On Mon, 15 Apr 2024, Bart Van Assche wrote:
On 4/15/24 12:56, Mikulas Patocka wrote:
I am wondering how should QUEUE_FLAG_SAME_COMP work for bio-based devices.
I grepped the kernel for QUEUE_FLAG_SAME_COMP and it is tested in block/blk-mq.c in blk_mq_complete_need_ipi (this code path is taken only for request-based devices) and in block/blk-sysfs.c in queue_rq_affinity_show (this just displays the value in sysfs). There are no other places where QUEUE_FLAG_SAME_COMP is tested, so I don't see what effect is it supposed to have.
I think the answer depends on whether or not the underlying device defines the .submit_bio() callback. From block/blk-core.c:
static void __submit_bio(struct bio *bio) { if (unlikely(!blk_crypto_bio_prep(&bio))) return;
if (!bio->bi_bdev->bd_has_submit_bio) { blk_mq_submit_bio(bio); } else if (likely(bio_queue_enter(bio) == 0)) { struct gendisk *disk = bio->bi_bdev->bd_disk;
disk->fops->submit_bio(bio); blk_queue_exit(disk->queue);} }
In other words, if the .submit_bio() callback is defined, that function is called. If it is not defined, blk_mq_submit_bio() converts the bio into a request. QUEUE_FLAG_SAME_COMP affects the request completion path. On my test setup there are multiple dm instances defined on top of SCSI devices. The SCSI core does not implement the .submit_bio() callback.
Thanks,
Bart.
Yes, setting the flag QUEUE_FLAG_SAME_COMP for bio-based drivers (those that define .submit_bio) has no effect. So, I think that you patch doesn't have any effect too.
Mikulas