ext4: fix a possible ABBA deadlock due to busy PA
authorChunguang Xu <brookxu@tencent.com>
Tue, 23 Nov 2021 01:17:57 +0000 (09:17 +0800)
committerTheodore Ts'o <tytso@mit.edu>
Mon, 10 Jan 2022 18:25:55 +0000 (13:25 -0500)
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 <brookxu@tencent.com>
Reported-by: kernel test robot <lkp@intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/1637630277-23496-1-git-send-email-brookxu.cn@gmail.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@kernel.org
fs/ext4/mballoc.c

index 215b706..3dd9b9e 100644 (file)
@@ -4814,7 +4814,7 @@ ext4_mb_release_group_pa(struct ext4_buddy *e4b,
  */
 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 *busy)
 {
        struct ext4_group_info *grp = ext4_get_group_info(sb, group);
        struct buffer_head *bitmap_bh = NULL;
@@ -4822,8 +4822,7 @@ ext4_mb_discard_group_preallocations(struct super_block *sb,
        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))
@@ -4846,19 +4845,14 @@ ext4_mb_discard_group_preallocations(struct super_block *sb,
                goto out_dbg;
        }
 
-       if (needed == 0)
-               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 +4892,13 @@ repeat:
                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 +5440,24 @@ 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);
+
+       if (needed == 0)
+               needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
+ repeat:
        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, &busy);
                freed += ret;
                needed -= ret;
+               cond_resched();
+       }
+
+       if (needed > 0 && busy && ++retry < 3) {
+               busy = 0;
+               goto repeat;
        }
 
        return freed;