(2011/06/02 16:00), Dave Chinner wrote:
> From: Dave Chinner <dchinner@xxxxxxxxxx>
> When a shrinker returns -1 to shrink_slab() to indicate it cannot do
> any work given the current memory reclaim requirements, it adds the
> entire total_scan count to shrinker->nr. The idea ehind this is that
> whenteh shrinker is next called and can do work, it will do the work
> of the previously aborted shrinker call as well.
> However, if a filesystem is doing lots of allocation with GFP_NOFS
> set, then we get many, many more aborts from the shrinkers than we
> do successful calls. The result is that shrinker->nr winds up to
> it's maximum permissible value (twice the current cache size) and
> then when the next shrinker call that can do work is issued, it
> has enough scan count built up to free the entire cache twice over.
> This manifests itself in the cache going from full to empty in a
> matter of seconds, even when only a small part of the cache is
> needed to be emptied to free sufficient memory.
> Under metadata intensive workloads on ext4 and XFS, I'm seeing the
> VFS caches increase memory consumption up to 75% of memory (no page
> cache pressure) over a period of 30-60s, and then the shrinker
> empties them down to zero in the space of 2-3s. This cycle repeats
> over and over again, with the shrinker completely trashing the іnode
> and dentry caches every minute or so the workload continues.
> This behaviour was made obvious by the shrink_slab tracepoints added
> earlier in the series, and made worse by the patch that corrected
> the concurrent accounting of shrinker->nr.
> To avoid this problem, stop repeated small increments of the total
> scan value from winding shrinker->nr up to a value that can cause
> the entire cache to be freed. We still need to allow it to wind up,
> so use the delta as the "large scan" threshold check - if the delta
> is more than a quarter of the entire cache size, then it is a large
> scan and allowed to cause lots of windup because we are clearly
> needing to free lots of memory.
> If it isn't a large scan then limit the total scan to half the size
> of the cache so that windup never increases to consume the whole
> cache. Reducing the total scan limit further does not allow enough
> wind-up to maintain the current levels of performance, whilst a
> higher threshold does not prevent the windup from freeing the entire
> cache under sustained workloads.
> Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx>
> mm/vmscan.c | 14 ++++++++++++++
> 1 files changed, 14 insertions(+), 0 deletions(-)
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index dce2767..3688f47 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -277,6 +277,20 @@ unsigned long shrink_slab(struct shrink_control *shrink,
> + * Avoid excessive windup on fielsystem shrinkers due to large
> + * numbers of GFP_NOFS allocations causing the shrinkers to
> + * return -1 all the time. This results in a large nr being
> + * built up so when a shrink that can do some work comes along
> + * it empties the entire cache due to nr >>> max_pass. This is
> + * bad for sustaining a working set in memory.
> + *
> + * Hence only allow nr to go large when a large delta is
> + * calculated.
> + */
> + if (delta < max_pass / 4)
> + total_scan = min(total_scan, max_pass / 2);
> + /*
> * Avoid risking looping forever due to too large nr value:
> * never try to free more than twice the estimate number of
> * freeable entries.
I guess "max_pass/4" and "min(total_scan, max_pass / 2)" are your heuristic
If so, please write your benchmark name and its result into the description. I
currently some mm folks plan to enhance shrinker. So, sharing benchmark may
help to avoid
an accidental regression.
I mean, your code itself looks pretty good to me.