From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: Failure to send fragmented IP packet in case of missing ARP entry Date: Mon, 10 Sep 2012 11:42:51 +0200 Message-ID: <1347270171.1234.1353.camel@edumazet-glaptop> References: <504DAC02.8040808@cogentembedded.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Cc: netdev To: Andrei Dolnikov Return-path: Received: from mail-ey0-f174.google.com ([209.85.215.174]:55028 "EHLO mail-ey0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751034Ab2IJJm5 (ORCPT ); Mon, 10 Sep 2012 05:42:57 -0400 Received: by eaac11 with SMTP id c11so780628eaa.19 for ; Mon, 10 Sep 2012 02:42:56 -0700 (PDT) In-Reply-To: <504DAC02.8040808@cogentembedded.com> Sender: netdev-owner@vger.kernel.org List-ID: On Mon, 2012-09-10 at 12:59 +0400, Andrei Dolnikov wrote: > Hello all, > > The following issue is observed on most Linux distributions: > Transmission of fragmented IP packets in case of missing ARP entry for > destination IP fails. > Actually ARP request is sent, and, once ARP response is received, only > few queued fragments are transmitted. Remaining fragments are lost. > It can be easily reproduced as follows: > # arp -d > # ping -s 65000 -c 1 > Ping result is: "1 packets transmitted, 0 received, 100% packet loss, > time 0ms". > > The latest kernel version I tried was 3.5.0-1 x86_64, but I also was > able to reproduce it with 3.2.x, 3.0.x and 2.6.32. > It doesn't depend on hardware: was able to reproduce with VMWare Player, > Intel based laptop, Intel Atom and ARM based custom boards. > As I'm not a networking standards expert I'm not sure if it's a real bug > or acceptable behaviour, but decided to raise the issue here as I can't > reproduce this anomaly with the Windows 7 PC. > > Thanks, > Andrei. > -- Its a bit better with linux-3.3, with commit 8b5c171bb3dc0686b2647a84e990199c5faa9ef8 (neigh: new unresolved queue limits) +neigh/default/unres_qlen_bytes - INTEGER + The maximum number of bytes which may be used by packets + queued for each unresolved address by other network layers. + (added in linux 3.3) + +neigh/default/unres_qlen - INTEGER + The maximum number of packets which may be queued for each + unresolved address by other network layers. + (deprecated in linux 3.3) : use unres_qlen_bytes instead. Problem is : unres_qlen_bytes default value is 65536, so its a bit too small once you take into account truesize overhead I guess following patch would be needed : diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c index 4780045..3395bb6 100644 --- a/net/ipv4/arp.c +++ b/net/ipv4/arp.c @@ -171,7 +171,7 @@ struct neigh_table arp_tbl = { .gc_staletime = 60 * HZ, .reachable_time = 30 * HZ, .delay_probe_time = 5 * HZ, - .queue_len_bytes = 64*1024, + .queue_len_bytes = 64 * SKB_TRUESIZE(1024), .ucast_probes = 3, .mcast_probes = 3, .anycast_delay = 1 * HZ,