mm, vmscan: do not special-case slab reclaim when watermarks are boosted

Dave Chinner reported a problem pointing a finger at commit 1c30844d2dfe
("mm: reclaim small amounts of memory when an external fragmentation
event occurs").

The report is extensive:

and it's worth recording the most relevant parts (colorful language and
typos included).

	When running a simple, steady state 4kB file creation test to
	simulate extracting tarballs larger than memory full of small
	files into the filesystem, I noticed that once memory fills up
	the cache balance goes to hell.

	The workload is creating one dirty cached inode for every dirty
	page, both of which should require a single IO each to clean and
	reclaim, and creation of inodes is throttled by the rate at which
	dirty writeback runs at (via balance dirty pages). Hence the ingest
	rate of new cached inodes and page cache pages is identical and
	steady. As a result, memory reclaim should quickly find a steady
	balance between page cache and inode caches.

	The moment memory fills, the page cache is reclaimed at a much
	faster rate than the inode cache, and evidence suggests that
	the inode cache shrinker is not being called when large batches
	of pages are being reclaimed. In roughly the same time period
	that it takes to fill memory with 50% pages and 50% slab caches,
	memory reclaim reduces the page cache down to just dirty pages
	and slab caches fill the entirety of memory.

	The LRU is largely full of dirty pages, and we're getting spikes
	of random writeback from memory reclaim so it's all going to shit.
	Behaviour never recovers, the page cache remains pinned at just
	dirty pages, and nothing I could tune would make any difference.
	vfs_cache_pressure makes no difference - I would set it so high
	it should trim the entire inode caches in a single pass, yet it
	didn't do anything. It was clear from tracing and live telemetry
	that the shrinkers were pretty much not running except when
	there was absolutely no memory free at all, and then they did
	the minimum necessary to free memory to make progress.

	So I went looking at the code, trying to find places where pages
	got reclaimed and the shrinkers weren't called. There's only one
	- kswapd doing boosted reclaim as per commit 1c30844d2dfe ("mm:
	reclaim small amounts of memory when an external fragmentation
	event occurs").

The watermark boosting introduced by the commit is triggered in response
to an allocation "fragmentation event".  The boosting was not intended
to target THP specifically and triggers even if THP is disabled.
However, with Dave's perfectly reasonable workload, fragmentation events
can be very common given the ratio of slab to page cache allocations so
boosting remains active for long periods of time.

As high-order allocations might use compaction and compaction cannot
move slab pages the decision was made in the commit to special-case
kswapd when watermarks are boosted -- kswapd avoids reclaiming slab as
reclaiming slab does not directly help compaction.

As Dave notes, this decision means that slab can be artificially
protected for long periods of time and messes up the balance with slab
and page caches.

Removing the special casing can still indirectly help avoid
fragmentation by avoiding fragmentation-causing events due to slab
allocation as pages from a slab pageblock will have some slab objects
freed.  Furthermore, with the special casing, reclaim behaviour is
unpredictable as kswapd sometimes examines slab and sometimes does not
in a manner that is tricky to tune or analyse.

This patch removes the special casing.  The downside is that this is not
a universal performance win.  Some benchmarks that depend on the
residency of data when rereading metadata may see a regression when slab
reclaim is restored to its original behaviour.  Similarly, some
benchmarks that only read-once or write-once may perform better when
page reclaim is too aggressive.  The primary upside is that slab
shrinker is less surprising (arguably more sane but that's a matter of
opinion), behaves consistently regardless of the fragmentation state of
the system and properly obeys VM sysctls.

A fsmark benchmark configuration was constructed similar to what Dave
reported and is codified by the mmtest configuration
config-io-fsmark-small-file-stream.  It was evaluated on a 1-socket
machine to avoid dealing with NUMA-related issues and the timing of
reclaim.  The storage was an SSD Samsung Evo and a fresh trimmed XFS
filesystem was used for the test data.

This is not an exact replication of Dave's setup.  The configuration
scales its parameters depending on the memory size of the SUT to behave
similarly across machines.  The parameters mean the first sample
reported by fs_mark is using 50% of RAM which will barely be throttled
and look like a big outlier.  Dave used fake NUMA to have multiple
kswapd instances which I didn't replicate.  Finally, the number of
iterations differ from Dave's test as the target disk was not large
enough.  While not identical, it should be representative.

                                     5.3.0-rc3              5.3.0-rc3
                                       vanilla          shrinker-v1r1
  Min       1-files/sec     4444.80 (   0.00%)     4765.60 (   7.22%)
  1st-qrtle 1-files/sec     5005.10 (   0.00%)     5091.70 (   1.73%)
  2nd-qrtle 1-files/sec     4917.80 (   0.00%)     4855.60 (  -1.26%)
  3rd-qrtle 1-files/sec     4667.40 (   0.00%)     4831.20 (   3.51%)
  Max-1     1-files/sec    11421.50 (   0.00%)     9999.30 ( -12.45%)
  Max-5     1-files/sec    11421.50 (   0.00%)     9999.30 ( -12.45%)
  Max-10    1-files/sec    11421.50 (   0.00%)     9999.30 ( -12.45%)
  Max-90    1-files/sec     4649.60 (   0.00%)     4780.70 (   2.82%)
  Max-95    1-files/sec     4491.00 (   0.00%)     4768.20 (   6.17%)
  Max-99    1-files/sec     4491.00 (   0.00%)     4768.20 (   6.17%)
  Max       1-files/sec    11421.50 (   0.00%)     9999.30 ( -12.45%)
  Hmean     1-files/sec     5004.75 (   0.00%)     5075.96 (   1.42%)
  Stddev    1-files/sec     1778.70 (   0.00%)     1369.66 (  23.00%)
  CoeffVar  1-files/sec       33.70 (   0.00%)       26.05 (  22.71%)
  BHmean-99 1-files/sec     5053.72 (   0.00%)     5101.52 (   0.95%)
  BHmean-95 1-files/sec     5053.72 (   0.00%)     5101.52 (   0.95%)
  BHmean-90 1-files/sec     5107.05 (   0.00%)     5131.41 (   0.48%)
  BHmean-75 1-files/sec     5208.45 (   0.00%)     5206.68 (  -0.03%)
  BHmean-50 1-files/sec     5405.53 (   0.00%)     5381.62 (  -0.44%)
  BHmean-25 1-files/sec     6179.75 (   0.00%)     6095.14 (  -1.37%)

                     5.3.0-rc3   5.3.0-rc3
  Duration User         501.82      497.29
  Duration System      4401.44     4424.08
  Duration Elapsed     8124.76     8358.05

This is showing a slight skew for the max result representing a large
outlier for the 1st, 2nd and 3rd quartile are similar indicating that
the bulk of the results show little difference.  Note that an earlier
version of the fsmark configuration showed a regression but that
included more samples taken while memory was still filling.

Note that the elapsed time is higher.  Part of this is that the
configuration included time to delete all the test files when the test
completes -- the test automation handles the possibility of testing
fsmark with multiple thread counts.  Without the patch, many of these
objects would be memory resident which is part of what the patch is

There are other important observations that justify the patch.

1. With the vanilla kernel, the number of dirty pages in the system is
   very low for much of the test. With this patch, dirty pages is
   generally kept at 10% which matches vm.dirty_background_ratio which
   is normal expected historical behaviour.

2. With the vanilla kernel, the ratio of Slab/Pagecache is close to
   0.95 for much of the test i.e. Slab is being left alone and
   dominating memory consumption. With the patch applied, the ratio
   varies between 0.35 and 0.45 with the bulk of the measured ratios
   roughly half way between those values. This is a different balance to
   what Dave reported but it was at least consistent.

3. Slabs are scanned throughout the entire test with the patch applied.
   The vanille kernel has periods with no scan activity and then
   relatively massive spikes.

4. Without the patch, kswapd scan rates are very variable. With the
   patch, the scan rates remain quite steady.

4. Overall vmstats are closer to normal expectations

	                                5.3.0-rc3      5.3.0-rc3
	                                  vanilla  shrinker-v1r1
    Ops Direct pages scanned             99388.00      328410.00
    Ops Kswapd pages scanned          45382917.00    33451026.00
    Ops Kswapd pages reclaimed        30869570.00    25239655.00
    Ops Direct pages reclaimed           74131.00        5830.00
    Ops Kswapd efficiency %                 68.02          75.45
    Ops Kswapd velocity                   5585.75        4002.25
    Ops Page reclaim immediate         1179721.00      430927.00
    Ops Slabs scanned                 62367361.00    73581394.00
    Ops Direct inode steals               2103.00        1002.00
    Ops Kswapd inode steals             570180.00     5183206.00

	o Vanilla kernel is hitting direct reclaim more frequently,
	  not very much in absolute terms but the fact the patch
	  reduces it is interesting
	o "Page reclaim immediate" in the vanilla kernel indicates
	  dirty pages are being encountered at the tail of the LRU.
	  This is generally bad and means in this case that the LRU
	  is not long enough for dirty pages to be cleaned by the
	  background flush in time. This is much reduced by the
	o With the patch, kswapd is reclaiming 10 times more slab
	  pages than with the vanilla kernel. This is indicative
	  of the watermark boosting over-protecting slab

A more complete set of tests were run that were part of the basis for
introducing boosting and while there are some differences, they are well
within tolerances.

Bottom line, the special casing kswapd to avoid slab behaviour is
unpredictable and can lead to abnormal results for normal workloads.

This patch restores the expected behaviour that slab and page cache is
balanced consistently for a workload with a steady allocation ratio of
slab/pagecache pages.  It also means that if there are workloads that
favour the preservation of slab over pagecache that it can be tuned via
vm.vfs_cache_pressure where as the vanilla kernel effectively ignores
the parameter when boosting is active.

Fixes: 1c30844d2dfe ("mm: reclaim small amounts of memory when an external fragmentation event occurs")
Signed-off-by: Mel Gorman <>
Reviewed-by: Dave Chinner <>
Acked-by: Vlastimil Babka <>
Cc: Michal Hocko <>
Cc: <>	[5.0+]
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
1 file changed