Re: [f2fs-dev] [RFC PATCH v2 00/10] f2fs: support & optimize large folios for writable files

[Chao Yu]

Nanzhe,

I can see a lots of testcases failure in xfstests, can you please address all
of them?

On 6/23/26 00:08, Nanzhe Zhao wrote:
> From: Nanzhe Zhao <nzzhao@xxxxxxx>
> 
> Make some clean up based on v1, and support minimum folio orders 
> beyond zero.
> 
> This RFC series supports large folios for most readable/writable files in
> buffered I/O paths, including normal files, block-layer encrypted files,
> and atomic files. Compressed files are still excluded.
> 
> Atomic files need explicit support here because Android enables atomic
> writes through ioctls, which mark the inode as an atomic file
> (FI_ATOMIC_FILE). Once large-folio mapping is enabled for such a file,
> the atomic buffered write path also needs to handle large folios
> correctly.
> 
> In the write path, allocating f2fs_folio_state for large folios starts
> to conflict with f2fs page-private flags. This RFC series extends
> f2fs_folio_state so that it can also store f2fs private flags, and
> updates the existing PAGE_PRIVATE helpers to work correctly with
> f2fs_folio_state.
> 
> In addition, fio results with max large-folio order set to 2 showed
> that 4K read/write performance did not improve much.
> Analysis showed that one important reason was the extra spinlock traffic
> from incrementing read_pages_pending once per 4K subpage.
> This RFC series therefore adds two optimizations to
> f2fs_read_data_large_folio:
> 
> 1. batch read_pages_pending updates by the mapped block count instead of
>    incrementing it once per 4K subpage;
> 
> 2. skip f2fs_folio_state allocation when a single block mapping / BIO
>    covers the whole folio, because folio_end_read() can complete such a
>    folio without extra per-folio state.
> 
> In the benchmark tables below, "skip=1" means enabling the second
> optimization above, i.e. skipping f2fs_folio_state allocation for the
> whole-folio-in-one-bio case.
> The "batch" optimization refers to updating read_pages_pending in
> and add to bio in larger chunks instead of once per subpage.
> 
> Test environment:
> - Device: Pixel 6 (device1A, 1A071FDF600053)
> - Filesystem: f2fs on dm-49, inlinecrypt enabled
> - File size: 256MB
> - Repetitions: 10
> - Prepare: end_fsync + sync + drop_caches
> - Fio: psync, direct=0, iodepth=1
> - Max folio order: 2
> 
> Table 1: HOLE_READ (10 repeats)
> ------------------------------------------------------------
> All bandwidth numbers are in MiB/s. Non-baseline entries show the
> absolute value followed by the percentage delta relative to the
> order=0 baseline in parentheses.
> 
> | bs  | order=0 | order=2 |
> |-----|---------|---------|
> | 4k  | 469.6   | 521.9 (+11.1%) |
> | 64k | 668.1   | 852.4 (+27.6%) |
> | 1M  | 653.0   | 867.2 (+32.8%) |
> 
> Table 2: DATA_READ (10 repeats)
> ----------------------------------------
> | bs  | order=0 | batch=0 | batch=1,skip=0 | batch=1,skip=1 |
> |-----|---------|---------|----------------|----------------|
> | 4k  | 441.6   | 456.5 (+3.4%) | 499.7 (+13.2%) | 544.4 (+23.3%) |
> | 64k | 632.8   | 697.0 (+10.1%) | 837.8 (+32.4%) | 990.9 (+56.6%) |
> | 1M  | 601.5   | 733.0 (+21.9%) | 927.5 (+54.2%) | 963.4 (+60.2%) |
> 
> Table 3: WRITE (10 reps)
> ----------------------------------------------------------
> O = overwrite (N = new write, Y = overwrite)
> S = sync / fsync (Y = fsync enabled, N = no fsync)
> W = writeback (Y = background writeback, N = no writeback)
> 
> | O,S,W | bs  | order=0 | order=2 |
> |-------|-----|---------|---------|
> | N,N,Y | 4k  | 263.4   | 286.3 (+8.7%) |
> | N,N,Y | 64k | 683.5   | 1199.6 (+75.5%) |
> | N,N,Y | 1M  | 767.4   | 1383.8 (+80.3%) |
> | N,Y,N | 4k  | 10.8    | 9.1 (-15.7%) |
> | N,Y,N | 64k | 69.3    | 50.3 (-27.4%) |
> | N,Y,N | 1M  | 103.5   | 157.8 (+52.5%) |
> | Y,N,Y | 4k  | 301.6   | 344.1 (+14.1%) |
> | Y,N,Y | 64k | 691.3   | 865.9 (+25.3%) |
> | Y,N,Y | 1M  | 742.3   | 969.2 (+30.6%) |
> | Y,Y,N | 4k  | 9.5     | 17.1 (+80.0%) |
> | Y,Y,N | 64k | 43.5    | 108.2 (+148.7%) |
> | Y,Y,N | 1M  | 140.9   | 146.6 (+4.0%) |
> 
> Nanzhe (9):
>   f2fs: extend folio state for large folio write path
>   f2fs: carry subpage offset and count in write IO
>   f2fs: support regular file buffered writes on large folios
>   f2fs: support atomic file large folios buffered write
>   f2fs: support large folio writeback
>   f2fs: prepare mmap write faults for large folios
>   f2fs: make GC migration large-folio aware
>   f2fs: allow large folio support to writeable files
>   f2fs: optimize small block size large folio read
> 
>  fs/f2fs/compress.c |    2 +
>  fs/f2fs/data.c     | 1015 +++++++++++++++++++++++++++++++++++++++-----
>  fs/f2fs/f2fs.h     |   75 +++-
>  fs/f2fs/file.c     |   81 ++--
>  fs/f2fs/gc.c       |   30 +-
>  fs/f2fs/inode.c    |    6 +-
>  fs/f2fs/segment.c  |    4 +-
>  7 files changed, 1064 insertions(+), 149 deletions(-)
>