[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <ebb26a4a8a80292423c8cfc965c7b16e2aa4e201.camel@redhat.com>
Date: Thu, 23 Nov 2023 19:16:12 +0100
From: Paolo Abeni <pabeni@...hat.com>
To: Eric Dumazet <edumazet@...gle.com>, Neal Cardwell
<ncardwell@...gle.com>, Wei Wang <weiwan@...gle.com>
Cc: netdev@...r.kernel.org, "David S. Miller" <davem@...emloft.net>, David
Ahern <dsahern@...nel.org>, Jakub Kicinski <kuba@...nel.org>, Neil Spring
<ntspring@...com>, David Gibson <david@...son.dropbear.id.au>
Subject: Re: [PATCH net] tcp: fix mid stream window clamp.
On Thu, 2023-11-23 at 18:10 +0100, Eric Dumazet wrote:
> CC Neal and Wei
>
> On Thu, Nov 23, 2023 at 4:25 PM Paolo Abeni <pabeni@...hat.com> wrote:
> >
> > After the blamed commit below, if the user-space application performs
> > window clamping when tp->rcv_wnd is 0, the TCP socket will never be
> > able to announce a non 0 receive window, even after completely emptying
> > the receive buffer and re-setting the window clamp to higher values.
> >
> > Refactor tcp_set_window_clamp() to address the issue: when the user
> > decreases the current clamp value, set rcv_ssthresh according to the
> > same logic used at buffer initialization time.
> > When increasing the clamp value, give the rcv_ssthresh a chance to grow
> > according to previously implemented heuristic.
> >
> > Fixes: 3aa7857fe1d7 ("tcp: enable mid stream window clamp")
> > Reported-by: David Gibson <david@...son.dropbear.id.au>
> > Reported-by: Stefano Brivio <sbrivio@...hat.com>
> > Reviewed-by: Stefano Brivio <sbrivio@...hat.com>
> > Tested-by: Stefano Brivio <sbrivio@...hat.com>
> > Signed-off-by: Paolo Abeni <pabeni@...hat.com>
> > ---
> > net/ipv4/tcp.c | 19 ++++++++++++++++---
> > 1 file changed, 16 insertions(+), 3 deletions(-)
> >
> > diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> > index 53bcc17c91e4..1a9b9064e080 100644
> > --- a/net/ipv4/tcp.c
> > +++ b/net/ipv4/tcp.c
> > @@ -3368,9 +3368,22 @@ int tcp_set_window_clamp(struct sock *sk, int val)
> > return -EINVAL;
> > tp->window_clamp = 0;
> > } else {
> > - tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
> > - SOCK_MIN_RCVBUF / 2 : val;
> > - tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
> > + u32 new_rcv_ssthresh, old_window_clamp = tp->window_clamp;
> > + u32 new_window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
> > + SOCK_MIN_RCVBUF / 2 : val;
> > +
> > + if (new_window_clamp == old_window_clamp)
> > + return 0;
> > +
> > + tp->window_clamp = new_window_clamp;
> > + if (new_window_clamp < old_window_clamp) {
> > + tp->rcv_ssthresh = min(tp->rcv_ssthresh,
> > + new_window_clamp);
> > + } else {
> > + new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
> > + tp->rcv_ssthresh = max(new_rcv_ssthresh,
> > + tp->rcv_ssthresh);
> > + }
> > }
> > return 0;
> > }
>
> It seems there is no provision for SO_RESERVE_MEM
Indeed I did take that in account.
> I wonder if tcp_adjust_rcv_ssthresh() could help here ?
I don't know how to fit it into the above. tcp_adjust_rcv_ssthresh()
tends to shrink rcv_ssthresh to low values when no memory is reserved.
Dealing directly with SO_RESERVE_MEM when shrinking the threshold feels
easier to me, something alike:
if (new_window_clamp == old_window_clamp)
return 0;
tp->window_clamp = new_window_clamp;
if (new_window_clamp < old_window_clamp) {
int unused_mem = sk_unused_reserved_mem(sk);
tp->rcv_ssthresh = min(tp->rcv_ssthresh,
new_window_clamp);
if (unused_mem)
tp->rcv_ssthresh = max_t(u32, tp->rcv_ssthresh,
tcp_win_from_space(sk, unused_mem));
} else {
new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
tp->rcv_ssthresh = max(new_rcv_ssthresh,
tp->rcv_ssthresh);
}
Possibly the bits shared with tcp_adjust_rcv_ssthresh() could be
factored out in a common helper.
> Have you considered reverting 3aa7857fe1d7 ("tcp: enable mid stream
> window clamp") ?
That would work, too and will be simpler.
The issue at hand was noted with an application that really wants to
limit the announced window:
https://gitlab.com/dgibson/passt
I guess touching rcv_ssthresh would be a bit more effective.
Not much more in the end, as both window_clamp and rcv_ssthresh can
later grow due to rcv buf auto-tune. Ideally we would like to prevent
tcp_rcv_space_adjust() from touching window_clamp after
TCP_WINDOW_CLAMP - but that is another matter/patch.
Thanks!
Paolo
Powered by blists - more mailing lists