On Mon, Jun 30, 2025 at 09:58:52PM +0800, Zhang Yi wrote:
On 2025/6/30 21:13, Sasha Levin wrote:When large folios are enabled, the blocks-per-folio calculation in ext4_da_writepages_trans_blocks() can overflow the journal transaction limits, causing the writeback path to fail with errors like: JBD2: kworker/u8:0 wants too many credits credits:416 rsv_credits:21 max:334 This occurs with small block sizes (1KB) and large folios (32MB), where the calculation results in 32768 blocks per folio. The transaction credit calculation then requests more credits than the journal can handle, leading to the following warning and writeback failure: WARNING: CPU: 1 PID: 43 at fs/jbd2/transaction.c:334 start_this_handle+0x4c0/0x4e0 EXT4-fs (loop0): ext4_do_writepages: jbd2_start: 9223372036854775807 pages, ino 14; err -28 Call trace leading to the issue: ext4_do_writepages() ext4_da_writepages_trans_blocks() bpp = ext4_journal_blocks_per_folio() // Returns 32768 for 32MB folio with 1KB blocks ext4_meta_trans_blocks(inode, MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp) // With bpp=32768, lblocks=34815, pextents=32768 // Returns credits=415, but with overhead becomes 416 > max 334 ext4_journal_start_with_reserve() jbd2_journal_start_reserved() start_this_handle() // Fails with warning when credits:416 > max:334 The issue was introduced by commit d6bf294773a47 ("ext4/jbd2: convert jbd2_journal_blocks_per_page() to support large folio"), which added support for large folios but didn't account for the journal credit limits. Fix this by capping the blocks-per-folio value at 8192 in the writeback path. This is the value we'd get with 32MB folios and 4KB blocks, or 8MB folios with 1KB blocks, which is reasonable and safe for typical journal configurations. Fixes: d6bf294773a4 ("ext4/jbd2: convert jbd2_journal_blocks_per_page() to support large folio") Cc: stable@xxxxxxxxxxxxxxx Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>Hi, Sasha! Thank you for the fix. However, simply limiting the credits is not enough, as this may result in a scenario where there are not enough credits available to map a large, non-contiguous folio. I've been working on this issue[1] and I'll release v3 tomorrow if my tests looks fine. [1] https://lore.kernel.org/linux-ext4/20250611111625.1668035-1-yi.zhang@xxxxxxxxxxxxxxx/
Ah perfect, I haven't seen your work, thank you for that. Please ignore my patch. -- Thanks, Sasha
