[PATCH 0/11] Reduce compaction-related stalls and improve asynchronous migration of dirty pages v6

Short summary: There are severe stalls when a USB stick using VFAT
is used with THP enabled that are reduced by this series. If you are
experiencing this problem, please test and report back and considering
I have seen complaints from openSUSE and Fedora users on this as well
as a few private mails, I'm guessing it's a widespread issue. This
is a new type of USB-related stall because it is due to synchronous
compaction writing where as in the past the big problem was dirty
pages reaching the end of the LRU and being written by reclaim.

Am cc'ing Andrew this time and this series would replace
mm-do-not-stall-in-synchronous-compaction-for-thp-allocations.patch.
I'm also cc'ing Dave Jones as he might have merged that patch to Fedora
for wider testing and ideally it would be reverted and replaced by
this series.

That said, the later patches could really do with some review. If this
series is not the answer then a new direction needs to be discussed
because as it is, the stalls are unacceptable as the results in this
leader show.

For testers that try backporting this to 3.1, it won't work because
there is a non-obvious dependency on not writing back pages in direct
reclaim so you need those patches too.

Changelog since V5
o Rebase to 3.2-rc5
o Tidy up the changelogs a bit

Changelog since V4
o Added reviewed-bys, credited Andrea properly for sync-light
o Allow dirty pages without mappings to be considered for migration
o Bound the number of pages freed for compaction
o Isolate PageReclaim pages on their own LRU list

This is against 3.2-rc5 and follows on from discussions on "mm: Do
not stall in synchronous compaction for THP allocations" and "[RFC
PATCH 0/5] Reduce compaction-related stalls". Initially, the proposed
patch eliminated stalls due to compaction which sometimes resulted in
user-visible interactivity problems on browsers by simply never using
sync compaction. The downside was that THP success allocation rates
were lower because dirty pages were not being migrated as reported by
Andrea. His approach at fixing this was nacked on the grounds that
it reverted fixes from Rik merged that reduced the amount of pages
reclaimed as it severely impacted his workloads performance.

This series attempts to reconcile the requirements of maximising THP
usage, without stalling in a user-visible fashion due to compaction
or cheating by reclaiming an excessive number of pages.

Patch 1 partially reverts commit 39deaf85 to allow migration to isolate
dirty pages. This is because migration can move some dirty
pages without blocking.

Patch 2 notes that the /proc/sys/vm/compact_memory handler is not using
synchronous compaction when it should be. This is unrelated
to the reported stalls but is worth fixing.

Patch 3 checks if we isolated a compound page during lumpy scan and
account for it properly. For the most part, this affects
tracing so it's unrelated to the stalls but worth fixing.

Patch 4 notes that it is possible to abort reclaim early for compaction
and return 0 to the page allocator potentially entering the
"may oom" path. This has not been observed in practice but
the rest of the series potentially makes it easier to happen.

Patch 5 adds a sync parameter to the migratepage callback and gives
the callback responsibility for migrating the page without
blocking if sync==false. For example, fallback_migrate_page
will not call writepage if sync==false. This increases the
number of pages that can be handled by asynchronous compaction
thereby reducing stalls.

Patch 6 restores filter-awareness to isolate_lru_page for migration.
In practice, it means that pages under writeback and pages
without a ->migratepage callback will not be isolated
for migration.

Patch 7 avoids calling direct reclaim if compaction is deferred but
makes sure that compaction is only deferred if sync
compaction was used.

Patch 8 introduces a sync-light migration mechanism that sync compaction
uses. The objective is to allow some stalls but to not call
->writepage which can lead to significant user-visible stalls.

Patch 9 notes that while we want to abort reclaim ASAP to allow
compation to go ahead that we leave a very small window of
opportunity for compaction to run. This patch allows more pages
to be freed by reclaim but bounds the number to a reasonable
level based on the high watermark on each zone.

Patch 10 allows slabs to be shrunk even after compaction_ready() is
true for one zone. This is to avoid a problem whereby a single
small zone can abort reclaim even though no pages have been
reclaimed and no suitably large zone is in a usable state.

Patch 11 fixes a problem with the rate of page scanning. As reclaim is
rarely stalling on pages under writeback it means that scan
rates are very high. This is particularly true for direct
reclaim which is not calling writepage. The vmstat figures
implied that much of this was busy work with PageReclaim pages
marked for immediate reclaim. This patch is a prototype that
moves these pages to their own LRU list.

This has been tested and other than 2 USB keys getting trashed,
nothing horrible fell out. That said, I am a bit unhappy with the
rescue logic in patch 11 but did not find a better way around it. It
does significantly reduce scan rates and System CPU time indicating
it is the right direction to take.

What is of critical importance is that stalls due to compaction
are massively reduced even though sync compaction was still
allowed. Testing from people complaining about stalls copying to USBs
with THP enabled are particularly welcome.

The following tests all involve THP usage and USB keys in some
way. Each test follows this type of pattern

1. Read from some fast fast storage, be it raw device or file. Each time
the copy finishes, start again until the test ends
2. Write a large file to a filesystem on a USB stick. Each time the copy
finishes, start again until the test ends
3. When memory is low, start an alloc process that creates a mapping
the size of physical memory to stress THP allocation. This is the
"real" part of the test and the part that is meant to trigger
stalls when THP is enabled. Copying continues in the background.
4. Record the CPU usage and time to execute of the alloc process
5. Record the number of THP allocs and fallbacks as well as the number of THP
pages in use a the end of the test just before alloc exited
6. Run the test 5 times to get an idea of variability
7. Between each run, sync is run and caches dropped and the test
waits until nr_dirty is a small number to avoid interference
or caching between iterations that would skew the figures.

The individual tests were then

writebackCPDeviceBasevfat
Disable THP, read from a raw device (sda), vfat on USB stick
writebackCPDeviceBaseext4
Disable THP, read from a raw device (sda), ext4 on USB stick
writebackCPDevicevfat
THP enabled, read from a raw device (sda), vfat on USB stick
writebackCPDeviceext4
THP enabled, read from a raw device (sda), ext4 on USB stick
writebackCPFilevfat
THP enabled, read from a file on fast storage and USB, both vfat
writebackCPFileext4
THP enabled, read from a file on fast storage and USB, both ext4

The kernels tested were

3.1 3.1
vanilla 3.2-rc5
freemore Patches 1-10
immediate Patches 1-11
andrea The 8 patches Andrea posted as a basis of comparison

The results are very long unfortunately. I'll start with the case
where we are not using THP at all

writebackCPDeviceBasevfat
3.1.0-vanilla rc5-vanilla freemore-v6r1 isolate-v6r1 andrea-v2r1
System Time 1.28 ( 0.00%) 54.49 (-4143.46%) 48.63 (-3687.69%) 4.69 ( -265.11%) 51.88 (-3940.81%)
+/- 0.06 ( 0.00%) 2.45 (-4305.55%) 4.75 (-8430.57%) 7.46 (-13282.76%) 4.76 (-8440.70%)
User Time 0.09 ( 0.00%) 0.05 ( 40.91%) 0.06 ( 29.55%) 0.07 ( 15.91%) 0.06 ( 27.27%)
+/- 0.02 ( 0.00%) 0.01 ( 45.39%) 0.02 ( 25.07%) 0.00 ( 77.06%) 0.01 ( 52.24%)
Elapsed Time 110.27 ( 0.00%) 56.38 ( 48.87%) 49.95 ( 54.70%) 11.77 ( 89.33%) 53.43 ( 51.54%)
+/- 7.33 ( 0.00%) 3.77 ( 48.61%) 4.94 ( 32.63%) 6.71 ( 8.50%) 4.76 ( 35.03%)
THP Active 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
+/- 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault Alloc 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
+/- 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault Fallback 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
+/- 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)

The THP figures are obviously all 0 because THP was enabled. The
main thing to watch is the elapsed times and how they compare to
times when THP is enabled later. It's also important to note that
elapsed time is improved by this series as System CPu time is much
reduced.

writebackCPDevicevfat

3.1.0-vanilla rc5-vanilla freemore-v6r1 isolate-v6r1 andrea-v2r1
System Time 1.22 ( 0.00%) 13.89 (-1040.72%) 46.40 (-3709.20%) 4.44 ( -264.37%) 47.37 (-3789.33%)
+/- 0.06 ( 0.00%) 22.82 (-37635.56%) 3.84 (-6249.44%) 6.48 (-10618.92%) 6.60
(-10818.53%)
User Time 0.06 ( 0.00%) 0.06 ( -6.90%) 0.05 ( 17.24%) 0.05 ( 13.79%) 0.04 ( 31.03%)
+/- 0.01 ( 0.00%) 0.01 ( 33.33%) 0.01 ( 33.33%) 0.01 ( 39.14%) 0.01 ( 25.46%)
Elapsed Time 10445.54 ( 0.00%) 2249.92 ( 78.46%) 70.06 ( 99.33%) 16.59 ( 99.84%) 472.43 (
95.48%)
+/- 643.98 ( 0.00%) 811.62 ( -26.03%) 10.02 ( 98.44%) 7.03 ( 98.91%) 59.99 ( 90.68%)
THP Active 15.60 ( 0.00%) 35.20 ( 225.64%) 65.00 ( 416.67%) 70.80 ( 453.85%) 62.20 ( 398.72%)
+/- 18.48 ( 0.00%) 51.29 ( 277.59%) 15.99 ( 86.52%) 37.91 ( 205.18%) 22.02 ( 119.18%)
Fault Alloc 121.80 ( 0.00%) 76.60 ( 62.89%) 155.40 ( 127.59%) 181.20 ( 148.77%) 286.60 ( 235.30%)
+/- 73.51 ( 0.00%) 61.11 ( 83.12%) 34.89 ( 47.46%) 31.88 ( 43.36%) 68.13 ( 92.68%)
Fault Fallback 881.20 ( 0.00%) 926.60 ( -5.15%) 847.60 ( 3.81%) 822.00 ( 6.72%) 716.60 ( 18.68%)
+/- 73.51 ( 0.00%) 61.26 ( 16.67%) 34.89 ( 52.54%) 31.65 ( 56.94%) 67.75 ( 7.84%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 3540.88 1945.37 716.04 64.97 1937.03
Total Elapsed Time (seconds) 52417.33 11425.90 501.02 230.95 2520.28

The first thing to note is the "Elapsed Time" for the vanilla kernels
of 2249 seconds versus 56 with THP disabled which might explain the
reports of USB stalls with THP enabled. Applying the patches brings
performance in line with THP-disabled performance while isolating
pages for immediate reclaim from the LRU cuts down System CPU time.

The "Fault Alloc" success rate figures are also improved. The vanilla
kernel only managed to allocate 76.6 pages on average over the course
of 5 iterations where as applying the series allocated 181.20 on
average albeit it is well within variance. It's worth noting that
applies the series at least descreases the amount of variance which
implies an improvement.

Andrea's series had a higher success rate for THP allocations but
at a severe cost to elapsed time which is still better than vanilla
but still much worse than disabling THP altogether. One can bring my
series close to Andrea's by removing this check

/*
* If compaction is deferred for high-order allocations, it is because
* sync compaction recently failed. In this is the case and the caller
* has requested the system not be heavily disrupted, fail the
* allocation now instead of entering direct reclaim
*/
if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
goto nopage;

I didn't include a patch that removed the above check because hurting
overall performance to improve the THP figure is not what the average
user wants. It's something to consider though if someone really wants
to maximise THP usage no matter what it does to the workload initially.

This is summary of vmstat figures from the same test.

3.1.0-vanilla rc5-vanilla freemore-v6r1 isolate-v6r1 andrea-v2r1
Page Ins 3257266139 1111844061 17263623 10901575 161423219
Page Outs 81054922 30364312 3626530 3657687 8753730
Swap Ins 3294 2851 6560 4964 4592
Swap Outs 390073 528094 620197 790912 698285
Direct pages scanned 1077581700 3024951463 1764930052 115140570 5901188831
Kswapd pages scanned 34826043 7112868 2131265 1686942 1893966
Kswapd pages reclaimed 28950067 4911036 1246044 966475 1497726
Direct pages reclaimed 805148398 280167837 3623473 2215044 40809360
Kswapd efficiency 83% 69% 58% 57% 79%
Kswapd velocity 664.399 622.521 4253.852 7304.360 751.490
Direct efficiency 74% 9% 0% 1% 0%
Direct velocity 20557.737 264745.137 3522673.849 498551.938 2341481.435
Percentage direct scans 96% 99% 99% 98% 99%
Page writes by reclaim 722646 529174 620319 791018 699198
Page writes file 332573 1080 122 106 913
Page writes anon 390073 528094 620197 790912 698285
Page reclaim immediate 0 2552514720 1635858848 111281140 5478375032
Page rescued immediate 0 0 0 87848 0
Slabs scanned 23552 23552 9216 8192 9216
Direct inode steals 231 0 0 0 0
Kswapd inode steals 0 0 0 0 0
Kswapd skipped wait 28076 786 0 61 6
THP fault alloc 609 383 753 906 1433
THP collapse alloc 12 6 0 0 6
THP splits 536 211 456 593 1136
THP fault fallback 4406 4633 4263 4110 3583
THP collapse fail 120 127 0 0 4
Compaction stalls 1810 728 623 779 3200
Compaction success 196 53 60 80 123
Compaction failures 1614 675 563 699 3077
Compaction pages moved 193158 53545 243185 333457 226688
Compaction move failure 9952 9396 16424 23676 45070


The main things to look at are

1. Page In/out figures are much reduced by the series.

2. Direct page scanning is incredibly high (264745.137 pages scanned
per second on the vanilla kernel) but isolating PageReclaim pages
on their own list reduces the number of pages scanned significantly.

3. The fact that "Page rescued immediate" is a positive number implies
that we sometimes race removing pages from the LRU_IMMEDIATE list
that need to be put back on a normal LRU but it happens only for
0.07% of the pages marked for immediate reclaim.

writebackCPDeviceext4
3.1.0-vanilla rc5-vanilla freemore-v6r1 isolate-v6r1 andrea-v2r1
System Time 1.51 ( 0.00%) 1.77 ( -17.66%) 1.46 ( 2.92%) 1.15 ( 23.77%) 1.89 ( -25.63%)
+/- 0.27 ( 0.00%) 0.67 ( -148.52%) 0.33 ( -22.76%) 0.30 ( -11.15%) 0.19 ( 30.16%)
User Time 0.03 ( 0.00%) 0.04 ( -37.50%) 0.05 ( -62.50%) 0.07 ( -112.50%) 0.04 ( -18.75%)
+/- 0.01 ( 0.00%) 0.02 ( -146.64%) 0.02 ( -97.91%) 0.02 ( -75.59%) 0.02 ( -63.30%)
Elapsed Time 124.93 ( 0.00%) 114.49 ( 8.36%) 96.77 ( 22.55%) 27.48 ( 78.00%) 205.70 ( -64.65%)
+/- 20.20 ( 0.00%) 74.39 ( -268.34%) 59.88 ( -196.48%) 7.72 ( 61.79%) 25.03 ( -23.95%)
THP Active 161.80 ( 0.00%) 83.60 ( 51.67%) 141.20 ( 87.27%) 84.60 ( 52.29%) 82.60 ( 51.05%)
+/- 71.95 ( 0.00%) 43.80 ( 60.88%) 26.91 ( 37.40%) 59.02 ( 82.03%) 52.13 ( 72.45%)
Fault Alloc 471.40 ( 0.00%) 228.60 ( 48.49%) 282.20 ( 59.86%) 225.20 ( 47.77%) 388.40 ( 82.39%)
+/- 88.07 ( 0.00%) 87.42 ( 99.26%) 73.79 ( 83.78%) 109.62 ( 124.47%) 82.62 ( 93.81%)
Fault Fallback 531.60 ( 0.00%) 774.60 ( -45.71%) 720.80 ( -35.59%) 777.80 ( -46.31%) 614.80 ( -15.65%)
+/- 88.07 ( 0.00%) 87.26 ( 0.92%) 73.79 ( 16.22%) 109.62 ( -24.47%) 82.29 ( 6.56%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 50.22 33.76 30.65 24.14 128.45
Total Elapsed Time (seconds) 1113.73 1132.19 1029.45 759.49 1707.26

Similar test but the USB stick is using ext4 instead of vfat. As
ext4 does not use writepage for migration, the large stalls due to
compaction when THP is enabled are not observed. Still, isolating
PageReclaim pages on their own list helped completion time largely
by reducing the number of pages scanned by direct reclaim although
time spend in congestion_wait could also be a factor.

Again, Andrea's series had far higher success rates for THP allocation
at the cost of elapsed time. I didn't look too closely but a quick
look at the vmstat figures tells me kswapd reclaimed 8 times more pages
than the patch series and direct reclaim reclaimed roughly three times
as many pages. It follows that if memory is aggressively reclaimed,
there will be more available for THP.

writebackCPFilevfat
3.1.0-vanilla rc5-vanilla freemore-v6r1 isolate-v6r1 andrea-v2r1
System Time 1.76 ( 0.00%) 29.10 (-1555.52%) 46.01 (-2517.18%) 4.79 ( -172.35%) 54.89 (-3022.53%)
+/- 0.14 ( 0.00%) 25.61 (-18185.17%) 2.15 (-1434.83%) 6.60 (-4610.03%) 9.75
(-6863.76%)
User Time 0.05 ( 0.00%) 0.07 ( -45.83%) 0.05 ( -4.17%) 0.06 ( -29.17%) 0.06 ( -16.67%)
+/- 0.02 ( 0.00%) 0.02 ( 20.11%) 0.02 ( -3.14%) 0.01 ( 31.58%) 0.01 ( 47.41%)
Elapsed Time 22520.79 ( 0.00%) 1082.85 ( 95.19%) 73.30 ( 99.67%) 32.43 ( 99.86%) 291.84 ( 98.70%)
+/- 7277.23 ( 0.00%) 706.29 ( 90.29%) 19.05 ( 99.74%) 17.05 ( 99.77%) 125.55 ( 98.27%)
THP Active 83.80 ( 0.00%) 12.80 ( 15.27%) 15.60 ( 18.62%) 13.00 ( 15.51%) 0.80 ( 0.95%)
+/- 66.81 ( 0.00%) 20.19 ( 30.22%) 5.92 ( 8.86%) 15.06 ( 22.54%) 1.17 ( 1.75%)
Fault Alloc 171.00 ( 0.00%) 67.80 ( 39.65%) 97.40 ( 56.96%) 125.60 ( 73.45%) 133.00 ( 77.78%)
+/- 82.91 ( 0.00%) 30.69 ( 37.02%) 53.91 ( 65.02%) 55.05 ( 66.40%) 21.19 ( 25.56%)
Fault Fallback 832.00 ( 0.00%) 935.20 ( -12.40%) 906.00 ( -8.89%) 877.40 ( -5.46%) 870.20 ( -4.59%)
+/- 82.91 ( 0.00%) 30.69 ( 62.98%) 54.01 ( 34.86%) 55.05 ( 33.60%) 20.91 ( 74.78%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 7229.81 928.42 704.52 80.68 1330.76
Total Elapsed Time (seconds) 112849.04 5618.69 571.11 360.54 1664.28

In this case, the test is reading/writing only from filesystems but as
it's vfat, it's slow due to calling writepage during compaction. Little
to observe really - the time to complete the test goes way down
with the series applied and THP allocation success rates go up in
comparison to 3.2-rc5. The success rates are lower than 3.1.0 but
the elapsed time for that kernel is abysmal so it is not really a
sensible comparison.

As before, Andrea's series allocates more THPs at the cost of overall
performance.

writebackCPFileext4
3.1.0-vanilla rc5-vanilla freemore-v6r1 isolate-v6r1 andrea-v2r1
System Time 1.51 ( 0.00%) 1.77 ( -17.66%) 1.46 ( 2.92%) 1.15 ( 23.77%) 1.89 ( -25.63%)
+/- 0.27 ( 0.00%) 0.67 ( -148.52%) 0.33 ( -22.76%) 0.30 ( -11.15%) 0.19 ( 30.16%)
User Time 0.03 ( 0.00%) 0.04 ( -37.50%) 0.05 ( -62.50%) 0.07 ( -112.50%) 0.04 ( -18.75%)
+/- 0.01 ( 0.00%) 0.02 ( -146.64%) 0.02 ( -97.91%) 0.02 ( -75.59%) 0.02 ( -63.30%)
Elapsed Time 124.93 ( 0.00%) 114.49 ( 8.36%) 96.77 ( 22.55%) 27.48 ( 78.00%) 205.70 ( -64.65%)
+/- 20.20 ( 0.00%) 74.39 ( -268.34%) 59.88 ( -196.48%) 7.72 ( 61.79%) 25.03 ( -23.95%)
THP Active 161.80 ( 0.00%) 83.60 ( 51.67%) 141.20 ( 87.27%) 84.60 ( 52.29%) 82.60 ( 51.05%)
+/- 71.95 ( 0.00%) 43.80 ( 60.88%) 26.91 ( 37.40%) 59.02 ( 82.03%) 52.13 ( 72.45%)
Fault Alloc 471.40 ( 0.00%) 228.60 ( 48.49%) 282.20 ( 59.86%) 225.20 ( 47.77%) 388.40 ( 82.39%)
+/- 88.07 ( 0.00%) 87.42 ( 99.26%) 73.79 ( 83.78%) 109.62 ( 124.47%) 82.62 ( 93.81%)
Fault Fallback 531.60 ( 0.00%) 774.60 ( -45.71%) 720.80 ( -35.59%) 777.80 ( -46.31%) 614.80 ( -15.65%)
+/- 88.07 ( 0.00%) 87.26 ( 0.92%) 73.79 ( 16.22%) 109.62 ( -24.47%) 82.29 ( 6.56%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 50.22 33.76 30.65 24.14 128.45
Total Elapsed Time (seconds) 1113.73 1132.19 1029.45 759.49 1707.26

Same type of story - elapsed times go down. In this case, allocation
success rates are roughtly the same. As before, Andrea's has higher
success rates but takes a lot longer.

Overall the series does reduce latencies and while the tests are
inherency racy as alloc competes with the cp processes, the variability
was included. The THP allocation rates are not as high as they could
be but that is because we would have to be more aggressive about
reclaim and compaction impacting overall performance.

Comments?

fs/btrfs/disk-io.c | 5 +-
fs/hugetlbfs/inode.c | 3 +-
fs/nfs/internal.h | 2 +-
fs/nfs/write.c | 4 +-
include/linux/fs.h | 11 ++-
include/linux/migrate.h | 23 +++++-
include/linux/mmzone.h | 4 +
include/linux/vm_event_item.h | 1 +
mm/compaction.c | 5 +-
mm/memory-failure.c | 2 +-
mm/memory_hotplug.c | 2 +-
mm/mempolicy.c | 2 +-
mm/migrate.c | 171 ++++++++++++++++++++++++++++-
mm/page_alloc.c | 50 +++++++++
mm/swap.c | 74 +++++++++++++++++-
mm/vmscan.c | 114 ++++++++++++++++++++++++-
mm/vmstat.c | 2 +
17 files changed, 371 insertions(+), 104 deletions(-)

1.7.3.4

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/