[XFS] Fix racy access to pb_flags. pagebuf_rele() modified pb_flags after

the pagebuf had been unlocked if the buffer was delwri. At high load, this
could result in a race when the superblock was being synced that would
result the flags being incorrect and the iodone functions being executed
incorrectly. This then leads to iclog callback failures or AIL list
corruptions resulting in filesystem shutdowns.

SGI-PV: 923981
SGI-Modid: xfs-linux:xfs-kern:23616a

Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Nathan Scott <nathans@sgi.com>

2 files changed, 51 insertions, 18 deletions

diff --git a/fs/xfs/linux-2.6/xfs_buf.c b/fs/xfs/linux-2.6/xfs_buf.c
index f43689754dae..e6340906342c 100644
--- a/fs/xfs/linux-2.6/xfs_buf.c
+++ b/fs/xfs/linux-2.6/xfs_buf.c
@@ -590,8 +590,10 @@ found:
                 PB_SET_OWNER(pb);
         }
 
-        if (pb->pb_flags & PBF_STALE)
+        if (pb->pb_flags & PBF_STALE) {
+                ASSERT((pb->pb_flags & _PBF_DELWRI_Q) == 0);
                 pb->pb_flags &= PBF_MAPPED;
+        }
         PB_TRACE(pb, "got_lock", 0);
         XFS_STATS_INC(pb_get_locked);
         return (pb);
@@ -872,6 +874,17 @@ pagebuf_rele(
 
         PB_TRACE(pb, "rele", pb->pb_relse);
 
+        /*
+         * pagebuf_lookup buffers are not hashed, not delayed write,
+         * and don't have their own release routines.  Special case.
+         */
+        if (unlikely(!hash)) {
+                ASSERT(!pb->pb_relse);
+                if (atomic_dec_and_test(&pb->pb_hold))
+                        xfs_buf_free(pb);
+                return;
+        }
+
         if (atomic_dec_and_lock(&pb->pb_hold, &hash->bh_lock)) {
                 int             do_free = 1;
 
@@ -883,22 +896,23 @@ pagebuf_rele(
                         do_free = 0;
                 }
 
-                if (pb->pb_flags & PBF_DELWRI) {
-                        pb->pb_flags |= PBF_ASYNC;
-                        atomic_inc(&pb->pb_hold);
-                        pagebuf_delwri_queue(pb, 0);
-                        do_free = 0;
-                } else if (pb->pb_flags & PBF_FS_MANAGED) {
+                if (pb->pb_flags & PBF_FS_MANAGED) {
                         do_free = 0;
                 }
 
                 if (do_free) {
+                        ASSERT((pb->pb_flags & (PBF_DELWRI|_PBF_DELWRI_Q)) == 0);
                         list_del_init(&pb->pb_hash_list);
                         spin_unlock(&hash->bh_lock);
                         pagebuf_free(pb);
                 } else {
                         spin_unlock(&hash->bh_lock);
                 }
+        } else {
+                /*
+                 * Catch reference count leaks
+                 */
+                ASSERT(atomic_read(&pb->pb_hold) >= 0);
         }
 }
 
@@ -976,13 +990,24 @@ pagebuf_lock(
  *      pagebuf_unlock
  *
  *      pagebuf_unlock releases the lock on the buffer object created by
- *      pagebuf_lock or pagebuf_cond_lock (not any
- *      pinning of underlying pages created by pagebuf_pin).
+ *      pagebuf_lock or pagebuf_cond_lock (not any pinning of underlying pages
+ *      created by pagebuf_pin).
+ *
+ *      If the buffer is marked delwri but is not queued, do so before we
+ *      unlock the buffer as we need to set flags correctly. We also need to
+ *      take a reference for the delwri queue because the unlocker is going to
+ *      drop their's and they don't know we just queued it.
  */
 void
 pagebuf_unlock(                         /* unlock buffer                */
         xfs_buf_t               *pb)    /* buffer to unlock             */
 {
+        if ((pb->pb_flags & (PBF_DELWRI|_PBF_DELWRI_Q)) == PBF_DELWRI) {
+                atomic_inc(&pb->pb_hold);
+                pb->pb_flags |= PBF_ASYNC;
+                pagebuf_delwri_queue(pb, 0);
+        }
+
         PB_CLEAR_OWNER(pb);
         up(&pb->pb_sema);
         PB_TRACE(pb, "unlock", 0);
@@ -1486,6 +1511,11 @@ again:
                         ASSERT(btp == bp->pb_target);
                         if (!(bp->pb_flags & PBF_FS_MANAGED)) {
                                 spin_unlock(&hash->bh_lock);
+                                /*
+                                 * Catch superblock reference count leaks
+                                 * immediately
+                                 */
+                                BUG_ON(bp->pb_bn == 0);
                                 delay(100);
                                 goto again;
                         }
@@ -1661,17 +1691,20 @@ pagebuf_delwri_queue(
         int                     unlock)
 {
         PB_TRACE(pb, "delwri_q", (long)unlock);
-        ASSERT(pb->pb_flags & PBF_DELWRI);
+        ASSERT((pb->pb_flags & (PBF_DELWRI|PBF_ASYNC)) ==
+                                        (PBF_DELWRI|PBF_ASYNC));
 
         spin_lock(&pbd_delwrite_lock);
         /* If already in the queue, dequeue and place at tail */
         if (!list_empty(&pb->pb_list)) {
+                ASSERT(pb->pb_flags & _PBF_DELWRI_Q);
                 if (unlock) {
                         atomic_dec(&pb->pb_hold);
                 }
                 list_del(&pb->pb_list);
         }
 
+        pb->pb_flags |= _PBF_DELWRI_Q;
         list_add_tail(&pb->pb_list, &pbd_delwrite_queue);
         pb->pb_queuetime = jiffies;
         spin_unlock(&pbd_delwrite_lock);
@@ -1688,10 +1721,11 @@ pagebuf_delwri_dequeue(
 
         spin_lock(&pbd_delwrite_lock);
         if ((pb->pb_flags & PBF_DELWRI) && !list_empty(&pb->pb_list)) {
+                ASSERT(pb->pb_flags & _PBF_DELWRI_Q);
                 list_del_init(&pb->pb_list);
                 dequeued = 1;
         }
-        pb->pb_flags &= ~PBF_DELWRI;
+        pb->pb_flags &= ~(PBF_DELWRI|_PBF_DELWRI_Q);
         spin_unlock(&pbd_delwrite_lock);
 
         if (dequeued)
@@ -1770,7 +1804,7 @@ xfsbufd(
                                         break;
                                 }
 
-                                pb->pb_flags &= ~PBF_DELWRI;
+                                pb->pb_flags &= ~(PBF_DELWRI|_PBF_DELWRI_Q);
                                 pb->pb_flags |= PBF_WRITE;
                                 list_move(&pb->pb_list, &tmp);
                         }
@@ -1820,15 +1854,13 @@ xfs_flush_buftarg(
                 if (pb->pb_target != target)
                         continue;
 
-                ASSERT(pb->pb_flags & PBF_DELWRI);
+                ASSERT(pb->pb_flags & (PBF_DELWRI|_PBF_DELWRI_Q));
                 PB_TRACE(pb, "walkq2", (long)pagebuf_ispin(pb));
                 if (pagebuf_ispin(pb)) {
                         pincount++;
                         continue;
                 }
 
-                pb->pb_flags &= ~PBF_DELWRI;
-                pb->pb_flags |= PBF_WRITE;
                 list_move(&pb->pb_list, &tmp);
         }
         spin_unlock(&pbd_delwrite_lock);
@@ -1837,12 +1869,14 @@ xfs_flush_buftarg(
          * Dropped the delayed write list lock, now walk the temporary list
          */
         list_for_each_entry_safe(pb, n, &tmp, pb_list) {
+                pagebuf_lock(pb);
+                pb->pb_flags &= ~(PBF_DELWRI|_PBF_DELWRI_Q);
+                pb->pb_flags |= PBF_WRITE;
                 if (wait)
                         pb->pb_flags &= ~PBF_ASYNC;
                 else
                         list_del_init(&pb->pb_list);
 
-                pagebuf_lock(pb);
                 pagebuf_iostrategy(pb);
         }
 
diff --git a/fs/xfs/linux-2.6/xfs_buf.h b/fs/xfs/linux-2.6/xfs_buf.h
index c322e9d71f3d..4b7fe3b5e460 100644
--- a/fs/xfs/linux-2.6/xfs_buf.h
+++ b/fs/xfs/linux-2.6/xfs_buf.h
@@ -89,6 +89,7 @@ typedef enum page_buf_flags_e {		/* pb_flags values */
         _PBF_PAGE_CACHE = (1 << 17),/* backed by pagecache                 */
         _PBF_KMEM_ALLOC = (1 << 18),/* backed by kmem_alloc()              */
         _PBF_RUN_QUEUES = (1 << 19),/* run block device task queue         */
+        _PBF_DELWRI_Q = (1 << 21),   /* buffer on delwri queue             */
 } page_buf_flags_t;
 
 #define PBF_UPDATE (PBF_READ | PBF_WRITE)
@@ -337,8 +338,6 @@ extern void pagebuf_trace(
 
 
 
-
-
 /* These are just for xfs_syncsub... it sets an internal variable
  * then passes it to VOP_FLUSH_PAGES or adds the flags to a newly gotten buf_t
  */