On 2022/08/12 23:41, Christoph Hellwig wrote: > On Fri, Aug 12, 2022 at 02:03:55PM -0700, Bart Van Assche wrote: >> When submitting a REQ_OP_WRITE | REQ_FUA request to a zoned storage >> device, these requests must be passed to the (mq-deadline) I/O scheduler >> to ensure that these happen at the write pointer. > > Yes. > > But maybe I'm stupid, but how is the patch related to fixing that? > blk_mq_plug_issue_direct is called from blk_mq_flush_plug_list for > only the !has_elevator case. How does that change a thing? And writes to zoned drives never get plugged in the first place, scheduler present or not. > > Also please include a description of why these changes are otherwise > good and won't regress other cases. > >> + blk_mq_sched_insert_request(rq, /*at_head=*/false, >> + /*run_queue=*/last, /*async=*/false); > > I find thise comment style very hard to read. Yes, maybe the three > bools here should become flags, but this is even worse than just > passing the arguments. -- Damien Le Moal Western Digital Research
