@@ -1589,7 +1589,15 @@ static blk_qc_t __split_and_process_bio(struct mapped_device *md,
ci.sector_count = bio_sectors(bio);
while (ci.sector_count && !error) {
error = __split_and_process_non_flush(&ci);
- if (current->bio_list && ci.sector_count && !error) {
+ /*
+ * Depth-first splitting to fix the potential deadlock caused by
+ * reordering of handling bios because of bio_list. We should enable
+ * this feature only when bio_list enabled, but the check of
+ * current->bio_list is not needed any more since all entries into
+ * __split_and_process_bio (dm_wq_work() from thread and submit_bio()
+ * from application) are under bio_list now.
+ */
+ if (ci.sector_count && !error) {
/*
* Remainder must be passed to submit_bio_noacct()
* so that it gets handled *after* bios already submitted
The depth-first splitting is introduced in commit 18a25da84354 ("dm: ensure bio submission follows a depth-first tree walk"), which is used to fix the potential deadlock in case of the misordering handling of bios caused by bio_list. There're two paths submitting split bios, dm_wq_work() from worker thread and submit_bio() from application. Back upon that time, dm_wq_work() thread calls __split_and_process_bio() directly and thus will not trigger this issue since bio_list doesn't exist here. So this issue will only be triggered from application calling submit_bio(), and the fix has to check if current->bio_list is non-NULL to distinguish this case. However since commit 0c2915b8c6db1 ("dm: fix missing imposition of queue_limits from dm_wq_work() thread"), dm_wq_work() thread calls submit_bio_noacct() and thus also uses bio_list. Since then all entries into __split_and_process_bio() are under protection of bio_list, and thus the checking of current->bio_list when determinning if the depth-first principle should be used, seems kind of nonsense. After all the checking always succeeds now. But the check should be there from the perspective of design, to make it a complete fix. So all the details and notes are explained thoroughly in comments. Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com> --- drivers/md/dm.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)