CIFS: Fix VFS lock usage for oplocked files
Pavel Shilovsky [Tue, 27 Mar 2012 11:36:15 +0000 (15:36 +0400)]
We can deadlock if we have a write oplock and two processes
use the same file handle. In this case the first process can't
unlock its lock if another process blocked on the lock in the
same time.

Fix this by removing lock_mutex protection from waiting on a
blocked lock and protect only posix_lock_file call.

Cc: stable@kernel.org
Signed-off-by: Pavel Shilovsky <piastry@etersoft.ru>
Acked-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: Steve French <sfrench@us.ibm.com>

fs/cifs/file.c

index 460d87b..0a11dbb 100644 (file)
@@ -671,6 +671,21 @@ cifs_del_lock_waiters(struct cifsLockInfo *lock)
        }
 }
 
+/*
+ * Copied from fs/locks.c with small changes.
+ * Remove waiter from blocker's block list.
+ * When blocker ends up pointing to itself then the list is empty.
+ */
+static void
+cifs_locks_delete_block(struct file_lock *waiter)
+{
+       lock_flocks();
+       list_del_init(&waiter->fl_block);
+       list_del_init(&waiter->fl_link);
+       waiter->fl_next = NULL;
+       unlock_flocks();
+}
+
 static bool
 __cifs_find_lock_conflict(struct cifsInodeInfo *cinode, __u64 offset,
                        __u64 length, __u8 type, __u16 netfid,
@@ -820,6 +835,39 @@ cifs_posix_lock_test(struct file *file, struct file_lock *flock)
        return rc;
 }
 
+/* Called with locked lock_mutex, return with unlocked. */
+static int
+cifs_posix_lock_file_wait_locked(struct file *file, struct file_lock *flock)
+{
+       struct cifsInodeInfo *cinode = CIFS_I(file->f_path.dentry->d_inode);
+       int rc;
+
+       while (true) {
+               rc = posix_lock_file(file, flock, NULL);
+               mutex_unlock(&cinode->lock_mutex);
+               if (rc != FILE_LOCK_DEFERRED)
+                       break;
+               rc = wait_event_interruptible(flock->fl_wait, !flock->fl_next);
+               if (!rc) {
+                       mutex_lock(&cinode->lock_mutex);
+                       continue;
+               }
+               cifs_locks_delete_block(flock);
+               break;
+       }
+       return rc;
+}
+
+static int
+cifs_posix_lock_file_wait(struct file *file, struct file_lock *flock)
+{
+       struct cifsInodeInfo *cinode = CIFS_I(file->f_path.dentry->d_inode);
+
+       mutex_lock(&cinode->lock_mutex);
+       /* lock_mutex will be released by the function below */
+       return cifs_posix_lock_file_wait_locked(file, flock);
+}
+
 /*
  * Set the byte-range lock (posix style). Returns:
  * 1) 0, if we set the lock and don't need to request to the server;
@@ -840,9 +888,9 @@ cifs_posix_lock_set(struct file *file, struct file_lock *flock)
                mutex_unlock(&cinode->lock_mutex);
                return rc;
        }
-       rc = posix_lock_file_wait(file, flock);
-       mutex_unlock(&cinode->lock_mutex);
-       return rc;
+
+       /* lock_mutex will be released by the function below */
+       return cifs_posix_lock_file_wait_locked(file, flock);
 }
 
 static int
@@ -1338,7 +1386,7 @@ cifs_setlk(struct file *file,  struct file_lock *flock, __u8 type,
 
 out:
        if (flock->fl_flags & FL_POSIX)
-               posix_lock_file_wait(file, flock);
+               cifs_posix_lock_file_wait(file, flock);
        return rc;
 }