lustre/kernel_patches/patches/dm-fix-handle-BLK_MQ_RQ_QUEUE_DEV_BUSY-rhel7.6.patch

   1 It looks like what's happening is when dm_dispatch_clone_request
   2 dispatches the "clone" I/O request to the underlying (real) device
   3 from the multipath device, the scsi driver can (often under load)
   4 return BLK_MQ_RQ_QUEUE_DEV_BUSY. dm_dispatch_clone_request doesn't
   5 have that as an exception the way it does BLK_MQ_RQ_QUEUE_BUSY and
   6 so it calls dm_complete_request which propagates
   7 the BLK_MQ_RQ_QUEUE_DEV_BUSY error code up the stack resulting
   8 in multipath_end_io calling fail_path and failing the path because
   9 there is an error value set.
  10
  11 diff --git a/drivers/md/dm-rq.c b/drivers/md/dm-rq.c
  12 index 02da1e65..e4f58472 100644
  13 --- a/drivers/md/dm-rq.c
  14 +++ b/drivers/md/dm-rq.c
  15 @@ -477,7 +477,8 @@ static int dm_dispatch_clone_request(struct request *clone, struct request *rq)
  16
  17         clone->start_time = jiffies;
  18         r = blk_insert_cloned_request(clone->q, clone);
  19 -       if (r != BLK_MQ_RQ_QUEUE_OK && r != BLK_MQ_RQ_QUEUE_BUSY)
  20 +       if (r != BLK_MQ_RQ_QUEUE_OK && r != BLK_MQ_RQ_QUEUE_BUSY &&
  21 +           r != BLK_MQ_RQ_QUEUE_DEV_BUSY)
  22                 /* must complete clone in terms of original request */
  23                 dm_complete_request(rq, r);
  24         return r;
  25 @@ -661,7 +662,7 @@ check_again:
  26                 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
  27                                      blk_rq_pos(rq));
  28                 ret = dm_dispatch_clone_request(clone, rq);
  29 -               if (ret == BLK_MQ_RQ_QUEUE_BUSY) {
  30 +               if (ret == BLK_MQ_RQ_QUEUE_BUSY || ret == BLK_MQ_RQ_QUEUE_DEV_BUSY) {
  31                         blk_rq_unprep_clone(clone);
  32                         tio->ti->type->release_clone_rq(clone);
  33                         tio->clone = NULL;