zram: switch to crypto compress API
authorSergey Senozhatsky <sergey.senozhatsky@gmail.com>
Tue, 26 Jul 2016 22:22:45 +0000 (15:22 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 26 Jul 2016 23:19:19 +0000 (16:19 -0700)
commitebaf9ab56d9d5f350969bd1ea8f47234623c9684
treeb7c5218f90d3e29ca246023c10ca55b556bf21d7
parent2aea8493d326bdf15446768333e1d2c91b040b5c
zram: switch to crypto compress API

We don't have an idle zstreams list anymore and our write path now works
absolutely differently, preventing preemption during compression.  This
removes possibilities of read paths preempting writes at wrong places
(which could badly affect the performance of both paths) and at the same
time opens the door for a move from custom LZO/LZ4 compression backends
implementation to a more generic one, using crypto compress API.

Joonsoo Kim [1] attempted to do this a while ago, but faced with the
need of introducing a new crypto API interface.  The root cause was the
fact that crypto API compression algorithms require a compression stream
structure (in zram terminology) for both compression and decompression
ops, while in reality only several of compression algorithms really need
it.  This resulted in a concept of context-less crypto API compression
backends [2].  Both write and read paths, though, would have been
executed with the preemption enabled, which in the worst case could have
resulted in a decreased worst-case performance, e.g.  consider the
following case:

CPU0

zram_write()
  spin_lock()
    take the last idle stream
  spin_unlock()

<< preempted >>

zram_read()
  spin_lock()
   no idle streams
  spin_unlock()
  schedule()

resuming zram_write compression()

but it took me some time to realize that, and it took even longer to
evolve zram and to make it ready for crypto API.  The key turned out to be
-- drop the idle streams list entirely.  Without the idle streams list we
are free to use compression algorithms that require compression stream for
decompression (read), because streams are now placed in per-cpu data and
each write path has to disable preemption for compression op, almost
completely eliminating the aforementioned case (technically, we still have
a small chance, because write path has a fast and a slow paths and the
slow path is executed with the preemption enabled; but the frequency of
failed fast path is too low).

TEST
====

- 4 CPUs, x86_64 system
- 3G zram, lzo
- fio tests: read, randread, write, randwrite, rw, randrw

test script [3] command:
 ZRAM_SIZE=3G LOG_SUFFIX=XXXX FIO_LOOPS=5 ./zram-fio-test.sh

                   BASE           PATCHED
jobs1
READ:           2527.2MB/s  2482.7MB/s
READ:           2102.7MB/s  2045.0MB/s
WRITE:          1284.3MB/s  1324.3MB/s
WRITE:          1080.7MB/s  1101.9MB/s
READ:           430125KB/s  437498KB/s
WRITE:          430538KB/s  437919KB/s
READ:           399593KB/s  403987KB/s
WRITE:          399910KB/s  404308KB/s
jobs2
READ:           8133.5MB/s  7854.8MB/s
READ:           7086.6MB/s  6912.8MB/s
WRITE:          3177.2MB/s  3298.3MB/s
WRITE:          2810.2MB/s  2871.4MB/s
READ:           1017.6MB/s  1023.4MB/s
WRITE:          1018.2MB/s  1023.1MB/s
READ:           977836KB/s  984205KB/s
WRITE:          979435KB/s  985814KB/s
jobs3
READ:           13557MB/s  13391MB/s
READ:           11876MB/s  11752MB/s
WRITE:          4641.5MB/s  4682.1MB/s
WRITE:          4164.9MB/s  4179.3MB/s
READ:           1453.8MB/s  1455.1MB/s
WRITE:          1455.1MB/s  1458.2MB/s
READ:           1387.7MB/s  1395.7MB/s
WRITE:          1386.1MB/s  1394.9MB/s
jobs4
READ:           20271MB/s  20078MB/s
READ:           18033MB/s  17928MB/s
WRITE:          6176.8MB/s  6180.5MB/s
WRITE:          5686.3MB/s  5705.3MB/s
READ:           2009.4MB/s  2006.7MB/s
WRITE:          2007.5MB/s  2004.9MB/s
READ:           1929.7MB/s  1935.6MB/s
WRITE:          1926.8MB/s  1932.6MB/s
jobs5
READ:           18823MB/s  19024MB/s
READ:           18968MB/s  19071MB/s
WRITE:          6191.6MB/s  6372.1MB/s
WRITE:          5818.7MB/s  5787.1MB/s
READ:           2011.7MB/s  1981.3MB/s
WRITE:          2011.4MB/s  1980.1MB/s
READ:           1949.3MB/s  1935.7MB/s
WRITE:          1940.4MB/s  1926.1MB/s
jobs6
READ:           21870MB/s  21715MB/s
READ:           19957MB/s  19879MB/s
WRITE:          6528.4MB/s  6537.6MB/s
WRITE:          6098.9MB/s  6073.6MB/s
READ:           2048.6MB/s  2049.9MB/s
WRITE:          2041.7MB/s  2042.9MB/s
READ:           2013.4MB/s  1990.4MB/s
WRITE:          2009.4MB/s  1986.5MB/s
jobs7
READ:           21359MB/s  21124MB/s
READ:           19746MB/s  19293MB/s
WRITE:          6660.4MB/s  6518.8MB/s
WRITE:          6211.6MB/s  6193.1MB/s
READ:           2089.7MB/s  2080.6MB/s
WRITE:          2085.8MB/s  2076.5MB/s
READ:           2041.2MB/s  2052.5MB/s
WRITE:          2037.5MB/s  2048.8MB/s
jobs8
READ:           20477MB/s  19974MB/s
READ:           18922MB/s  18576MB/s
WRITE:          6851.9MB/s  6788.3MB/s
WRITE:          6407.7MB/s  6347.5MB/s
READ:           2134.8MB/s  2136.1MB/s
WRITE:          2132.8MB/s  2134.4MB/s
READ:           2074.2MB/s  2069.6MB/s
WRITE:          2087.3MB/s  2082.4MB/s
jobs9
READ:           19797MB/s  19994MB/s
READ:           18806MB/s  18581MB/s
WRITE:          6878.7MB/s  6822.7MB/s
WRITE:          6456.8MB/s  6447.2MB/s
READ:           2141.1MB/s  2154.7MB/s
WRITE:          2144.4MB/s  2157.3MB/s
READ:           2084.1MB/s  2085.1MB/s
WRITE:          2091.5MB/s  2092.5MB/s
jobs10
READ:           19794MB/s  19784MB/s
READ:           18794MB/s  18745MB/s
WRITE:          6984.4MB/s  6676.3MB/s
WRITE:          6532.3MB/s  6342.7MB/s
READ:           2150.6MB/s  2155.4MB/s
WRITE:          2156.8MB/s  2161.5MB/s
READ:           2106.4MB/s  2095.6MB/s
WRITE:          2109.7MB/s  2098.4MB/s

                                    BASE                       PATCHED
jobs1                              perfstat
stalled-cycles-frontend     102,480,595,419 (  41.53%)   114,508,864,804 (  46.92%)
stalled-cycles-backend       51,941,417,832 (  21.05%)    46,836,112,388 (  19.19%)
instructions                283,612,054,215 (    1.15)   283,918,134,959 (    1.16)
branches                     56,372,560,385 ( 724.923)    56,449,814,753 ( 733.766)
branch-misses                   374,826,000 (   0.66%)       326,935,859 (   0.58%)
jobs2                              perfstat
stalled-cycles-frontend     155,142,745,777 (  40.99%)   164,170,979,198 (  43.82%)
stalled-cycles-backend       70,813,866,387 (  18.71%)    66,456,858,165 (  17.74%)
instructions                463,436,648,173 (    1.22)   464,221,890,191 (    1.24)
branches                     91,088,733,902 ( 760.088)    91,278,144,546 ( 769.133)
branch-misses                   504,460,363 (   0.55%)       394,033,842 (   0.43%)
jobs3                              perfstat
stalled-cycles-frontend     201,300,397,212 (  39.84%)   223,969,902,257 (  44.44%)
stalled-cycles-backend       87,712,593,974 (  17.36%)    81,618,888,712 (  16.19%)
instructions                642,869,545,023 (    1.27)   644,677,354,132 (    1.28)
branches                    125,724,560,594 ( 690.682)   126,133,159,521 ( 694.542)
branch-misses                   527,941,798 (   0.42%)       444,782,220 (   0.35%)
jobs4                              perfstat
stalled-cycles-frontend     246,701,197,429 (  38.12%)   280,076,030,886 (  43.29%)
stalled-cycles-backend      119,050,341,112 (  18.40%)   110,955,641,671 (  17.15%)
instructions                822,716,962,127 (    1.27)   825,536,969,320 (    1.28)
branches                    160,590,028,545 ( 688.614)   161,152,996,915 ( 691.068)
branch-misses                   650,295,287 (   0.40%)       550,229,113 (   0.34%)
jobs5                              perfstat
stalled-cycles-frontend     298,958,462,516 (  38.30%)   344,852,200,358 (  44.16%)
stalled-cycles-backend      137,558,742,122 (  17.62%)   129,465,067,102 (  16.58%)
instructions              1,005,714,688,752 (    1.29) 1,007,657,999,432 (    1.29)
branches                    195,988,773,962 ( 697.730)   196,446,873,984 ( 700.319)
branch-misses                   695,818,940 (   0.36%)       624,823,263 (   0.32%)
jobs6                              perfstat
stalled-cycles-frontend     334,497,602,856 (  36.71%)   387,590,419,779 (  42.38%)
stalled-cycles-backend      163,539,365,335 (  17.95%)   152,640,193,639 (  16.69%)
instructions              1,184,738,177,851 (    1.30) 1,187,396,281,677 (    1.30)
branches                    230,592,915,640 ( 702.902)   231,253,802,882 ( 702.356)
branch-misses                   747,934,786 (   0.32%)       643,902,424 (   0.28%)
jobs7                              perfstat
stalled-cycles-frontend     396,724,684,187 (  37.71%)   460,705,858,952 (  43.84%)
stalled-cycles-backend      188,096,616,496 (  17.88%)   175,785,787,036 (  16.73%)
instructions              1,364,041,136,608 (    1.30) 1,366,689,075,112 (    1.30)
branches                    265,253,096,936 ( 700.078)   265,890,524,883 ( 702.839)
branch-misses                   784,991,589 (   0.30%)       729,196,689 (   0.27%)
jobs8                              perfstat
stalled-cycles-frontend     440,248,299,870 (  36.92%)   509,554,793,816 (  42.46%)
stalled-cycles-backend      222,575,930,616 (  18.67%)   213,401,248,432 (  17.78%)
instructions              1,542,262,045,114 (    1.29) 1,545,233,932,257 (    1.29)
branches                    299,775,178,439 ( 697.666)   300,528,458,505 ( 694.769)
branch-misses                   847,496,084 (   0.28%)       748,794,308 (   0.25%)
jobs9                              perfstat
stalled-cycles-frontend     506,269,882,480 (  37.86%)   592,798,032,820 (  44.43%)
stalled-cycles-backend      253,192,498,861 (  18.93%)   233,727,666,185 (  17.52%)
instructions              1,721,985,080,913 (    1.29) 1,724,666,236,005 (    1.29)
branches                    334,517,360,255 ( 694.134)   335,199,758,164 ( 697.131)
branch-misses                   873,496,730 (   0.26%)       815,379,236 (   0.24%)
jobs10                             perfstat
stalled-cycles-frontend     549,063,363,749 (  37.18%)   651,302,376,662 (  43.61%)
stalled-cycles-backend      281,680,986,810 (  19.07%)   277,005,235,582 (  18.55%)
instructions              1,901,859,271,180 (    1.29) 1,906,311,064,230 (    1.28)
branches                    369,398,536,153 ( 694.004)   370,527,696,358 ( 688.409)
branch-misses                   967,929,335 (   0.26%)       890,125,056 (   0.24%)

                            BASE           PATCHED
seconds elapsed        79.421641008 78.735285546
seconds elapsed        61.471246133 60.869085949
seconds elapsed        62.317058173 62.224188495
seconds elapsed        60.030739363 60.081102518
seconds elapsed        74.070398362 74.317582865
seconds elapsed        84.985953007 85.414364176
seconds elapsed        97.724553255 98.173311344
seconds elapsed        109.488066758 110.268399318
seconds elapsed        122.768189405 122.967164498
seconds elapsed        135.130035105 136.934770801

On my other system (8 x86_64 CPUs, short version of test results):

                            BASE           PATCHED
seconds elapsed        19.518065994 19.806320662
seconds elapsed        15.172772749 15.594718291
seconds elapsed        13.820925970 13.821708564
seconds elapsed        13.293097816 14.585206405
seconds elapsed        16.207284118 16.064431606
seconds elapsed        17.958376158 17.771825767
seconds elapsed        19.478009164 19.602961508
seconds elapsed        21.347152811 21.352318709
seconds elapsed        24.478121126 24.171088735
seconds elapsed        26.865057442 26.767327618

So performance-wise the numbers are quite similar.

Also update zcomp interface to be more aligned with the crypto API.

[1] http://marc.info/?l=linux-kernel&m=144480832108927&w=2
[2] http://marc.info/?l=linux-kernel&m=145379613507518&w=2
[3] https://github.com/sergey-senozhatsky/zram-perf-test

Link: http://lkml.kernel.org/r/20160531122017.2878-3-sergey.senozhatsky@gmail.com
Signed-off-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Suggested-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/block/zram/Kconfig
drivers/block/zram/zcomp.c
drivers/block/zram/zcomp.h
drivers/block/zram/zram_drv.c