releases/2.6.32.12/dm-mpath-fix-stall-when-requeueing-io.patch

   1 From knikanth@suse.de  Mon Apr 19 11:23:07 2010
   2 From: Nikanth Karthikesan <knikanth@suse.de>
   3 Date: Fri, 26 Mar 2010 12:03:13 +0530
   4 Subject: dm mpath: fix stall when requeueing io
   5 To: stable@kernel.org
   6 Cc: "Jun'ichi Nomura" <j-nomura@ce.jp.nec.com>, Kiyoshi Ueda <k-ueda@ct.jp.nec.com>, Alasdair G Kergon <agk@redhat.com>
   7 Message-ID: <201003261203.13503.knikanth@suse.de>
   8
   9 From: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
  10
  11 upstream commit 9eef87da2a8ea4920e0d913ff977cac064b68ee0 backported to
  12 2.6.32.10 by Nikanth Karthikesan <knikanth@suse.de>
  13
  14 This patch fixes the problem that system may stall if target's ->map_rq
  15 returns DM_MAPIO_REQUEUE in map_request().
  16 E.g. stall happens on 1 CPU box when a dm-mpath device with queue_if_no_path
  17      bounces between all-paths-down and paths-up on I/O load.
  18
  19 When target's ->map_rq returns DM_MAPIO_REQUEUE, map_request() requeues
  20 the request and returns to dm_request_fn().  Then, dm_request_fn()
  21 doesn't exit the I/O dispatching loop and continues processing
  22 the requeued request again.
  23 This map and requeue loop can be done with interrupt disabled,
  24 so 1 CPU system can be stalled if this situation happens.
  25
  26 For example, commands below can stall my 1 CPU box within 1 minute or so:
  27   # dmsetup table mp
  28   mp: 0 2097152 multipath 1 queue_if_no_path 0 1 1 service-time 0 1 2 8:144 1 1
  29   # while true; do dd if=/dev/mapper/mp of=/dev/null bs=1M count=100; done &
  30   # while true; do \
  31   > dmsetup message mp 0 "fail_path 8:144" \
  32   > dmsetup suspend --noflush mp \
  33   > dmsetup resume mp \
  34   > dmsetup message mp 0 "reinstate_path 8:144" \
  35   > done
  36
  37 To fix the problem above, this patch changes dm_request_fn() to exit
  38 the I/O dispatching loop once if a request is requeued in map_request().
  39
  40 Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
  41 Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
  42 Signed-off-by: Alasdair G Kergon <agk@redhat.com>
  43 Signed-off-by: Nikanth Karthikesan <knikanth@suse.de>
  44 Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
  45
  46 ---
  47  drivers/md/dm.c |   21 +++++++++++++++++----
  48  1 file changed, 17 insertions(+), 4 deletions(-)
  49
  50 --- a/drivers/md/dm.c
  51 +++ b/drivers/md/dm.c
  52 @@ -1487,10 +1487,15 @@ static int dm_prep_fn(struct request_que
  53         return BLKPREP_OK;
  54  }
  55
  56 -static void map_request(struct dm_target *ti, struct request *rq,
  57 -                       struct mapped_device *md)
  58 +/*
  59 + * Returns:
  60 + * 0  : the request has been processed (not requeued)
  61 + * !0 : the request has been requeued
  62 + */
  63 +static int map_request(struct dm_target *ti, struct request *rq,
  64 +                      struct mapped_device *md)
  65  {
  66 -       int r;
  67 +       int r, requeued = 0;
  68         struct request *clone = rq->special;
  69         struct dm_rq_target_io *tio = clone->end_io_data;
  70
  71 @@ -1516,6 +1521,7 @@ static void map_request(struct dm_target
  72         case DM_MAPIO_REQUEUE:
  73                 /* The target wants to requeue the I/O */
  74                 dm_requeue_unmapped_request(clone);
  75 +               requeued = 1;
  76                 break;
  77         default:
  78                 if (r > 0) {
  79 @@ -1527,6 +1533,8 @@ static void map_request(struct dm_target
  80                 dm_kill_unmapped_request(clone, r);
  81                 break;
  82         }
  83 +
  84 +       return requeued;
  85  }
  86
  87  /*
  88 @@ -1568,12 +1576,17 @@ static void dm_request_fn(struct request
  89
  90                 blk_start_request(rq);
  91                 spin_unlock(q->queue_lock);
  92 -               map_request(ti, rq, md);
  93 +               if (map_request(ti, rq, md))
  94 +                       goto requeued;
  95 +
  96                 spin_lock_irq(q->queue_lock);
  97         }
  98
  99         goto out;
 100
 101 +requeued:
 102 +       spin_lock_irq(q->queue_lock);
 103 +
 104  plug_and_out:
 105         if (!elv_queue_empty(q))
 106                 /* Some requests still remain, retry later */