Re: TCP receive failure

On Tue, 2020-03-10 at 12:36 -0400, Neal Cardwell wrote:

On Tue, Mar 10, 2020 at 12:19 PM Eric Dumazet <eric.dumazet@gmail.com

quoted

wrote:

On 3/10/20 2:40 AM, David Woodhouse wrote:

quoted

I'm chasing a problem which was reported to me as an OpenConnect

packet

quoted

loss, with downloads stalling until curl times out and aborts.

I can't see a transport problem though; I think I see TCP on the
receive side misbehaving. This is an Ubuntu 5.3.x client kernel
(5.3.0-40-generic #32~18.04.1-Ubuntu) which I think is 5.3.18?

The test is just downloading a large file full of zeroes. The

problem

quoted

starts with a bit of packet loss and a 40ms time warp:

David, would you be able to post a binary tcpdump pcap file
(including SYN/SYNACK, to get the wscale option) at a public HTTP/S
URL somewhere? A time/sequence plot may shed some light on this, but
the tools that generate those generally want a pcap file as input.

I also like Eric's idea to reproduce and take periodic
ss/nstat dumps. That might be extremely informative.

Thanks. I do have the nstat part of that, taken each second during
Martin's last repro. But Willy was right — it was conntrack dropping
those packets, locally, before TCP ever saw them.

I've suggested that Martin set nf_conntrack_log_invalid so we get a
better look at why. And also perhaps nf_conntrack_tcp_be_liberal to see
if it goes away.