ocfs2: avoid blocking in ocfs2_mark_lockres_freeing() in downconvert thread

If we are dropping last inode reference from downconvert thread, we will
end up calling ocfs2_mark_lockres_freeing() which can block if the lock
we are freeing is queued thus creating an A-A deadlock.  Luckily, since
we are the downconvert thread, we can immediately dequeue the lock and
thus avoid waiting in this case.

Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Reviewed-by: Srinivas Eeda <srinivas.eeda@oracle.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>

3 files changed, 47 insertions, 7 deletions

diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
index 19986959d149..6bd690b5a061 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -3144,22 +3144,60 @@ out:
         return 0;
 }
 
+static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
+                                       struct ocfs2_lock_res *lockres);
+
 /* Mark the lockres as being dropped. It will no longer be
  * queued if blocking, but we still may have to wait on it
  * being dequeued from the downconvert thread before we can consider
  * it safe to drop.
  *
  * You can *not* attempt to call cluster_lock on this lockres anymore. */
-void ocfs2_mark_lockres_freeing(struct ocfs2_lock_res *lockres)
+void ocfs2_mark_lockres_freeing(struct ocfs2_super *osb,
+                                struct ocfs2_lock_res *lockres)
 {
         int status;
         struct ocfs2_mask_waiter mw;
-        unsigned long flags;
+        unsigned long flags, flags2;
 
         ocfs2_init_mask_waiter(&mw);
 
         spin_lock_irqsave(&lockres->l_lock, flags);
         lockres->l_flags |= OCFS2_LOCK_FREEING;
+        if (lockres->l_flags & OCFS2_LOCK_QUEUED && current == osb->dc_task) {
+                /*
+                 * We know the downconvert is queued but not in progress
+                 * because we are the downconvert thread and processing
+                 * different lock. So we can just remove the lock from the
+                 * queue. This is not only an optimization but also a way
+                 * to avoid the following deadlock:
+                 *   ocfs2_dentry_post_unlock()
+                 *     ocfs2_dentry_lock_put()
+                 *       ocfs2_drop_dentry_lock()
+                 *         iput()
+                 *           ocfs2_evict_inode()
+                 *             ocfs2_clear_inode()
+                 *               ocfs2_mark_lockres_freeing()
+                 *                 ... blocks waiting for OCFS2_LOCK_QUEUED
+                 *                 since we are the downconvert thread which
+                 *                 should clear the flag.
+                 */
+                spin_unlock_irqrestore(&lockres->l_lock, flags);
+                spin_lock_irqsave(&osb->dc_task_lock, flags2);
+                list_del_init(&lockres->l_blocked_list);
+                osb->blocked_lock_count--;
+                spin_unlock_irqrestore(&osb->dc_task_lock, flags2);
+                /*
+                 * Warn if we recurse into another post_unlock call.  Strictly
+                 * speaking it isn't a problem but we need to be careful if
+                 * that happens (stack overflow, deadlocks, ...) so warn if
+                 * ocfs2 grows a path for which this can happen.
+                 */
+                WARN_ON_ONCE(lockres->l_ops->post_unlock);
+                /* Since the lock is freeing we don't do much in the fn below */
+                ocfs2_process_blocked_lock(osb, lockres);
+                return;
+        }
         while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
                 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
                 spin_unlock_irqrestore(&lockres->l_lock, flags);
@@ -3180,7 +3218,7 @@ void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
 {
         int ret;
 
-        ocfs2_mark_lockres_freeing(lockres);
+        ocfs2_mark_lockres_freeing(osb, lockres);
         ret = ocfs2_drop_lock(osb, lockres);
         if (ret)
                 mlog_errno(ret);
diff --git a/fs/ocfs2/dlmglue.h b/fs/ocfs2/dlmglue.h
index 1d596d8c4a4a..d293a22c32c5 100644
--- a/fs/ocfs2/dlmglue.h
+++ b/fs/ocfs2/dlmglue.h
@@ -157,7 +157,8 @@ int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex);
 void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex);
 
 
-void ocfs2_mark_lockres_freeing(struct ocfs2_lock_res *lockres);
+void ocfs2_mark_lockres_freeing(struct ocfs2_super *osb,
+                                struct ocfs2_lock_res *lockres);
 void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
                                struct ocfs2_lock_res *lockres);
 
diff --git a/fs/ocfs2/inode.c b/fs/ocfs2/inode.c
index 809b5d57a6b8..d437f3ba90b0 100644
--- a/fs/ocfs2/inode.c
+++ b/fs/ocfs2/inode.c
@@ -1080,6 +1080,7 @@ static void ocfs2_clear_inode(struct inode *inode)
 {
         int status;
         struct ocfs2_inode_info *oi = OCFS2_I(inode);
+        struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 
         clear_inode(inode);
         trace_ocfs2_clear_inode((unsigned long long)oi->ip_blkno,
@@ -1096,9 +1097,9 @@ static void ocfs2_clear_inode(struct inode *inode)
 
         /* Do these before all the other work so that we don't bounce
          * the downconvert thread while waiting to destroy the locks. */
-        ocfs2_mark_lockres_freeing(&oi->ip_rw_lockres);
-        ocfs2_mark_lockres_freeing(&oi->ip_inode_lockres);
-        ocfs2_mark_lockres_freeing(&oi->ip_open_lockres);
+        ocfs2_mark_lockres_freeing(osb, &oi->ip_rw_lockres);
+        ocfs2_mark_lockres_freeing(osb, &oi->ip_inode_lockres);
+        ocfs2_mark_lockres_freeing(osb, &oi->ip_open_lockres);
 
         ocfs2_resv_discard(&OCFS2_SB(inode->i_sb)->osb_la_resmap,
                            &oi->ip_la_data_resv);