md/raid5: ensure sync and DISCARD don't happen at the same time.
authorNeilBrown <neilb@suse.de>
Tue, 12 Mar 2013 01:18:06 +0000 (12:18 +1100)
committerNeilBrown <neilb@suse.de>
Wed, 20 Mar 2013 02:20:59 +0000 (13:20 +1100)
A number of problems can occur due to races between
resync/recovery and discard.

- if sync_request calls handle_stripe() while a discard is
  happening on the stripe, it might call handle_stripe_clean_event
  before all of the individual discard requests have completed
  (so some devices are still locked, but not all).
  Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65
     md/raid5: Make sure we clear R5_Discard when discard is finished.
  this will cause R5_Discard to be cleared for the parity device,
  so handle_stripe_clean_event() will not be called when the other
  devices do become unlocked, so their ->written will not be cleared.
  This ultimately leads to a WARN_ON in init_stripe and a lock-up.

- If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward
  time for resync, it can lead to s->uptodate being less than disks
  in handle_parity_checks5(), which triggers a BUG (because it is
  one).

So:
 - keep R5_Discard on the parity device until all other devices have
   completed their discard request
 - make sure we don't try to have a 'discard' and a 'sync' action at
   the same time.
   This involves a new stripe flag to we know when a 'discard' is
   happening, and the use of R5_Overlap on the parity disk so when a
   discard is wanted while a sync is active, so we know to wake up
   the discard at the appropriate time.

Discard support for RAID5 was added in 3.7, so this is suitable for
any -stable kernel since 3.7.

Cc: stable@vger.kernel.org (v3.7+)
Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com>
Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid5.c
drivers/md/raid5.h

index 52ba88a..42a8997 100644 (file)
@@ -2576,6 +2576,8 @@ handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
        int i;
 
        clear_bit(STRIPE_SYNCING, &sh->state);
+       if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
+               wake_up(&conf->wait_for_overlap);
        s->syncing = 0;
        s->replacing = 0;
        /* There is nothing more to do for sync/check/repair.
@@ -2749,6 +2751,7 @@ static void handle_stripe_clean_event(struct r5conf *conf,
 {
        int i;
        struct r5dev *dev;
+       int discard_pending = 0;
 
        for (i = disks; i--; )
                if (sh->dev[i].written) {
@@ -2777,9 +2780,23 @@ static void handle_stripe_clean_event(struct r5conf *conf,
                                                STRIPE_SECTORS,
                                         !test_bit(STRIPE_DEGRADED, &sh->state),
                                                0);
-                       }
-               } else if (test_bit(R5_Discard, &sh->dev[i].flags))
-                       clear_bit(R5_Discard, &sh->dev[i].flags);
+                       } else if (test_bit(R5_Discard, &dev->flags))
+                               discard_pending = 1;
+               }
+       if (!discard_pending &&
+           test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
+               clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
+               clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
+               if (sh->qd_idx >= 0) {
+                       clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
+                       clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
+               }
+               /* now that discard is done we can proceed with any sync */
+               clear_bit(STRIPE_DISCARD, &sh->state);
+               if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
+                       set_bit(STRIPE_HANDLE, &sh->state);
+
+       }
 
        if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
                if (atomic_dec_and_test(&conf->pending_full_writes))
@@ -3431,9 +3448,15 @@ static void handle_stripe(struct stripe_head *sh)
                return;
        }
 
-       if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
-               set_bit(STRIPE_SYNCING, &sh->state);
-               clear_bit(STRIPE_INSYNC, &sh->state);
+       if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
+               spin_lock(&sh->stripe_lock);
+               /* Cannot process 'sync' concurrently with 'discard' */
+               if (!test_bit(STRIPE_DISCARD, &sh->state) &&
+                   test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
+                       set_bit(STRIPE_SYNCING, &sh->state);
+                       clear_bit(STRIPE_INSYNC, &sh->state);
+               }
+               spin_unlock(&sh->stripe_lock);
        }
        clear_bit(STRIPE_DELAYED, &sh->state);
 
@@ -3593,6 +3616,8 @@ static void handle_stripe(struct stripe_head *sh)
            test_bit(STRIPE_INSYNC, &sh->state)) {
                md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
                clear_bit(STRIPE_SYNCING, &sh->state);
+               if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
+                       wake_up(&conf->wait_for_overlap);
        }
 
        /* If the failed drives are just a ReadError, then we might need
@@ -4159,6 +4184,13 @@ static void make_discard_request(struct mddev *mddev, struct bio *bi)
                sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
                prepare_to_wait(&conf->wait_for_overlap, &w,
                                TASK_UNINTERRUPTIBLE);
+               set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
+               if (test_bit(STRIPE_SYNCING, &sh->state)) {
+                       release_stripe(sh);
+                       schedule();
+                       goto again;
+               }
+               clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
                spin_lock_irq(&sh->stripe_lock);
                for (d = 0; d < conf->raid_disks; d++) {
                        if (d == sh->pd_idx || d == sh->qd_idx)
@@ -4171,6 +4203,7 @@ static void make_discard_request(struct mddev *mddev, struct bio *bi)
                                goto again;
                        }
                }
+               set_bit(STRIPE_DISCARD, &sh->state);
                finish_wait(&conf->wait_for_overlap, &w);
                for (d = 0; d < conf->raid_disks; d++) {
                        if (d == sh->pd_idx || d == sh->qd_idx)
index 18b2c4a..050a334 100644 (file)
@@ -323,6 +323,7 @@ enum {
        STRIPE_COMPUTE_RUN,
        STRIPE_OPS_REQ_PENDING,
        STRIPE_ON_UNPLUG_LIST,
+       STRIPE_DISCARD,
 };
 
 /*