On Wed, Feb 23, 2022 at 07:42:26AM +0100, Christoph Hellwig wrote:

On Wed, Feb 23, 2022 at 10:08:20AM +0800, Ming Lei wrote:

quoted

quoted

-	return (rq->rq_flags & RQF_IO_STAT) && rq->q->disk;
+	return (rq->rq_flags & RQF_IO_STAT) && !blk_rq_is_passthrough(rq);

I guess this way may cause regression for workloads with lots of userspace IO
from user viewpoint?

I'd say it fixes it as the accounting right now is completely bogus.

There are small amount of in-kernel passthrough requests(admin, or driver
private) which shouldn't be accounted, but passthrough RW IO requests from
userspace can be lots of, and user may rely on diskstat to account them.


Thanks,
Ming