Commit 312045ee authored Dec 21, 2015 by NeilBrown

md: remove check for MD_RECOVERY_NEEDED in action_store.



md currently doesn't allow a 'sync_action' such as 'reshape' to be set
while MD_RECOVERY_NEEDED is set.

This s a problem, particularly since commit 738a2738 as that can
cause ->check_shape to call mddev_resume() which sets
MD_RECOVERY_NEEDED.  So by the time we come to start 'reshape' it is
very likely that MD_RECOVERY_NEEDED is still set.

Testing for this flag is not really needed and is in any case very
racy as it can be set at any moment - asynchronously.  Any race
between setting a sync_action and setting MD_RECOVERY_NEEDED must
already be handled properly in some locked code, probably
md_check_recovery(), so remove the test here.

The test on MD_RECOVERY_RUNNING is also racy in the 'reshape' case
so we should test it again after getting mddev_lock().

As this fixes a race and a regression which can cause 'reshape' to
fail, it is suitable for -stable kernels since 4.1

Reported-by: Xiao Ni <xni@redhat.com>
Fixes: 738a2738 ("md/raid5: fix allocation of 'scribble' array.")
Cc: stable@vger.kernel.org (v4.1+)
Signed-off-by: NeilBrown <neilb@suse.com>

parent 9f7e4327

drivers/md/md.c

+7 −4

Original line number	Diff line number	Diff line
		@@ -4326,8 +4326,7 @@ action_store(struct mddev mddev, const char page, size_t len)
		}
		mddev_unlock(mddev);
		}
		} else if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) \|\|
		test_bit(MD_RECOVERY_NEEDED, &mddev->recovery))
		} else if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
		return -EBUSY;
		else if (cmd_match(page, "resync"))
		clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
		@@ -4340,8 +4339,12 @@ action_store(struct mddev mddev, const char page, size_t len)
		return -EINVAL;
		err = mddev_lock(mddev);
		if (!err) {
		if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
		err = -EBUSY;
		else {
		clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
		err = mddev->pers->start_reshape(mddev);
		}
		mddev_unlock(mddev);
		}
		if (err)

Admin message