blk-mq: avoid to iterate over stale request

blk-mq can't run allocating driver tag and updating ->rqs[tag]
atomically, meantime blk-mq doesn't clear ->rqs[tag] after the driver
tag is released.

So there is chance to iterating over one stale request just after the
tag is allocated and before updating ->rqs[tag].

scsi_host_busy_iter() calls scsi_host_check_in_flight() to count scsi
in-flight requests after scsi host is blocked, so no new scsi command can
be marked as SCMD_STATE_INFLIGHT. However, driver tag allocation still can
be run by blk-mq core. One request is marked as SCMD_STATE_INFLIGHT,
but this request may have been kept in another slot of ->rqs[], meantime
the slot can be allocated out but ->rqs[] isn't updated yet. Then this
in-flight request is counted twice as SCMD_STATE_INFLIGHT. This way causes
trouble in handling scsi error.

Fixes the issue by not iterating over stale request.

Cc: linux-scsi@vger.kernel.org
Cc: "Martin K. Petersen" <martin.petersen@oracle.com>
Reported-by: luojiaxing <luojiaxing@huawei.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20210906065003.439019-1-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 86f87346232a6b9a7f704dc4e0d41378ee7bd056..ff5caeb825429d91c3f6bf58af68d91fc50b222d 100644 (file)
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -208,7 +208,7 @@ static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
 
        spin_lock_irqsave(&tags->lock, flags);
        rq = tags->rqs[bitnr];
-       if (!rq || !refcount_inc_not_zero(&rq->ref))
+       if (!rq || rq->tag != bitnr || !refcount_inc_not_zero(&rq->ref))
                rq = NULL;
        spin_unlock_irqrestore(&tags->lock, flags);
        return rq;