On 2026/6/12 11:32, Yifan Zhao wrote:
EROFS chunked files may place adjacent logical chunks on different
devices. The physical block numbers are per-device, so two neighboring
chunks can still look sector-contiguous to the generic iomap buffered
read code.
For example:
logical file offset
0 8K 16K 24K
+----------+----------+----------+
| chunk 0 | chunk 1 | chunk 2 |
+----------+----------+----------+
| | |
v v v
dev 1 dev 3 dev 3
sector 8 sector 24 sector 40
The transition from chunk 0 to chunk 1 crosses a device boundary, but
iomap can still treat sector-contiguous bios as mergeable without
checking whether they belong to the same device.
The pending bio, however, is still bound to the previous block device:
bio->bi_bdev = dev 1
file 0..8K -> dev 1, sector 8
file 8..16K -> dev 3, sector 24
(must not be added here)
If the second range is added to the same bio, it will be submitted to
dev 1 and read from the wrong backing device, which is easy to trigger
with readahead.
This only affects paths using erofs_aops, where buffered reads go
through iomap bio helpers.
Fix by install EROFS-specific iomap read ops and split the pending
buffered read bio whenever the next mapped range belongs to a different
bdev. After the split, fall back to the generic iomap bio read helper
for the normal sector-based merge checks.
Reported-by: Kelu Ye <[email protected]>
Assisted-by: Codex:GPT-5.5
Signed-off-by: Yifan Zhao <[email protected]>
I think it's an iomap bug instead, see:
iomap_bio_read_folio_range(), we should fix iomap instead.
Thanks,
Gao Xiang
