linux-mm.kvack.org archive mirror
 help / color / mirror / Atom feed
From: Wang Yugui <wangyugui@e16-tech.com>
To: Dennis Zhou <dennis@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>,
	linux-mm@kvack.org, linux-btrfs@vger.kernel.org
Subject: Re: unexpected -ENOMEM from percpu_counter_init()
Date: Mon, 12 Apr 2021 13:24:20 +0800	[thread overview]
Message-ID: <20210412132419.2BE8.409509F4@e16-tech.com> (raw)
In-Reply-To: <YHPGdVDc6/4jxggg@google.com>

Hi,

> On Sun, Apr 11, 2021 at 11:20:00PM +0800, Wang Yugui wrote:
> > Hi, Dennis Zhou
> > 
> > > Hi,
> > > 
> > > > On Sat, Apr 10, 2021 at 11:29:17PM +0800, Wang Yugui wrote:
> > > > > Hi, Dennis Zhou 
> > > > > 
> > > > > Thanks for your ncie answer.
> > > > > but still a few questions.
> > > > > 
> > > > > > Percpu is not really cheap memory to allocate because it has a
> > > > > > amplification factor of NR_CPUS. As a result, percpu on the critical
> > > > > > path is really not something that is expected to be high throughput.
> > > > > 
> > > > > > Ideally things like btrfs snapshots should preallocate a number of these
> > > > > > and not try to do atomic allocations because that in theory could fail
> > > > > > because even after we go to the page allocator in the future we can't
> > > > > > get enough pages due to needing to go into reclaim.
> > > > > 
> > > > > pre-allocate in module such as mempool_t is just used in a few place in
> > > > > linux/fs.  so most people like system wide pre-allocate, because it is
> > > > > more easy to use?
> > > > > 
> > > > > can we add more chance to management the system wide pre-alloc
> > > > > just like this?
> > > > > 
> > > > > diff --git a/include/linux/sched/mm.h b/include/linux/sched/mm.h
> > > > > index dc1f4dc..eb3f592 100644
> > > > > --- a/include/linux/sched/mm.h
> > > > > +++ b/include/linux/sched/mm.h
> > > > > @@ -226,6 +226,11 @@ static inline void memalloc_noio_restore(unsigned int flags)
> > > > >  static inline unsigned int memalloc_nofs_save(void)
> > > > >  {
> > > > >  	unsigned int flags = current->flags & PF_MEMALLOC_NOFS;
> > > > > +
> > > > > +	// just like slab_pre_alloc_hook
> > > > > +	fs_reclaim_acquire(current->flags & gfp_allowed_mask);
> > > > > +	fs_reclaim_release(current->flags & gfp_allowed_mask);
> > > > > +
> > > > >  	current->flags |= PF_MEMALLOC_NOFS;
> > > > >  	return flags;
> > > > >  }
> > > > > 
> > > > > 
> > > > > > The workqueue approach has been good enough so far. Technically there is
> > > > > > a higher priority workqueue that this work could be scheduled on, but
> > > > > > save for this miss on my part, the system workqueue has worked out fine.
> > > > > 
> > > > > > In the future as I mentioned above. It would be good to support actually
> > > > > > getting pages, but it's work that needs to be tackled with a bit of
> > > > > > care. I might target the work for v5.14.
> > > > > > 
> > > > > > > this is our application pipeline.
> > > > > > > 	file_pre_process |
> > > > > > > 	bwa.nipt xx |
> > > > > > > 	samtools.nipt sort xx |
> > > > > > > 	file_post_process
> > > > > > > 
> > > > > > > file_pre_process/file_post_process is fast, so often are blocked by
> > > > > > > pipe input/output.
> > > > > > > 
> > > > > > > 'bwa.nipt xx' is a high-cpu-load, almost all of CPU cores.
> > > > > > > 
> > > > > > > 'samtools.nipt sort xx' is a high-mem-load, it keep the input in memory.
> > > > > > > if the memory is not enough, it will save all the buffer to temp file,
> > > > > > > so it is sometimes high-IO-load too(write 60G or more to file).
> > > > > > > 
> > > > > > > 
> > > > > > > xfstests(generic/476) is just high-IO-load, cpu/memory load is NOT high.
> > > > > > > so xfstests(generic/476) maybe easy than our application pipeline.
> > > > > > > 
> > > > > > > Although there is yet not a simple reproducer for another problem
> > > > > > > happend here, but there is a little high chance that something is wrong
> > > > > > > in btrfs/mm/fs-buffer.
> > > > > > > > but another problem(os freezed without call trace, PANIC without OOPS?,
> > > > > > > > the reason is yet unkown) still happen.
> > > > > > 
> > > > > > I do not have an answer for this. I would recommend looking into kdump.
> > > > > 
> > > > > percpu ENOMEM problem blocked many heavy load test a little long time?
> > > > > I still guess this problem of system freeze is a mm/btrfs problem.
> > > > > OOM not work, OOPS not work too.
> > > > > 
> > > > 
> > > > I don't follow. Is this still a problem after the patch?
> > > 
> > > 
> > > After the patch for percpu ENOMEM,  the problem of system freeze have a high
> > > frequecy (>75%) to be triggered by our user-space application.
> > > 
> > > The problem of system freeze maybe not caused by the percpu ENOMEM patch.
> > > 
> > > percpu ENOMEM problem maybe more easy to happen than the problem of
> > > system freeze.
> > 
> > After highmem zone +80% / otherzone +40% of WMARK_MIN/ WMARK_LOW/
> > WMARK_HIGH, we walked around or reduced the reproduce frequency of the
> > problem of system freeze.
> > 
> > so this is a problem of linux-mm.
> > 
> > the user case of our user-space application.
> > 1)  write the files with the total size > 3 * memory size.
> >      the memory size > 128G
> > 2)  btrfs with SSD/SAS, SSD/SATA, or btrfs RAID6 hdd
> >     SSD/NVMe maybe too fast, so difficult to reproduce.
> > 3) some CPU load, and some memory load.
> > 
> 
> To me it just sounds like writeback is slow. It's hard to debug a system
> without actually observing it as well. You might want to limit the
> memory allotted to the workload cgroup possibly memory.high. This may
> help kick reclaim in earlier.

It is  system panic (but NO OOPS info).
some info is outputed to IPMI/dell iDRAC
   2 | 04/03/2021 | 11:35:01 | OS Critical Stop #0x46 | Run-time critical stop () | Asserted
   3 | Linux kernel panic: Fatal excep
   4 | Linux kernel panic: tion

Now I use vm.watermark_scale_factor=40 (default 10) to help kick reclaim
in earlier,  the reproduce frequency is reduced but still >25%.

I will start a new mail thread because this is not related to
percpu ENOMEM.

> > btrfs and other fs seem not like mempool_t wiht pre-alloc, so difficult
> > job is left to the system-wide reclaim/pre-alloc of linux-mm.
> > 
> > maye memalloc_nofs_save() or memalloc_nofs_restore() is a good place to
> >  add some sync/aysnc memory reclaim/pre-alloc operations for WMARK_MIN/
> > WMARK_LOW/WMARK_HIGH and percpu PCPU_EMPTY_POP_PAGES_LOW.
> > 
> 
> It's not that simple. Memory reclaim is a balancing act and these places
> mark where reclaim cannot trigger writeback and thus oom-killer is the
> only way out. I'm sorry, but beyond the above, I don't really have any
> additional advice besides retuning your workload to use less memory and
> give the system more headroom.

Thanks for this info.

> I appreciate the bug report though and if its anything percpu related I
> will always be available.

The patch that fixed percpu ENOMEM already helps a lot. Thanks.

Best Regards
Wang Yugui (wangyugui@e16-tech.com)
2021/04/12



  reply	other threads:[~2021-04-12  5:24 UTC|newest]

Thread overview: 27+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-04-01 10:51 Wang Yugui
2021-04-02  1:49 ` Wang Yugui
2021-04-07 12:35 ` Vlastimil Babka
2021-04-07 13:09   ` Wang Yugui
2021-04-07 14:56     ` Dennis Zhou
2021-04-07 23:28       ` Wang Yugui
2021-04-08  2:44         ` Dennis Zhou
2021-04-08  9:20           ` Wang Yugui
2021-04-08 13:48             ` Dennis Zhou
2021-04-08 14:28               ` Filipe Manana
2021-04-08 15:02                 ` Dennis Zhou
2021-04-09 11:39                   ` Filipe Manana
2021-04-09 13:39                     ` Dennis Zhou
2021-04-09 13:42                       ` Filipe Manana
2021-04-09  0:08               ` Wang Yugui
2021-04-09  2:14                 ` Dennis Zhou
2021-04-09  4:02                   ` Wang Yugui
2021-04-09  7:36                     ` Wang Yugui
2021-04-09  7:48                       ` Wang Yugui
2021-04-09 13:56                       ` Dennis Zhou
2021-04-10 15:29                         ` Wang Yugui
2021-04-10 15:52                           ` Dennis Zhou
2021-04-10 16:08                             ` Wang Yugui
2021-04-11 15:20                               ` Wang Yugui
2021-04-12  4:03                                 ` Dennis Zhou
2021-04-12  5:24                                   ` Wang Yugui [this message]
2021-04-09  9:52   ` Wang Yugui

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210412132419.2BE8.409509F4@e16-tech.com \
    --to=wangyugui@e16-tech.com \
    --cc=dennis@kernel.org \
    --cc=linux-btrfs@vger.kernel.org \
    --cc=linux-mm@kvack.org \
    --cc=vbabka@suse.cz \
    --subject='Re: unexpected -ENOMEM from percpu_counter_init()' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
on how to clone and mirror all data and code used for this inbox