cluster/afr: Reset attempted count before attempting blocking lock

Problem:
internal_lock->lk_attempted_count keeps track of the number of blocking
locks attempted. lk_expected_count keeps track of the number locks expected.
Here are the sequence of steps that happen which lead to the illution that
a full file lock is achieved, even without attempting any lock.

2 mounts are doing dd on same file. Both of them witness a brick going
down and coming back up again. Both of the mounts issue self-heal
1) Both mount-1, mount-2 attempt full file locks in self-heal domain.
lets say mount-1 got the lock, mount-2 attempts blocking lock.

2) mount-1 attempts full file lock in data domain. It goes into blocking
mode because some other writes are in progress. Eventually it gets the lock.
But this results in lk_attempted_count to be still as 2 and will not be reset.
It completes syncing the data.

3) mount-1 before unlocking final small range lock attempts full file lock in
data domain to figure out the source/sink. This will be put into blocked mode
again because some other writes are in progress. But this time seeing the
stale value of lk_attempted_count being equal to lk_expected_count, blocking_lock
phase thinks it completed locking without acquiring a single lock :-O.

4) mount-1 reads xattrs without any lock but since it does not modify the xattrs,
no harm is done by this phase. It tries to do unlocks and the unlocks will fail
because the locks are never taken in data domain. mount-1 also unlocks
self-heal domain locks.

Our beloved mount-2 now gets the chance to cause horror :-(.

5) mount-2 gets the full range blocking lock in self-heal domain.
Please note that this sets lk_attempted_count to 2.

6) mount-2 attempts full range lock in data domain, since there are still
writes on going, it switches to blocking mode. But since lk_attempted_count is 2
which is same as lk_expected_count, blocking phase locks thinks it actually got
the full range locks even though not a single lock request went out the wire.

7) mount-2 reads the change-log xattrs, which would give the number of operations
in progress (lets call this 'X'). It does the syncing and at the end of the sync
decrements the changelog by 'X'. But since that 'X' was introduced by 'X' number
of transactions that are in progress, they also decrement the changelog by 'X'.
Effectively for 'X' operations 'X' number of pre-ops are done but 2 times 'X'
number of post-ops are done resulting in -ve changelog numbers.

Fix:
Reset the lk_attempted_count and inode locks array that is used to remember locks
that are granted.

Change-Id: Ic0a79cd16f32392ea7c790511343c73592bbe6bd
BUG: 1002698
Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>
Reviewed-on: http://review.gluster.org/5736
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Anand Avati <avati@redhat.com>

1 files changed, 3 insertions, 0 deletions

diff --git a/xlators/cluster/afr/src/afr-lk-common.c b/xlators/cluster/afr/src/afr-lk-common.c
index 4d19a0f83..e2136089b 100644
--- a/xlators/cluster/afr/src/afr-lk-common.c
+++ b/xlators/cluster/afr/src/afr-lk-common.c
@@ -568,11 +568,14 @@ initialize_inodelk_variables (call_frame_t *frame, xlator_t *this)
         inodelk = afr_get_inodelk (int_lock, int_lock->domain);
 
         inodelk->lock_count    = 0;
+        int_lock->lk_attempted_count = 0;
         int_lock->lock_op_ret   = -1;
         int_lock->lock_op_errno = 0;
 
         memset (inodelk->locked_nodes, 0,
                 sizeof (*inodelk->locked_nodes) * priv->child_count);
+        memset (int_lock->locked_nodes, 0,
+                sizeof (*int_lock->locked_nodes) * priv->child_count);
 
         return 0;
 }