Message-Id: <20121219131316.7d13fcb1.akpm@linux-foundation.org>
Date:	Wed, 19 Dec 2012 13:13:16 -0800
From:	Andrew Morton <akpm@...ux-foundation.org>
To:	Michal Hocko <mhocko@...e.cz>
Cc:	linux-mm@...ck.org, linux-kernel@...r.kernel.org,
	Mel Gorman <mgorman@...e.de>, Rik van Riel <riel@...hat.com>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>
Subject: Re: [PATCH v2] mm: limit mmu_gather batching to fix soft lockups on
 !CONFIG_PREEMPT

On Wed, 19 Dec 2012 16:04:37 +0100
Michal Hocko <mhocko@...e.cz> wrote:

> Since e303297 (mm: extended batches for generic mmu_gather) we are batching
> pages to be freed until either tlb_next_batch cannot allocate a new batch or we
> are done.
> 
> This works just fine most of the time but we can get in troubles with
> non-preemptible kernel (CONFIG_PREEMPT_NONE or CONFIG_PREEMPT_VOLUNTARY)
> on large machines where too aggressive batching might lead to soft
> lockups during process exit path (exit_mmap) because there are no
> scheduling points down the free_pages_and_swap_cache path and so the
> freeing can take long enough to trigger the soft lockup.
> 
> The lockup is harmless except when the system is setup to panic on
> softlockup which is not that unusual.
> 
> The simplest way to work around this issue is to limit the maximum
> number of batches in a single mmu_gather for !CONFIG_PREEMPT kernels.
> Let's use 1G of resident memory for the limit for now. This shouldn't
> make the batching less effective and it shouldn't trigger lockups as
> well because freeing 262144 should be OK.
> 
> ...
>
> diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h
> index ed6642a..5843f59 100644
> --- a/include/asm-generic/tlb.h
> +++ b/include/asm-generic/tlb.h
> @@ -78,6 +78,19 @@ struct mmu_gather_batch {
>  #define MAX_GATHER_BATCH	\
>  	((PAGE_SIZE - sizeof(struct mmu_gather_batch)) / sizeof(void *))
>  
> +/*
> + * Limit the maximum number of mmu_gather batches for non-preemptible kernels
> + * to reduce a risk of soft lockups on huge machines when a lot of memory is
> + * zapped during unmapping.
> + * 1GB of resident memory should be safe to free up at once even without
> + * explicit preemption point.
> + */
> +#if defined(CONFIG_PREEMPT_COUNT)
> +#define MAX_GATHER_BATCH_COUNT	(UINT_MAX)
> +#else
> +#define MAX_GATHER_BATCH_COUNT	(((1UL<<(30-PAGE_SHIFT))/MAX_GATHER_BATCH))

Geeze.  I spent waaaaay too long staring at that expression trying to
work out "how many pages is in a batch" and gave up.

Realistically, I don't think we need to worry about CONFIG_PREEMPT here
- if we just limit the thing to, say, 64k pages per batch then that
will be OK for preemptible and non-preemptible kernels.  The
performance difference between "64k" and "infinite" will be miniscule
and unmeasurable.

Also, the batch count should be independent of PAGE_SIZE.  Because
PAGE_SIZE can vary by a factor of 16 and you don't want to fix the
problem on 4k page size but leave it broken on 64k page size.

Also, while the patch might prevent softlockup warnings, the kernel
will still exhibit large latency glitches and those are undesirable.

Also, does this patch actually work?  It doesn't add a scheduling
point.  It assumes that by returning zero from tlb_next_batch(), the
process will back out to some point where it hits a cond_resched()?


So I'm thinking that to address both the softlockup-detector problem
and the large-latency-glitch problem we should do something like:

	if (need_resched() && tlb->batch_count > 64k)
		return 0;

and then ensure that there's a cond_resched() at a safe point between
batches?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists