (83.32%)
120,485,917,953 stalled-cycles-backend # 78.51% backend cycles idle (83.24%)
36,803,672,106 instructions # 0.24 insn per cycle
# 3.27 stalled cycles per insn (83.18%)
5,947,266,275 branches # 121417383.427 M/sec (83.64%)
87,984,616 branch-misses # 1.48% of all branches (83.43%)
23.281200256 seconds time elapsed
TcpInSegs 1434706 0.0
TcpOutSegs 58883378 0.0 # Average of 41 4K segments per TSO packet
TcpExtTCPDelivered 58878971 0.0
TcpExtTCPDeliveredCE 9664 0.0
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Link: https://lore.kernel.org/r/20220309015757.2532973-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We have a number of cases where function returns drop/no drop
decision as a boolean. Now that we want to report the reason
code as well we have to pass extra output arguments.
We can make the reason code evaluate correctly as bool.
I believe we're good to reorder the reasons as they are
reported to user space as strings.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The functions do essentially the same work to verify TCP-MD5 sign.
Code can be merged into one family-independent function in order to
reduce copy'n'paste and generated code.
Later with TCP-AO option added, this will allow to create one function
that's responsible for segment verification, that will have all the
different checks for MD5/AO/non-signed packets, which in turn will help
to see checks for all corner-cases in one function, rather than spread
around different families and functions.
Cc: Eric Dumazet <edumazet@google.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/r/20220223175740.452397-1-dima@arista.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace kfree_skb() used in tcp_v4_do_rcv() and tcp_v6_do_rcv() with
kfree_skb_reason().
Reviewed-by: Mengen Sun <mengensun@tencent.com>
Reviewed-by: Hao Peng <flyingpeng@tencent.com>
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pass the address of drop_reason to tcp_add_backlog() to store the
reasons for skb drops when fails. Following drop reasons are
introduced:
SKB_DROP_REASON_SOCKET_BACKLOG
Reviewed-by: Mengen Sun <mengensun@tencent.com>
Reviewed-by: Hao Peng <flyingpeng@tencent.com>
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pass the address of drop reason to tcp_v4_inbound_md5_hash() and
tcp_v6_inbound_md5_hash() to store the reasons for skb drops when this
function fails. Therefore, the drop reason can be passed to
kfree_skb_reason() when the skb needs to be freed.
Following drop reasons are added:
SKB_DROP_REASON_TCP_MD5NOTFOUND
SKB_DROP_REASON_TCP_MD5UNEXPECTED
SKB_DROP_REASON_TCP_MD5FAILURE
SKB_DROP_REASON_TCP_MD5* above correspond to LINUX_MIB_TCPMD5*
Reviewed-by: Mengen Sun <mengensun@tencent.com>
Reviewed-by: Hao Peng <flyingpeng@tencent.com>
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use kfree_skb_reason() for some path in tcp_v4_rcv() that missed before,
including:
SKB_DROP_REASON_SOCKET_FILTER
SKB_DROP_REASON_XFRM_POLICY
Reviewed-by: Mengen Sun <mengensun@tencent.com>
Reviewed-by: Hao Peng <flyingpeng@tencent.com>
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current release - new code bugs:
- tcp: add a missing sk_defer_free_flush() in tcp_splice_read()
- tcp: add a stub for sk_defer_free_flush(), fix CONFIG_INET=n
- nf_tables: set last expression in register tracking area
- nft_connlimit: fix memleak if nf_ct_netns_get() fails
- mptcp: fix removing ids bitmap setting
- bonding: use rcu_dereference_rtnl when getting active slave
- fix three cases of sleep in atomic context in drivers: lan966x, gve
- handful of build fixes for esoteric drivers after netdev->dev_addr
was made const
Previous releases - regressions:
- revert "ipv6: Honor all IPv6 PIO Valid Lifetime values", it broke
Linux compatibility with USGv6 tests
- procfs: show net device bound packet types
- ipv4: fix ip option filtering for locally generated fragments
- phy: broadcom: hook up soft_reset for BCM54616S
Previous releases - always broken:
- ipv4: raw: lock the socket in raw_bind()
- ipv4: decrease the use of shared IPID generator to decrease the
chance of attackers guessing the values
- procfs: fix cross-netns information leakage in /proc/net/ptype
- ethtool: fix link extended state for big endian
- bridge: vlan: fix single net device option dumping
- ping: fix the sk_bound_dev_if match in ping_lookup
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
-----BEGIN PGP SIGNATURE-----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=mwR3
-----END PGP SIGNATURE-----
Merge tag 'net-5.17-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net
Pull networking fixes from Jakub Kicinski:
"Including fixes from netfilter and can.
Current release - new code bugs:
- tcp: add a missing sk_defer_free_flush() in tcp_splice_read()
- tcp: add a stub for sk_defer_free_flush(), fix CONFIG_INET=n
- nf_tables: set last expression in register tracking area
- nft_connlimit: fix memleak if nf_ct_netns_get() fails
- mptcp: fix removing ids bitmap setting
- bonding: use rcu_dereference_rtnl when getting active slave
- fix three cases of sleep in atomic context in drivers: lan966x, gve
- handful of build fixes for esoteric drivers after netdev->dev_addr
was made const
Previous releases - regressions:
- revert "ipv6: Honor all IPv6 PIO Valid Lifetime values", it broke
Linux compatibility with USGv6 tests
- procfs: show net device bound packet types
- ipv4: fix ip option filtering for locally generated fragments
- phy: broadcom: hook up soft_reset for BCM54616S
Previous releases - always broken:
- ipv4: raw: lock the socket in raw_bind()
- ipv4: decrease the use of shared IPID generator to decrease the
chance of attackers guessing the values
- procfs: fix cross-netns information leakage in /proc/net/ptype
- ethtool: fix link extended state for big endian
- bridge: vlan: fix single net device option dumping
- ping: fix the sk_bound_dev_if match in ping_lookup"
* tag 'net-5.17-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (86 commits)
net: bridge: vlan: fix memory leak in __allowed_ingress
net: socket: rename SKB_DROP_REASON_SOCKET_FILTER
ipv4: remove sparse error in ip_neigh_gw4()
ipv4: avoid using shared IP generator for connected sockets
ipv4: tcp: send zero IPID in SYNACK messages
ipv4: raw: lock the socket in raw_bind()
MAINTAINERS: add missing IPv4/IPv6 header paths
MAINTAINERS: add more files to eth PHY
net: stmmac: dwmac-sun8i: use return val of readl_poll_timeout()
net: bridge: vlan: fix single net device option dumping
net: stmmac: skip only stmmac_ptp_register when resume from suspend
net: stmmac: configure PTP clock source prior to PTP initialization
Revert "ipv6: Honor all IPv6 PIO Valid Lifetime values"
connector/cn_proc: Use task_is_in_init_pid_ns()
pid: Introduce helper task_is_in_init_pid_ns()
gve: Fix GFP flags when allocing pages
net: lan966x: Fix sleep in atomic context when updating MAC table
net: lan966x: Fix sleep in atomic context when injecting frames
ethernet: seeq/ether3: don't write directly to netdev->dev_addr
ethernet: 8390/etherh: don't write directly to netdev->dev_addr
...
Rename SKB_DROP_REASON_SOCKET_FILTER, which is used
as the reason of skb drop out of socket filter before
it's part of a released kernel. It will be used for
more protocols than just TCP in future series.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/all/20220127091308.91401-2-imagedong@tencent.com/
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
TCP ipv4 uses per-cpu/per-netns ctl sockets in order to send
RST and some ACK packets (on behalf of TIMEWAIT sockets).
This adds memory and cpu costs, which do not seem needed.
Now typical servers have 256 or more cores, this adds considerable
tax to netns users.
tcp sockets are used from BH context, are not receiving packets,
and do not store any persistent state but the 'struct net' pointer
in order to be able to use IPv4 output functions.
Note that I attempted a related change in the past, that had
to be hot-fixed in commit bdbbb8527b ("ipv4: tcp: get rid of ugly unicast_sock")
This patch could very well surface old bugs, on layers not
taking care of sk->sk_kern_sock properly.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Prior patches in the series made sure tw_timer_handler()
can be fired after netns has been dismantled/freed.
We no longer have to scan a potentially big TCP ehash
table at netns dismantle.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove PDE_DATA() completely and replace it with pde_data().
[akpm@linux-foundation.org: fix naming clash in drivers/nubus/proc.c]
[akpm@linux-foundation.org: now fix it properly]
Link: https://lkml.kernel.org/r/20211124081956.87711-2-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Christian Brauner <christian.brauner@ubuntu.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Alexey Gladkov <gladkov.alexey@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Replace kfree_skb() with kfree_skb_reason() in tcp_v4_rcv(). Following
drop reasons are added:
SKB_DROP_REASON_NO_SOCKET
SKB_DROP_REASON_PKT_TOO_SMALL
SKB_DROP_REASON_TCP_CSUM
SKB_DROP_REASON_TCP_FILTER
After this patch, 'kfree_skb' event will print message like this:
$ TASK-PID CPU# ||||| TIMESTAMP FUNCTION
$ | | | ||||| | |
<idle>-0 [000] ..s1. 36.113438: kfree_skb: skbaddr=(____ptrval____) protocol=2048 location=(____ptrval____) reason: NO_SOCKET
The reason of skb drop is printed too.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The return value of BPF_CGROUP_RUN_PROG_INET{4,6}_POST_BIND() in
__inet_bind() is not handled properly. While the return value
is non-zero, it will set inet_saddr and inet_rcv_saddr to 0 and
exit:
err = BPF_CGROUP_RUN_PROG_INET4_POST_BIND(sk);
if (err) {
inet->inet_saddr = inet->inet_rcv_saddr = 0;
goto out_release_sock;
}
Let's take UDP for example and see what will happen. For UDP
socket, it will be added to 'udp_prot.h.udp_table->hash' and
'udp_prot.h.udp_table->hash2' after the sk->sk_prot->get_port()
called success. If 'inet->inet_rcv_saddr' is specified here,
then 'sk' will be in the 'hslot2' of 'hash2' that it don't belong
to (because inet_saddr is changed to 0), and UDP packet received
will not be passed to this sock. If 'inet->inet_rcv_saddr' is not
specified here, the sock will work fine, as it can receive packet
properly, which is wired, as the 'bind()' is already failed.
To undo the get_port() operation, introduce the 'put_port' field
for 'struct proto'. For TCP proto, it is inet_put_port(); For UDP
proto, it is udp_lib_unhash(); For icmp proto, it is
ping_unhash().
Therefore, after sys_bind() fail caused by
BPF_CGROUP_RUN_PROG_INET4_POST_BIND(), it will be unbinded, which
means that it can try to be binded to another port.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20220106132022.3470772-2-imagedong@tencent.com
tcp recvmsg() (or rx zerocopy) spends a fair amount of time
freeing skbs after their payload has been consumed.
A typical ~64KB GRO packet has to release ~45 page
references, eventually going to page allocator
for each of them.
Currently, this freeing is performed while socket lock
is held, meaning that there is a high chance that
BH handler has to queue incoming packets to tcp socket backlog.
This can cause additional latencies, because the user
thread has to process the backlog at release_sock() time,
and while doing so, additional frames can be added
by BH handler.
This patch adds logic to defer these frees after socket
lock is released, or directly from BH handler if possible.
Being able to free these skbs from BH handler helps a lot,
because this avoids the usual alloc/free assymetry,
when BH handler and user thread do not run on same cpu or
NUMA node.
One cpu can now be fully utilized for the kernel->user copy,
and another cpu is handling BH processing and skb/page
allocs/frees (assuming RFS is not forcing use of a single CPU)
Tested:
100Gbit NIC
Max throughput for one TCP_STREAM flow, over 10 runs
MTU : 1500
Before: 55 Gbit
After: 66 Gbit
MTU : 4096+(headers)
Before: 82 Gbit
After: 95 Gbit
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of using a full netdev_features_t, we can use a single bit,
as sk_route_nocaps is only used to remove NETIF_F_GSO_MASK from
sk->sk_route_cap.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If packet is going to be coalesced, sk_sndbuf/sk_rcvbuf values
are not used. Defer their access to the point we need them.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
RFC 5082 IP_MINTTL option is rarely used on hosts.
Add a static key to remove from TCP fast path useless code,
and potential cache line miss to fetch inet_sk(sk)->min_ttl
Note that once ip4_min_ttl static key has been enabled,
it stays enabled until next boot.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
No report yet from KCSAN, yet worth documenting the races.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Increase cache locality by moving rx_dst_ifindex next to sk->sk_rx_dst
This is part of an effort to reduce cache line misses in TCP fast path.
This removes one cache line miss in early demux.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
