Blob Blame History Raw
autofs-5.0.6 - fix recursive mount deadlock

From: Ian Kent <raven@themaw.net>

Prior to the vfs-automount changes that went into 2.6.38
and were finalized in 3.1 it was not possible to block
path walks into multi-mounts whose root was covered by
another mount. To deal with that a write lock was used
to ensure the mount tree construction was completed. This
restricts the types of recursively defined mount maps that
can be used and can lead to a deadlock during lookup.

Now that we can prevent processes walking into multi-mounts
that are under construction we no longer need to use a
write lock.

Also, in the patch below, a cache writelock is changed to
a read lock because a write lock isn't needed since the
map cache entry isn't being updated.
---

 CHANGELOG       |    1 +
 daemon/direct.c |   14 ++++++++++++--
 2 files changed, 13 insertions(+), 2 deletions(-)


diff --git a/CHANGELOG b/CHANGELOG
index 936c9ab..9cdad6e 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -12,6 +12,7 @@
 - configure.in: allow cross compilation.
 - README: update mailing list subscription info.
 - allow non root user to check status.
+- fix recursive mount deadlock.
 
 25/07/2012 autofs-5.0.7
 =======================
diff --git a/daemon/direct.c b/daemon/direct.c
index 7e2f0d7..3e09c5d 100644
--- a/daemon/direct.c
+++ b/daemon/direct.c
@@ -1285,6 +1285,8 @@ int handle_packet_missing_direct(struct autofs_point *ap, autofs_packet_missing_
 	struct timespec wait;
 	struct timeval now;
 	int ioctlfd, len, state;
+	unsigned int kver_major = get_kver_major();
+	unsigned int kver_minor = get_kver_minor();
 
 	pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &state);
 
@@ -1297,8 +1299,16 @@ int handle_packet_missing_direct(struct autofs_point *ap, autofs_packet_missing_
 	 * cache entry we will not be able to find the mapent. So
 	 * we must take the source writelock to ensure the parent
 	 * has mount is complete before we look for the entry.
+	 *
+	 * Since the vfs-automount kernel changes we can now block
+	 * on covered mounts during mount tree construction so a
+	 * write lock is no longer needed. So we now can handle a
+	 * wider class of recursively define mount lookups.
 	 */
-	master_source_writelock(ap->entry);
+	if (kver_major > 5 || (kver_major == 5 && kver_minor > 1))
+		master_source_readlock(ap->entry);
+	else
+		master_source_writelock(ap->entry);
 	map = ap->entry->maps;
 	while (map) {
 		/*
@@ -1311,7 +1321,7 @@ int handle_packet_missing_direct(struct autofs_point *ap, autofs_packet_missing_
 		}
 
 		mc = map->mc;
-		cache_writelock(mc);
+		cache_readlock(mc);
 		me = cache_lookup_ino(mc, pkt->dev, pkt->ino);
 		if (me)
 			break;