ocfs2: do not set OCFS2_LOCK_UPCONVERT_FINISHING if nonblocking lock can not be granted at once
ocfs2_readpages() use nonblocking flag to avoid page lock inversion. It
will trigger cluster hang because that flag OCFS2_LOCK_UPCONVERT_FINISHING
is not cleared if nonblocking lock cannot be granted at once. The flag
would prevent dc thread from downconverting. So other nodes cannot
acheive this lockres for ever.
So we should not set OCFS2_LOCK_UPCONVERT_FINISHING when receiving ast if
nonblocking lock had already returned.
Signed-off-by: joyce.xue <xuejiufei@huawei.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Mark Fasheh <mfasheh@suse.com>
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/dlmglue.c b/fs/ocfs2/dlmglue.c
index 21262f2..953f4eb 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -861,8 +861,13 @@
* We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
* the OCFS2_LOCK_BUSY flag to prevent the dc thread from
* downconverting the lock before the upconvert has fully completed.
+ * Do not prevent the dc thread from downconverting if NONBLOCK lock
+ * had already returned.
*/
- lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
+ if (!(lockres->l_flags & OCFS2_LOCK_NONBLOCK_FINISHED))
+ lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
+ else
+ lockres_clear_flags(lockres, OCFS2_LOCK_NONBLOCK_FINISHED);
lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
}
@@ -1324,13 +1329,12 @@
/* returns 0 if the mw that was removed was already satisfied, -EBUSY
* if the mask still hadn't reached its goal */
-static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
+static int __lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
struct ocfs2_mask_waiter *mw)
{
- unsigned long flags;
int ret = 0;
- spin_lock_irqsave(&lockres->l_lock, flags);
+ assert_spin_locked(&lockres->l_lock);
if (!list_empty(&mw->mw_item)) {
if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
ret = -EBUSY;
@@ -1338,6 +1342,18 @@
list_del_init(&mw->mw_item);
init_completion(&mw->mw_complete);
}
+
+ return ret;
+}
+
+static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
+ struct ocfs2_mask_waiter *mw)
+{
+ unsigned long flags;
+ int ret = 0;
+
+ spin_lock_irqsave(&lockres->l_lock, flags);
+ ret = __lockres_remove_mask_waiter(lockres, mw);
spin_unlock_irqrestore(&lockres->l_lock, flags);
return ret;
@@ -1373,6 +1389,7 @@
unsigned long flags;
unsigned int gen;
int noqueue_attempted = 0;
+ int dlm_locked = 0;
ocfs2_init_mask_waiter(&mw);
@@ -1481,6 +1498,7 @@
ocfs2_recover_from_dlm_error(lockres, 1);
goto out;
}
+ dlm_locked = 1;
mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
lockres->l_name);
@@ -1514,10 +1532,17 @@
if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
wait = 0;
- if (lockres_remove_mask_waiter(lockres, &mw))
+ spin_lock_irqsave(&lockres->l_lock, flags);
+ if (__lockres_remove_mask_waiter(lockres, &mw)) {
+ if (dlm_locked)
+ lockres_or_flags(lockres,
+ OCFS2_LOCK_NONBLOCK_FINISHED);
+ spin_unlock_irqrestore(&lockres->l_lock, flags);
ret = -EAGAIN;
- else
+ } else {
+ spin_unlock_irqrestore(&lockres->l_lock, flags);
goto again;
+ }
}
if (wait) {
ret = ocfs2_wait_for_mask(&mw);