Re: [net,v2] tcp: correct handling of extreme memory squeeze

[Stefano Brivio <sbrivio@redhat.com>]

On Mon, 27 Jan 2025 21:37:23 +0800
Menglong Dong <menglong8.dong@gmail.com> wrote:

> On Mon, Jan 27, 2025 at 6:32 PM Stefano Brivio <sbrivio@redhat.com> wrote:
> >
> > On Mon, 27 Jan 2025 18:17:28 +0800
> > Jason Xing <kerneljasonxing@gmail.com> wrote:
> >  
> > > I'm not that sure if it's a bug belonging to the Linux kernel.  
> >
> > It is, because for at least 20-25 years (before that it's a bit hard to
> > understand from history) a non-zero window would be announced, as
> > obviously expected, once there's again space in the receive window.  
> 
> Sorry for the late reply. I think the key of this problem is
> what should we do when we receive a tcp packet and we are
> out of memory.
> 
> The RFC doesn't define such a thing,

Why not? RFC 9293, 3.8.6:

  There is an assumption that this is related to the data buffer space
  currently available for this connection.

That is, out-of-memory -> zero window.

> so in the commit
> e2142825c120 ("net: tcp: send zero-window ACK when no memory"),
> I reply with a zero-window ACK to the peer.

Your patch is fundamentally correct, nobody is disputing that. The
problem is that it introduces a side effect because it gets the notion
of "current window" out of sync by sending a one-off packet with a
zero-window, without recording that.

> And the peer will keep
> probing the window by retransmitting the packet that we dropped if
> the peer is a LINUX SYSTEM.
> 
> As I said, the RFC doesn't define such a case, so the behavior of
> the peer is undefined if it is not a LINUX SYSTEM. If the peer doesn't
> keep retransmitting the packet, it will hang the connection, just like
> the problem that described in this commit log.

It's not undefined. RFC 9293 3.8.6.1 (just like RFC 1122 4.2.2.17,
RFC 793 3.7) requires zero-window probes.

But keeping the window closed indefinitely if there's no zero-window
probe is a regression anyway:

- a retransmission timeout must elapse (RFC 9293 3.8.1) before the
  zero-window probe is sent, so relying on zero-window probes means
  introducing an unnecessary delay

- if the peer (as it was the case here) fails to send a zero-window
  probe for whatever reason, things break. This is a userspace
  breakage, regardless of the fact that the peer should send a
  zero-window probe

> However, we can make some optimization to make it more
> adaptable. We can send a ACK with the right window to the
> peer when the memory is available, and __tcp_cleanup_rbuf()
> is a good choice.
> 
> Generally speaking, I think this patch makes sense. However,
> I'm not sure if there is any other influence if we make
> "tp->rcv_wnd=0", but it can trigger a ACK in __tcp_cleanup_rbuf().

I don't understand what's your concern with the patch that was proposed
(and tested quite thoroughly, by the way).

> Following is the code that I thought before to optimize this
> case (the code is totally not tested):
>
> [...]

-- 
Stefano