ocfs2: fix deadlock when two nodes are converting same lock from PR to EX and idletimeout closes conn
Orabug: 18639535
Two node cluster and both nodes hold a lock at PR level and both want to
convert to EX at the same time. Master node 1 has sent BAST and then
closes the connection due to idletime out. Node 0 receives BAST, sends
unlock req with cancel flag but gets error -ENOTCONN. The problem is
this error is ignored in dlm_send_remote_unlock_request() on the
**incorrect** assumption that the master is dead. See NOTE in comment
why it returns DLM_NORMAL. Upon getting DLM_NORMAL, node 0 proceeds to
sends convert (without cancel flg) which fails with -ENOTCONN. waits 5
sec and resends.
This time gets DLM_IVLOCKID from the master since lock not found in
grant, it had been moved to converting queue in response to conv PR->EX
req. No way out.
Node 1 (master) Node 0
============== ======
lock mode PR PR
convert PR -> EX
mv grant -> convert and que BAST
...
<-------- convert PR -> EX
convert que looks like this: ((node 1, PR -> EX) (node 0, PR -> EX))
...
BAST (want PR -> NL)
------------------>
...
idle timout, conn closed
...
In response to BAST,
sends unlock with cancel convert flag
gets -ENOTCONN. Ignores and
sends remote convert request
gets -ENOTCONN, waits 5 Sec, retries
...
reconnects
<----------------- convert req goes through on next try
does not find lock on grant que
status DLM_IVLOCKID
------------------>
...
No way out. Fix is to keep retrying unlock with cancel flag until it
succeeds or the master dies.
Signed-off-by: Tariq Saeed <tariq.x.saeed@oracle.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/fs/ocfs2/dlm/dlmunlock.c b/fs/ocfs2/dlm/dlmunlock.c
index 5698b52..2e3c9db 100644
--- a/fs/ocfs2/dlm/dlmunlock.c
+++ b/fs/ocfs2/dlm/dlmunlock.c
@@ -191,7 +191,9 @@
DLM_UNLOCK_CLEAR_CONVERT_TYPE);
} else if (status == DLM_RECOVERING ||
status == DLM_MIGRATING ||
- status == DLM_FORWARD) {
+ status == DLM_FORWARD ||
+ status == DLM_NOLOCKMGR
+ ) {
/* must clear the actions because this unlock
* is about to be retried. cannot free or do
* any list manipulation. */
@@ -200,7 +202,8 @@
res->lockname.name,
status==DLM_RECOVERING?"recovering":
(status==DLM_MIGRATING?"migrating":
- "forward"));
+ (status == DLM_FORWARD ? "forward" :
+ "nolockmanager")));
actions = 0;
}
if (flags & LKM_CANCEL)
@@ -364,7 +367,10 @@
* updated state to the recovery master. this thread
* just needs to finish out the operation and call
* the unlockast. */
- ret = DLM_NORMAL;
+ if (dlm_is_node_dead(dlm, owner))
+ ret = DLM_NORMAL;
+ else
+ ret = DLM_NOLOCKMGR;
} else {
/* something bad. this will BUG in ocfs2 */
ret = dlm_err_to_dlm_status(tmpret);
@@ -638,7 +644,9 @@
if (status == DLM_RECOVERING ||
status == DLM_MIGRATING ||
- status == DLM_FORWARD) {
+ status == DLM_FORWARD ||
+ status == DLM_NOLOCKMGR) {
+
/* We want to go away for a tiny bit to allow recovery
* / migration to complete on this resource. I don't
* know of any wait queue we could sleep on as this
@@ -650,7 +658,7 @@
msleep(50);
mlog(0, "retrying unlock due to pending recovery/"
- "migration/in-progress\n");
+ "migration/in-progress/reconnect\n");
goto retry;
}