[PATCH v5 0/2] f2fs: enable buffered RWF_DONTCACHE

[Wenjie Qi]

This series enables buffered RWF_DONTCACHE on F2FS for sustained one-pass
streaming writes, where retaining the written data can displace more useful
cache.

Patch 1 marks dropbehind write bios with BIO_COMPLETE_IN_TASK. Normal and
dropbehind folios may share otherwise mergeable IPU and OPU bios; the flag
is set monotonically when a dropbehind folio joins an existing bio. The
block layer owns deferral from unsafe completion contexts.

Patch 2 passes FGP_DONTCACHE to the F2FS buffered write folio lookup,
advertises FOP_DONTCACHE, and handles internal write_begin callers that
pass a NULL kiocb.

Tests were run on a Xiaomi phone with 10.7 GiB of kernel-visible memory,
running Android 16 and Linux 6.12.69 with 4 KiB pages. /data used F2FS.

The performance test wrote exactly 64 GiB per run at 4 KiB,
8 KiB, 16 KiB, 32 KiB, 64 KiB, 128 KiB, 256 KiB, 512 KiB, and 1 MiB.
Two counterbalanced rounds ran ascending normal-first and descending
dontcache-first. Values below are equal-weight means of both runs; N=2.
The pwritev2() writer models the streaming workload; it does not show that
an unchanged Android application already issues RWF_DONTCACHE.

Android remained active with displays off. Each run started after a cache
reset and at least 120 seconds of cooldown. Throughput and one-second
kswapd0/global-memory samples cover the write loop.

Write-loop throughput was:

          normal MiB/s              dontcache MiB/s
 I/O      r1       r2      mean      r1       r2      mean     change
 4K      946.28   935.74   941.01   291.36   309.31   300.33   -68.08%
 8K     1077.05  1105.84  1091.45   477.79   479.79   478.79   -56.13%
 16K    1126.84  1118.49  1122.67   643.40   652.06   647.73   -42.30%
 32K    1150.62  1036.33  1093.48   762.24   751.45   756.84   -30.79%
 64K    1144.80  1163.82  1154.31   852.11   851.19   851.65   -26.22%
 128K   1166.29  1162.84  1164.57   867.47   865.05   866.26   -25.61%
 256K   1153.61  1172.78  1163.19   895.53   885.33   890.43   -23.45%
 512K   1173.61  1197.34  1185.48   903.09   903.01   903.05   -23.82%
 1M     1126.22  1154.59  1140.41   850.74   894.60   872.67   -23.48%

Average kswapd0 CPU and average global Cached were:

 I/O      kswapd0 CPU, normal/DC        Cached MiB, normal/DC
 4K             18.45% / 0%                 4830.15 / 686.56
 8K             21.60% / 0%                 4862.35 / 591.48
 16K            22.75% / 0%                 4905.47 / 625.07
 32K            22.05% / 0%                 4945.82 / 561.59
 64K            23.46% / 0%                 4920.77 / 639.78
 128K           23.16% / 0%                 4971.37 / 693.26
 256K           23.68% / 0%                 4956.93 / 668.60
 512K           24.25% / 0%                 4972.22 / 663.92
 1M             22.01% / 0%                 5001.09 / 705.30

Other global memory means were:

          MemAvailable MiB       Dirty MiB          Writeback MiB
 I/O       normal / DC          normal / DC          normal / DC
 4K      6513.08 / 6382.31     640.48 / 33.93       37.94 / 0.09
 8K      6560.77 / 6529.33     690.89 / 43.57       40.84 / 0.54
 16K     6587.46 / 6519.94     789.16 / 62.67       61.58 / 4.64
 32K     6571.95 / 6566.74     850.48 / 69.14       67.69 / 9.24
 64K     6600.64 / 6559.89    856.97 / 134.28      59.60 / 16.19
 128K    6627.03 / 6465.10    873.68 / 137.57      60.57 / 41.50
 256K    6615.58 / 6541.09    885.98 / 158.15      61.25 / 29.57
 512K    6625.52 / 6534.41    900.65 / 139.07      63.65 / 30.35
 1M      6677.09 / 6539.89    909.70 / 187.56      51.58 / 33.46

          Active(file) MiB      Inactive(file) MiB
 I/O       normal / DC             normal / DC
 4K       279.57 / 264.13         4426.38 / 183.12
 8K       262.32 / 260.97         4472.62 / 182.19
 16K      392.09 / 252.95         4392.78 / 195.82
 32K      254.30 / 248.44         4554.18 / 187.34
 64K      252.73 / 244.94         4545.40 / 267.28
 128K     322.02 / 243.95         4530.63 / 298.91
 256K     245.47 / 240.02         4586.64 / 303.92
 512K     254.75 / 232.05         4591.17 / 288.27
 1M       238.88 / 233.64         4638.81 / 341.00

Dontcache left zero target-file pages resident at every size. Normal
retained about 1.19--1.24 million pages. Normal runs incurred roughly
15.6 million kswapd page scans and steals per run, while dontcache recorded
zero. Direct scan and allocation-stall deltas were zero in both modes.

A controlled explicit-dontcache model issued 64 KiB writes for 120 seconds
at 64, 128, and 256 MiB/s. Both modes sustained all three rates in both
rounds with no final schedule overrun. Between 0.02% and 0.41% of writes
completed late, with a maximum schedule lag of 3.3--6.0 ms. Dontcache left
zero target pages resident. This was a controlled model, not an unchanged
Xiaomi application.

Buffered read throughput was:

 I/O       normal MiB/s   dontcache MiB/s   change
 4K          1825.50          1695.10        -7.14%
 8K          1901.46          1847.96        -2.81%
 16K         1941.60          1872.55        -3.56%
 32K         1960.95          1905.92        -2.81%
 64K         1951.35          1886.54        -3.32%
 128K        1960.03          1915.82        -2.26%
 256K        1976.01          1900.41        -3.83%
 512K        1973.31          1911.24        -3.15%
 1M          2230.91          2007.82       -10.00%

Dontcache left zero source pages resident in all measured read runs.

The normal-I/O control showed read-throughput differences of +0.67%,
-0.90%, and -2.58%, and write-throughput differences of -1.75%, +0.33%,
and +2.33%, at 4 KiB, 64 KiB, and 1 MiB respectively.

As a separate memory-pressure supplement, I tested F2FS on x86-64 QEMU
with the F2FS images backed by the host NVMe. Each run issued a 1-GiB
homogeneous sequential write using 8-KiB or 1-MiB I/O. Normal and
dontcache each ran in three counterbalanced rounds with
memory.max=255852544, memory.swap.max=0, and formal tracing was disabled.
Values below are medians of three completed runs.

 I/O    N done  D done   MiB/s N/D        memcg MiB N/D
 8 KiB  3/3    3/3      236.259/227.634  160.867/2.075
 1 MiB  3/3    3/3      222.232/859.805  155.178/9.098

 I/O    scan N/D            steal N/D
 8 KiB  270,821/0           201,227/0
 1 MiB  344,267/0           201,013/0

Scan is the number of pages examined by reclaim, while steal is the number
of pages successfully reclaimed. memcg MiB is average memory.current.

At 8 KiB, dontcache throughput was 3.65% lower. The paired changes were
-12.7%, +65.8%, and -3.65%, so the direction was not stable. At 1 MiB,
dontcache throughput was 286.90% higher, and all three paired changes were
positive: +96.1%, +257.8%, and +295.8%. Average memory.current decreased by
98.71% at 8 KiB and 94.14% at 1 MiB. Dontcache recorded zero scan and steal
in both profiles.

Changes since v4:

- allow normal and dropbehind folios to share write bios, following iomap,
  and set BIO_COMPLETE_IN_TASK when a dropbehind folio joins an existing
  bio;
- handle internal write_begin callers that pass a NULL kiocb;
- add QEMU memory-pressure and mixed-bio measurements.

Wenjie Qi (2):
  f2fs: complete dropbehind write bios in task context
  f2fs: enable buffered RWF_DONTCACHE

 fs/f2fs/data.c | 15 ++++++++++++---
 fs/f2fs/file.c |  2 +-
 2 files changed, 13 insertions(+), 4 deletions(-)

-- 
2.43.0