| Age | Commit message (Collapse) | Author |
|---|
|
icsk->icsk_ack.timeout can be replaced by icsk->csk_delack_timer.expires
This saves 8 bytes in TCP/DCCP sockets and helps for better cache locality.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250324203607.703850-3-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
icsk->icsk_timeout can be replaced by icsk->icsk_retransmit_timer.expires
This saves 8 bytes in TCP/DCCP sockets and helps for better cache locality.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250324203607.703850-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
inet_connection_sock_af_ops.addr2sockaddr() hasn't been used at all
in the git era.
$ git grep addr2sockaddr $(git rev-list HEAD | tail -n 1)
Let's remove it.
Note that there was a 4 bytes hole after sockaddr_len and now it's
6 bytes, so the binary layout is not changed.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250318060112.3729-1-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
As a followup of my presentation in Zagreb for netdev 0x19:
icsk_clean_acked is only used by TCP when/if CONFIG_TLS_DEVICE
is enabled from tcp_ack().
Rename it to tcp_clean_acked, move it to tcp_sock structure
in the tcp_sock_read_rx for better cache locality in TCP
fast path.
Define this field only when CONFIG_TLS_DEVICE is enabled
saving 8 bytes on configs not using it.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Sabrina Dubroca <sd@queasysnail.net>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250317085313.2023214-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
tcp_in_quickack_mode() is called from input path for small packets.
It calls __sk_dst_get() which reads sk->sk_dst_cache which has been
put in sock_read_tx group (for good reasons).
Then dst_metric(dst, RTAX_QUICKACK) also needs extra cache line misses.
Cache RTAX_QUICKACK in icsk->icsk_ack.dst_quick_ack to no longer pull
these cache lines for the cases a delayed ACK is scheduled.
After this patch TCP receive path does not longer access sock_read_tx
group.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250312083907.1931644-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
Currently, TCP stack uses a constant (120 seconds)
to limit the RTO value exponential growth.
Some applications want to set a lower value.
Add TCP_RTO_MAX_MS socket option to set a value (in ms)
between 1 and 120 seconds.
It is discouraged to change the socket rto max on a live
socket, as it might lead to unexpected disconnects.
Following patch is adding a netns sysctl to control the
default value at socket creation time.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
inet_csk_delete_keepalive_timer() and inet_csk_reset_keepalive_timer()
are only used from core TCP, there is no need to export them.
Replace their prefix by tcp.
Move them to net/ipv4/tcp_timer.c and make tcp_delete_keepalive_timer()
static.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250206094605.2694118-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
If the backlog of listen() is set to zero, sk_acceptq_is_full() allows
one connection to be made, but inet_csk_reqsk_queue_is_full() does not.
When the net.ipv4.tcp_syncookies is zero, inet_csk_reqsk_queue_is_full()
will cause an immediate drop before the sk_acceptq_is_full() check in
tcp_conn_request(), resulting in no connection can be made.
This patch tries to keep consistent with 64a146513f8f ("[NET]: Revert
incorrect accept queue backlog changes.").
Link: https://lore.kernel.org/netdev/20250102080258.53858-1-kuniyu@amazon.com/
Fixes: ef547f2ac16b ("tcp: remove max_qlen_log")
Signed-off-by: Zhongqiu Duan <dzq.aishenghu0@gmail.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250102171426.915276-1-dzq.aishenghu0@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
delack timer is not stopped from inet_csk_clear_xmit_timer()
because we do not define INET_CSK_CLEAR_TIMERS.
This is a conscious choice : inet_csk_clear_xmit_timer()
is often called from another cpu. Calling del_timer()
would cause false sharing and lock contention.
This means that very often, tcp_delack_timer() is called
at the timer expiration, while there is no ACK to transmit.
This can be detected very early, avoiding the socket spinlock.
Notes:
- test about tp->compressed_ack is racy,
but in the unlikely case there is a race, the dedicated
compressed_ack_timer hrtimer would close it.
- Even if the fast path is not taken, reading
icsk->icsk_ack.pending and tp->compressed_ack
before acquiring the socket spinlock reduces
acquisition time and chances of contention.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20241002173042.917928-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
icsk->icsk_pending can be read locklessly already.
Following patch in the series will add another lockless read.
Add smp_load_acquire() and smp_store_release() annotations
because following patch will add a test in tcp_write_timer(),
and READ_ONCE()/WRITE_ONCE() alone would possibly lead to races.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20241002173042.917928-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When bonding is configured in BOND_MODE_BROADCAST mode, if two identical
SYN packets are received at the same time and processed on different CPUs,
it can potentially create the same sk (sock) but two different reqsk
(request_sock) in tcp_conn_request().
These two different reqsk will respond with two SYNACK packets, and since
the generation of the seq (ISN) incorporates a timestamp, the final two
SYNACK packets will have different seq values.
The consequence is that when the Client receives and replies with an ACK
to the earlier SYNACK packet, we will reset(RST) it.
========================================================================
This behavior is consistently reproducible in my local setup,
which comprises:
| NETA1 ------ NETB1 |
PC_A --- bond --- | | --- bond --- PC_B
| NETA2 ------ NETB2 |
- PC_A is the Server and has two network cards, NETA1 and NETA2. I have
bonded these two cards using BOND_MODE_BROADCAST mode and configured
them to be handled by different CPU.
- PC_B is the Client, also equipped with two network cards, NETB1 and
NETB2, which are also bonded and configured in BOND_MODE_BROADCAST mode.
If the client attempts a TCP connection to the server, it might encounter
a failure. Capturing packets from the server side reveals:
10.10.10.10.45182 > localhost: Flags [S], seq 320236027,
10.10.10.10.45182 > localhost: Flags [S], seq 320236027,
localhost > 10.10.10.10.45182: Flags [S.], seq 2967855116,
localhost > 10.10.10.10.45182: Flags [S.], seq 2967855123, <==
10.10.10.10.45182 > localhost: Flags [.], ack 4294967290,
10.10.10.10.45182 > localhost: Flags [.], ack 4294967290,
localhost > 10.10.10.10.45182: Flags [R], seq 2967855117, <==
localhost > 10.10.10.10.45182: Flags [R], seq 2967855117,
Two SYNACKs with different seq numbers are sent by localhost,
resulting in an anomaly.
========================================================================
The attempted solution is as follows:
Add a return value to inet_csk_reqsk_queue_hash_add() to confirm if the
ehash insertion is successful (Up to now, the reason for unsuccessful
insertion is that a reqsk for the same connection has already been
inserted). If the insertion fails, release the reqsk.
Due to the refcnt, Kuniyuki suggests also adding a return value check
for the DCCP module; if ehash insertion fails, indicating a successful
insertion of the same connection, simply release the reqsk as well.
Simultaneously, In the reqsk_queue_hash_req(), the start of the
req->rsk_timer is adjusted to be after successful insertion.
Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: luoxuanqiang <luoxuanqiang@kylinos.cn>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240621013929.1386815-1-luoxuanqiang@kylinos.cn
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
Rather than pass in flags, error pointer, and whether this is a kernel
invocation or not, add a struct proto_accept_arg struct as the argument.
This then holds all of these arguments, and prepares accept for being
able to pass back more information.
No functional changes in this patch.
Acked-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Since commit 099ecf59f05b ("net: annotate lockless accesses to
sk->sk_max_ack_backlog") decided to handle the sk_max_ack_backlog
locklessly, there is one more function mostly called in TCP/DCCP
cases. So this patch completes it:)
Signed-off-by: Jason Xing <kernelxing@tencent.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240331090521.71965-1-kerneljasonxing@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
We can change inet_csk() to propagate its argument const qualifier,
thanks to container_of_const().
We have to fix few places that had mistakes, like tcp_bound_rto().
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240329144931.295800-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
We had various syzbot reports about tcp timers firing after
the corresponding netns has been dismantled.
Fortunately Josef Bacik could trigger the issue more often,
and could test a patch I wrote two years ago.
When TCP sockets are closed, we call inet_csk_clear_xmit_timers()
to 'stop' the timers.
inet_csk_clear_xmit_timers() can be called from any context,
including when socket lock is held.
This is the reason it uses sk_stop_timer(), aka del_timer().
This means that ongoing timers might finish much later.
For user sockets, this is fine because each running timer
holds a reference on the socket, and the user socket holds
a reference on the netns.
For kernel sockets, we risk that the netns is freed before
timer can complete, because kernel sockets do not hold
reference on the netns.
This patch adds inet_csk_clear_xmit_timers_sync() function
that using sk_stop_timer_sync() to make sure all timers
are terminated before the kernel socket is released.
Modules using kernel sockets close them in their netns exit()
handler.
Also add sock_not_owned_by_me() helper to get LOCKDEP
support : inet_csk_clear_xmit_timers_sync() must not be called
while socket lock is held.
It is very possible we can revert in the future commit
3a58f13a881e ("net: rds: acquire refcount on TCP sockets")
which attempted to solve the issue in rds only.
(net/smc/af_smc.c and net/mptcp/subflow.c have similar code)
We probably can remove the check_net() tests from
tcp_out_of_resources() and __tcp_close() in the future.
Reported-by: Josef Bacik <josef@toxicpanda.com>
Closes: https://lore.kernel.org/netdev/20240314210740.GA2823176@perftesting/
Fixes: 26abe14379f8 ("net: Modify sk_alloc to not reference count the netns of kernel sockets.")
Fixes: 8a68173691f0 ("net: sk_clone_lock() should only do get_net() if the parent is not a kernel socket")
Link: https://lore.kernel.org/bpf/CANn89i+484ffqb93aQm1N-tjxxvb3WDKX0EbD7318RwRgsatjw@mail.gmail.com/
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Josef Bacik <josef@toxicpanda.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Link: https://lore.kernel.org/r/20240322135732.1535772-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When I run syz's reproduction C program locally, it causes the following
issue:
pvqspinlock: lock 0xffff9d181cd5c660 has corrupted value 0x0!
WARNING: CPU: 19 PID: 21160 at __pv_queued_spin_unlock_slowpath (kernel/locking/qspinlock_paravirt.h:508)
Hardware name: Red Hat KVM, BIOS 0.5.1 01/01/2011
RIP: 0010:__pv_queued_spin_unlock_slowpath (kernel/locking/qspinlock_paravirt.h:508)
Code: 73 56 3a ff 90 c3 cc cc cc cc 8b 05 bb 1f 48 01 85 c0 74 05 c3 cc cc cc cc 8b 17 48 89 fe 48 c7 c7
30 20 ce 8f e8 ad 56 42 ff <0f> 0b c3 cc cc cc cc 0f 0b 0f 1f 40 00 90 90 90 90 90 90 90 90 90
RSP: 0018:ffffa8d200604cb8 EFLAGS: 00010282
RAX: 0000000000000000 RBX: 0000000000000000 RCX: ffff9d1ef60e0908
RDX: 00000000ffffffd8 RSI: 0000000000000027 RDI: ffff9d1ef60e0900
RBP: ffff9d181cd5c280 R08: 0000000000000000 R09: 00000000ffff7fff
R10: ffffa8d200604b68 R11: ffffffff907dcdc8 R12: 0000000000000000
R13: ffff9d181cd5c660 R14: ffff9d1813a3f330 R15: 0000000000001000
FS: 00007fa110184640(0000) GS:ffff9d1ef60c0000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000020000000 CR3: 000000011f65e000 CR4: 00000000000006f0
Call Trace:
<IRQ>
_raw_spin_unlock (kernel/locking/spinlock.c:186)
inet_csk_reqsk_queue_add (net/ipv4/inet_connection_sock.c:1321)
inet_csk_complete_hashdance (net/ipv4/inet_connection_sock.c:1358)
tcp_check_req (net/ipv4/tcp_minisocks.c:868)
tcp_v4_rcv (net/ipv4/tcp_ipv4.c:2260)
ip_protocol_deliver_rcu (net/ipv4/ip_input.c:205)
ip_local_deliver_finish (net/ipv4/ip_input.c:234)
__netif_receive_skb_one_core (net/core/dev.c:5529)
process_backlog (./include/linux/rcupdate.h:779)
__napi_poll (net/core/dev.c:6533)
net_rx_action (net/core/dev.c:6604)
__do_softirq (./arch/x86/include/asm/jump_label.h:27)
do_softirq (kernel/softirq.c:454 kernel/softirq.c:441)
</IRQ>
<TASK>
__local_bh_enable_ip (kernel/softirq.c:381)
__dev_queue_xmit (net/core/dev.c:4374)
ip_finish_output2 (./include/net/neighbour.h:540 net/ipv4/ip_output.c:235)
__ip_queue_xmit (net/ipv4/ip_output.c:535)
__tcp_transmit_skb (net/ipv4/tcp_output.c:1462)
tcp_rcv_synsent_state_process (net/ipv4/tcp_input.c:6469)
tcp_rcv_state_process (net/ipv4/tcp_input.c:6657)
tcp_v4_do_rcv (net/ipv4/tcp_ipv4.c:1929)
__release_sock (./include/net/sock.h:1121 net/core/sock.c:2968)
release_sock (net/core/sock.c:3536)
inet_wait_for_connect (net/ipv4/af_inet.c:609)
__inet_stream_connect (net/ipv4/af_inet.c:702)
inet_stream_connect (net/ipv4/af_inet.c:748)
__sys_connect (./include/linux/file.h:45 net/socket.c:2064)
__x64_sys_connect (net/socket.c:2073 net/socket.c:2070 net/socket.c:2070)
do_syscall_64 (arch/x86/entry/common.c:51 arch/x86/entry/common.c:82)
entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:129)
RIP: 0033:0x7fa10ff05a3d
Code: 5b 41 5c c3 66 0f 1f 84 00 00 00 00 00 f3 0f 1e fa 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89
c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d ab a3 0e 00 f7 d8 64 89 01 48
RSP: 002b:00007fa110183de8 EFLAGS: 00000202 ORIG_RAX: 000000000000002a
RAX: ffffffffffffffda RBX: 0000000020000054 RCX: 00007fa10ff05a3d
RDX: 000000000000001c RSI: 0000000020000040 RDI: 0000000000000003
RBP: 00007fa110183e20 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000202 R12: 00007fa110184640
R13: 0000000000000000 R14: 00007fa10fe8b060 R15: 00007fff73e23b20
</TASK>
The issue triggering process is analyzed as follows:
Thread A Thread B
tcp_v4_rcv //receive ack TCP packet inet_shutdown
tcp_check_req tcp_disconnect //disconnect sock
... tcp_set_state(sk, TCP_CLOSE)
inet_csk_complete_hashdance ...
inet_csk_reqsk_queue_add inet_listen //start listen
spin_lock(&queue->rskq_lock) inet_csk_listen_start
... reqsk_queue_alloc
... spin_lock_init
spin_unlock(&queue->rskq_lock) //warning
When the socket receives the ACK packet during the three-way handshake,
it will hold spinlock. And then the user actively shutdowns the socket
and listens to the socket immediately, the spinlock will be initialized.
When the socket is going to release the spinlock, a warning is generated.
Also the same issue to fastopenq.lock.
Move init spinlock to inet_create and inet_accept to make sure init the
accept_queue's spinlocks once.
Fixes: fff1f3001cc5 ("tcp: add a spinlock to protect struct request_sock_queue")
Fixes: 168a8f58059a ("tcp: TCP Fast Open Server - main code path")
Reported-by: Ming Shu <sming56@aliyun.com>
Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240118012019.1751966-1-shaozhengchao@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
RTAX_FEATURE_ALLFRAG was added before the first git commit:
https://www.mail-archive.com/bk-commits-head@vger.kernel.org/msg03399.html
The feature would send packets to the fragmentation path if a box
receives a PMTU value with less than 1280 byte. However, since commit
9d289715eb5c ("ipv6: stop sending PTB packets for MTU < 1280"), such
message would be simply discarded. The feature flag is neither supported
in iproute2 utility. In theory one can still manipulate it with direct
netlink message, but it is not ideal because it was based on obsoleted
guidance of RFC-2460 (replaced by RFC-8200).
The feature would always test false at the moment, so remove related
code or mark them as unused.
Signed-off-by: Yan Zhai <yan@cloudflare.com>
Reviewed-by: Florian Westphal <fw@strlen.de>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/d78e44dcd9968a252143ffe78460446476a472a1.1698156966.git.yan@cloudflare.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
TCP pingpong threshold is 1 by default. But some applications, like SQL DB
may prefer a higher pingpong threshold to activate delayed acks in quick
ack mode for better performance.
The pingpong threshold and related code were changed to 3 in the year
2019 in:
commit 4a41f453bedf ("tcp: change pingpong threshold to 3")
And reverted to 1 in the year 2022 in:
commit 4d8f24eeedc5 ("Revert "tcp: change pingpong threshold to 3"")
There is no single value that fits all applications.
Add net.ipv4.tcp_pingpong_thresh sysctl tunable, so it can be tuned for
optimal performance based on the application needs.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/1697056244-21888-1-git-send-email-haiyangz@microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
In order to better estimate whether a data packet has been
retransmitted or is the result of a TLP, we save the last received
ipv6 flowlabel.
To make space for this field we resize the "ato" field in
inet_connection_sock as the current value of TCP_DELACK_MAX can be
fully contained in 8 bits and add a compile_time_assert ensuring this
field is the required size.
v2: addressed kernel bot feedback about dccp_delack_timer()
v3: addressed build error introduced by commit bbf80d713fe7 ("tcp:
derive delack_max from rto_min")
Signed-off-by: David Morley <morleyd@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Tested-by: David Morley <morleyd@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
We move single bit fields to inet->inet_flags to avoid races.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
For now, skb will be dropped when no memory, which makes client keep
retrans util timeout and it's not friendly to the users.
In this patch, we reply an ACK with zero-window in this case to update
the snd_wnd of the sender to 0. Therefore, the sender won't timeout the
connection and will probe the zero-window with the retransmits.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
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>
|
|
This reverts commit 4a41f453bedfd5e9cd040bad509d9da49feb3e2c.
This to-be-reverted commit was meant to apply a stricter rule for the
stack to enter pingpong mode. However, the condition used to check for
interactive session "before(tp->lsndtime, icsk->icsk_ack.lrcvtime)" is
jiffy based and might be too coarse, which delays the stack entering
pingpong mode.
We revert this patch so that we no longer use the above condition to
determine interactive session, and also reduce pingpong threshold to 1.
Fixes: 4a41f453bedf ("tcp: change pingpong threshold to 3")
Reported-by: LemmyHuang <hlm3280@163.com>
Suggested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20220721204404.388396-1-weiwan@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
This reverts:
commit d5a42de8bdbe ("net: Add a second bind table hashed by port and address")
commit 538aaf9b2383 ("selftests: Add test for timing a bind request to a port with a populated bhash entry")
Link: https://lore.kernel.org/netdev/20220520001834.2247810-1-kuba@kernel.org/
There are a few things that need to be fixed here:
* Updating bhash2 in cases where the socket's rcv saddr changes
* Adding bhash2 hashbucket locks
Links to syzbot reports:
https://lore.kernel.org/netdev/00000000000022208805e0df247a@google.com/
https://lore.kernel.org/netdev/0000000000003f33bc05dfaf44fe@google.com/
Fixes: d5a42de8bdbe ("net: Add a second bind table hashed by port and address")
Reported-by: syzbot+015d756bbd1f8b5c8f09@syzkaller.appspotmail.com
Reported-by: syzbot+98fd2d1422063b0f8c44@syzkaller.appspotmail.com
Reported-by: syzbot+0a847a982613c6438fba@syzkaller.appspotmail.com
Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
Link: https://lore.kernel.org/r/20220615193213.2419568-1-joannelkoong@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
We currently have one tcp bind table (bhash) which hashes by port
number only. In the socket bind path, we check for bind conflicts by
traversing the specified port's inet_bind2_bucket while holding the
bucket'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, checking for a bind conflict 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 proposes adding a second bind table, bhash2, that hashes by
port and ip address. Searching the bhash2 table leads to significantly
faster conflict resolution and less time holding the spinlock.
Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The listen sk is currently stored in two hash tables,
listening_hash (hashed by port) and lhash2 (hashed by port and address).
After commit 0ee58dad5b06 ("net: tcp6: prefer listeners bound to an address")
and commit d9fbc7f6431f ("net: tcp: prefer listeners bound to an address"),
the TCP-SYN lookup fast path does not use listening_hash.
The commit 05c0b35709c5 ("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>
|
|
When setting RTO through BPF program, some SYN ACK packets were unaffected
and continued to use TCP_TIMEOUT_INIT constant. This patch adds timeout
option to struct request_sock. Option is initialized with TCP_TIMEOUT_INIT
and is reassigned through BPF using tcp_timeout_init call. SYN ACK
retransmits now use newly added timeout option.
Signed-off-by: Akhmat Karakotov <hmukos@yandex-team.ru>
Acked-by: Martin KaFai Lau <kafai@fb.com>
v2:
- Add timeout option to struct request_sock. Do not call
tcp_timeout_init on every syn ack retransmit.
v3:
- Use unsigned long for min. Bound tcp_timeout_init to TCP_RTO_MAX.
v4:
- Refactor duplicate code by adding reqsk_timeout function.
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The commit 1295e2cf3065 ("inet: minor optimization for backlog setting in
listen(2)") added change so that sk_max_ack_backlog is initialised earlier
in inet_dccp_listen() and inet_listen(). Since then, we no longer use
backlog in inet_csk_listen_start(), so let's remove it.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
Acked-by: Yafang Shao <laoar.shao@gmail.com>
Reviewed-by: Richard Sailer <richard_siegfried@systemli.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Use of percpu_counter structure to track count of orphaned
sockets is causing problems on modern hosts with 256 cpus
or more.
Stefan Bach reported a serious spinlock contention in real workloads,
that I was able to reproduce with a netfilter rule dropping
incoming FIN packets.
53.56% server [kernel.kallsyms] [k] queued_spin_lock_slowpath
|
---queued_spin_lock_slowpath
|
--53.51%--_raw_spin_lock_irqsave
|
--53.51%--__percpu_counter_sum
tcp_check_oom
|
|--39.03%--__tcp_close
| tcp_close
| inet_release
| inet6_release
| sock_close
| __fput
| ____fput
| task_work_run
| exit_to_usermode_loop
| do_syscall_64
| entry_SYSCALL_64_after_hwframe
| __GI___libc_close
|
--14.48%--tcp_out_of_resources
tcp_write_timeout
tcp_retransmit_timer
tcp_write_timer_handler
tcp_write_timer
call_timer_fn
expire_timers
__run_timers
run_timer_softirq
__softirqentry_text_start
As explained in commit cf86a086a180 ("net/dst: use a smaller percpu_counter
batch for dst entries accounting"), default batch size is too big
for the default value of tcp_max_orphans (262144).
But even if we reduce batch sizes, there would still be cases
where the estimated count of orphans is beyond the limit,
and where tcp_too_many_orphans() has to call the expensive
percpu_counter_sum_positive().
One solution is to use plain per-cpu counters, and have
a timer to periodically refresh this cache.
Updating this cache every 100ms seems about right, tcp pressure
state is not radically changing over shorter periods.
percpu_counter was nice 15 years ago while hosts had less
than 16 cpus, not anymore by current standards.
v2: Fix the build issue for CONFIG_CRYPTO_DEV_CHELSIO_TLS=m,
reported by kernel test robot <lkp@intel.com>
Remove unused socket argument from tcp_too_many_orphans()
Fixes: dd24c00191d5 ("net: Use a percpu_counter for orphan_count")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Stefan Bach <sfb@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Instead of a magic number (13 currently) and having
to change it every other year, use sizeof_field() macro.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Currently tcp_check_req can be called with obsolete req socket for which big
socket have been already created (because of CPU race or early demux
assigning req socket to multiple packets in gro batch).
Commit e0f9759f530bf789e984 ("tcp: try to keep packet if SYN_RCV race
is lost") added retry in case when tcp_check_req is called for PSH|ACK packet.
But if client sends RST+ACK immediatly after connection being
established (it is performing healthcheck, for example) retry does not
occur. In that case tcp_check_req tries to close req socket,
leaving big socket active.
Fixes: e0f9759f530 ("tcp: try to keep packet if SYN_RCV race is lost")
Signed-off-by: Alexander Ovechkin <ovov@yandex-team.ru>
Reported-by: Oleg Senin <olegsenin@yandex-team.ru>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This commit shrinks inet_connection_sock by 4 bytes, by shrinking
icsk_mtup.enabled from 32 bits to 1 bit, and shrinking
icsk_mtup.probe_size from s32 to an unsuigned 31 bit field.
This is to save space to compensate for the recent introduction of a
new u32 in inet_connection_sock, icsk_probes_tstamp, in the recent bug
fix commit 9d9b1ee0b2d1 ("tcp: fix TCP_USER_TIMEOUT with zero window").
This should not change functionality, since icsk_mtup.enabled is only
ever set to 0 or 1, and icsk_mtup.probe_size can only be either 0
or a positive MTU value returned by tcp_mss_to_mtu()
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20210129185438.1813237-1-ncardwell.kernel@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Since the early retransmit has been removed by
commit bec41a11dd3d ("tcp: remove early retransmit"),
we also remove the unused ICSK_TIME_EARLY_RETRANS macro.
Signed-off-by: Pengcheng Yang <yangpc@wangsu.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/1611239473-27304-1-git-send-email-yangpc@wangsu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The TCP session does not terminate with TCP_USER_TIMEOUT when data
remain untransmitted due to zero window.
The number of unanswered zero-window probes (tcp_probes_out) is
reset to zero with incoming acks irrespective of the window size,
as described in tcp_probe_timer():
RFC 1122 4.2.2.17 requires the sender to stay open indefinitely
as long as the receiver continues to respond probes. We support
this by default and reset icsk_probes_out with incoming ACKs.
This counter, however, is the wrong one to be used in calculating the
duration that the window remains closed and data remain untransmitted.
Thanks to Jonathan Maxwell <jmaxwell37@gmail.com> for diagnosing the
actual issue.
In this patch a new timestamp is introduced for the socket in order to
track the elapsed time for the zero-window probes that have not been
answered with any non-zero window ack.
Fixes: 9721e709fa68 ("tcp: simplify window probe aborting on USER_TIMEOUT")
Reported-by: William McCall <william.mccall@gmail.com>
Co-developed-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Enke Chen <enchen@paloaltonetworks.com>
Reviewed-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20210115223058.GA39267@localhost.localdomain
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Whenever host is under very high memory pressure,
__tcp_send_ack() skb allocation fails, and we setup
a 200 ms (TCP_DELACK_MAX) timer before retrying.
On hosts with high number of TCP sockets, we can spend
considerable amount of cpu cycles in these attempts,
add high pressure on various spinlocks in mm-layer,
ultimately blocking threads attempting to free space
from making any progress.
This patch adds standard exponential backoff to avoid
adding fuel to the fire.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
TCP has been using it to work around the possibility of tcp_delack_timer()
finding the socket owned by user.
After commit 6f458dfb4092 ("tcp: improve latencies of timer triggered events")
we added TCP_DELACK_TIMER_DEFERRED atomic bit for more immediate recovery,
so we can get rid of icsk_ack.blocked
This frees space that following patch will reuse.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Change tcp_init_transfer() to only initialize congestion control if it
has not been initialized already.
With this new approach, we can arrange things so that if the EBPF code
sets the congestion control by calling setsockopt(TCP_CONGESTION) then
tcp_init_transfer() will not re-initialize the CC module.
This is an approach that has the following beneficial properties:
(1) This allows CC module customizations made by the EBPF called in
tcp_init_transfer() to persist, and not be wiped out by a later
call to tcp_init_congestion_control() in tcp_init_transfer().
(2) Does not flip the order of EBPF and CC init, to avoid causing bugs
for existing code upstream that depends on the current order.
(3) Does not cause 2 initializations for for CC in the case where the
EBPF called in tcp_init_transfer() wants to set the CC to a new CC
algorithm.
(4) Allows follow-on simplifications to the code in net/core/filter.c
and net/ipv4/tcp_cong.c, which currently both have some complexity
to special-case CC initialization to avoid double CC
initialization if EBPF sets the CC.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Kevin Yang <yyd@google.com>
Cc: Lawrence Brakmo <brakmo@fb.com>
|
|
This patch adds bpf_setsockopt(TCP_BPF_RTO_MIN) to allow bpf prog
to set the min rto of a connection. It could be used together
with the earlier patch which has added bpf_setsockopt(TCP_BPF_DELACK_MAX).
A later selftest patch will communicate the max delay ack in a
bpf tcp header option and then the receiving side can use
bpf_setsockopt(TCP_BPF_RTO_MIN) to set a shorter rto.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20200820190027.2884170-1-kafai@fb.com
|
|
This change is mostly from an internal patch and adapts it from sysctl
config to the bpf_setsockopt setup.
The bpf_prog can set the max delay ack by using
bpf_setsockopt(TCP_BPF_DELACK_MAX). This max delay ack can be communicated
to its peer through bpf header option. The receiving peer can then use
this max delay ack and set a potentially lower rto by using
bpf_setsockopt(TCP_BPF_RTO_MIN) which will be introduced
in the next patch.
Another later selftest patch will also use it like the above to show
how to write and parse bpf tcp header option.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20200820190021.2884000-1-kafai@fb.com
|
|
Refactor the fastreuse update code in inet_csk_get_port into a small
helper function that can be called from other places.
Acked-by: Matthieu Baerts <matthieu.baerts@tessares.net>
Signed-off-by: Tim Froidcoeur <tim.froidcoeur@tessares.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Rework the remaining setsockopt code to pass a sockptr_t instead of a
plain user pointer. This removes the last remaining set_fs(KERNEL_DS)
outside of architecture specific code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Stefan Schmidt <stefan@datenfreihafen.org> [ieee802154]
Acked-by: Matthieu Baerts <matthieu.baerts@tessares.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Handle the few cases that need special treatment in-line using
in_compat_syscall(). This also removes all the now unused
compat_{get,set}sockopt methods.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Clearing the 'inet_num' field is necessary and safe if and
only if the socket is not bound. The MPTCP protocol calls
the destroy helper on bound sockets, as tcp_v{4,6}_syn_recv_sock
completed successfully.
Move the clearing of such field out of the common code, otherwise
the MPTCP MP_JOIN error path will find the wrong 'inet_num' value
on socket disposal, __inet_put_port() will acquire the wrong lock
and bind_node removal could race with other modifiers possibly
corrupting the bind hash table.
Reported-and-tested-by: Christoph Paasch <cpaasch@apple.com>
Fixes: 729cd6436f35 ("mptcp: cope better with MP_JOIN failure")
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Move the steps to prepare an inet_connection_sock for
forced disposal inside a separate helper. No functional
changes inteded, this will just simplify the next patch.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Christoph Paasch <cpaasch@apple.com>
Reviewed-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The sock map code checks that a socket does not have an active upper
layer protocol before inserting it into the map. This requires casting
via inet_csk, which isn't valid for UDP sockets.
Guard checks for ULP by checking inet_sk(sk)->is_icsk first.
Signed-off-by: Lorenz Bauer <lmb@cloudflare.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20200309111243.6982-2-lmb@cloudflare.com
|
|
We need to make sure context does not get freed while diag
code is interrogating it. Free struct tls_context with
kfree_rcu().
We add the __rcu annotation directly in icsk, and cast it
away in the datapath accessor. Presumably all ULPs will
do a similar thing.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license as published by
the free software foundation either version 2 of the license or at
your option any later version
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 3029 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190527070032.746973796@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
|
|
In order to be more confident about an on-going interactive session, we
increment pingpong count by 1 for every interactive transaction and we
adjust TCP_PINGPONG_THRESH to 3.
This means, we only consider a session in pingpong mode after we see 3
interactive transactions, and start to activate delayed acks in quick
ack mode.
And in order to not over-count the credits, we only increase pingpong
count for the first packet sent in response for the previous received
packet.
This is mainly to prevent delaying the ack immediately after some
handshake protocol but no real interactive traffic pattern afterwards.
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Instead of using pingpong as a single bit information, we refactor the
code to treat it as a counter. When interactive session is detected,
we set pingpong count to TCP_PINGPONG_THRESH. And when pingpong count
is >= TCP_PINGPONG_THRESH, we consider the session in pingpong mode.
This patch is a pure refactor and sets foundation for the next patch.
This patch itself does not change any pingpong logic.
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Aggregation effects are extremely common with wifi, cellular, and cable
modem link technologies, ACK decimation in middleboxes, and LRO and GRO
in receiving hosts. The aggregation can happen in either direction,
data or ACKs, but in either case the aggregation effect is visible
to the sender in the ACK stream.
Previously BBR's sending was often limited by cwnd under severe ACK
aggregation/decimation because BBR sized the cwnd at 2*BDP. If packets
were acked in bursts after long delays (e.g. one ACK acking 5*BDP after
5*RTT), BBR's sending was halted after sending 2*BDP over 2*RTT, leaving
the bottleneck idle for potentially long periods. Note that loss-based
congestion control does not have this issue because when facing
aggregation it continues increasing cwnd after bursts of ACKs, growing
cwnd until the buffer is full.
To achieve good throughput in the presence of aggregation effects, this
algorithm allows the BBR sender to put extra data in flight to keep the
bottleneck utilized during silences in the ACK stream that it has evidence
to suggest were caused by aggregation.
A summary of the algorithm: when a burst of packets are acked by a
stretched ACK or a burst of ACKs or both, BBR first estimates the expected
amount of data that should have been acked, based on its estimated
bandwidth. Then the surplus ("extra_acked") is recorded in a windowed-max
filter to estimate the recent level of observed ACK aggregation. Then cwnd
is increased by the ACK aggregation estimate. The larger cwnd avoids BBR
being cwnd-limited in the face of ACK silences that recent history suggests
were caused by aggregation. As a sanity check, the ACK aggregation degree
is upper-bounded by the cwnd (at the time of measurement) and a global max
of BW * 100ms. The algorithm is further described by the following
presentation:
https://datatracker.ietf.org/meeting/101/materials/slides-101-iccrg-an-update-on-bbr-work-at-google-00
In our internal testing, we observed a significant increase in BBR
throughput (measured using netperf), in a basic wifi setup.
- Host1 (sender on ethernet) -> AP -> Host2 (receiver on wifi)
- 2.4 GHz -> BBR before: ~73 Mbps; BBR after: ~102 Mbps; CUBIC: ~100 Mbps
- 5.0 GHz -> BBR before: ~362 Mbps; BBR after: ~593 Mbps; CUBIC: ~601 Mbps
Also, this code is running globally on YouTube TCP connections and produced
significant bandwidth increases for YouTube traffic.
This is based on Ian Swett's max_ack_height_ algorithm from the
QUIC BBR implementation.
Signed-off-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
