Message-ID: <CANn89iK0qqpt_aLhf+HN8V_Kvwt6sEFVbAiJhpe9_F+H+SUDNQ@mail.gmail.com>
Date: Fri, 24 Oct 2025 08:25:20 -0700
From: Eric Dumazet <edumazet@...gle.com>
To: Dragos Tatulea <dtatulea@...dia.com>
Cc: "David S . Miller" <davem@...emloft.net>, Jakub Kicinski <kuba@...nel.org>, 
	Paolo Abeni <pabeni@...hat.com>, Simon Horman <horms@...nel.org>, 
	Neal Cardwell <ncardwell@...gle.com>, Willem de Bruijn <willemb@...gle.com>, 
	Kuniyuki Iwashima <kuniyu@...gle.com>, Matthieu Baerts <matttbe@...nel.org>, 
	Mat Martineau <martineau@...nel.org>, Geliang Tang <geliang@...nel.org>, netdev@...r.kernel.org, 
	eric.dumazet@...il.com
Subject: Re: [PATCH net-next 3/3] tcp: fix too slow tcp_rcvbuf_grow() action

On Fri, Oct 24, 2025 at 8:13 AM Dragos Tatulea <dtatulea@...dia.com> wrote:
>
> On 24.10.25 09:50, Eric Dumazet wrote:
> > While the blamed commits apparently avoided an overshoot,
> > they also limited how fast a sender can increase BDP at each RTT.
> >
> > This is not exactly a revert, we do not add the 16 * tp->advmss
> > cushion we had, and we are keeping the out_of_order_queue
> > contribution.
> >
> > Do the same in mptcp_rcvbuf_grow().
> >
> > Tested:
> >
> > emulated 50ms rtt (tcp_stream --tcp-tx-delay 50000), cubic 20 second flow.
> > net.ipv4.tcp_rmem set to "4096 131072 67000000"
> >
> > perf record -a -e tcp:tcp_rcvbuf_grow sleep 20
> > perf script
> >
> > Before:
> >
> > We can see we fail to roughly double RWIN at each RTT.
> > Sender is RWIN limited while CWND is ramping up (before getting tcp_wmem limited)
> >
> > tcp_stream 33793 [010]  825.717525: tcp:tcp_rcvbuf_grow: time=100869 rtt_us=50428 copied=49152 inq=0 space=40960 ooo=0 scaling_ratio=219 rcvbuf=131072 rcv_ssthresh=103970 window_clamp=112128 rcv_wnd=106496
> > tcp_stream 33793 [010]  825.768966: tcp:tcp_rcvbuf_grow: time=51447 rtt_us=50362 copied=86016 inq=0 space=49152 ooo=0 scaling_ratio=219 rcvbuf=131072 rcv_ssthresh=107474 window_clamp=112128 rcv_wnd=106496
> > tcp_stream 33793 [010]  825.821539: tcp:tcp_rcvbuf_grow: time=52577 rtt_us=50243 copied=114688 inq=0 space=86016 ooo=0 scaling_ratio=219 rcvbuf=201096 rcv_ssthresh=167377 window_clamp=172031 rcv_wnd=167936
> > tcp_stream 33793 [010]  825.871781: tcp:tcp_rcvbuf_grow: time=50248 rtt_us=50237 copied=167936 inq=0 space=114688 ooo=0 scaling_ratio=219 rcvbuf=268129 rcv_ssthresh=224722 window_clamp=229375 rcv_wnd=225280
> > tcp_stream 33793 [010]  825.922475: tcp:tcp_rcvbuf_grow: time=50698 rtt_us=50183 copied=241664 inq=0 space=167936 ooo=0 scaling_ratio=219 rcvbuf=392617 rcv_ssthresh=331217 window_clamp=335871 rcv_wnd=323584
> > tcp_stream 33793 [010]  825.973326: tcp:tcp_rcvbuf_grow: time=50855 rtt_us=50213 copied=339968 inq=0 space=241664 ooo=0 scaling_ratio=219 rcvbuf=564986 rcv_ssthresh=478674 window_clamp=483327 rcv_wnd=462848
> > tcp_stream 33793 [010]  826.023970: tcp:tcp_rcvbuf_grow: time=50647 rtt_us=50248 copied=491520 inq=0 space=339968 ooo=0 scaling_ratio=219 rcvbuf=794811 rcv_ssthresh=671778 window_clamp=679935 rcv_wnd=651264
> > tcp_stream 33793 [010]  826.074612: tcp:tcp_rcvbuf_grow: time=50648 rtt_us=50227 copied=700416 inq=0 space=491520 ooo=0 scaling_ratio=219 rcvbuf=1149124 rcv_ssthresh=974881 window_clamp=983039 rcv_wnd=942080
> > tcp_stream 33793 [010]  826.125452: tcp:tcp_rcvbuf_grow: time=50845 rtt_us=50225 copied=987136 inq=8192 space=700416 ooo=0 scaling_ratio=219 rcvbuf=1637502 rcv_ssthresh=1392674 window_clamp=1400831 rcv_wnd=1339392
> > tcp_stream 33793 [010]  826.175698: tcp:tcp_rcvbuf_grow: time=50250 rtt_us=50198 copied=1347584 inq=0 space=978944 ooo=0 scaling_ratio=219 rcvbuf=2288672 rcv_ssthresh=1949729 window_clamp=1957887 rcv_wnd=1945600
> > tcp_stream 33793 [010]  826.225947: tcp:tcp_rcvbuf_grow: time=50252 rtt_us=50240 copied=1945600 inq=0 space=1347584 ooo=0 scaling_ratio=219 rcvbuf=3150516 rcv_ssthresh=2687010 window_clamp=2695167 rcv_wnd=2691072
> > tcp_stream 33793 [010]  826.276175: tcp:tcp_rcvbuf_grow: time=50233 rtt_us=50224 copied=2691072 inq=0 space=1945600 ooo=0 scaling_ratio=219 rcvbuf=4548617 rcv_ssthresh=3883041 window_clamp=3891199 rcv_wnd=3887104
> > tcp_stream 33793 [010]  826.326403: tcp:tcp_rcvbuf_grow: time=50233 rtt_us=50229 copied=3887104 inq=0 space=2691072 ooo=0 scaling_ratio=219 rcvbuf=6291456 rcv_ssthresh=5370482 window_clamp=5382144 rcv_wnd=5373952
> > tcp_stream 33793 [010]  826.376723: tcp:tcp_rcvbuf_grow: time=50323 rtt_us=50218 copied=5373952 inq=0 space=3887104 ooo=0 scaling_ratio=219 rcvbuf=9087658 rcv_ssthresh=7755537 window_clamp=7774207 rcv_wnd=7757824
> > tcp_stream 33793 [010]  826.426991: tcp:tcp_rcvbuf_grow: time=50274 rtt_us=50196 copied=7757824 inq=180224 space=5373952 ooo=0 scaling_ratio=219 rcvbuf=12563759 rcv_ssthresh=10729233 window_clamp=10747903 rcv_wnd=10575872
> > tcp_stream 33793 [010]  826.477229: tcp:tcp_rcvbuf_grow: time=50241 rtt_us=50078 copied=10731520 inq=180224 space=7577600 ooo=0 scaling_ratio=219 rcvbuf=17715667 rcv_ssthresh=15136529 window_clamp=15155199 rcv_wnd=14983168
> > tcp_stream 33793 [010]  826.527482: tcp:tcp_rcvbuf_grow: time=50258 rtt_us=50153 copied=15138816 inq=360448 space=10551296 ooo=0 scaling_ratio=219 rcvbuf=24667870 rcv_ssthresh=21073410 window_clamp=21102591 rcv_wnd=20766720
> > tcp_stream 33793 [010]  826.577712: tcp:tcp_rcvbuf_grow: time=50234 rtt_us=50228 copied=21073920 inq=0 space=14778368 ooo=0 scaling_ratio=219 rcvbuf=34550339 rcv_ssthresh=29517041 window_clamp=29556735 rcv_wnd=29519872
> > tcp_stream 33793 [010]  826.627982: tcp:tcp_rcvbuf_grow: time=50275 rtt_us=50220 copied=29519872 inq=540672 space=21073920 ooo=0 scaling_ratio=219 rcvbuf=49268707 rcv_ssthresh=42090625 window_clamp=42147839 rcv_wnd=41627648
> > tcp_stream 33793 [010]  826.678274: tcp:tcp_rcvbuf_grow: time=50296 rtt_us=50185 copied=42053632 inq=761856 space=28979200 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57238168 window_clamp=57316406 rcv_wnd=56606720
> > tcp_stream 33793 [010]  826.728627: tcp:tcp_rcvbuf_grow: time=50357 rtt_us=50128 copied=43913216 inq=851968 space=41291776 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57290728 window_clamp=57316406 rcv_wnd=56524800
> > tcp_stream 33793 [010]  827.131364: tcp:tcp_rcvbuf_grow: time=50239 rtt_us=50127 copied=43843584 inq=655360 space=43061248 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57290728 window_clamp=57316406 rcv_wnd=56696832
> > tcp_stream 33793 [010]  827.181613: tcp:tcp_rcvbuf_grow: time=50254 rtt_us=50115 copied=43843584 inq=524288 space=43188224 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57290728 window_clamp=57316406 rcv_wnd=56807424
> > tcp_stream 33793 [010]  828.339635: tcp:tcp_rcvbuf_grow: time=50283 rtt_us=50110 copied=43843584 inq=458752 space=43319296 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57290728 window_clamp=57316406 rcv_wnd=56864768
> > tcp_stream 33793 [010]  828.440350: tcp:tcp_rcvbuf_grow: time=50404 rtt_us=50099 copied=43843584 inq=393216 space=43384832 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57290728 window_clamp=57316406 rcv_wnd=56922112
> > tcp_stream 33793 [010]  829.195106: tcp:tcp_rcvbuf_grow: time=50154 rtt_us=50077 copied=43843584 inq=196608 space=43450368 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57290728 window_clamp=57316406 rcv_wnd=57090048
> >
> > After:
> >
> > It takes few steps to increase RWIN. Sender is no longer RWIN limited.
> >
> > tcp_stream 50826 [010]  935.634212: tcp:tcp_rcvbuf_grow: time=100788 rtt_us=50315 copied=49152 inq=0 space=40960 ooo=0 scaling_ratio=219 rcvbuf=131072 rcv_ssthresh=103970 window_clamp=112128 rcv_wnd=106496
> > tcp_stream 50826 [010]  935.685642: tcp:tcp_rcvbuf_grow: time=51437 rtt_us=50361 copied=86016 inq=0 space=49152 ooo=0 scaling_ratio=219 rcvbuf=160875 rcv_ssthresh=132969 window_clamp=137623 rcv_wnd=131072
> > tcp_stream 50826 [010]  935.738299: tcp:tcp_rcvbuf_grow: time=52660 rtt_us=50256 copied=139264 inq=0 space=86016 ooo=0 scaling_ratio=219 rcvbuf=502741 rcv_ssthresh=411497 window_clamp=430079 rcv_wnd=413696
> > tcp_stream 50826 [010]  935.788544: tcp:tcp_rcvbuf_grow: time=50249 rtt_us=50233 copied=307200 inq=0 space=139264 ooo=0 scaling_ratio=219 rcvbuf=728690 rcv_ssthresh=618717 window_clamp=623371 rcv_wnd=618496
> > tcp_stream 50826 [010]  935.838796: tcp:tcp_rcvbuf_grow: time=50258 rtt_us=50202 copied=618496 inq=0 space=307200 ooo=0 scaling_ratio=219 rcvbuf=2450338 rcv_ssthresh=1855709 window_clamp=2096187 rcv_wnd=1859584
> > tcp_stream 50826 [010]  935.889140: tcp:tcp_rcvbuf_grow: time=50347 rtt_us=50166 copied=1261568 inq=0 space=618496 ooo=0 scaling_ratio=219 rcvbuf=4376503 rcv_ssthresh=3725291 window_clamp=3743961 rcv_wnd=3706880
> > tcp_stream 50826 [010]  935.939435: tcp:tcp_rcvbuf_grow: time=50300 rtt_us=50185 copied=2478080 inq=24576 space=1261568 ooo=0 scaling_ratio=219 rcvbuf=9082648 rcv_ssthresh=7733731 window_clamp=7769921 rcv_wnd=7692288
> > tcp_stream 50826 [010]  935.989681: tcp:tcp_rcvbuf_grow: time=50251 rtt_us=50221 copied=4915200 inq=114688 space=2453504 ooo=0 scaling_ratio=219 rcvbuf=16574936 rcv_ssthresh=14108110 window_clamp=14179339 rcv_wnd=14024704
> > tcp_stream 50826 [010]  936.039967: tcp:tcp_rcvbuf_grow: time=50289 rtt_us=50279 copied=9830400 inq=114688 space=4800512 ooo=0 scaling_ratio=219 rcvbuf=32695050 rcv_ssthresh=27896187 window_clamp=27969593 rcv_wnd=27815936
> > tcp_stream 50826 [010]  936.090172: tcp:tcp_rcvbuf_grow: time=50211 rtt_us=50200 copied=19841024 inq=114688 space=9715712 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57245176 window_clamp=57316406 rcv_wnd=57163776
> > tcp_stream 50826 [010]  936.140430: tcp:tcp_rcvbuf_grow: time=50262 rtt_us=50197 copied=39501824 inq=114688 space=19726336 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57245176 window_clamp=57316406 rcv_wnd=57163776
> > tcp_stream 50826 [010]  936.190527: tcp:tcp_rcvbuf_grow: time=50101 rtt_us=50071 copied=43655168 inq=262144 space=39387136 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57259192 window_clamp=57316406 rcv_wnd=57032704
> > tcp_stream 50826 [010]  936.240719: tcp:tcp_rcvbuf_grow: time=50197 rtt_us=50057 copied=43843584 inq=262144 space=43393024 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57259192 window_clamp=57316406 rcv_wnd=57032704
> > tcp_stream 50826 [010]  936.341271: tcp:tcp_rcvbuf_grow: time=50297 rtt_us=50123 copied=43843584 inq=131072 space=43581440 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57259192 window_clamp=57316406 rcv_wnd=57147392
> > tcp_stream 50826 [010]  936.642503: tcp:tcp_rcvbuf_grow: time=50131 rtt_us=50084 copied=43843584 inq=0 space=43712512 ooo=0 scaling_ratio=219 rcvbuf=67000000 rcv_ssthresh=57259192 window_clamp=57316406 rcv_wnd=57262080
> >
> > Fixes: 65c5287892e9 ("tcp: fix sk_rcvbuf overshoot")
> > Fixes: e118cdc34dd1 ("mptcp: rcvbuf auto-tuning improvement")
> > Reported-by: Neal Cardwell <ncardwell@...gle.com>
> > Signed-off-by: Eric Dumazet <edumazet@...gle.com>
> > ---
> >  net/ipv4/tcp_input.c | 8 +++++++-
> >  net/mptcp/protocol.c | 7 +++++++
> >  2 files changed, 14 insertions(+), 1 deletion(-)
> >
> > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> > index c8cfd700990f28a8bc64e4353a2c78a82bb6bcb2..f004072654a4c50da14b9dafc46133feb71f12cd 100644
> > --- a/net/ipv4/tcp_input.c
> > +++ b/net/ipv4/tcp_input.c
> > @@ -896,6 +896,7 @@ void tcp_rcvbuf_grow(struct sock *sk, u32 newval)
> >       const struct net *net = sock_net(sk);
> >       struct tcp_sock *tp = tcp_sk(sk);
> >       u32 rcvwin, rcvbuf, cap, oldval;
> > +     u64 grow;
> >
> >       oldval = tp->rcvq_space.space;
> >       tp->rcvq_space.space = newval;
> > @@ -904,9 +905,14 @@ void tcp_rcvbuf_grow(struct sock *sk, u32 newval)
> >           (sk->sk_userlocks & SOCK_RCVBUF_LOCK))
> >               return;
> >
> > -     /* slow start: allow the sender to double its rate. */
> > +     /* DRS is always one RTT late. */
> >       rcvwin = newval << 1;
> >
> > +     /* slow start: allow the sender to double its rate. */
> > +     grow = (u64)rcvwin * (newval - oldval);
> > +     do_div(grow, oldval);
> > +     rcvwin += grow << 1;
> > +
> >       if (!RB_EMPTY_ROOT(&tp->out_of_order_queue))
> >               rcvwin += TCP_SKB_CB(tp->ooo_last_skb)->end_seq - tp->rcv_nxt;
> >
> Hi Eric,
>
> When applying this series I see a regression in a simple 25G iperf test:
> retransmissions are seen due to packet drops (out of buffer) on the
> server side.
>
> The test:
> - server: iperf3 -s -A 5
> - client: iperf3 -c 1.1.1.1 -B 25G
> - Configuration:
>   - Server has a single queue with affinity set on CPU 5.
>   - Ring size: 1K (4K ring size seems ok)
>   - MTU: 1500
>   - Client uses TSO, server uses SW GRO.
>
> Before series (includes first patch):
> <...>-2192  [005]   162.451893: tcp_rcvbuf_grow: time=1622 rtt_us=1596 copied=76781 inq=30408 space=14480 ooo=0 scaling_ratio=188 rcvbuf=131072 rcv_ssthresh=91990 window_clamp=96256 rcv_wnd=66560
> <...>-2192  [005]   162.451998: tcp_rcvbuf_grow: time=106 rtt_us=105 copied=158720 inq=0 space=46373 ooo=0 scaling_ratio=188 rcvbuf=131072 rcv_ssthresh=91990 window_clamp=96256 rcv_wnd=92160
> <...>-2192  [005]   162.453254: tcp_rcvbuf_grow: time=142 rtt_us=44 copied=292496 inq=91512 space=158720 ooo=0 scaling_ratio=188 rcvbuf=432258 rcv_ssthresh=270533 window_clamp=317439 rcv_wnd=253952
> <...>-2192  [005]   162.454446: tcp_rcvbuf_grow: time=113 rtt_us=44 copied=343176 inq=127424 space=200984 ooo=0 scaling_ratio=188 rcvbuf=547360 rcv_ssthresh=349656 window_clamp=401967 rcv_wnd=345088
> <...>-2192  [005]   162.455726: tcp_rcvbuf_grow: time=52 rtt_us=44 copied=264464 inq=40544 space=215752 ooo=0 scaling_ratio=188 rcvbuf=587579 rcv_ssthresh=391036 window_clamp=431503 rcv_wnd=194560
> <...>-2192  [005]   162.456444: tcp_rcvbuf_grow: time=37 rtt_us=36 copied=322560 inq=0 space=223920 ooo=0 scaling_ratio=188 rcvbuf=609824 rcv_ssthresh=391036 window_clamp=447839 rcv_wnd=323584
> <...>-2192  [005]   162.456865: tcp_rcvbuf_grow: time=40 rtt_us=36 copied=421840 inq=73848 space=322560 ooo=0 scaling_ratio=188 rcvbuf=878461 rcv_ssthresh=581105 window_clamp=645119 rcv_wnd=515072
> <...>-2192  [005]   162.457762: tcp_rcvbuf_grow: time=38 rtt_us=36 copied=430176 inq=65160 space=347992 ooo=0 scaling_ratio=188 rcvbuf=947722 rcv_ssthresh=631969 window_clamp=695983 rcv_wnd=467968
> <...>-2192  [005]   162.463191: tcp_rcvbuf_grow: time=35 rtt_us=34 copied=411336 inq=0 space=365016 ooo=0 scaling_ratio=188 rcvbuf=994086 rcv_ssthresh=666017 window_clamp=730031 rcv_wnd=354304
> <...>-2192  [005]   162.469069: tcp_rcvbuf_grow: time=38 rtt_us=34 copied=444520 inq=0 space=411336 ooo=0 scaling_ratio=188 rcvbuf=1120234 rcv_ssthresh=783379 window_clamp=822671 rcv_wnd=679936
>
> After series:
> <...>-2585  [005]  1061.768676: tcp_rcvbuf_grow: time=623 rtt_us=600 copied=72437 inq=28960 space=14480 ooo=0 scaling_ratio=188 rcvbuf=131072 rcv_ssthresh=81968 window_clamp=96256 rcv_wnd=82944
> <...>-2585  [005]  1061.769859: tcp_rcvbuf_grow: time=89 rtt_us=55 copied=250560 inq=46336 space=43477 ooo=0 scaling_ratio=188 rcvbuf=592631 rcv_ssthresh=302062 window_clamp=435213 rcv_wnd=230400
> <...>-2585  [005]  1061.775618: tcp_rcvbuf_grow: time=56 rtt_us=55 copied=405296 inq=140016 space=204224 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=1927847 window_clamp=3428745 rcv_wnd=1928192
> <...>-2585  [005]  1061.777324: tcp_rcvbuf_grow: time=57 rtt_us=55 copied=450664 inq=131072 space=265280 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=3106743 window_clamp=3428745 rcv_wnd=3006464
> <...>-2585  [005]  1061.783411: tcp_rcvbuf_grow: time=58 rtt_us=55 copied=521280 inq=41160 space=319592 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=3364731 window_clamp=3428745 rcv_wnd=2086912
> <...>-2585  [005]  1061.790393: tcp_rcvbuf_grow: time=55 rtt_us=55 copied=524288 inq=0 space=480120 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=3364731 window_clamp=3428745 rcv_wnd=2492416
> <...>-2585  [005]  1061.935387: tcp_rcvbuf_grow: time=55 rtt_us=55 copied=537824 inq=0 space=524288 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=3364731 window_clamp=3428745 rcv_wnd=2258944
> <...>-2585  [005]  1062.977374: tcp_rcvbuf_grow: time=57 rtt_us=55 copied=545064 inq=0 space=537824 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=3428745 window_clamp=3428745 rcv_wnd=2223104
> <...>-2585  [005]  1064.873376: tcp_rcvbuf_grow: time=57 rtt_us=55 copied=549408 inq=0 space=545064 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=3428745 window_clamp=3428745 rcv_wnd=2509824
> <...>-2585  [005]  1065.984340: tcp_rcvbuf_grow: time=59 rtt_us=55 copied=574024 inq=0 space=549408 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=3428745 window_clamp=3428745 rcv_wnd=2336768
> <...>-2585  [005]  1066.210718: tcp_rcvbuf_grow: time=410 rtt_us=55 copied=589448 inq=0 space=574024 ooo=0 scaling_ratio=188 rcvbuf=4668930 rcv_ssthresh=3428745 window_clamp=3428745 rcv_wnd=3364864
>
> Is this expected?

Typical cubic reaction I would say, with pfifo_fast ?

I do not think that artificially slowing down senders by not growing
rwin fast enough is the right answer.

If you have drops, that is because the sender is sending too fast or
with too large burts.

Try using the fq packet scheduler and see if it helps.

Powered by blists - more mailing lists