Re: [f2fs-dev] [PATCH v2 06/20] f2fs: support subpage blocks in data read path and GC
From: Jaegeuk Kim
Date: Tue Nov 17 2026 - 11:51:41 EST
On 11/17, Chao Yu wrote:
> On 11/16/26 12:08, Daeho Jeong wrote:
> > From: Daeho Jeong <daehojeong@xxxxxxxxxx>
> >
> > With subpage blocks, a folio can be partially uptodate, so reads and GC
> > have to work per block.
> >
> > - Read only the non-uptodate blocks of a multi-block folio, and send
> > every inode to this path in subpage mode.
> > - Do not use the folio state when a single bio reads the whole folio.
> > - Implement ->is_partially_uptodate.
> > - Convert block indices to folio indices in the data folio getters and
> > GC, and prefetch the whole folio in GC so that its blocks are read
> > together.
> > - Fix error handling in f2fs_read_data_large_folio().
> > - Zero the blocks past EOF in f2fs_read_data_large_folio() instead of
> > reading them, as f2fs_read_single_page() does. With subpage blocks,
> > they share a page with the last block of the file and can be seen
> > through mmap.
> > - Test the per-block uptodate state in f2fs_find_data_folio() under the
> > folio lock, as ->release_folio can free the folio state otherwise. A
> > fully uptodate folio is still returned without the lock.
> > - Keep the folio state of a folio that is not fully uptodate in
> > ->release_folio while the folio has other users. Otherwise a block
> > that a lockless reader, e.g. a dentry lookup, is using can be read
> > from disk again.
> > - Get sbi from the folio in f2fs_finish_read_bio() only after the
> > compressed page check, as nothing pins the inode of a compressed page
> > once its cluster is unlocked.
> > - Rename helpers that now take block indices, e.g.
> > f2fs_submit_page_read() to f2fs_submit_block_read().
> >
> > The read changes also apply to large folios when the block size equals
> > PAGE_SIZE.
> >
> > Signed-off-by: Daeho Jeong <daehojeong@xxxxxxxxxx>
>
> Reviewed-by: Chao Yu <chao@xxxxxxxxxx>
>
> > @@ -5636,6 +5684,16 @@ bool f2fs_release_folio(struct folio *folio, gfp_t wait)
> > if (folio_test_dirty(folio))
> > return false;
> >
> > + /*
> > + * The folio state holds the uptodate state of the blocks of a folio
> > + * that is not fully uptodate. Keep it while the folio has users other
> > + * than the caller, e.g. a dentry lookup that parses a block without
> > + * the folio lock, so that the block is not read again under them.
> > + */
> > + if (f2fs_folio_has_ffs(folio) && !folio_test_uptodate(folio) &&
> > + folio_ref_count(folio) > folio_expected_ref_count(folio) + 1)
> > + return false;
> > +
> > ffs_detach_free(folio);
> > return true;
> > }
>
> BTW, we should notice this side effect, dio may be blocked due to it can not
> invalidate folios which is referenced by other buffered reader.
Not from my end. Can we add a bug_on to detect this first with a reproducer?
>
> Thread A (buffered read) Thread B (DIO write, same range)
> - filemap_read
> - filemap_get_pages
> : folio ref taken (batch), folio
> partially uptodate, ffs present
> - copy_folio_to_iter
> - iomap_dio_complete
> - kiocb_invalidate_post_direct_write
> - invalidate_inode_pages2_range
> - folio_unmap_invalidate
> - f2fs_release_folio
> : !uptodate && ref > expected + 1
> -> false
> : -EBUSY
> - dio_warn_stale_pagecache
> : pr_crit, wb_err = -EIO
>
> Thanks,