From: Chen Cheng <[email protected]> Hi,
dm only splits discards at a chunk boundary for targets that ask for it with ti->max_discard_granularity, and dm-snapshot never did. So a discard can arrive starting in the middle of a chunk, or covering several chunks at once. zero_exception() zeroes the COW chunk using bio->bi_iter.bi_sector, which remap_exception() has already turned into "COW chunk start + offset inside the chunk". The zeroing window is shifted by that offset, so the head of the chunk we were asked to discard keeps its old data, and the tail spills into the next COW chunk, which belongs to a different exception. Patch 1 sets ti->max_discard_granularity, so dm hands dm-snapshot discards that are split on chunk boundaries. Patch 2 makes io_overlaps_chunk() also require the discard to start on a chunk boundary, and completes the discards that don't cover a whole chunk with bio_endio() instead of sending them to the COW device. Thanks, Chen Cheng Chen Cheng (2): dm snapshot: split discards on chunk boundaries dm snapshot: fix discard writing into the next COW chunk drivers/md/dm-snap.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) -- 2.56.0
