Remove dependency on congestion_wait in mm/
Subsystems (10)
v4 cover Remove dependency on congestion_wait in mm/
Changelog since v3
o Count writeback completions for NR_THROTTLED_WRITTEN only
o Use IRQ-safe inc_node_page_state
o Remove redundant throttling
This series is also available at
git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git mm-reclaimcongest-v4r2
This series that removes all calls to congestion_wait
in mm/ and deletes wait_iff_congested. It's not a clever implementation but congestion_wait has been broken for a long time
(https://lore.kernel.org/linux-mm/[email protected]/).
Even if congestion throttling worked, it was never a great idea. While excessive dirty/writeback pages at the tail of the LRU is one possibility that reclaim may be slow, there is also the problem of too many pages being isolated and reclaim failing for other reasons (elevated references, too many pages isolated, excessive LRU contention etc).
This series replaces the "congestion" throttling with 3 different types.
o If there are too many dirty/writeback pages, sleep until a timeout or enough pages get cleaned
o If too many pages are isolated, sleep until enough isolated pages are either reclaimed or put back on the LRU
o If no progress is being made, direct reclaim tasks sleep until another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger corner cases that no longer work. A new test case was created called "stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly created XFS filesystem. Note that it may be necessary to increase the timeout of ssh if executing remotely as ssh itself can get throttled and the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4 to check the impact as the number of direct reclaimers increase. It has four types of worker.
o One "anon latency" worker creates small mappings with mmap() and times how long it takes to fault the mapping reading it 4K at a time
o X file writers which is fio randomly writing X files where the total size of the files add up to the allowed dirty_ratio. fio is allowed to run for a warmup period to allow some file-backed pages to accumulate. The duration of the warmup is based on the best-case linear write speed of the storage.
o Y file readers which is fio randomly reading small files
o Z anon memory hogs which continually map (100-dirty_ratio)% of memory
o Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon memory where some anonymous memory must be swapped and there is a high likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress dirty/writeback pages. The results below are based on having zero readers.
The short summary of the results is that the series works and stalls until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be treated as one block that replaces a broken throttling mechanism with a working one.
Finally, three machines were tested but I'm reporting the worst set of results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v4r2
Amean mmap-4 31.4003 ( 0.00%) 176.8729 (-463.28%)
Amean mmap-7 38.1641 ( 0.00%) 605.8866 (-1487.58%)
Amean mmap-12 60.0981 ( 0.00%) 2866.8561 (-4670.29%)
Amean mmap-21 161.2699 ( 0.00%) 118.6587 ( 26.42%)
Amean mmap-30 174.5589 ( 0.00%) 3729.2263 (-2036.37%)
Amean mmap-48 8106.8160 ( 0.00%) 1463.7815 ( 81.94%)
Stddev mmap-4 41.3455 ( 0.00%) 5847.5425 (-14043.13%)
Stddev mmap-7 53.5556 ( 0.00%) 12091.9011 (-22478.20%)
Stddev mmap-12 171.3897 ( 0.00%) 28785.9881 (-16695.63%)
Stddev mmap-21 1506.6752 ( 0.00%) 1609.0361 ( -6.79%)
Stddev mmap-30 557.5806 ( 0.00%) 32712.2440 (-5766.82%)
Stddev mmap-48 61681.5718 ( 0.00%) 15971.4654 ( 74.11%)
Max-90 mmap-4 31.4243 ( 0.00%) 30.1957 ( 3.91%)
Max-90 mmap-7 41.0410 ( 0.00%) 36.7782 ( 10.39%)
Max-90 mmap-12 66.5255 ( 0.00%) 121.8574 ( -83.17%)
Max-90 mmap-21 146.7479 ( 0.00%) 132.2327 ( 9.89%)
Max-90 mmap-30 193.9513 ( 0.00%) 61.6135 ( 68.23%)
Max-90 mmap-48 277.9137 ( 0.00%) 593.7413 (-113.64%)
Max mmap-4 1913.8009 ( 0.00%) 239690.5578 (-12424.32%)
Max mmap-7 2423.9665 ( 0.00%) 270122.1751 (-11043.81%)
Max mmap-12 6845.6573 ( 0.00%) 308761.7416 (-4410.33%)
Max mmap-21 56278.6508 ( 0.00%) 79286.8553 ( -40.88%)
Max mmap-30 19716.2990 ( 0.00%) 306793.2333 (-1456.04%)
Max mmap-48 477923.9400 ( 0.00%) 229791.8793 ( 51.92%)For most thread counts, the time to mmap() is unfortunately increased. In earlier versions of the series, this was lower but a large number of throttling events were reaching their timeout increasing the amount of inefficient scanning of the LRU. There is no prioritisation of reclaim tasks making progress based on each tasks rate of page allocation versus progress of reclaim. The variance is also impacted for high worker counts but in all cases, the differences in latency are not statistically significant due to very large maximum outliers. Max-90 shows that 90% of the stalls are comparable but the Max results show the massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the test design, reclaim is difficult so allocations stall and there are variances depending on whether THPs can be allocated or not. The amount of memory will affect exactly how bad the corner cases are and how often they trigger. The warmup period calculation is not ideal as it's based on linear writes where as fio is randomly writing multiple files from multiple tasks so the start state of the test is variable. For example, these are the latencies on a single-socket machine that had more memory
Amean mmap-4 20.5437 ( 0.00%) 19.1772 * 6.65%* Amean mmap-6 39.2860 ( 0.00%) 69.4987 ( -76.90%) Amean mmap-8 2476.1950 ( 0.00%) 151.7673 ( 93.87%) Amean mmap-12 178.0936 ( 0.00%) 209.1427 ( -17.43%) Amean mmap-18 3238.9125 ( 0.00%) 262.5806 ( 91.89%) Amean mmap-24 7922.7016 ( 0.00%) 322.9738 ( 95.92%) Amean mmap-30 1766.8392 ( 0.00%) 405.8898 ( 77.03%) Amean mmap-32 7542.2844 ( 0.00%) 555.6236 ( 92.63%) Amean mmap-32 7542.2844 ( 0.00%) 512.1812 ( 93.21%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v4r2
Duration User 6989.03 2368.70
Duration System 7308.12 843.35
Duration Elapsed 2277.67 2131.77The patches reduce system CPU usage by 88% as the vanilla kernel is rarely stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v4r2
Ops Direct pages scanned 1056608451.00 76886196.00
Ops Kswapd pages scanned 109795048.00 82179688.00
Ops Kswapd pages reclaimed 63269243.00 27410157.00Ops Direct pages reclaimed 10803973.00 8016444.00
Ops Kswapd efficiency % 57.62 33.35
Ops Kswapd velocity 48204.98 38549.98
Ops Direct efficiency % 1.02 10.43
Ops Direct velocity 463898.83 36066.83
Kswapd scanned lesspages but the detailed pattern is different. The vanilla kernel scans slowly over time where as the patches exhibits burst patterns of scan activity. Direct reclaim scanning is reduced by 92% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels exhibit spikes but the vanilla kernel when reclaiming shows pages being reclaimed over a period of time where as the patches tend to reclaim in spikes. The difference is that vanilla is not throttling and instead scanning constantly finding some pages over time where as the patched kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 48.34
For direct reclaim, vanilla scanned 90.59% of pages where as with the patches, 48.34% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1882533.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 2266749.00
And there is slightly less swapping.
Ops Page reclaim immediate 996248528.00 29230920.00
The number of pages encountered at the tail of the LRU tagged for immediate reclaim but still dirty/writeback is reduced by 97%.
Ops Slabs scanned 164284.00 166646.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000 82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What is likely happening is that cond_resched() reschedules the task for a short period when the BDI is not registering congestion (which it never will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000congestion_wait if called always exceeds the timeout as there is no trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for immediate reclaim at the tail of the LRU
1 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK 1 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK 1 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK 1 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK 1 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK 2 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK 8 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
23 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK 52 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK 61 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBAC
The majority of events did not stall or stalled for a short period. Roughly 16% of stalls reached the timeout before expiry. For direct reclaim, the number of times stalled for each reason were
13594 reason=VMSCAN_THROTTLE_ISOLATED
72247 reason=VMSCAN_THROTTLE_WRITEBACK
77203 reason=VMSCAN_THROTTLE_NOPROGRESS
The most common reason to stall was due to a failure to make forward progress followed closely by excessive pages tagged for immediate reclaim at the tail of the LRU. A relatively small number were due to too many pages isolated from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
3 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED 8 usec_timeout=20000 usect_delayed=8000 reason=VMSCAN_THROTTLE_ISOLATED 9 usec_timeout=20000 usect_delayed=12000 reason=VMSCAN_THROTTLE_ISOLATED
18 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED 69 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
1946 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all or for a short period. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over the map
1 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=276000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=388000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=236000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=396000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=336000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
15 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
24 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
24 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
39 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
47 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
47 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
52 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
54 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
63 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
68 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
75 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
80 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
83 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
97 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
99 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
99 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
102 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
149 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
187 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
196 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
245 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
322 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
406 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
588 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
843 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS 1299 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS 2839 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
10111 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS 21492 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS 36441 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at all. The remainder slept a little allowing other reclaim tasks to make progress.
While this timeout could be further increased, it could also negatively impact worst-case behaviour when there is no prioritisation of what task should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
8 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
9 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
13 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
14 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
14 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
21 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
25 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
26 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
45 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
50 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
60 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
74 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
122 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
134 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
310 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
568 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK 2038 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK 7061 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
61547 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a sizable number did not stall at all. This is very different to kswapd where only a tiny percentage of stalls due to writeback reached the timeout.
Bottom line, the throttling appears to work and the wakeup events may limit worst case stalls. There might be some grounds for adjusting timeouts but it's likely futile as the worst-case scenarios depend on the workload, memory size and the speed of the storage. A better approach to improve the series further would be to prioritise tasks based on their rate of allocation with the caveat that it may be very expensive to track.
Signature · 1 line
2.31.1
PATCH v4 1/8 mm/vmscan: Throttle reclaim until some writeback completes if congested
Page reclaim throttles on wait_iff_congested under the following conditions
o kswapd is encountering pages under writeback and marked for immediate reclaim implying that pages are cycling through the LRU faster than pages can be cleaned.
o Direct reclaim will stall if all dirty pages are backed by congested inodes.
wait_iff_congested is almost completely broken with few exceptions. This patch adds a new node-based workqueue and tracks the number of throttled tasks and pages written back since throttling started. If enough pages belonging to the node are written back then the throttled tasks will wake early. If not, the throttled tasks sleeps until the timeout expires.
--- include/linux/backing-dev.h | 1 - include/linux/mmzone.h | 9 ++++ include/trace/events/vmscan.h | 34 +++++++++++++ include/trace/events/writeback.h | 7 --- mm/backing-dev.c | 48 ------------------- mm/filemap.c | 1 + mm/internal.h | 11 +++++ mm/page_alloc.c | 1 + mm/vmscan.c | 82 +++++++++++++++++++++++++++----- mm/vmstat.c | 1 + 10 files changed, 127 insertions(+), 68 deletions(-)
include/linux/backing-dev.h +0 -1
@@ -154,7 +154,6 @@ static inline int wb_congested(struct bdi_writeback *wb, int cong_bits) }long congestion_wait (int sync,long timeout); -long wait_iff_congested (int sync,long timeout);static inline bool mapping_can_writeback (struct address_space *mapping) {
include/linux/mmzone.h +9 -0
@@ -199,6 +199,7 @@ enum node_stat_item {NR_VMSCAN_IMMEDIATE ,/* Prioritise for reclaim when writeback ends */ NR_DIRTIED ,/* page dirtyings since bootup */ NR_WRITTEN ,/* page writings since bootup */ +NR_THROTTLED_WRITTEN ,/* NR_WRITTEN while reclaim throttled */ NR_KERNEL_MISC_RECLAIMABLE ,/* reclaimable non-slab kernel pages */ NR_FOLL_PIN_ACQUIRED ,/* via: pin_user_page(), gup flag: FOLL_PIN */ NR_FOLL_PIN_RELEASED, /* pages returned via unpin_user_page() */
@@ -272,6 +273,10 @@ enum lru_list { NR_LRU_LISTS }; +enum vmscan_throttle_state { +VMSCAN_THROTTLE_WRITEBACK , +}; +#define for_each_lru (lru) for (lru = 0; lru < NR_LRU_LISTS; lru++)#define for_each_evictable_lru (lru) for (lru = 0; lru <= LRU_ACTIVE_FILE; lru++)
@@ -841,6 +846,10 @@ typedef struct pglist_data {int node_id;wait_queue_head_t kswapd_wait;wait_queue_head_t pfmemalloc_wait; +wait_queue_head_t reclaim_wait;/* wq for throttling reclaim */ +atomic_t nr_reclaim_throttled;/* nr of throtted tasks */ +unsigned long nr_reclaim_start;/* nr pages written while throttled +* when throttling started. */ struct task_struct *kswapd;/* Protected by mem_hotplug_begin/end() */ int kswapd_order;
include/trace/events/vmscan.h +34 -0
@@ -27,6 +27,14 @@ {RECLAIM_WB_ASYNC, "RECLAIM_WB_ASYNC"} \ ) : "RECLAIM_WB_NONE" +#define _VMSCAN_THROTTLE_WRITEBACK (1 << VMSCAN_THROTTLE_WRITEBACK) + +#define show_throttle_flags (flags) \ + (flags) ? __print_flags(flags, "|", \ + {_VMSCAN_THROTTLE_WRITEBACK, "VMSCAN_THROTTLE_WRITEBACK"} \ + ) : "VMSCAN_THROTTLE_NONE" + +#define trace_reclaim_flags (file) ( \ (file ? RECLAIM_WB_FILE : RECLAIM_WB_ANON) | \ (RECLAIM_WB_ASYNC) \
@@ -454,6 +462,32 @@ DEFINE_EVENT(mm_vmscan_direct_reclaim_end_template, mm_vmscan_node_reclaim_end,TP_ARGS (nr_reclaimed) ); +TRACE_EVENT (mm_vmscan_throttled, + + TP_PROTO(int nid, int usec_timeout, int usec_delayed, int reason), + +TP_ARGS (nid, usec_timeout, usec_delayed, reason), + +TP_STRUCT__entry ( + __field(int, nid) + __field(int, usec_timeout) + __field(int, usec_delayed) +__field (int, reason) + ), + +TP_fast_assign ( + __entry->nid = nid; + __entry->usec_timeout = usec_timeout; + __entry->usec_delayed = usec_delayed; + __entry->reason = 1U << reason; + ), + +TP_printk ("nid=%d usec_timeout=%d usect_delayed=%d reason=%s" , + __entry->nid, + __entry->usec_timeout, + __entry->usec_delayed, +show_throttle_flags (__entry->reason)) +);#endif /* _TRACE_VMSCAN_H */ /* This part must be outside protection */
include/trace/events/writeback.h +0 -7
@@ -763,13 +763,6 @@ DEFINE_EVENT(writeback_congest_waited_template, writeback_congestion_wait,TP_ARGS (usec_timeout, usec_delayed) ); -DEFINE_EVENT (writeback_congest_waited_template, writeback_wait_iff_congested, - - TP_PROTO(unsigned int usec_timeout, unsigned int usec_delayed), - -TP_ARGS (usec_timeout, usec_delayed) -); - DECLARE_EVENT_CLASS(writeback_single_inode_template, TP_PROTO(struct inode *inode,
mm/backing-dev.c +0 -48
51 lines not on this page · raw message
mm/filemap.c +1 -0
@@ -1605,6 +1605,7 @@ void end_page_writeback(struct page *page)smp_mb__after_atomic ();wake_up_page (page, PG_writeback); +acct_reclaim_writeback (page);put_page (page); }EXPORT_SYMBOL (end_page_writeback);
mm/internal.h +11 -0
@@ -34,6 +34,17 @@void page_writeback_init (void ); +void __acct_reclaim_writeback (pg_data_t *pgdat,struct page *page, +int nr_throttled); +static inline void acct_reclaim_writeback (struct page *page) +{ +pg_data_t *pgdat =page_pgdat (page); +int nr_throttled =atomic_read (&pgdat->nr_reclaim_throttled); + +if (nr_throttled) +__acct_reclaim_writeback (pgdat, page, nr_throttled); +} +vm_fault_t do_swap_page (struct vm_fault *vmf); void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *start_vma,
mm/page_alloc.c +1 -0
@@ -7396,6 +7396,7 @@ static void __meminit pgdat_init_internals(struct pglist_data *pgdat)init_waitqueue_head (&pgdat->kswapd_wait);init_waitqueue_head (&pgdat->pfmemalloc_wait); +init_waitqueue_head (&pgdat->reclaim_wait);pgdat_page_ext_init (pgdat);lruvec_init (&pgdat->__lruvec);
mm/vmscan.c +70 -12
125 lines not on this page · raw message
mm/vmstat.c +1 -0
@@ -1225,6 +1225,7 @@ const char * const vmstat_text[] = {"nr_vmscan_immediate_reclaim" ,"nr_dirtied" ,"nr_written" , +"nr_throttled_written" ,"nr_kernel_misc_reclaimable" ,"nr_foll_pin_acquired" , "nr_foll_pin_released",
Signature · 1 line
2.31.1
PATCH v4 2/8 mm/vmscan: Throttle reclaim and compaction when too may pages are isolated
In reply to Mel GormanPage reclaim throttles on congestion if too many parallel reclaim instances have isolated too many pages. This makes no sense, excessive parallelisation has nothing to do with writeback or congestion.
This patch creates an additional workqueue to sleep on when too many pages are isolated. The throttled tasks are woken when the number of isolated pages is reduced or a timeout occurs. There may be some false positive wakeups for GFP_NOIO/GFP_NOFS callers but the tasks will throttle again if necessary.
--- include/linux/mmzone.h | 6 ++++-- include/trace/events/vmscan.h | 4 +++- mm/compaction.c | 10 ++++++++-- mm/internal.h | 13 ++++++++++++- mm/page_alloc.c | 6 +++++- mm/vmscan.c | 28 +++++++++++++++++++--------- 6 files changed, 51 insertions(+), 16 deletions(-)
include/linux/mmzone.h +4 -2
@@ -275,6 +275,8 @@ enum lru_list {enum vmscan_throttle_state {VMSCAN_THROTTLE_WRITEBACK , +VMSCAN_THROTTLE_ISOLATED , +NR_VMSCAN_THROTTLE , };#define for_each_lru (lru) for (lru = 0; lru < NR_LRU_LISTS; lru++)
@@ -846,8 +848,8 @@ typedef struct pglist_data {int node_id;wait_queue_head_t kswapd_wait;wait_queue_head_t pfmemalloc_wait; -wait_queue_head_t reclaim_wait;/* wq for throttling reclaim */ -atomic_t nr_reclaim_throttled;/* nr of throtted tasks */ +wait_queue_head_t reclaim_wait[NR_VMSCAN_THROTTLE ]; +atomic_t nr_writeback_throttled;/* nr of writeback-throttled tasks */ unsigned long nr_reclaim_start;/* nr pages written while throttled * when throttling started. */ struct task_struct *kswapd;/* Protected by
include/trace/events/vmscan.h +3 -1
@@ -28,10 +28,12 @@ ) : "RECLAIM_WB_NONE"#define _VMSCAN_THROTTLE_WRITEBACK (1 << VMSCAN_THROTTLE_WRITEBACK) +#define _VMSCAN_THROTTLE_ISOLATED (1 << VMSCAN_THROTTLE_ISOLATED)#define show_throttle_flags (flags) \ (flags) ? __print_flags(flags, "|", \ - {_VMSCAN_THROTTLE_WRITEBACK, "VMSCAN_THROTTLE_WRITEBACK"} \ + {_VMSCAN_THROTTLE_WRITEBACK, "VMSCAN_THROTTLE_WRITEBACK"}, \ + {_VMSCAN_THROTTLE_ISOLATED, "VMSCAN_THROTTLE_ISOLATED"} \ ) : "VMSCAN_THROTTLE_NONE"
mm/compaction.c +8 -2
@@ -761,6 +761,8 @@ isolate_freepages_range(struct compact_control *cc,/* Similar to reclaim, but different enough that they don't share logic */ static bool too_many_isolated (pg_data_t *pgdat) { +bool too_many; +unsigned long active, inactive, isolated; inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
@@ -770,7 +772,11 @@ static bool too_many_isolated(pg_data_t *pgdat) isolated =node_page_state (pgdat,NR_ISOLATED_FILE ) +node_page_state (pgdat,NR_ISOLATED_ANON ); -return isolated > (inactive + active) /2 ; + too_many = isolated > (inactive + active) /2 ; +if (!too_many) +wake_throttle_isolated (pgdat); + +return too_many; }/**
@@ -822,7 +828,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,if (cc->mode ==MIGRATE_ASYNC )return -EAGAIN ; -congestion_wait (BLK_RW_ASYNC ,HZ /10 ); +reclaim_throttle (pgdat,VMSCAN_THROTTLE_ISOLATED ,HZ /10 );if (fatal_signal_pending (current))return -EINTR ;
mm/internal.h +12 -1
@@ -39,12 +39,21 @@ void __acct_reclaim_writeback(pg_data_t *pgdat, struct page *page,static inline void acct_reclaim_writeback (struct page *page) {pg_data_t *pgdat =page_pgdat (page); -int nr_throttled =atomic_read (&pgdat->nr_reclaim_throttled); +int nr_throttled =atomic_read (&pgdat->nr_writeback_throttled);if (nr_throttled)__acct_reclaim_writeback (pgdat, page, nr_throttled); } +static inline void wake_throttle_isolated (pg_data_t *pgdat) +{ +wait_queue_head_t *wqh; + + wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_ISOLATED ]; +if (waitqueue_active (wqh)) +wake_up_all (wqh); +} +vm_fault_t do_swap_page (struct vm_fault *vmf); void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *start_vma,
@@ -120,6 +129,8 @@ extern unsigned long highest_memmap_pfn;*/ extern int isolate_lru_page (struct page *page);extern void putback_lru_page (struct page *page); +extern void reclaim_throttle (pg_data_t *pgdat,enum vmscan_throttle_state reason, +long timeout);/* * in mm/rmap.c:
mm/page_alloc.c +5 -1
@@ -7389,6 +7389,8 @@ static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}static void __meminit pgdat_init_internals (struct pglist_data *pgdat) { +int i; +pgdat_resize_init (pgdat);pgdat_init_split_queue (pgdat);
@@ -7396,7 +7398,9 @@ static void __meminit pgdat_init_internals(struct pglist_data *pgdat)init_waitqueue_head (&pgdat->kswapd_wait);init_waitqueue_head (&pgdat->pfmemalloc_wait); -init_waitqueue_head (&pgdat->reclaim_wait); + +for (i =0 ; i <NR_VMSCAN_THROTTLE ; i++) +init_waitqueue_head (&pgdat->reclaim_wait[i]);pgdat_page_ext_init (pgdat);lruvec_init (&pgdat->__lruvec);
mm/vmscan.c +19 -9
@@ -1006,12 +1006,12 @@ static void handle_write_error(struct address_space *mapping,unlock_page (page); } -static void -reclaim_throttle (pg_data_t *pgdat,enum vmscan_throttle_state reason, +void reclaim_throttle (pg_data_t *pgdat,enum vmscan_throttle_state reason,long timeout) { -wait_queue_head_t *wqh = &pgdat->reclaim_wait; +wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];long ret; +bool acct_writeback = (reason ==VMSCAN_THROTTLE_WRITEBACK );DEFINE_WAIT (wait);/*
@@ -1023,7 +1023,8 @@ reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason, current->flags & (PF_IO_WORKER|PF_KTHREAD))return ; -if (atomic_inc_return (&pgdat->nr_reclaim_throttled) ==1 ) { +if (acct_writeback && +atomic_inc_return (&pgdat->nr_writeback_throttled) ==1 ) {WRITE_ONCE (pgdat->nr_reclaim_start,node_page_state (pgdat,NR_THROTTLED_WRITTEN )); }
@@ -1031,7 +1032,9 @@ reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason,prepare_to_wait (wqh, &wait,TASK_UNINTERRUPTIBLE ); ret =schedule_timeout (timeout);finish_wait (wqh, &wait); -atomic_dec (&pgdat->nr_reclaim_throttled); + +if (acct_writeback) +atomic_dec (&pgdat->nr_writeback_throttled); trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout), jiffies_to_usecs(timeout - ret),
@@ -1061,7 +1064,7 @@ void __acct_reclaim_writeback(pg_data_t *pgdat, struct page *page,READ_ONCE (pgdat->nr_reclaim_start);if (nr_written >SWAP_CLUSTER_MAX * nr_throttled) -wake_up_all (&pgdat->reclaim_wait); +wake_up_all (&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK ]); }/* possible outcome of pageout() */
@@ -2176,6 +2179,7 @@ static int too_many_isolated(struct pglist_data *pgdat, int file, struct scan_control *sc) {unsigned long inactive, isolated; +bool too_many;if (current_is_kswapd ())return 0 ;
@@ -2199,7 +2203,13 @@ static int too_many_isolated(struct pglist_data *pgdat, int file,if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) inactive >>=3 ; -return isolated > inactive; + too_many = isolated > inactive; + +/* Wake up tasks throttled due to too_many_isolated. */ +if (!too_many) +wake_throttle_isolated (pgdat); + +return too_many; }/*
@@ -2308,8 +2318,8 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,return 0 ;/* wait a bit for the reclaimer. */ -msleep (100 ); stalled = true; +reclaim_throttle (pgdat,VMSCAN_THROTTLE_ISOLATED ,HZ /10 );/* We are about to die and free our memory. Return now. */ if (fatal_signal_pending (current))
@@ -4343,7 +4353,7 @@ static int kswapd(void *p)WRITE_ONCE (pgdat->kswapd_order,0 );WRITE_ONCE (pgdat->kswapd_highest_zoneidx,MAX_NR_ZONES ); -atomic_set (&pgdat->nr_reclaim_throttled,0 ); +atomic_set (&pgdat->nr_writeback_throttled,0 );for ( ; ; ) {bool ret;
Signature · 1 line
2.31.1
PATCH v4 3/8 mm/vmscan: Throttle reclaim when no progress is being made
In reply to Mel GormanMemcg reclaim throttles on congestion if no reclaim progress is made. This makes little sense, it might be due to writeback or a host of other factors.
For !memcg reclaim, it's messy. Direct reclaim primarily is throttled in the page allocator if it is failing to make progress. Kswapd throttles if too many pages are under writeback and marked for immediate reclaim.
This patch explicitly throttles if reclaim is failing to make progress.
--- include/linux/mmzone.h | 1 + include/trace/events/vmscan.h | 4 +++- mm/memcontrol.c | 10 +--------- mm/vmscan.c | 28 ++++++++++++++++++++++++++++ 4 files changed, 33 insertions(+), 10 deletions(-)
include/linux/mmzone.h +1 -0
@@ -276,6 +276,7 @@ enum lru_list {enum vmscan_throttle_state {VMSCAN_THROTTLE_WRITEBACK ,VMSCAN_THROTTLE_ISOLATED , +VMSCAN_THROTTLE_NOPROGRESS ,NR_VMSCAN_THROTTLE , };
include/trace/events/vmscan.h +3 -1
@@ -29,11 +29,13 @@#define _VMSCAN_THROTTLE_WRITEBACK (1 << VMSCAN_THROTTLE_WRITEBACK)#define _VMSCAN_THROTTLE_ISOLATED (1 << VMSCAN_THROTTLE_ISOLATED) +#define _VMSCAN_THROTTLE_NOPROGRESS (1 << VMSCAN_THROTTLE_NOPROGRESS)#define show_throttle_flags (flags) \ (flags) ? __print_flags(flags, "|", \ {_VMSCAN_THROTTLE_WRITEBACK, "VMSCAN_THROTTLE_WRITEBACK"}, \ - {_VMSCAN_THROTTLE_ISOLATED, "VMSCAN_THROTTLE_ISOLATED"} \ + {_VMSCAN_THROTTLE_ISOLATED, "VMSCAN_THROTTLE_ISOLATED"}, \ + {_VMSCAN_THROTTLE_NOPROGRESS, "VMSCAN_THROTTLE_NOPROGRESS"} \ ) : "VMSCAN_THROTTLE_NONE"
mm/memcontrol.c +1 -9
@@ -3465,19 +3465,11 @@ static int mem_cgroup_force_empty(struct mem_cgroup *memcg)/* try to free all pages in this cgroup */ while (nr_retries &&page_counter_read (&memcg->memory)) { -int progress; -if (signal_pending (current))return -EINTR ; - progress =try_to_free_mem_cgroup_pages (memcg,1 , -GFP_KERNEL , true); -if (!progress) { +if (!try_to_free_mem_cgroup_pages (memcg,1 ,GFP_KERNEL , true)) nr_retries--; -/* maybe some writeback is necessary */ -congestion_wait (BLK_RW_ASYNC ,HZ /10 ); - } - }return 0 ;
mm/vmscan.c +28 -0
@@ -3323,6 +3323,33 @@ static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)return zone_watermark_ok_safe (zone,0 , watermark, sc->reclaim_idx); } +static void consider_reclaim_throttle (pg_data_t *pgdat,struct scan_control *sc) +{ +/* If reclaim is making progress, wake any throttled tasks. */ +if (sc->nr_reclaimed) { +wait_queue_head_t *wqh; + + wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS ]; +if (waitqueue_active (wqh)) +wake_up_all (wqh); + +return ; + } + +/* +* Do not throttle kswapd on NOPROGRESS as it will throttle on +* VMSCAN_THROTTLE_WRITEBACK if there are too many pages under +* writeback and marked for immediate reclaim at the tail of +* the LRU. +*/ +if (current_is_kswapd ()) +return ; + +/* Throttle if making no progress at high prioities. */ +if (sc->priority <DEF_PRIORITY -2 ) +reclaim_throttle (pgdat,VMSCAN_THROTTLE_NOPROGRESS ,HZ /10 ); +} +/* * This is the direct reclaim path, for page-allocating processes. We only * try to reclaim pages from zones which will satisfy the caller's allocation
@@ -3407,6 +3434,7 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)continue ; last_pgdat = zone->zone_pgdat;shrink_node (zone->zone_pgdat, sc); +consider_reclaim_throttle (zone->zone_pgdat, sc); }/*
Signature · 1 line
2.31.1
PATCH v4 4/8 mm/writeback: Throttle based on page writeback instead of congestion
In reply to Mel Gormando_writepages throttles on congestion if the writepages() fails due to a lack of memory but congestion_wait() is partially broken as the congestion state is not updated for all BDIs.
This patch stalls waiting for a number of pages to complete writeback that located on the local node. The main weakness is that there is no correlation between the location of the inode's pages and locality but that is still better than congestion_wait.
--- mm/page-writeback.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)
mm/page-writeback.c +9 -2
@@ -2366,8 +2366,15 @@ int do_writepages(struct address_space *mapping, struct writeback_control *wbc) ret =generic_writepages (mapping, wbc);if ((ret != -ENOMEM ) || (wbc->sync_mode !=WB_SYNC_ALL ))break ; -cond_resched (); -congestion_wait (BLK_RW_ASYNC ,HZ /50 ); + +/* +* Lacking an allocation context or the locality or writeback +* state of any of the inode's pages, throttle based on +* writeback activity on the local node. It's as good a +* guess as any. +*/ +reclaim_throttle (NODE_DATA (numa_node_id ()), +VMSCAN_THROTTLE_WRITEBACK ,HZ /50 ); }/* * Usually few pages are written by now from those we've just submitted
Signature · 1 line
2.31.1
PATCH v4 5/8 mm/page_alloc: Remove the throttling logic from the page allocator
In reply to Mel GormanThe page allocator stalls based on the number of pages that are waiting for writeback to start but this should now be redundant. shrink_inactive_list() will wake flusher threads if the LRU tail are unqueued dirty pages so the flusher should be active. If it fails to make progress due to pages under writeback not being completed quickly then it should stall on VMSCAN_THROTTLE_WRITEBACK.
--- mm/page_alloc.c | 21 +-------------------- 1 file changed, 1 insertion(+), 20 deletions(-)
mm/page_alloc.c +1 -20
@@ -4795,30 +4795,11 @@ should_reclaim_retry(gfp_t gfp_mask, unsigned order,trace_reclaim_retry_zone (z, order, reclaimable, available, min_wmark, *no_progress_loops, wmark);if (wmark) { -/* -* If we didn't make any progress and have a lot of -* dirty + writeback pages then we should wait for -* an IO to complete to slow down the reclaim and -* prevent from pre mature OOM -*/ -if (!did_some_progress) { -unsigned long write_pending; - - write_pending =zone_page_state_snapshot (zone, -NR_ZONE_WRITE_PENDING ); - -if (2 * write_pending > reclaimable) { -congestion_wait (BLK_RW_ASYNC ,HZ /10 ); -return true; - } - } - ret = true; - goto out; +break ; } } -out:/* * Memory allocation/reclaim might be called from a WQ context and the * current implementation of the WQ concurrency control doesn't
Signature · 1 line
2.31.1
PATCH v4 6/8 mm/vmscan: Centralise timeout values for reclaim_throttle
In reply to Mel GormanNeil Brown raised concerns about callers of reclaim_throttle specifying a timeout value. The original timeout values to congestion_wait() were probably pulled out of thin air or copy&pasted from somewhere else. This patch centralises the timeout values and selects a timeout based on the reason for reclaim throttling. These figures are also pulled out of the same thin air but better values may be derived
Running a workload that is throttling for inappropriate periods and tracing mm_vmscan_throttled can be used to pick a more appropriate value. Excessive throttling would pick a lower timeout where as excessive CPU usage in reclaim context would select a larger timeout. Ideally a large value would always be used and the wakeups would occur before a timeout but that requires careful testing.
--- mm/compaction.c | 2 +- mm/internal.h | 3 +-- mm/page-writeback.c | 2 +- mm/vmscan.c | 48 +++++++++++++++++++++++++++++++++------------ 4 files changed, 38 insertions(+), 17 deletions(-)
mm/compaction.c +1 -1
@@ -828,7 +828,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,if (cc->mode ==MIGRATE_ASYNC )return -EAGAIN ; -reclaim_throttle (pgdat,VMSCAN_THROTTLE_ISOLATED ,HZ /10 ); +reclaim_throttle (pgdat,VMSCAN_THROTTLE_ISOLATED );if (fatal_signal_pending (current))return -EINTR ;
mm/internal.h +1 -2
@@ -129,8 +129,7 @@ extern unsigned long highest_memmap_pfn;*/ extern int isolate_lru_page (struct page *page);extern void putback_lru_page (struct page *page); -extern void reclaim_throttle (pg_data_t *pgdat,enum vmscan_throttle_state reason, -long timeout); +extern void reclaim_throttle (pg_data_t *pgdat,enum vmscan_throttle_state reason);/* * in mm/rmap.c:
mm/page-writeback.c +1 -1
@@ -2374,7 +2374,7 @@ int do_writepages(struct address_space *mapping, struct writeback_control *wbc)* guess as any. */ reclaim_throttle (NODE_DATA (numa_node_id ()), -VMSCAN_THROTTLE_WRITEBACK ,HZ /50 ); +VMSCAN_THROTTLE_WRITEBACK ); }/* * Usually few pages are written by now from those we've just submitted
mm/vmscan.c +35 -13
92 lines not on this page · raw message
Signature · 1 line
2.31.1
PATCH v4 7/8 mm/vmscan: Increase the timeout if page reclaim is not making progress
In reply to Mel GormanTracing of the stutterp workload showed the following delays
1 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=536000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=544000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=556000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=624000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=716000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=772000 reason=VMSCAN_THROTTLE_NOPROGRESS 2 usect_delayed=512000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS 53 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
116 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
5907 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
71741 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
All the throttling hit the full timeout and then there was wakeup delays meaning that the wakeups are premature as no other reclaimer such as kswapd has made progress. This patch increases the maximum timeout.
--- mm/vmscan.c | 2 ++ 1 file changed, 2 insertions(+)
mm/vmscan.c +2 -0
@@ -1033,6 +1033,8 @@ void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)*/ switch (reason) {case VMSCAN_THROTTLE_NOPROGRESS : + timeout =HZ /2 ; +break ;case VMSCAN_THROTTLE_WRITEBACK : timeout =HZ /10 ;
Signature · 1 line
2.31.1
PATCH v4 8/8 mm/vmscan: Delay waking of tasks throttled on NOPROGRESS
In reply to Mel GormanTracing indicates that tasks throttled on NOPROGRESS are woken prematurely resulting in occasional massive spikes in direct reclaim activity. This patch wakes tasks throttled on NOPROGRESS if reclaim efficiency is at least 12%.
--- mm/vmscan.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)
mm/vmscan.c +6 -3
@@ -1057,7 +1057,7 @@ void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason) ret =schedule_timeout (timeout);finish_wait (wqh, &wait); -if (reason ==VMSCAN_THROTTLE_ISOLATED ) +if (reason ==VMSCAN_THROTTLE_WRITEBACK )atomic_dec (&pgdat->nr_writeback_throttled); trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
@@ -3349,8 +3349,11 @@ static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)static void consider_reclaim_throttle (pg_data_t *pgdat,struct scan_control *sc) { -/* If reclaim is making progress, wake any throttled tasks. */ -if (sc->nr_reclaimed) { +/* +* If reclaim is making progress greater than 12% efficiency then +* wake all the NOPROGRESS throttled tasks. +*/ +if (sc->nr_reclaimed > (sc->nr_scanned >>3 )) {wait_queue_head_t *wqh; wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS ];
Signature · 1 line
2.31.1
Re: [PATCH 5/8] mm/page_alloc: Remove the throttling logic from the page allocator
In reply to Mel GormanQuoted 8 lines · Mel Gorman
The page allocator stalls based on the number of pages that are waiting for writeback to start but this should now be redundant.
shrink_inactive_list()will wake flusher threads if the LRU tail are unqueued dirty pages so the flusher should be active. If it fails to make progress due to pages under writeback not being completed quickly then it should stall onVMSCAN_THROTTLE_WRITEBACK.Signed-off-by Mel Gorman <[email protected]>
(did in v3 already)
Re: [PATCH 2/8] mm/vmscan: Throttle reclaim and compaction when too may pages are isolated
In reply to Mel GormanQuoted 5 lines · Mel Gorman
Page reclaim throttles on congestion if too many parallel reclaim instances have isolated too many pages. This makes no sense, excessive parallelisation has nothing to do with writeback or congestion.
Quoted 7 lines
This patch creates an additional workqueue to sleep on when too many pages are isolated. The throttled tasks are woken when the number of isolated pages is reduced or a timeout occurs. There may be some false positive wakeups for GFP_NOIO/GFP_NOFS callers but the tasks will throttle again if necessary.
[[email protected]: Wake up from compaction context]
Quoted 245 lines
[[email protected]: Account number of throttled tasks only for writeback]Signed-off-by Mel Gorman <[email protected]>Acked-by Vlastimil Babka <[email protected]>--- include/linux/mmzone.h | 6 ++++-- include/trace/events/vmscan.h | 4 +++- mm/compaction.c | 10 ++++++++-- mm/internal.h | 13 ++++++++++++- mm/page_alloc.c | 6 +++++- mm/vmscan.c | 28 +++++++++++++++++++--------- 6 files changed, 51 insertions(+), 16 deletions(-)include/linux/mmzone.h +4
-2@@ -275,6 +275,8 @@ enum lru_list {enum vmscan_throttle_state {VMSCAN_THROTTLE_WRITEBACK , +VMSCAN_THROTTLE_ISOLATED , +NR_VMSCAN_THROTTLE , };#define for_each_lru (lru) for (lru = 0; lru < NR_LRU_LISTS; lru++)@@ -846,8 +848,8 @@ typedef struct pglist_data {int node_id;wait_queue_head_t kswapd_wait;wait_queue_head_t pfmemalloc_wait; -wait_queue_head_t reclaim_wait;/* wq for throttling reclaim */ -atomic_t nr_reclaim_throttled;/* nr of throtted tasks */ +wait_queue_head_t reclaim_wait[NR_VMSCAN_THROTTLE ]; +atomic_t nr_writeback_throttled;/* nr of writeback-throttled tasks */ unsigned long nr_reclaim_start;/* nr pages written while throttled * when throttling started. */ struct task_struct *kswapd;/* Protected by include/trace/events/vmscan.h +3
-1@@ -28,10 +28,12 @@ ) : "RECLAIM_WB_NONE"#define _VMSCAN_THROTTLE_WRITEBACK (1 << VMSCAN_THROTTLE_WRITEBACK) +#define _VMSCAN_THROTTLE_ISOLATED (1 << VMSCAN_THROTTLE_ISOLATED)#define show_throttle_flags (flags) \ (flags) ? __print_flags(flags, "|", \ - {_VMSCAN_THROTTLE_WRITEBACK, "VMSCAN_THROTTLE_WRITEBACK"} \ + {_VMSCAN_THROTTLE_WRITEBACK, "VMSCAN_THROTTLE_WRITEBACK"}, \ + {_VMSCAN_THROTTLE_ISOLATED, "VMSCAN_THROTTLE_ISOLATED"} \ ) : "VMSCAN_THROTTLE_NONE"mm/compaction.c +8
-2@@ -761,6 +761,8 @@ isolate_freepages_range(struct compact_control *cc,/* Similar to reclaim, but different enough that they don't share logic */ static bool too_many_isolated (pg_data_t *pgdat) { +bool too_many; +unsigned long active, inactive, isolated; inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +@@ -770,7 +772,11 @@ static bool too_many_isolated(pg_data_t *pgdat) isolated =node_page_state (pgdat,NR_ISOLATED_FILE ) +node_page_state (pgdat,NR_ISOLATED_ANON ); -return isolated > (inactive + active) /2 ; + too_many = isolated > (inactive + active) /2 ; +if (!too_many) +wake_throttle_isolated (pgdat); + +return too_many; }/** @@ -822,7 +828,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,if (cc->mode ==MIGRATE_ASYNC )return -EAGAIN ; -congestion_wait (BLK_RW_ASYNC ,HZ /10 ); +reclaim_throttle (pgdat,VMSCAN_THROTTLE_ISOLATED ,HZ /10 );if (fatal_signal_pending (current))return -EINTR ;mm/internal.h +12
-1@@ -39,12 +39,21 @@ void __acct_reclaim_writeback(pg_data_t *pgdat, struct page *page,static inline void acct_reclaim_writeback (struct page *page) {pg_data_t *pgdat =page_pgdat (page); -int nr_throttled =atomic_read (&pgdat->nr_reclaim_throttled); +int nr_throttled =atomic_read (&pgdat->nr_writeback_throttled);if (nr_throttled)__acct_reclaim_writeback (pgdat, page, nr_throttled); } +static inline void wake_throttle_isolated (pg_data_t *pgdat) +{ +wait_queue_head_t *wqh; + + wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_ISOLATED ]; +if (waitqueue_active (wqh)) +wake_up_all (wqh); +} +vm_fault_t do_swap_page (struct vm_fault *vmf); void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *start_vma,@@ -120,6 +129,8 @@ extern unsigned long highest_memmap_pfn;*/ extern int isolate_lru_page (struct page *page);extern void putback_lru_page (struct page *page); +extern void reclaim_throttle (pg_data_t *pgdat,enum vmscan_throttle_state reason, +long timeout);/* * in mm/rmap.c: mm/page_alloc.c +5
-1@@ -7389,6 +7389,8 @@ static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}static void __meminit pgdat_init_internals (struct pglist_data *pgdat) { +int i; +pgdat_resize_init (pgdat);pgdat_init_split_queue (pgdat);@@ -7396,7 +7398,9 @@ static void __meminit pgdat_init_internals(struct pglist_data *pgdat)init_waitqueue_head (&pgdat->kswapd_wait);init_waitqueue_head (&pgdat->pfmemalloc_wait); -init_waitqueue_head (&pgdat->reclaim_wait); + +for (i =0 ; i <NR_VMSCAN_THROTTLE ; i++) +init_waitqueue_head (&pgdat->reclaim_wait[i]);pgdat_page_ext_init (pgdat);lruvec_init (&pgdat->__lruvec);mm/vmscan.c +19
-10@@ -1006,12 +1006,12 @@ static void handle_write_error(struct address_space *mapping,unlock_page (page); } -static void -reclaim_throttle (pg_data_t *pgdat,enum vmscan_throttle_state reason, +void reclaim_throttle (pg_data_t *pgdat,enum vmscan_throttle_state reason,long timeout) { -wait_queue_head_t *wqh = &pgdat->reclaim_wait; +wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];long ret; +bool acct_writeback = (reason ==VMSCAN_THROTTLE_WRITEBACK );DEFINE_WAIT (wait);/* @@ -1023,7 +1023,8 @@ reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason, current->flags & (PF_IO_WORKER|PF_KTHREAD))return ; -if (atomic_inc_return (&pgdat->nr_reclaim_throttled) ==1 ) { +if (acct_writeback && +atomic_inc_return (&pgdat->nr_writeback_throttled) ==1 ) {WRITE_ONCE (pgdat->nr_reclaim_start,node_page_state (pgdat,NR_THROTTLED_WRITTEN )); }@@ -1031,7 +1032,9 @@ reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason,prepare_to_wait (wqh, &wait,TASK_UNINTERRUPTIBLE ); ret =schedule_timeout (timeout);finish_wait (wqh, &wait); -atomic_dec (&pgdat->nr_reclaim_throttled); + +if (acct_writeback) +atomic_dec (&pgdat->nr_writeback_throttled); trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout), jiffies_to_usecs(timeout - ret),@@ -1061,7 +1064,7 @@ void __acct_reclaim_writeback(pg_data_t *pgdat, struct page *page,READ_ONCE (pgdat->nr_reclaim_start);if (nr_written >SWAP_CLUSTER_MAX * nr_throttled) -wake_up_all (&pgdat->reclaim_wait); +wake_up_all (&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK ]); }/* possible outcome of pageout() */ @@ -2176,6 +2179,7 @@ static int too_many_isolated(struct pglist_data *pgdat, int file, struct scan_control *sc) {unsigned long inactive, isolated; +bool too_many;if (current_is_kswapd ())return 0 ;@@ -2199,7 +2203,13 @@ static int too_many_isolated(struct pglist_data *pgdat, int file,if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) inactive >>=3 ; -return isolated > inactive; + too_many = isolated > inactive; + +/* Wake up tasks throttled due to too_many_isolated. */ +if (!too_many) +wake_throttle_isolated (pgdat); + +return too_many; }/* @@ -2308,8 +2318,8 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,return 0 ;/* wait a bit for the reclaimer. */ -msleep (100 ); stalled = true; +reclaim_throttle (pgdat,VMSCAN_THROTTLE_ISOLATED ,HZ /10 );/* We are about to die and free our memory. Return now. */ if (fatal_signal_pending (current))@@ -4343,7 +4353,7 @@ static int kswapd(void *p)WRITE_ONCE (pgdat->kswapd_order,0 );WRITE_ONCE (pgdat->kswapd_highest_zoneidx,MAX_NR_ZONES ); -atomic_set (&pgdat->nr_reclaim_throttled,0 ); +atomic_set (&pgdat->nr_writeback_throttled,0 );for ( ; ; ) {bool ret; --2.31.1
Andrew Morton
maintainerQuoted 5 lines · On Tue, 19 Oct 2021 10:01:00 +0100 Mel Gorman <[email protected]>…
Changelog since v3
o Count writeback completions forNR_THROTTLED_WRITTENonly
o Use IRQ-safeinc_node_page_state
o Remove redundant throttling
Quoted 8 lines
This series is also available at
git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.gitmm-reclaimcongest-v4r2This series that removes all calls to
congestion_waitinmm/and deleteswait_iff_congested. It's not a clever implementation butcongestion_waithas been broken for a long time
(https://lore.kernel.org/linux-mm/[email protected]/).
The block layer doesn't call clear_bdi_congested() at all. I never knew this until recent discussions :(
So this means that congestion_wait() will always time out, yes?
Quoted 1 line
Even if congestion throttling worked, it was never a great idea.
Well. It was a good idea until things like isolation got added!
Quoted 15 lines
While
excessive dirty/writeback pages at the tail of the LRU is one possibility that reclaim may be slow, there is also the problem of too many pages being isolated and reclaim failing for other reasons (elevated references, too many pages isolated, excessive LRU contention etc).This series replaces the "congestion" throttling with 3 different types.
o If there are too many dirty/writeback pages, sleep until a timeout or enough pages get cleaned
o If too many pages are isolated, sleep until enough isolated pages are either reclaimed or put back on the LRU
o If no progress is being made, direct reclaim tasks sleep until another task makes progress with acceptable efficiency.
Quoted 2 lines
This was initially tested with a mix of workloads that used to trigger corner cases that no longer work.
Mix of workloads is nice, but a mix of devices is more important here. I trust some testing was done on plain old spinning disks? And USB storage, please! And NFS plays with BDI congestion. Ceph and FUSE also.
We've had complaints about this stuff forever. Usually of the form of interactive tasks getting horridly stalled by writeout/swap activity.
Quoted 16 lines · Andrew Morton
Quoted 5 lines · On Tue, 19 Oct 2021 10:01:00 +0100 Mel Gorman <[email protected]>…
Changelog since v3
o Count writeback completions forNR_THROTTLED_WRITTENonly
o Use IRQ-safeinc_node_page_state
o Remove redundant throttlingQuoted 8 lines
This series is also available at
git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.gitmm-reclaimcongest-v4r2This series that removes all calls to
congestion_waitinmm/and deleteswait_iff_congested. It's not a clever implementation butcongestion_waithas been broken for a long time
(https://lore.kernel.org/linux-mm/[email protected]/).
Quoted 5 lines
The block layer doesn't call
clear_bdi_congested()at all. I never knew this until recent discussions :(So this means that
congestion_wait()will always time out, yes?
Unfortunately, yes except for filesystems that call [set_clear]_bdi_congested. For the test case in the series leader, congestion_wait always hit the full timeout.
Quoted 4 lines
Quoted 1 line
Even if congestion throttling worked, it was never a great idea.
Well. It was a good idea until things like isolation got added!
Well, true to an extent although it was always true that reclaim could fail to make progress for reasons other than pages under writeback. But you're right, saying it was "never a great idea" is overkill. congestion_wait used to work and I expect it was particularly helpful before IO-less write throttling, accurate dirty page tracking and immediate reclaim existed.
Quoted 15 lines
Quoted 15 lines
While
excessive dirty/writeback pages at the tail of the LRU is one possibility that reclaim may be slow, there is also the problem of too many pages being isolated and reclaim failing for other reasons (elevated references, too many pages isolated, excessive LRU contention etc).This series replaces the "congestion" throttling with 3 different types.
o If there are too many dirty/writeback pages, sleep until a timeout or enough pages get cleaned
o If too many pages are isolated, sleep until enough isolated pages are either reclaimed or put back on the LRU
o If no progress is being made, direct reclaim tasks sleep until another task makes progress with acceptable efficiency.
Quoted 4 lines
Quoted 2 lines
This was initially tested with a mix of workloads that used to trigger corner cases that no longer work.
Mix of workloads is nice, but a mix of devices is more important here.
I tested as much as I could but as well as storage devices, different memory sizes are also relevant.
Quoted 3 lines
I trust some testing was done on plain old spinning disks? And USB storage, please! And NFS plays with BDI congestion. Ceph and FUSE also.
Plain old spinning disk was tested. Basic USB testing didn't show many problems although given it was my desktop machine, it might have had too memory as no amount of IO to the USB key triggered a problem where reclaim failed to make progress and get throttled. There was basic NFS testing although I didn't try running stutterp over NFS. Given the original thread motivating this was NFS-related and they are cc'd, I'm hoping they'll give it a realistic kick. I don't have a realistic setup for ceph and didn't try fuse.
Quoted 2 lines
We've had complaints about this stuff forever. Usually of the form of interactive tasks getting horridly stalled by writeout/swap activity.
I know and there is no guarantee it won't happen again. The main problem I was trying to solve was that congestion-based throttling is not suitable for mm/.
From reclaim context, there isn't a good way of detecting "interactive" tasks. At best, under reclaim pressure we could try tracking allocation rates and throttle heavy allocators more than light allocators but I didn't want to introduce complexity prematurely.
Signature · 2 lines
Mel Gorman SUSE Labs
Re: [PATCH 6/8] mm/vmscan: Centralise timeout values for reclaim_throttle
In reply to Mel GormanOn Tue, 19 Oct 2021, Mel Gorman wrote:
...
mm/vmscan.c void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason, · line 1041
Quoted 8 lines
mm/vmscan.c +8
-0+switch (reason) { +case VMSCAN_THROTTLE_NOPROGRESS : +case VMSCAN_THROTTLE_WRITEBACK : + timeout =HZ /10 ; + +if (atomic_inc_return (&pgdat->nr_writeback_throttled) ==1 ) { +WRITE_ONCE (pgdat->nr_reclaim_start, +node_page_state (pgdat,NR_THROTTLED_WRITTEN ));
You have introduced a behaviour change that wasn't flagged in the commit message.
Previously nr_writeback_throttled was only incremented for VMSCAN_THROTTLE_WRITEBACK, now it is incremented for VMSCAN_THROTTLE_NOPROGRESS as well.
Some justification would be good.
Quoted 8 lines
mm/vmscan.c +8
-0+ } + +break ; +case VMSCAN_THROTTLE_ISOLATED : + timeout =HZ /50 ; +break ; +default : +WARN_ON_ONCE (1 );
mm/vmscan.c void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason, · line 1058
Quoted 10 lines
mm/vmscan.c +3
-1+ timeout =HZ ; +break ; }prepare_to_wait (wqh, &wait,TASK_UNINTERRUPTIBLE ); ret =schedule_timeout (timeout);finish_wait (wqh, &wait); -if (acct_writeback) +if (reason ==VMSCAN_THROTTLE_ISOLATED )
(defect) I think you want "!=" there.
While the numbers a still magic, they are now well documented and all in one place - a definite improvement!
Thanks,
NeilBrown
Re: [PATCH 7/8] mm/vmscan: Increase the timeout if page reclaim is not making progress
In reply to Mel GormanQuoted 18 lines · Mel Gorman
Tracing of the stutterp workload showed the following delays
1 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=536000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=544000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=556000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=624000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=716000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=772000 reason=VMSCAN_THROTTLE_NOPROGRESS 2 usect_delayed=512000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS 53 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
116 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
5907 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
71741 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
Quoted 3 lines
All the throttling hit the full timeout and then there was wakeup delays meaning that the wakeups are premature as no other reclaimer such as kswapd has made progress. This patch increases the maximum timeout.
Would love to see the comparable tracing results for after the patch.
Thanks,
NeilBrown
Quoted 24 lines
Signed-off-by Mel Gorman <[email protected]>Acked-by Vlastimil Babka <[email protected]>--- mm/vmscan.c | 2 ++ 1 file changed, 2 insertions(+)mm/vmscan.c +2
-0@@ -1033,6 +1033,8 @@ void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)*/ switch (reason) {case VMSCAN_THROTTLE_NOPROGRESS : + timeout =HZ /2 ; +break ;case VMSCAN_THROTTLE_WRITEBACK : timeout =HZ /10 ;Signature · 1 line
2.31.1
Quoted 11 lines · Mel Gorman
Changelog since v3
o Count writeback completions forNR_THROTTLED_WRITTENonly
o Use IRQ-safeinc_node_page_state
o Remove redundant throttlingThis series is also available at
git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.gitmm-reclaimcongest-v4r2This series that removes all calls to
congestion_waitinmm/and deleteswait_iff_congested.
Thanks for this.
I don't have sufficient expertise for a positive review, but it seems to make sense with one exception which I have commented on separately.
In general, I still don't like the use of wake_up_all(), though it won't cause incorrect behaviour.
I would prefer the first patch would:
- define NR_VMSCAN_THROTTLE
- make reclaim_wait an array
- spelled nr_reclaim_throttled as nr_writeback_throttled
rather than leaving those changes for the second patch. I think that would make review easier.
But these are minor and I'll not mention them again.
Thanks,
NeilBrown
Re: [PATCH 6/8] mm/vmscan: Centralise timeout values for reclaim_throttle
In reply to NeilBrownQuoted 17 lines · NeilBrown
On Tue, 19 Oct 2021, Mel Gorman wrote:
...Review
Quoted 8 lines
(quoted diff) +8
-0+ switch(reason) { + case VMSCAN_THROTTLE_NOPROGRESS: + case VMSCAN_THROTTLE_WRITEBACK: + timeout = HZ/10; + + if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) { + WRITE_ONCE(pgdat->nr_reclaim_start, + node_page_state(pgdat, NR_THROTTLED_WRITTEN));You have introduced a behaviour change that wasn't flagged in the commit message.
Previouslynr_writeback_throttledwas only incremented forVMSCAN_THROTTLE_WRITEBACK, now it is incremented forVMSCAN_THROTTLE_NOPROGRESSas well.
Quoted 2 lines
Some justification would be good.
This is the result of rebase near the end of a day going sideways. There is no justification, it's just wrong.
I'm rerunning the entire series, will update the leader and resend the series.
--8<--
mm/vmscan: Centralise timeout values for reclaim_throttle -fix
Neil Brown spotted the fallthrough-logic for reclaim_throttle was wrong -- only VMSCAN_THROTTLE_WRITEBACK affects pgdat->nr_writeback_throttled. This was the result of a rebase going sideways and only happens to sometimes work by co-incidence.
This is a fix to the mmotm patch mm-vmscan-centralise-timeout-values-for-reclaim_throttle.patch
--- mm/vmscan.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
mm/vmscan.c +4 -2
@@ -1032,7 +1032,6 @@ void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)* of the inactive LRU. */ switch (reason) { -case VMSCAN_THROTTLE_NOPROGRESS :case VMSCAN_THROTTLE_WRITEBACK : timeout =HZ /10 ;
@@ -1041,6 +1040,9 @@ void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason) node_page_state(pgdat, NR_THROTTLED_WRITTEN)); } +break ; +case VMSCAN_THROTTLE_NOPROGRESS : + timeout =HZ /10 ;break ;case VMSCAN_THROTTLE_ISOLATED : timeout =HZ /50 ;
@@ -1055,7 +1057,7 @@ void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason) ret =schedule_timeout (timeout);finish_wait (wqh, &wait); -if (reason ==VMSCAN_THROTTLE_ISOLATED ) +if (reason ==VMSCAN_THROTTLE_WRITEBACK )atomic_dec (&pgdat->nr_writeback_throttled); trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
Re: [PATCH 7/8] mm/vmscan: Increase the timeout if page reclaim is not making progress
In reply to NeilBrownQuoted 19 lines · NeilBrown
Quoted 18 lines · Mel Gorman
Tracing of the stutterp workload showed the following delays
1 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=536000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=544000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=556000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=624000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=716000 reason=VMSCAN_THROTTLE_NOPROGRESS 1 usect_delayed=772000 reason=VMSCAN_THROTTLE_NOPROGRESS 2 usect_delayed=512000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS 53 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
116 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
5907 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
71741 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
Quoted 6 lines
Quoted 3 lines
All the throttling hit the full timeout and then there was wakeup delays meaning that the wakeups are premature as no other reclaimer such as kswapd has made progress. This patch increases the maximum timeout.
Would love to see the comparable tracing results for after the patch.
They're in the leader. The trace figures in the changelog are the ones I had at the time the patch was developed and I didn't keep them up to date to reduce overall test time. At the last set of results, some throttling was still hitting the full timeout;
[....]
843 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
1299 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS 2839 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
10111 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS 21492 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS 36441 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
Signature · 2 lines
Mel Gorman SUSE Labs
Quoted 8 lines · NeilBrown
Quoted 7 lines · Mel Gorman
Changelog since v3
o Count writeback completions forNR_THROTTLED_WRITTENonly
o Use IRQ-safeinc_node_page_state
o Remove redundant throttlingThis series is also available at
Quoted 9 lines
Quoted 4 lines
git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.gitmm-reclaimcongest-v4r2This series that removes all calls to
congestion_waitinmm/and deleteswait_iff_congested.Thanks for this.
I don't have sufficient expertise for a positive review, but it seems to make sense with one exception which I have commented on separately.
A test battering NFS would still be nice!
Quoted 3 lines
In general, I still don't like the use of
wake_up_all(), though it won't cause incorrect behaviour.
Removing wake_up_all would be tricky. Ideally it would be prioritised but more importantly, some sort of guarantee should exist that enough wakeup events trigger to wake tasks before the timeout. That would need careful thinking about each reclaim reason. For example, if N tasks throttle on NOPROGRESS, there is no guarantee that N tasks are currently in reclaim that would wake each sleeping task as progress is made. It's similar for writeback, are enough pages under writeback to trigger each wakeup? A more subtle issue is if each reason should be strict if waking tasks one at a time. For example, a task sleeping on writeback might make progress for other reasons such as the working set changing during reclaim or a large task exiting. Of course the same concerns exist for the series as it stands but the worst case scenarios are mitigated by wake_up_all.
Quoted 8 lines
I would prefer the first patch would:
- defineNR_VMSCAN_THROTTLE
- makereclaim_waitan array
- spellednr_reclaim_throttledasnr_writeback_throttledrather than leaving those changes for the second patch. I think that would make review easier.
I can do this. Normally I try structure series from least-to-most controversial so that it can be cut at any point and still make sense so the array was defined in the second patch because that's when it is required. However, I already had defined the enum in patch 1 for the tracepoint so I might as well make it an array too.
Signature · 2 lines
Mel Gorman SUSE Labs
Quoted 7 lines · Mel Gorman
Quoted 3 lines · NeilBrown
In general, I still don't like the use of
wake_up_all(), though it won't cause incorrect behaviour.Removing
wake_up_allwould be tricky.
I think there is a misunderstanding. Removing wake_up_all() is as simple as
s/wake_up_all/wake_up/
If you used prepare_to_wait_exclusive(), then wake_up() would only wake one waiter, while wake_up_all() would wake all of them.
As you use prepare_to_wait(), wake_up() will wake all waiters - as will wake_up_all().
When I see "wake_up_all()" I assume it is an exclusive wait, and that for some reason this particular wake_up needs to wake up all waiters. That is not the case here.
I suspect it would be clearer if "wake_up" always woke everything, and "wake_up_one" was the special case - but unfortunately that isn't what we have.
There are other non-exclusive waiters which use wake_up_all(), but the vast majority of wakeups use wake_up(), and most of those are for non-exclusive waiters.
Thanks,
NeilBrown
Quoted 9 lines · NeilBrown
Quoted 7 lines · Mel Gorman
Quoted 3 lines · NeilBrown
In general, I still don't like the use of
wake_up_all(), though it won't cause incorrect behaviour.Removing
wake_up_allwould be tricky.
Quoted 9 lines
I think there is a misunderstanding. Removing
wake_up_all()is as simple as
s/wake_up_all/wake_up/If you used
prepare_to_wait_exclusive(), thenwake_up()would only wake one waiter, whilewake_up_all()would wake all of them.
As you useprepare_to_wait(),wake_up()will wake all waiters - as willwake_up_all().
Ok, yes, there was a misunderstanding. I thought you were suggesting a move to exclusive wakeups. I felt that the wake_up_all was explicit in terms of intent and that I really meant for all tasks to wake instead of one at a time.
Signature · 2 lines
Mel Gorman SUSE Labs
Quoted 20 lines · Mel Gorman
Quoted 9 lines · NeilBrown
Quoted 7 lines · Mel Gorman
Quoted 3 lines · NeilBrown
In general, I still don't like the use of
wake_up_all(), though it won't cause incorrect behaviour.Removing
wake_up_allwould be tricky.Quoted 9 lines
I think there is a misunderstanding. Removing
wake_up_all()is as simple as
s/wake_up_all/wake_up/If you used
prepare_to_wait_exclusive(), thenwake_up()would only wake one waiter, whilewake_up_all()would wake all of them.
As you useprepare_to_wait(),wake_up()will wake all waiters - as willwake_up_all().
Quoted 4 lines
Ok, yes, there was a misunderstanding. I thought you were suggesting a move to exclusive wakeups. I felt that the
wake_up_allwas explicit in terms of intent and that I really meant for all tasks to wake instead of one at a time.
Fair enough. Thanks for changing it :-)
But this prompts me to wonder if exclusive wakeups would be a good idea
- which is a useful springboard to try to understand the code better.
For VMSCAN_THROTTLE_ISOLATED they probably are.
One pattern for reliable exclusive wakeups is for any thread that received a wake-up to then consider sending a wake up.
Two places receive VMSCAN_THROTTLE_ISOLATED wakeups and both then call too_many_isolated() which - on success - sends another wakeup - before the caller has had a chance to isolate anything. If, instead, the wakeup was sent sometime later, after pages were isolated by before the caller (isoloate_migratepages_block() or shrink_inactive_list()) returned, then we would get an orderly progression of threads running through that code.
For VMSCAN_THROTTLE_WRITEBACK is a little less straight forward. There are two different places that wait for the wakeup, and a wake_up is sent to all waiters after a time proportional to the number of waiters. It might make sense to wake one thread per time unit? That might work well for do_writepages - every SWAP_CLUSTER_MAX writes triggers one wakeup.
I'm less sure that it would work for shrink_node(). Maybe the shrink_node() waiters could be non-exclusive so they get woken as soon a SWAP_CLUSTER_MAX writes complete, while do_writepages are exclusive and get woken one at a time.
For VMSCAN_THROTTLE_NOPROGRESS .... I don't understand.
If one zone isn't making "enough" progress, we throttle before moving on to the next zone. So we delay processing of the next zone, and only indirectly delay re-processing of the current congested zone. Maybe it make sense, but I don't see it yet. I note that the commit message says "it's messy". I can't argue with that!
I'll follow up with patches to clarify what I am thinking about the first two. I'm not proposing the patches, just presenting them as part of improving my understanding.
Thanks,
NeilBrown
Quoted 35 lines · NeilBrown
Quoted 20 lines · Mel Gorman
Quoted 9 lines · NeilBrown
Quoted 7 lines · Mel Gorman
Quoted 3 lines · NeilBrown
In general, I still don't like the use of
wake_up_all(), though it won't cause incorrect behaviour.Removing
wake_up_allwould be tricky.Quoted 9 lines
I think there is a misunderstanding. Removing
wake_up_all()is as simple as
s/wake_up_all/wake_up/If you used
prepare_to_wait_exclusive(), thenwake_up()would only wake one waiter, whilewake_up_all()would wake all of them.
As you useprepare_to_wait(),wake_up()will wake all waiters - as willwake_up_all().Quoted 4 lines
Ok, yes, there was a misunderstanding. I thought you were suggesting a move to exclusive wakeups. I felt that the
wake_up_allwas explicit in terms of intent and that I really meant for all tasks to wake instead of one at a time.Fair enough. Thanks for changing it :-)
But this prompts me to wonder if exclusive wakeups would be a good idea
- which is a useful springboard to try to understand the code better.For
VMSCAN_THROTTLE_ISOLATEDthey probably are.
One pattern for reliable exclusive wakeups is for any thread that received a wake-up to then consider sending a wake up.
Quoted 8 lines
Two places receive
VMSCAN_THROTTLE_ISOLATEDwakeups and both then calltoo_many_isolated()which - on success - sends another wakeup - before the caller has had a chance to isolate anything. If, instead, the wakeup was sent sometime later, after pages were isolated by before the caller (isoloate_migratepages_block()orshrink_inactive_list()) returned, then we would get an orderly progression of threads running through that code.
That should work as the throttling condition is straight-forward. It might even reduce a race condition where waking all throttled tasks all then trigger the same "too many isolated" condition.
Quoted 4 lines
For
VMSCAN_THROTTLE_WRITEBACKis a little less straight forward. There are two different places that wait for the wakeup, and awake_upis sent to all waiters after a time proportional to the number of waiters. It might make sense to wake one thread per time unit?
I'd avoid time as a wakeup condition other than the timeout which is there to guarantee forward progress. I assume you mean "one thread per SWAP_CLUSTER_MAX completions".
Quoted 7 lines
That might work well for
do_writepages- everySWAP_CLUSTER_MAXwrites triggers one wakeup.
I'm less sure that it would work forshrink_node(). Maybe theshrink_node()waiters could be non-exclusive so they get woken as soon aSWAP_CLUSTER_MAXwrites complete, whiledo_writepagesare exclusive and get woken one at a time.
It should work for either with the slight caveat that the last waiter may not see SWAP_CLUSTER_MAX completions.
Quoted 7 lines
For
VMSCAN_THROTTLE_NOPROGRESS.... I don't understand.
If one zone isn't making "enough" progress, we throttle before moving on to the next zone. So we delay processing of the next zone, and only indirectly delay re-processing of the current congested zone. Maybe it make sense, but I don't see it yet. I note that the commit message says "it's messy". I can't argue with that!
Yes, we delay the processing of the next zone when a given zone cannot make progress. The thinking is that circumstances that cause one zone to fail to make progress could spill over to other zones in the absense of any throttling. Where it might cause problems is where the preferred zone is very small. If a bug showed up like that, a potential fix would be to avoid throttling if the preferred zone is very small relative to the total amount of memory local to the node or total memory (preferably local node).
Quoted 4 lines
I'll follow up with patches to clarify what I am thinking about the first two. I'm not proposing the patches, just presenting them as part of improving my understanding.
If I'm cc'd, I'll review and if I think they're promising, I'll run them through the same tests and machines.
Signature · 2 lines
Mel Gorman SUSE Labs