Re: sk_page_frag_refill OOM killing spree

Re: sk_page_frag_refill OOM killing spree

[Eric Dumazet]

On Tue, 2013-05-21 at 14:28 +0200, Florian Westphal wrote:
> Hi Eric,
> 
> seems like sk_page_frag_refill() can cause oom-killer invocation:
> 
> postgres invoked oom-killer: gfp_mask=0x42d0, order=3, oom_score_adj=0
> Pid: 10551, comm: postgres Tainted: G           O 3.8.6-5.g613ca40-smp #1
> Call Trace:
>  [<c106dbd5>] ? dump_header+0x60/0x191
>  [<c1133d3a>] ? ___ratelimit+0xb2/0xc4
>  [<c106dfd3>] ? oom_kill_process+0x61/0x2d1
>  [<c1030042>] ? has_capability_noaudit+0x1c/0x23
>  [<c106df0f>] ? oom_badness+0x8c/0xef
>  [<c106e446>] ? out_of_memory+0x203/0x247
>  [<c107128a>] ? __alloc_pages_nodemask+0x42b/0x4c3
>  [<c11fa66a>] ? sk_page_frag_refill+0x6a/0xd2
>  [<c1233548>] ? tcp_sendmsg+0x3e8/0x7c6
>  [<c124f34b>] ? inet_sendmsg+0x6b/0x75
>  [<c11f74d8>] ? sock_sendmsg+0x8d/0xa6
>  [<c11f7b83>] ? sys_sendto+0x105/0x130
>  [<c1025927>] ? __kunmap_atomic+0x62/0x8a
>  [<c1025940>] ? __kunmap_atomic+0x7b/0x8a
>  [<c1073d78>] ? __lru_cache_add+0x18/0x47
>  [<c10812f9>] ? handle_pte_fault+0x745/0x751
>  [<c1025a2d>] ? kmap_atomic_prot+0xd3/0xf1
>  [<c10817da>] ? handle_mm_fault+0x112/0x121
>  [<c11f7be5>] ? sys_send+0x37/0x3b
> 
> The system is busy, so, order-3 alloc failure doesn't strike me as odd.
> 
> There are no allocation failures with order != 3.
> 
> Sometimes this can happen in very short sucession, i.e.
> and oom-killer did end up zapping 30 processes or so.
> 
> My question is, should sk_page_frag_refill use __GFP_NORETRY, at least
> for order 3 requests?

Yes, you may be right, but I prefer some mm expertise here, so I CC lkml

Re: sk_page_frag_refill OOM killing spree

[David Rientjes]

On Tue, 21 May 2013, Eric Dumazet wrote:

> On Tue, 2013-05-21 at 14:28 +0200, Florian Westphal wrote:
> > Hi Eric,
> > 
> > seems like sk_page_frag_refill() can cause oom-killer invocation:
> > 
> > postgres invoked oom-killer: gfp_mask=0x42d0, order=3, oom_score_adj=0
> > Pid: 10551, comm: postgres Tainted: G           O 3.8.6-5.g613ca40-smp #1
> > Call Trace:
> >  [<c106dbd5>] ? dump_header+0x60/0x191
> >  [<c1133d3a>] ? ___ratelimit+0xb2/0xc4
> >  [<c106dfd3>] ? oom_kill_process+0x61/0x2d1
> >  [<c1030042>] ? has_capability_noaudit+0x1c/0x23
> >  [<c106df0f>] ? oom_badness+0x8c/0xef
> >  [<c106e446>] ? out_of_memory+0x203/0x247
> >  [<c107128a>] ? __alloc_pages_nodemask+0x42b/0x4c3
> >  [<c11fa66a>] ? sk_page_frag_refill+0x6a/0xd2
> >  [<c1233548>] ? tcp_sendmsg+0x3e8/0x7c6
> >  [<c124f34b>] ? inet_sendmsg+0x6b/0x75
> >  [<c11f74d8>] ? sock_sendmsg+0x8d/0xa6
> >  [<c11f7b83>] ? sys_sendto+0x105/0x130
> >  [<c1025927>] ? __kunmap_atomic+0x62/0x8a
> >  [<c1025940>] ? __kunmap_atomic+0x7b/0x8a
> >  [<c1073d78>] ? __lru_cache_add+0x18/0x47
> >  [<c10812f9>] ? handle_pte_fault+0x745/0x751
> >  [<c1025a2d>] ? kmap_atomic_prot+0xd3/0xf1
> >  [<c10817da>] ? handle_mm_fault+0x112/0x121
> >  [<c11f7be5>] ? sys_send+0x37/0x3b
> > 
> > The system is busy, so, order-3 alloc failure doesn't strike me as odd.
> > 
> > There are no allocation failures with order != 3.
> > 
> > Sometimes this can happen in very short sucession, i.e.
> > and oom-killer did end up zapping 30 processes or so.
> > 

Aside from the __GFP_NORETRY issue, could you post the full oom killer log 
where it kills more than one process with /proc/sys/vm/oom_dump_tasks 
enabled?  That shouldn't happen unless you have a memory leak.

Re: sk_page_frag_refill OOM killing spree

[Florian Westphal]

David Rientjes <rientjes@google.com> wrote:
> > On Tue, 2013-05-21 at 14:28 +0200, Florian Westphal wrote:
> > > seems like sk_page_frag_refill() can cause oom-killer invocation:
> > > 
> > > postgres invoked oom-killer: gfp_mask=0x42d0, order=3, oom_score_adj=0
> > > Pid: 10551, comm: postgres Tainted: G           O 3.8.6-5.g613ca40-smp #1
> > > Call Trace:
> > >  [<c106dbd5>] ? dump_header+0x60/0x191
> > >  [<c1133d3a>] ? ___ratelimit+0xb2/0xc4
> > >  [<c106dfd3>] ? oom_kill_process+0x61/0x2d1
> > >  [<c1030042>] ? has_capability_noaudit+0x1c/0x23
> > >  [<c106df0f>] ? oom_badness+0x8c/0xef
> > >  [<c106e446>] ? out_of_memory+0x203/0x247
> > >  [<c107128a>] ? __alloc_pages_nodemask+0x42b/0x4c3
> > >  [<c11fa66a>] ? sk_page_frag_refill+0x6a/0xd2
> > >  [<c1233548>] ? tcp_sendmsg+0x3e8/0x7c6
> > >  [<c124f34b>] ? inet_sendmsg+0x6b/0x75
> > >  [<c11f74d8>] ? sock_sendmsg+0x8d/0xa6
> > >  [<c11f7b83>] ? sys_sendto+0x105/0x130
> > >  [<c1025927>] ? __kunmap_atomic+0x62/0x8a
> > >  [<c1025940>] ? __kunmap_atomic+0x7b/0x8a
> > >  [<c1073d78>] ? __lru_cache_add+0x18/0x47
> > >  [<c10812f9>] ? handle_pte_fault+0x745/0x751
> > >  [<c1025a2d>] ? kmap_atomic_prot+0xd3/0xf1
> > >  [<c10817da>] ? handle_mm_fault+0x112/0x121
> > >  [<c11f7be5>] ? sys_send+0x37/0x3b
> > > 
> > > The system is busy, so, order-3 alloc failure doesn't strike me as odd.
> > > 
> > > There are no allocation failures with order != 3.
> > > 
> > > Sometimes this can happen in very short sucession, i.e.
> > > and oom-killer did end up zapping 30 processes or so.
> 
> Aside from the __GFP_NORETRY issue, could you post the full oom killer log 
> where it kills more than one process with /proc/sys/vm/oom_dump_tasks 
> enabled?  That shouldn't happen unless you have a memory leak.

http://strlen.de/fw/oom.txt

HOWEVER, before you spend time on this:
I don't think there is an issue with oom killer, I only noticed
that after kernel update oom killer invocations are
 a) more frequent than before
 b) often show above backtrace (sk_page_frag_refill).

I'm not even saying "sk_page_frag_refill is broken", since I don't know if
adding GFP_NORETRY might add silent performance degradation, etc.

Its very well possible that everything is working as intended :-)