[GFS2] Fix a page lock / glock deadlock

We've previously been using a "try lock" in readpage on the basis that
it would prevent deadlocks due to the inverted lock ordering (our normal
lock ordering is glock first and then page lock). Unfortunately tests
have shown that this isn't enough. If the glock has a demote request
queued such that run_queue() in the glock code tries to do a demote when
its called under readpage then it will try and write out all the dirty
pages which requires locking them. This then deadlocks with the page
locked by readpage.

The solution is to always require two calls into readpage. The first
unlocks the page, gets the glock and returns AOP_TRUNCATED_PAGE, the
second does the actual readpage and unlocks the glock & page as
required.

Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>

1 files changed, 6 insertions, 7 deletions

diff --git a/fs/gfs2/glock.h b/fs/gfs2/glock.h
index ace5770760ce..cdad3e6f8150 100644
--- a/fs/gfs2/glock.h
+++ b/fs/gfs2/glock.h
@@ -32,24 +32,23 @@
 #define GLR_TRYFAILED           13
 #define GLR_CANCELED            14
 
-static inline int gfs2_glock_is_locked_by_me(struct gfs2_glock *gl)
+static inline struct gfs2_holder *gfs2_glock_is_locked_by_me(struct gfs2_glock *gl)
 {
         struct gfs2_holder *gh;
-        int locked = 0;
         struct pid *pid;
 
         /* Look in glock's list of holders for one with current task as owner */
         spin_lock(&gl->gl_spin);
         pid = task_pid(current);
         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
-                if (gh->gh_owner_pid == pid) {
-                        locked = 1;
-                        break;
-                }
+                if (gh->gh_owner_pid == pid)
+                        goto out;
         }
+        gh = NULL;
+out:
         spin_unlock(&gl->gl_spin);
 
-        return locked;
+        return gh;
 }
 
 static inline int gfs2_glock_is_held_excl(struct gfs2_glock *gl)