md/raid5: fix newly-broken locking in get_active_stripe.
authorNeilBrown <neilb@suse.de>
Wed, 27 Nov 2013 23:55:27 +0000 (10:55 +1100)
committerNeilBrown <neilb@suse.de>
Thu, 28 Nov 2013 00:00:15 +0000 (11:00 +1100)
commit 566c09c53455d7c4f1 raid5: relieve lock contention in get_active_stripe()

modified the locking in get_active_stripe() reducing the range
protected by the (highly contended) device_lock.
Unfortunately it reduced the range too much opening up some races.

One race can occur if get_priority_stripe runs between the
test on sh->count and device_lock being taken.
This will mean that sh->lru is not empty while get_active_stripe
thinks ->count is zero resulting in a 'BUG' firing.

Another race happens if __release_stripe is called immediately
after sh->count is tested and found to be non-zero.  If STRIPE_HANDLE
is not set, get_active_stripe should increment ->active_stripes
when it increments ->count from 0, but as it didn't think it was 0,
it doesn't.

Extending device_lock to cover the test on sh->count close these
races.

While we are here, fix the two BUG tests:
 -If count is zero, then lru really must not be empty, or we've
  lock the stripe_head somehow - no other tests are relevant.
 -STRIPE_ON_RELEASE_LIST is completely independent of ->lru so
  testing it is pointless.

Reported-and-tested-by: Brassow Jonathan <jbrassow@redhat.com>
Reviewed-by: Shaohua Li <shli@kernel.org>
Fixes: 566c09c53455d7c4f1
Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid5.c

index 676d8b7c51170b9a993fd87b2dc8c7f9c9a99ece..cc055da02e2a300706548041dc4d9cf63957978c 100644 (file)
@@ -678,26 +678,23 @@ get_active_stripe(struct r5conf *conf, sector_t sector,
                        } else
                                init_stripe(sh, sector, previous);
                } else {
+                       spin_lock(&conf->device_lock);
                        if (atomic_read(&sh->count)) {
                                BUG_ON(!list_empty(&sh->lru)
                                    && !test_bit(STRIPE_EXPANDING, &sh->state)
                                    && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state)
-                                   && !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
+                                       );
                        } else {
-                               spin_lock(&conf->device_lock);
                                if (!test_bit(STRIPE_HANDLE, &sh->state))
                                        atomic_inc(&conf->active_stripes);
-                               if (list_empty(&sh->lru) &&
-                                   !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state) &&
-                                   !test_bit(STRIPE_EXPANDING, &sh->state))
-                                       BUG();
+                               BUG_ON(list_empty(&sh->lru));
                                list_del_init(&sh->lru);
                                if (sh->group) {
                                        sh->group->stripes_cnt--;
                                        sh->group = NULL;
                                }
-                               spin_unlock(&conf->device_lock);
                        }
+                       spin_unlock(&conf->device_lock);
                }
        } while (sh == NULL);