From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: tcp hang when socket fills up ? Date: Sun, 15 Apr 2018 19:26:09 -0700 Message-ID: <38b6a690-e01d-471f-ce85-e6e6a8acd26d@gmail.com> References: <20180406090720.GA31845@nautica> <20180413163254.ucyvnkqpb4psluea@unicorn.suse.cz> <20180414010907.GA11526@nautica> <863dc88d-bc91-d3b6-70c6-5e279451f66d@gmail.com> <20180414015515.GA24798@nautica> <20180416014740.GA12245@nautica> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Cc: Michal Kubecek , netdev@vger.kernel.org To: Dominique Martinet , Eric Dumazet Return-path: Received: from mail-pf0-f173.google.com ([209.85.192.173]:44439 "EHLO mail-pf0-f173.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752552AbeDPC0L (ORCPT ); Sun, 15 Apr 2018 22:26:11 -0400 Received: by mail-pf0-f173.google.com with SMTP id p15so9800389pff.11 for ; Sun, 15 Apr 2018 19:26:11 -0700 (PDT) In-Reply-To: <20180416014740.GA12245@nautica> Content-Language: en-US Sender: netdev-owner@vger.kernel.org List-ID: On 04/15/2018 06:47 PM, Dominique Martinet wrote: > Also, here are the per-socket stats I could find (ss -i after having > reproduced hang): > reno wscale:7,7 rto:7456 backoff:5 rtt:32.924/1.41 ato:40 mss:1374 > pmtu:1500 rcvmss:1248 advmss:1448 cwnd:1 ssthresh:16 > bytes_acked:32004 bytes_received:4189 segs_out:85 segs_in:54 > data_segs_out:78 data_segs_in:18 send 333.9Kbps lastsnd:3912 > lastrcv:11464 lastack:11387 pacing_rate 21.4Mbps delivery_rate > 3.5Mbps busy:12188ms unacked:33 retrans:1/5 lost:33 rcv_rtt:37 > rcv_space:29200 rcv_ssthresh:39184 notsent:28796 minrtt:24.986 > ss -temoi might give us more info Really it looks like at some point, all incoming packets are shown by tcpdump but do not reach the TCP socket anymore. (segs_in: might be steady, look at the d0 counter shown by ss -temoi (dX : drop counters, sk->sk_drops) Are you sure you do not have some iptables/netfilter stuff ? While running your experiment, try on the server. perf record -a -g -e skb:kfree_skb sleep 30 perf report