lkml.org 
[lkml]   [2013]   [Jul]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH for 3.2] memcg: do not trap chargers with full callstack on OOM
On Tue, Jul 16, 2013 at 06:09:05PM +0200, Michal Hocko wrote:
> On Tue 16-07-13 11:35:44, Johannes Weiner wrote:
> > On Mon, Jul 15, 2013 at 06:00:06PM +0200, Michal Hocko wrote:
> > > On Mon 15-07-13 17:41:19, Michal Hocko wrote:
> > > > On Sun 14-07-13 01:51:12, azurIt wrote:
> > > > > > CC: "Johannes Weiner" <hannes@cmpxchg.org>, linux-kernel@vger.kernel.org, linux-mm@kvack.org, "cgroups mailinglist" <cgroups@vger.kernel.org>, "KAMEZAWA Hiroyuki" <kamezawa.hiroyu@jp.fujitsu.com>, righi.andrea@gmail.com
> > > > > >> CC: "Johannes Weiner" <hannes@cmpxchg.org>, linux-kernel@vger.kernel.org, linux-mm@kvack.org, "cgroups mailinglist" <cgroups@vger.kernel.org>, "KAMEZAWA Hiroyuki" <kamezawa.hiroyu@jp.fujitsu.com>, righi.andrea@gmail.com
> > > > > >>On Wed 10-07-13 18:25:06, azurIt wrote:
> > > > > >>> >> Now i realized that i forgot to remove UID from that cgroup before
> > > > > >>> >> trying to remove it, so cgroup cannot be removed anyway (we are using
> > > > > >>> >> third party cgroup called cgroup-uid from Andrea Righi, which is able
> > > > > >>> >> to associate all user's processes with target cgroup). Look here for
> > > > > >>> >> cgroup-uid patch:
> > > > > >>> >> https://www.develer.com/~arighi/linux/patches/cgroup-uid/cgroup-uid-v8.patch
> > > > > >>> >>
> > > > > >>> >> ANYWAY, i'm 101% sure that 'tasks' file was empty and 'under_oom' was
> > > > > >>> >> permanently '1'.
> > > > > >>> >
> > > > > >>> >This is really strange. Could you post the whole diff against stable
> > > > > >>> >tree you are using (except for grsecurity stuff and the above cgroup-uid
> > > > > >>> >patch)?
> > > > > >>>
> > > > > >>>
> > > > > >>> Here are all patches which i applied to kernel 3.2.48 in my last test:
> > > > > >>> http://watchdog.sk/lkml/patches3/
> > > > > >>
> > > > > >>The two patches from Johannes seem correct.
> > > > > >>
> > > > > >>From a quick look even grsecurity patchset shouldn't interfere as it
> > > > > >>doesn't seem to put any code between handle_mm_fault and mm_fault_error
> > > > > >>and there also doesn't seem to be any new handle_mm_fault call sites.
> > > > > >>
> > > > > >>But I cannot tell there aren't other code paths which would lead to a
> > > > > >>memcg charge, thus oom, without proper FAULT_FLAG_KERNEL handling.
> > > > > >
> > > > > >
> > > > > >Michal,
> > > > > >
> > > > > >now i can definitely confirm that problem with unremovable cgroups
> > > > > >persists. What info do you need from me? I applied also your little
> > > > > >'WARN_ON' patch.
> > > > >
> > > > > Ok, i think you want this:
> > > > > http://watchdog.sk/lkml/kern4.log
> > > >
> > > > Jul 14 01:11:39 server01 kernel: [ 593.589087] [ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name
> > > > Jul 14 01:11:39 server01 kernel: [ 593.589451] [12021] 1333 12021 172027 64723 4 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.589647] [12030] 1333 12030 172030 64748 2 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.589836] [12031] 1333 12031 172030 64749 3 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.590025] [12032] 1333 12032 170619 63428 3 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.590213] [12033] 1333 12033 167934 60524 2 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.590401] [12034] 1333 12034 170747 63496 4 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.590588] [12035] 1333 12035 169659 62451 1 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.590776] [12036] 1333 12036 167614 60384 3 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.590984] [12037] 1333 12037 166342 58964 3 0 0 apache2
> > > > Jul 14 01:11:39 server01 kernel: [ 593.591178] Memory cgroup out of memory: Kill process 12021 (apache2) score 847 or sacrifice child
> > > > Jul 14 01:11:39 server01 kernel: [ 593.591370] Killed process 12021 (apache2) total-vm:688108kB, anon-rss:255472kB, file-rss:3420kB
> > > > Jul 14 01:11:41 server01 kernel: [ 595.392920] ------------[ cut here ]------------
> > > > Jul 14 01:11:41 server01 kernel: [ 595.393096] WARNING: at kernel/exit.c:888 do_exit+0x7d0/0x870()
> > > > Jul 14 01:11:41 server01 kernel: [ 595.393256] Hardware name: S5000VSA
> > > > Jul 14 01:11:41 server01 kernel: [ 595.393415] Pid: 12037, comm: apache2 Not tainted 3.2.48-grsec #1
> > > > Jul 14 01:11:41 server01 kernel: [ 595.393577] Call Trace:
> > > > Jul 14 01:11:41 server01 kernel: [ 595.393737] [<ffffffff8105520a>] warn_slowpath_common+0x7a/0xb0
> > > > Jul 14 01:11:41 server01 kernel: [ 595.393903] [<ffffffff8105525a>] warn_slowpath_null+0x1a/0x20
> > > > Jul 14 01:11:41 server01 kernel: [ 595.394068] [<ffffffff81059c50>] do_exit+0x7d0/0x870
> > > > Jul 14 01:11:41 server01 kernel: [ 595.394231] [<ffffffff81050254>] ? thread_group_times+0x44/0xb0
> > > > Jul 14 01:11:41 server01 kernel: [ 595.394392] [<ffffffff81059d41>] do_group_exit+0x51/0xc0
> > > > Jul 14 01:11:41 server01 kernel: [ 595.394551] [<ffffffff81059dc7>] sys_exit_group+0x17/0x20
> > > > Jul 14 01:11:41 server01 kernel: [ 595.394714] [<ffffffff815caea6>] system_call_fastpath+0x18/0x1d
> > > > Jul 14 01:11:41 server01 kernel: [ 595.394921] ---[ end trace 738570e688acf099 ]---
> > > >
> > > > OK, so you had an OOM which has been handled by in-kernel oom handler
> > > > (it killed 12021) and 12037 was in the same group. The warning tells us
> > > > that it went through mem_cgroup_oom as well (otherwise it wouldn't have
> > > > memcg_oom.wait_on_memcg set and the warning wouldn't trigger) and then
> > > > it exited on the userspace request (by exit syscall).
> > > >
> > > > I do not see any way how, this could happen though. If mem_cgroup_oom
> > > > is called then we always return CHARGE_NOMEM which turns into ENOMEM
> > > > returned by __mem_cgroup_try_charge (invoke_oom must have been set to
> > > > true). So if nobody screwed the return value on the way up to page
> > > > fault handler then there is no way to escape.
> > > >
> > > > I will check the code.
> > >
> > > OK, I guess I found it:
> > > __do_fault
> > > fault = filemap_fault
> > > do_async_mmap_readahead
> > > page_cache_async_readahead
> > > ondemand_readahead
> > > __do_page_cache_readahead
> > > read_pages
> > > readpages = ext3_readpages
> > > mpage_readpages # Doesn't propagate ENOMEM
> > > add_to_page_cache_lru
> > > add_to_page_cache
> > > add_to_page_cache_locked
> > > mem_cgroup_cache_charge
> > >
> > > So the read ahead most probably. Again! Duhhh. I will try to think
> > > about a fix for this. One obvious place is mpage_readpages but
> > > __do_page_cache_readahead ignores read_pages return value as well and
> > > page_cache_async_readahead, even worse, is just void and exported as
> > > such.
> > >
> > > So this smells like a hard to fix bugger. One possible, and really ugly
> > > way would be calling mem_cgroup_oom_synchronize even if handle_mm_fault
> > > doesn't return VM_FAULT_ERROR, but that is a crude hack.
> >
> > Ouch, good spot.
> >
> > I don't think we need to handle an OOM from the readahead code. If
> > readahead does not produce the desired page, we retry synchroneously
> > in page_cache_read() and handle the OOM properly. We should not
> > signal an OOM for optional pages anyway.
> >
> > So either we pass a flag from the readahead code down to
> > add_to_page_cache and mem_cgroup_cache_charge that tells the charge
> > code to ignore OOM conditions and do not set up an OOM context.
>
> That was my previous attempt and it was sooo painful.
>
> > Or we DO call mem_cgroup_oom_synchronize() from the read_cache_pages,
> > with an argument that makes it only clean up the context and not wait.
>
> Yes, I was playing with this idea as well. I just do not like how
> fragile this is. We need some way to catch all possible places which
> might leak it.

I don't think this is necessary, but we could add a sanity check
in/near mem_cgroup_clear_userfault() that makes sure the OOM context
is only set up when an error is returned.

> > It would not be completely outlandish to place it there, since it's
> > right next to where an error from add_to_page_cache() is not further
> > propagated back through the fault stack.
> >
> > I'm travelling right now, I'll send a patch when I get back
> > (Thursday). Unless you beat me to it :)
>
> I can cook something up but there is quite a big pile on my desk
> currently (as always :/).

No worries, I'll send an update.


\
 
 \ /
  Last update: 2013-07-16 19:42    [W:0.873 / U:0.032 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site