Re: Failure to send fragmented IP packet in case of missing ARP entry

[Andrei Dolnikov <andrey.dolnikov@cogentembedded.com>]

Works for me.

Thank you!
Andrei.

On 09/10/2012 01:53 PM, Eric Dumazet wrote:
> On Mon, 2012-09-10 at 11:42 +0200, Eric Dumazet wrote:
>> 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 <dst IP>
>>>       # ping -s 65000 -c 1 <dst IP>
>>> 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,
> In the mean time, you can also do
>
> echo 50 >/proc/sys/net/ipv4/neigh/eth0/unres_qlen
>
> (change eth0 by the name of your interface)
>
>
>