ocfs2: Prevent a livelock in dlmglue
There is possibility of a livelock in __ocfs2_cluster_lock(). If a node were
to get an ast for an upconvert request, followed immediately by a bast,
there is a small window where the fs may downconvert the lock before the
process requesting the upconvert is able to take the lock.
This patch adds a new flag to indicate that the upconvert is still in
progress and that the dc thread should not downconvert it right now.
Wengang Wang <wen.gang.wang@oracle.com> and Joel Becker
<joel.becker@oracle.com> contributed heavily to this patch.
Reported-by: David Teigland <teigland@redhat.com>
Signed-off-by: Sunil Mushran <sunil.mushran@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
index 0cdf630..85d7c49 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -875,6 +875,14 @@
lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
lockres->l_level = lockres->l_requested;
+
+ /*
+ * 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.
+ */
+ lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
+
lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
mlog_exit_void();
@@ -1134,6 +1142,7 @@
mlog_entry_void();
spin_lock_irqsave(&lockres->l_lock, flags);
lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
+ lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
if (convert)
lockres->l_action = OCFS2_AST_INVALID;
else
@@ -1324,13 +1333,13 @@
again:
wait = 0;
+ spin_lock_irqsave(&lockres->l_lock, flags);
+
if (catch_signals && signal_pending(current)) {
ret = -ERESTARTSYS;
- goto out;
+ goto unlock;
}
- spin_lock_irqsave(&lockres->l_lock, flags);
-
mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
"Cluster lock called on freeing lockres %s! flags "
"0x%lx\n", lockres->l_name, lockres->l_flags);
@@ -1347,6 +1356,25 @@
goto unlock;
}
+ if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) {
+ /*
+ * We've upconverted. If the lock now has a level we can
+ * work with, we take it. If, however, the lock is not at the
+ * required level, we go thru the full cycle. One way this could
+ * happen is if a process requesting an upconvert to PR is
+ * closely followed by another requesting upconvert to an EX.
+ * If the process requesting EX lands here, we want it to
+ * continue attempting to upconvert and let the process
+ * requesting PR take the lock.
+ * If multiple processes request upconvert to PR, the first one
+ * here will take the lock. The others will have to go thru the
+ * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
+ * downconvert request.
+ */
+ if (level <= lockres->l_level)
+ goto update_holders;
+ }
+
if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
!ocfs2_may_continue_on_blocked_lock(lockres, level)) {
/* is the lock is currently blocked on behalf of
@@ -1417,11 +1445,14 @@
goto again;
}
+update_holders:
/* Ok, if we get here then we're good to go. */
ocfs2_inc_holders(lockres, level);
ret = 0;
unlock:
+ lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
+
spin_unlock_irqrestore(&lockres->l_lock, flags);
out:
/*
@@ -3402,6 +3433,18 @@
goto leave;
}
+ /*
+ * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
+ * set when the ast is received for an upconvert just before the
+ * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
+ * on the heels of the ast, we want to delay the downconvert just
+ * enough to allow the up requestor to do its task. Because this
+ * lock is in the blocked queue, the lock will be downconverted
+ * as soon as the requestor is done with the lock.
+ */
+ if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING)
+ goto leave_requeue;
+
/* if we're blocking an exclusive and we have *any* holders,
* then requeue. */
if ((lockres->l_blocking == DLM_LOCK_EX)