ext4: fix a possible ABBA deadlock dued to busy PA

From: Chunguang Xu <[email protected]>

We found on older kernel (3.10) that in the scenario of insufficient disk space, system may trigger an ABBA deadlock problem, it seems that this problem still exists in latest kernel, try to fix it here. The main process triggered by this problem is that task A occupies the PA and waits for the jbd2 transaction finish, the jbd2 transaction waits for the completion of task B's IO (plug_list), but task B waits for the release of PA by task A to finish discard, which indirectly forms an ABBA deadlock. The related calltrace is as follows:

    Task A
    vfs_write
    ext4_mb_new_blocks()
    ext4_mb_mark_diskspace_used()       JBD2
    jbd2_journal_get_write_access()  -> jbd2_journal_commit_transaction()
  ->schedule()                          filemap_fdatawait()
 |                                              |
 | Task B                                       |
 | do_unlinkat()                                |
 | ext4_evict_inode()                           |
 | jbd2_journal_begin_ordered_truncate()        |
 | filemap_fdatawrite_range()                   |
 | ext4_mb_new_blocks()                         |

-ext4_mb_discard_group_preallocations() <-----

Here, try to cancel ext4_mb_discard_group_preallocations() internal retry due to PA busy, and do a limited number of retries inside ext4_mb_discard_preallocations(), which can circumvent the above problems, but also has some advantages:

1. Since the PA is in a busy state, if other groups have free PAs, keeping the current PA may help to reduce fragmentation.
2. Continue to traverse forward instead of waiting for the current group PA to be released. In most scenarios, the PA discard time can be reduced.

However, in the case of smaller free space, if only a few groups have space, then due to multiple traversals of the group, it may increase CPU overhead. But in contrast, I feel that the overall benefit is better than the cost.

Signed-off-by Chunguang Xu <[email protected]>

---
v2: reset busy to zero before goto repeat.

 fs/ext4/mballoc.c | 35 ++++++++++++++++-------------------
 1 file changed, 16 insertions(+), 19 deletions(-)

fs/ext4/mballoc.c +16 -19

index 72bfac2..72de6c1 100644
@@ -4814,7 +4814,7 @@ static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
  */
 static noinline_for_stack int
 ext4_mb_discard_group_preallocations(struct super_block *sb,
-					ext4_group_t group, int needed)
+				     ext4_group_t group, int needed, int *busy)
 {
 	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 	struct buffer_head *bitmap_bh = NULL;

@@ -4822,8 +4822,7 @@ static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
 	struct list_head list;
 	struct ext4_buddy e4b;
 	int err;
-	int busy = 0;
-	int free, free_total = 0;
+	int free = 0;
 
 	mb_debug(sb, "discard preallocation for group %u\n", group);
 	if (list_empty(&grp->bb_prealloc_list))

@@ -4850,15 +4849,13 @@ static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
 		needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
 
 	INIT_LIST_HEAD(&list);
-repeat:
-	free = 0;
 	ext4_lock_group(sb, group);
 	list_for_each_entry_safe(pa, tmp,
 				&grp->bb_prealloc_list, pa_group_list) {
 		spin_lock(&pa->pa_lock);
 		if (atomic_read(&pa->pa_count)) {
 			spin_unlock(&pa->pa_lock);
-			busy = 1;
+			*busy = 1;
 			continue;
 		}
 		if (pa->pa_deleted) {

@@ -4898,22 +4895,13 @@ static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
 		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
 	}
 
-	free_total += free;
-
-	/* if we still need more blocks and some PAs were used, try again */
-	if (free_total < needed && busy) {
-		ext4_unlock_group(sb, group);
-		cond_resched();
-		busy = 0;
-		goto repeat;
-	}
 	ext4_unlock_group(sb, group);
 	ext4_mb_unload_buddy(&e4b);
 	put_bh(bitmap_bh);
 out_dbg:
 	mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
-		 free_total, group, grp->bb_free);
-	return free_total;
+		 free, group, grp->bb_free);
+	return free;
 }
 
 /*

@@ -5455,13 +5443,22 @@ static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
 {
 	ext4_group_t i, ngroups = ext4_get_groups_count(sb);
 	int ret;
-	int freed = 0;
+	int freed = 0, busy = 0;
+	int retry = 0;
 
 	trace_ext4_mb_discard_preallocations(sb, needed);
+ repeat:
+	retry++;
 	for (i = 0; i < ngroups && needed > 0; i++) {
-		ret = ext4_mb_discard_group_preallocations(sb, i, needed);
+		ret = ext4_mb_discard_group_preallocations(sb, i, needed, &busy);
 		freed += ret;
 		needed -= ret;
+		cond_resched();
+	}
+
+	if (needed > 0 && busy && retry < 3) {
+		busy = 0;
+		goto repeat;
 	}
 
 	return freed;

Signature · 1 line

1.8.3.1

Hello,

it seems this patch has fallen through the cracks.

Quoted 37 lines · brookxu

From: Chunguang Xu <[email protected]>

We found on older kernel (3.10) that in the scenario of insufficient disk space, system may trigger an ABBA deadlock problem, it seems that this problem still exists in latest kernel, try to fix it here. The main process triggered by this problem is that task A occupies the PA and waits for the jbd2 transaction finish, the jbd2 transaction waits for the completion of task B's IO (plug_list), but task B waits for the release of PA by task A to finish discard, which indirectly forms an ABBA deadlock. The related calltrace is as follows:

    Task A
    vfs_write
    ext4_mb_new_blocks()
    ext4_mb_mark_diskspace_used()       JBD2
    jbd2_journal_get_write_access()  -> jbd2_journal_commit_transaction()
  ->schedule()                          filemap_fdatawait()
 |                                              |
 | Task B                                       |
 | do_unlinkat()                                |
 | ext4_evict_inode()                           |
 | jbd2_journal_begin_ordered_truncate()        |
 | filemap_fdatawrite_range()                   |
 | ext4_mb_new_blocks()                         |

-ext4_mb_discard_group_preallocations() <-----

Here, try to cancel ext4_mb_discard_group_preallocations() internal retry due to PA busy, and do a limited number of retries inside ext4_mb_discard_preallocations(), which can circumvent the above problems, but also has some advantages:

1. Since the PA is in a busy state, if other groups have free PAs, keeping the current PA may help to reduce fragmentation.
2. Continue to traverse forward instead of waiting for the current group PA to be released. In most scenarios, the PA discard time can be reduced.

Quoted 6 lines

However, in the case of smaller free space, if only a few groups have space, then due to multiple traversals of the group, it may increase CPU overhead. But in contrast, I feel that the overall benefit is better than the cost.

Signed-off-by Chunguang Xu <[email protected]>

Thanks for the patch! I guess this is a reasonable middle-ground so feel free to add:

Reviewed-by Jan Kara <[email protected]>

I think Ritesh was last touching this logic so let's check he doesn't see anything wrong with this change. Ritesh?

Honza

Quoted 99 lines

---
v2: reset busy to zero before goto repeat.

 fs/ext4/mballoc.c | 35 ++++++++++++++++-------------------
 1 file changed, 16 insertions(+), 19 deletions(-)

fs/ext4/mballoc.c +16 -19

index 72bfac2..72de6c1 100644
@@ -4814,7 +4814,7 @@ static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
  */
 static noinline_for_stack int
 ext4_mb_discard_group_preallocations(struct super_block *sb,
-					ext4_group_t group, int needed)
+				     ext4_group_t group, int needed, int *busy)
 {
 	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 	struct buffer_head *bitmap_bh = NULL;

@@ -4822,8 +4822,7 @@ static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
 	struct list_head list;
 	struct ext4_buddy e4b;
 	int err;
-	int busy = 0;
-	int free, free_total = 0;
+	int free = 0;
 
 	mb_debug(sb, "discard preallocation for group %u\n", group);
 	if (list_empty(&grp->bb_prealloc_list))

@@ -4850,15 +4849,13 @@ static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
 		needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
 
 	INIT_LIST_HEAD(&list);
-repeat:
-	free = 0;
 	ext4_lock_group(sb, group);
 	list_for_each_entry_safe(pa, tmp,
 				&grp->bb_prealloc_list, pa_group_list) {
 		spin_lock(&pa->pa_lock);
 		if (atomic_read(&pa->pa_count)) {
 			spin_unlock(&pa->pa_lock);
-			busy = 1;
+			*busy = 1;
 			continue;
 		}
 		if (pa->pa_deleted) {

@@ -4898,22 +4895,13 @@ static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
 		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
 	}
 
-	free_total += free;
-
-	/* if we still need more blocks and some PAs were used, try again */
-	if (free_total < needed && busy) {
-		ext4_unlock_group(sb, group);
-		cond_resched();
-		busy = 0;
-		goto repeat;
-	}
 	ext4_unlock_group(sb, group);
 	ext4_mb_unload_buddy(&e4b);
 	put_bh(bitmap_bh);
 out_dbg:
 	mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
-		 free_total, group, grp->bb_free);
-	return free_total;
+		 free, group, grp->bb_free);
+	return free;
 }
 
 /*

@@ -5455,13 +5443,22 @@ static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
 {
 	ext4_group_t i, ngroups = ext4_get_groups_count(sb);
 	int ret;
-	int freed = 0;
+	int freed = 0, busy = 0;
+	int retry = 0;
 
 	trace_ext4_mb_discard_preallocations(sb, needed);
+ repeat:
+	retry++;
 	for (i = 0; i < ngroups && needed > 0; i++) {
-		ret = ext4_mb_discard_group_preallocations(sb, i, needed);
+		ret = ext4_mb_discard_group_preallocations(sb, i, needed, &busy);
 		freed += ret;
 		needed -= ret;
+		cond_resched();
+	}
+
+	if (needed > 0 && busy && retry < 3) {
+		busy = 0;
+		goto repeat;
 	}
 
 	return freed;

Signature · 1 line

1.8.3.1