On Wed, 2023-12-20 at 17:27 -0800, Song Liu wrote: > > External email : Please do not click links or open attachments until > you have verified the sender or the content. > A recent bug report [1] shows md is handling a flush from bcachefs > as read: > > bch2_journal_write=> > submit_bio=> > ... > md_handle_request => > raid5_make_request => > chunk_aligned_read => > raid5_read_one_chunk => > ... > > It appears md code only checks REQ_PREFLUSH for flush requests, which > doesn't cover all cases. OTOH, op_is_flush() doesn't check > REQ_OP_FLUSH > either. > > Fix this by: > 1) Check REQ_PREFLUSH in op_is_flush(); > 2) Use op_is_flush() in md code. > > Thanks, > Song > > [1] > https://urldefense.com/v3/__https://bugzilla.kernel.org/show_bug.cgi?id=218184__;!!CTRNKA9wMg0ARbw!gQbjtS_f5d3Du2prpIT8zUM4mkZf7qDleyaAuEfG8j5tMrDvw7cfJUB04VWl0uVAL4BJ4YWbVopp$ > REQ_OP_FLUSH is only used by the block layer's flush code, and the filesystem should use REQ_PREFLUSH with an empty write bio. If we want upper layer to be able to directly send REQ_OP_FLUSH bio, then we should retrieve all REQ_PREFLUSH to confirm. At least for now, it seems that REQ_OP_FLUSH without REQ_PREFLUSH in `blk_flush_policy` will directly return 0 and no flush operation will be sent to the driver. > > > Song Liu (2): > block: Check REQ_OP_FLUSH in op_is_flush() > md: Use op_is_flush() to check flush bio > > drivers/md/raid0.c | 2 +- > drivers/md/raid1.c | 2 +- > drivers/md/raid10.c | 2 +- > drivers/md/raid5.c | 2 +- > include/linux/blk_types.h | 3 ++- > 5 files changed, 6 insertions(+), 5 deletions(-) > > -- > 2.34.1
