On Wed, May 27 2015 at 12:14P -0400, Mike Snitzer <snitzer@xxxxxxxxxx> wrote: > On Wed, May 27 2015 at 11:33am -0400, > Bart Van Assche <bart.vanassche@xxxxxxxxxxx> wrote: > > > On 05/27/15 17:29, Bart Van Assche wrote: > > >On 05/27/15 14:57, Mike Snitzer wrote: > > >>Looks like Junichi likely fixed this issue you reported, please try this > > >>patch: https://patchwork.kernel.org/patch/6487321/ > > > > > >Hello Mike, > > > > > >On a setup on which an I/O verification test passes with > > >blk-mq/scsi-mq/dm-mq disabled, this is what fio reports after a few > > >minutes with scsi-mq and dm-mq enabled: > > > > > >test: Laying out IO file(s) (1 file(s) / 10MB) > > >fio: io_u error on file /mnt/test.0.0: Input/output error: write > > >offset=8327168, buflen=4096 > > >fio: io_u error on file /mnt/test.0.0: Input/output error: write > > >offset=9007104, buflen=4096 > > >fio: pid=4568, err=5/file:io_u.c:1564, func=io_u error, > > >error=Input/output error > > I'll look closer at this.. so NULL pointer is fixed but this test hits > IO errors. Further code inspection revealed an issue with dm-mq enabled but scsi-mq disabled (when requeuing the original request after clone_rq() failure DM core wasn't unwinding the dm_start_request() accounting). The following patch will fix this issue. I've also switched the dm-mq on scsi-mq case to return BLK_MQ_RQ_QUEUE_BUSY directly (like hch suggested last week). I have no idea if this would actually fix your case (would be surprising but worth a shot I suppose). Anyway, feel free to try this patch: diff --git a/drivers/md/dm.c b/drivers/md/dm.c index 85966ee..02e2d1f 100644 --- a/drivers/md/dm.c +++ b/drivers/md/dm.c @@ -1115,23 +1115,37 @@ static void old_requeue_request(struct request *rq) spin_unlock_irqrestore(q->queue_lock, flags); } -static void dm_requeue_original_request(struct mapped_device *md, - struct request *rq) +static void __dm_requeue_original_request(struct mapped_device *md, + struct request *rq, bool in_blk_mq_queue_rq) { int rw = rq_data_dir(rq); dm_unprep_request(rq); - if (!rq->q->mq_ops) - old_requeue_request(rq); - else { - blk_mq_requeue_request(rq); - blk_mq_kick_requeue_list(rq->q); + if (!in_blk_mq_queue_rq) { + if (!rq->q->mq_ops) + old_requeue_request(rq); + else { + blk_mq_requeue_request(rq); + blk_mq_kick_requeue_list(rq->q); + } } rq_completed(md, rw, false); } +static void dm_requeue_original_request(struct mapped_device *md, + struct request *rq) +{ + return __dm_requeue_original_request(md, rq, false); +} + +static void dm_unprep_before_requeuing_original_request(struct mapped_device *md, + struct request *rq) +{ + return __dm_requeue_original_request(md, rq, true); +} + static void old_stop_queue(struct request_queue *q) { unsigned long flags; @@ -2679,15 +2693,18 @@ static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx, /* clone request is allocated at the end of the pdu */ tio->clone = (void *)blk_mq_rq_to_pdu(rq) + sizeof(struct dm_rq_target_io); if (!clone_rq(rq, md, tio, GFP_ATOMIC)) - return BLK_MQ_RQ_QUEUE_BUSY; + goto out_requeue; queue_kthread_work(&md->kworker, &tio->work); } else { /* Direct call is fine since .queue_rq allows allocations */ if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) - dm_requeue_original_request(md, rq); + goto out_requeue; } return BLK_MQ_RQ_QUEUE_OK; +out_requeue: + dm_unprep_before_requeuing_original_request(md, rq); + return BLK_MQ_RQ_QUEUE_BUSY; } static struct blk_mq_ops dm_mq_ops = { -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
