btrfs: transaction: Avoid deadlock due to bad initialization timing of fs_info::journ...
authorQu Wenruo <wqu@suse.com>
Mon, 27 Apr 2020 06:50:14 +0000 (14:50 +0800)
committerDavid Sterba <dsterba@suse.com>
Mon, 27 Apr 2020 15:16:07 +0000 (17:16 +0200)
[BUG]
One run of btrfs/063 triggered the following lockdep warning:
  ============================================
  WARNING: possible recursive locking detected
  5.6.0-rc7-custom+ #48 Not tainted
  --------------------------------------------
  kworker/u24:0/7 is trying to acquire lock:
  ffff88817d3a46e0 (sb_internal#2){.+.+}, at: start_transaction+0x66c/0x890 [btrfs]

  but task is already holding lock:
  ffff88817d3a46e0 (sb_internal#2){.+.+}, at: start_transaction+0x66c/0x890 [btrfs]

  other info that might help us debug this:
   Possible unsafe locking scenario:

         CPU0
         ----
    lock(sb_internal#2);
    lock(sb_internal#2);

   *** DEADLOCK ***

   May be due to missing lock nesting notation

  4 locks held by kworker/u24:0/7:
   #0: ffff88817b495948 ((wq_completion)btrfs-endio-write){+.+.}, at: process_one_work+0x557/0xb80
   #1: ffff888189ea7db8 ((work_completion)(&work->normal_work)){+.+.}, at: process_one_work+0x557/0xb80
   #2: ffff88817d3a46e0 (sb_internal#2){.+.+}, at: start_transaction+0x66c/0x890 [btrfs]
   #3: ffff888174ca4da8 (&fs_info->reloc_mutex){+.+.}, at: btrfs_record_root_in_trans+0x83/0xd0 [btrfs]

  stack backtrace:
  CPU: 0 PID: 7 Comm: kworker/u24:0 Not tainted 5.6.0-rc7-custom+ #48
  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
  Workqueue: btrfs-endio-write btrfs_work_helper [btrfs]
  Call Trace:
   dump_stack+0xc2/0x11a
   __lock_acquire.cold+0xce/0x214
   lock_acquire+0xe6/0x210
   __sb_start_write+0x14e/0x290
   start_transaction+0x66c/0x890 [btrfs]
   btrfs_join_transaction+0x1d/0x20 [btrfs]
   find_free_extent+0x1504/0x1a50 [btrfs]
   btrfs_reserve_extent+0xd5/0x1f0 [btrfs]
   btrfs_alloc_tree_block+0x1ac/0x570 [btrfs]
   btrfs_copy_root+0x213/0x580 [btrfs]
   create_reloc_root+0x3bd/0x470 [btrfs]
   btrfs_init_reloc_root+0x2d2/0x310 [btrfs]
   record_root_in_trans+0x191/0x1d0 [btrfs]
   btrfs_record_root_in_trans+0x90/0xd0 [btrfs]
   start_transaction+0x16e/0x890 [btrfs]
   btrfs_join_transaction+0x1d/0x20 [btrfs]
   btrfs_finish_ordered_io+0x55d/0xcd0 [btrfs]
   finish_ordered_fn+0x15/0x20 [btrfs]
   btrfs_work_helper+0x116/0x9a0 [btrfs]
   process_one_work+0x632/0xb80
   worker_thread+0x80/0x690
   kthread+0x1a3/0x1f0
   ret_from_fork+0x27/0x50

It's pretty hard to reproduce, only one hit so far.

[CAUSE]
This is because we're calling btrfs_join_transaction() without re-using
the current running one:

btrfs_finish_ordered_io()
|- btrfs_join_transaction() <<< Call #1
   |- btrfs_record_root_in_trans()
      |- btrfs_reserve_extent()
 |- btrfs_join_transaction() <<< Call #2

Normally such btrfs_join_transaction() call should re-use the existing
one, without trying to re-start a transaction.

But the problem is, in btrfs_join_transaction() call #1, we call
btrfs_record_root_in_trans() before initializing current::journal_info.

And in btrfs_join_transaction() call #2, we're relying on
current::journal_info to avoid such deadlock.

[FIX]
Call btrfs_record_root_in_trans() after we have initialized
current::journal_info.

CC: stable@vger.kernel.org # 4.4+
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/transaction.c

index 8cede6e..2d54981 100644 (file)
@@ -662,10 +662,19 @@ again:
        }
 
 got_it:
-       btrfs_record_root_in_trans(h, root);
-
        if (!current->journal_info)
                current->journal_info = h;
+
+       /*
+        * btrfs_record_root_in_trans() needs to alloc new extents, and may
+        * call btrfs_join_transaction() while we're also starting a
+        * transaction.
+        *
+        * Thus it need to be called after current->journal_info initialized,
+        * or we can deadlock.
+        */
+       btrfs_record_root_in_trans(h, root);
+
        return h;
 
 join_fail: