]> www.infradead.org Git - users/dwmw2/linux.git/commit
tcp: fix potential xmit stalls caused by TCP_NOTSENT_LOWAT
authorEric Dumazet <edumazet@google.com>
Mon, 25 Apr 2022 00:34:07 +0000 (17:34 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 9 May 2022 07:05:04 +0000 (09:05 +0200)
commit8a9d6ca3608f76beac1c40a02073c3c52a572b10
tree4b131ab3c733104e75f9907e7ef29bee5f58f066
parent720b6ced8549c210d880aeaedf055e442d6edaff
tcp: fix potential xmit stalls caused by TCP_NOTSENT_LOWAT

[ Upstream commit 4bfe744ff1644fbc0a991a2677dc874475dd6776 ]

I had this bug sitting for too long in my pile, it is time to fix it.

Thanks to Doug Porter for reminding me of it!

We had various attempts in the past, including commit
0cbe6a8f089e ("tcp: remove SOCK_QUEUE_SHRUNK"),
but the issue is that TCP stack currently only generates
EPOLLOUT from input path, when tp->snd_una has advanced
and skb(s) cleaned from rtx queue.

If a flow has a big RTT, and/or receives SACKs, it is possible
that the notsent part (tp->write_seq - tp->snd_nxt) reaches 0
and no more data can be sent until tp->snd_una finally advances.

What is needed is to also check if POLLOUT needs to be generated
whenever tp->snd_nxt is advanced, from output path.

This bug triggers more often after an idle period, as
we do not receive ACK for at least one RTT. tcp_notsent_lowat
could be a fraction of what CWND and pacing rate would allow to
send during this RTT.

In a followup patch, I will remove the bogus call
to tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED)
from tcp_check_space(). Fact that we have decided to generate
an EPOLLOUT does not mean the application has immediately
refilled the transmit queue. This optimistic call
might have been the reason the bug seemed not too serious.

Tested:

200 ms rtt, 1% packet loss, 32 MB tcp_rmem[2] and tcp_wmem[2]

$ echo 500000 >/proc/sys/net/ipv4/tcp_notsent_lowat
$ cat bench_rr.sh
SUM=0
for i in {1..10}
do
 V=`netperf -H remote_host -l30 -t TCP_RR -- -r 10000000,10000 -o LOCAL_BYTES_SENT | egrep -v "MIGRATED|Bytes"`
 echo $V
 SUM=$(($SUM + $V))
done
echo SUM=$SUM

Before patch:
$ bench_rr.sh
130000000
80000000
140000000
140000000
140000000
140000000
130000000
40000000
90000000
110000000
SUM=1140000000

After patch:
$ bench_rr.sh
430000000
590000000
530000000
450000000
450000000
350000000
450000000
490000000
480000000
460000000
SUM=4680000000  # This is 410 % of the value before patch.

Fixes: c9bee3b7fdec ("tcp: TCP_NOTSENT_LOWAT socket option")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Doug Porter <dsp@fb.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
include/net/tcp.h
net/ipv4/tcp_input.c
net/ipv4/tcp_output.c