On Fri, 2017-09-22 at 09:35 +0800, Ming Lei wrote: > + /* > + * blk-mq's SCHED_RESTART can cover this requeue, so > + * we needn't to deal with it by DELAY_REQUEUE. More > + * importantly, we have to return DM_MAPIO_REQUEUE > + * so that blk-mq can get the queue busy feedback, > + * otherwise I/O merge can be hurt. > + */ > + if (q->mq_ops) > + return DM_MAPIO_REQUEUE; > + else > + return DM_MAPIO_DELAY_REQUEUE; > } This patch is inferior to what I posted because this patch does not avoid the delay if multiple LUNs are associated with the same SCSI host. Consider e.g. the following configuration: * A single SCSI host with two SCSI LUNs associated to that host, e.g. /dev/sda and /dev/sdb. * A dm-mpath instance has been created on top of /dev/sda. If all tags are in use by requests queued to /dev/sdb, no dm requests are in progress and a request is submitted against the dm-mpath device then the blk_get_request(q, GFP_ATOMIC) call will fail. The request will be requeued and the queue will be rerun after a delay. My patch does not introduce a delay in this case. Bart.
