Re: unexpected -ENOMEM from percpu_counter_init()

[Dennis Zhou <dennis@kernel.org>]

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.

> 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.

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

> Best Regards
> Wang Yugui (wangyugui@e16-tech.com)
> 2021/04/11
> 

Thanks,
Dennis