netdev - Re: [PATCHv 2] tcp: properly initialize tcp memory limits part 2 (fix nfs regression)

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <4F545AC3.50004@redhat.com>
Date:	Mon, 05 Mar 2012 14:18:43 +0800
From:	Jason Wang <jasowang@...hat.com>
To:	Sergei Trofimovich <slyich@...il.com>
CC:	Glauber Costa <glommer@...allels.com>,
	linux-kernel@...r.kernel.org, netdev@...r.kernel.org,
	"David S. Miller" <davem@...emloft.net>
Subject: Re: [PATCHv 2] tcp: properly initialize tcp memory limits part 2
 (fix nfs regression)

On 03/04/2012 05:14 PM, Sergei Trofimovich wrote:
> On Sat, 3 Mar 2012 20:27:17 -0300
> Glauber Costa<glommer@...allels.com>  wrote:
>
>> On 03/03/2012 11:43 AM, Sergei Trofimovich wrote:
>>> On Sat, 3 Mar 2012 11:16:41 -0300
>>> Glauber Costa<glommer@...allels.com>   wrote:
>>>
>>>> On 03/02/2012 02:50 PM, Sergei Trofimovich wrote:
>>>>>>>> The change looks like a typo (division flipped to multiplication):
>>>>>>>>> limit = nr_free_buffer_pages() / 8;
>>>>>>>>> limit = nr_free_buffer_pages()<<     (PAGE_SHIFT - 10);
>>>>>>> Hi, thanks for the reporting. It's not a typo. It was previously:
>>>>>>> sysctl_tcp_mem[1]<<    (PAGE_SHIFT -  7). Looks like we need to do the
>>>>>>> limit check before shift the value. Please try the following patch, thanks.
>>>>>> Still does not help. I test it by checking sha1sum of a large file over NFS
>>>>>> (small files seem to work simetimes):
>>>>>>
>>>>>>        $ strace sha1sum /gentoo/distfiles/gcc-4.6.2.tar.bz2
>>>>>>        ...
>>>>>>        open("/gentoo/distfiles/gcc-4.6.2.tar.bz2", O_RDONLY
>>>>>>        <HUNG>
Hi Sergei:

Looks like the client does not even start to read the file.
>>>>>> After a certain timeout dmesg gets odd spam:
>>>>>> [  314.848094] nfs: server vmhost not responding, still trying
>>>>>> [  314.848134] nfs: server vmhost not responding, still trying
>>>>>> [  314.848145] nfs: server vmhost not responding, still trying
>>>>>> [  314.957047] nfs: server vmhost not responding, still trying
>>>>>> [  314.957066] nfs: server vmhost not responding, still trying
>>>>>> [  314.957075] nfs: server vmhost not responding, still trying
>>>>>> [  314.957085] nfs: server vmhost not responding, still trying
>>>>>> [  314.957100] nfs: server vmhost not responding, still trying
>>>>>> [  314.958023] nfs: server vmhost not responding, still trying
>>>>>> [  314.958035] nfs: server vmhost not responding, still trying
>>>>>> [  314.958044] nfs: server vmhost not responding, still trying
>>>>>> [  314.958054] nfs: server vmhost not responding, still trying
>>>>>>
>>>>>> looks like bogus messages. Might be relevant to mishandled timings
>>>>>> somewhere else or a bug in nfs code.

  Did you use a virtual machine as your NFS server? Have you tried to 
bisect the server side code?
>>>>> And after 120 seconds hung tasks shows it might be an OOM issue
>>>>> Likely caused by patch, as it's a 2GB RAM +4GB swap amd64 box
>>>>> not running anything heavy:
>>>> That is a bit weird.
>>>>
>>>> First because with Jason's patch, we should end up with the very same
>>>> calculation, at the same exact order, as it was in older kernels.
>>>> Second, because by shifting<<   10, you should be ending up with very
>>>> small numbers, effectively having tcp_rmem[1] == tcp_rmem[2], and the
>>>> same for wmem.
>>>>
>>>> Can you share which numbers you end up with at
>>>> /proc/sys/net/ipv4/tcp_{r,w}mem ?
>>>>
>>> Sure:
>>>
>>>       $ cat /proc/sys/net/ipv4/tcp_{r,w}mem
>>>       4096    87380   1999072
>>>       4096    16384   1999072
>>>
>> Sergei,
>>
>> Sorry for not being clearer. I was expecting you'd post those values
>> both in the scenario in which you see the bug, and in the scenario you
>> don't.
> Ah, I see.  Sorry. Patches are on top of v3.3-rc5-166-g1f033c1. Buggy one:
>> -       limit = nr_free_buffer_pages()<<  (PAGE_SHIFT - 10);
>> -       limit = max(limit, 128UL);
>> +       limit = nr_free_buffer_pages() / 8;
>> +       limit = max(limit, 128UL)<<  (PAGE_SHIFT - 7);
>>          max_share = min(4UL*1024*1024, limit);
>> +       printk(KERN_INFO "TCP: max_share=%u\n", max_share);
>      $ cat /proc/sys/net/ipv4/tcp_{r,w}mem
>      4096    87380   1999072
>      4096    16384   1999072

Nothing strange to me.
> Working one:
>> -       limit = nr_free_buffer_pages()<<  (PAGE_SHIFT - 10);
>> +       limit = nr_free_buffer_pages()>>  (PAGE_SHIFT - 10);
>>          limit = max(limit, 128UL);
>>          max_share = min(4UL*1024*1024, limit);
>> +       printk(KERN_INFO "TCP: max_share=%u\n", max_share);
>      $ cat /proc/sys/net/ipv4/tcp_{r,w}mem
>      4096    87380   124942
>      4096    16384   124942

This one looks small to me, as the tcp_{r,w}mem were count by bytes and 
limit were count by number of pages, so we need to shift PAGE_SHIFT.

As I can't reproduce this locally, in order to narrow down the problem, 
could you please help to check whether the issue were 
introduced/eliminated by commit  4acb4190 or 3dc43e3?

Thanks
>>> Nothing special with NFS nere, so I guess it uses UDP.
>>> TCP works fine on machine (I do everything via SSH).
>> Can you confirm that? If you're using nfs through udp, it makes
>> even less sense that the default values of tcp sock mem will harm
>> you. So it might be a bug somewhere else...
> Rechecked with tcpdump. It uses TCP.
>

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html