btrfs: qgroup: Don't hold qgroup_ioctl_lock in btrfs_qgroup_inherit()
authorQu Wenruo <wqu@suse.com>
Thu, 13 Jun 2019 09:31:24 +0000 (17:31 +0800)
committerDavid Sterba <dsterba@suse.com>
Tue, 2 Jul 2019 10:30:48 +0000 (12:30 +0200)
[BUG]
Lockdep will report the following circular locking dependency:

  WARNING: possible circular locking dependency detected
  5.2.0-rc2-custom #24 Tainted: G           O
  ------------------------------------------------------
  btrfs/8631 is trying to acquire lock:
  000000002536438c (&fs_info->qgroup_ioctl_lock#2){+.+.}, at: btrfs_qgroup_inherit+0x40/0x620 [btrfs]

  but task is already holding lock:
  000000003d52cc23 (&fs_info->tree_log_mutex){+.+.}, at: create_pending_snapshot+0x8b6/0xe60 [btrfs]

  which lock already depends on the new lock.

  the existing dependency chain (in reverse order) is:

  -> #2 (&fs_info->tree_log_mutex){+.+.}:
         __mutex_lock+0x76/0x940
         mutex_lock_nested+0x1b/0x20
         btrfs_commit_transaction+0x475/0xa00 [btrfs]
         btrfs_commit_super+0x71/0x80 [btrfs]
         close_ctree+0x2bd/0x320 [btrfs]
         btrfs_put_super+0x15/0x20 [btrfs]
         generic_shutdown_super+0x72/0x110
         kill_anon_super+0x18/0x30
         btrfs_kill_super+0x16/0xa0 [btrfs]
         deactivate_locked_super+0x3a/0x80
         deactivate_super+0x51/0x60
         cleanup_mnt+0x3f/0x80
         __cleanup_mnt+0x12/0x20
         task_work_run+0x94/0xb0
         exit_to_usermode_loop+0xd8/0xe0
         do_syscall_64+0x210/0x240
         entry_SYSCALL_64_after_hwframe+0x49/0xbe

  -> #1 (&fs_info->reloc_mutex){+.+.}:
         __mutex_lock+0x76/0x940
         mutex_lock_nested+0x1b/0x20
         btrfs_commit_transaction+0x40d/0xa00 [btrfs]
         btrfs_quota_enable+0x2da/0x730 [btrfs]
         btrfs_ioctl+0x2691/0x2b40 [btrfs]
         do_vfs_ioctl+0xa9/0x6d0
         ksys_ioctl+0x67/0x90
         __x64_sys_ioctl+0x1a/0x20
         do_syscall_64+0x65/0x240
         entry_SYSCALL_64_after_hwframe+0x49/0xbe

  -> #0 (&fs_info->qgroup_ioctl_lock#2){+.+.}:
         lock_acquire+0xa7/0x190
         __mutex_lock+0x76/0x940
         mutex_lock_nested+0x1b/0x20
         btrfs_qgroup_inherit+0x40/0x620 [btrfs]
         create_pending_snapshot+0x9d7/0xe60 [btrfs]
         create_pending_snapshots+0x94/0xb0 [btrfs]
         btrfs_commit_transaction+0x415/0xa00 [btrfs]
         btrfs_mksubvol+0x496/0x4e0 [btrfs]
         btrfs_ioctl_snap_create_transid+0x174/0x180 [btrfs]
         btrfs_ioctl_snap_create_v2+0x11c/0x180 [btrfs]
         btrfs_ioctl+0xa90/0x2b40 [btrfs]
         do_vfs_ioctl+0xa9/0x6d0
         ksys_ioctl+0x67/0x90
         __x64_sys_ioctl+0x1a/0x20
         do_syscall_64+0x65/0x240
         entry_SYSCALL_64_after_hwframe+0x49/0xbe

  other info that might help us debug this:

  Chain exists of:
    &fs_info->qgroup_ioctl_lock#2 --> &fs_info->reloc_mutex --> &fs_info->tree_log_mutex

   Possible unsafe locking scenario:

         CPU0                    CPU1
         ----                    ----
    lock(&fs_info->tree_log_mutex);
                                 lock(&fs_info->reloc_mutex);
                                 lock(&fs_info->tree_log_mutex);
    lock(&fs_info->qgroup_ioctl_lock#2);

   *** DEADLOCK ***

  6 locks held by btrfs/8631:
   #0: 00000000ed8f23f6 (sb_writers#12){.+.+}, at: mnt_want_write_file+0x28/0x60
   #1: 000000009fb1597a (&type->i_mutex_dir_key#10/1){+.+.}, at: btrfs_mksubvol+0x70/0x4e0 [btrfs]
   #2: 0000000088c5ad88 (&fs_info->subvol_sem){++++}, at: btrfs_mksubvol+0x128/0x4e0 [btrfs]
   #3: 000000009606fc3e (sb_internal#2){.+.+}, at: start_transaction+0x37a/0x520 [btrfs]
   #4: 00000000f82bbdf5 (&fs_info->reloc_mutex){+.+.}, at: btrfs_commit_transaction+0x40d/0xa00 [btrfs]
   #5: 000000003d52cc23 (&fs_info->tree_log_mutex){+.+.}, at: create_pending_snapshot+0x8b6/0xe60 [btrfs]

[CAUSE]
Due to the delayed subvolume creation, we need to call
btrfs_qgroup_inherit() inside commit transaction code, with a lot of
other mutex hold.
This hell of lock chain can lead to above problem.

[FIX]
On the other hand, we don't really need to hold qgroup_ioctl_lock if
we're in the context of create_pending_snapshot().
As in that context, we're the only one being able to modify qgroup.

All other qgroup functions which needs qgroup_ioctl_lock are either
holding a transaction handle, or will start a new transaction:
  Functions will start a new transaction():
  * btrfs_quota_enable()
  * btrfs_quota_disable()
  Functions hold a transaction handler:
  * btrfs_add_qgroup_relation()
  * btrfs_del_qgroup_relation()
  * btrfs_create_qgroup()
  * btrfs_remove_qgroup()
  * btrfs_limit_qgroup()
  * btrfs_qgroup_inherit() call inside create_subvol()

So we have a higher level protection provided by transaction, thus we
don't need to always hold qgroup_ioctl_lock in btrfs_qgroup_inherit().

Only the btrfs_qgroup_inherit() call in create_subvol() needs to hold
qgroup_ioctl_lock, while the btrfs_qgroup_inherit() call in
create_pending_snapshot() is already protected by transaction.

So the fix is to detect the context by checking
trans->transaction->state.
If we're at TRANS_STATE_COMMIT_DOING, then we're in commit transaction
context and no need to get the mutex.

Reported-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/qgroup.c

index 3e6ffbb..f8a3c1b 100644 (file)
@@ -2614,6 +2614,7 @@ int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
        int ret = 0;
        int i;
        u64 *i_qgroups;
+       bool committing = false;
        struct btrfs_fs_info *fs_info = trans->fs_info;
        struct btrfs_root *quota_root;
        struct btrfs_qgroup *srcgroup;
@@ -2621,7 +2622,25 @@ int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
        u32 level_size = 0;
        u64 nums;
 
-       mutex_lock(&fs_info->qgroup_ioctl_lock);
+       /*
+        * There are only two callers of this function.
+        *
+        * One in create_subvol() in the ioctl context, which needs to hold
+        * the qgroup_ioctl_lock.
+        *
+        * The other one in create_pending_snapshot() where no other qgroup
+        * code can modify the fs as they all need to either start a new trans
+        * or hold a trans handler, thus we don't need to hold
+        * qgroup_ioctl_lock.
+        * This would avoid long and complex lock chain and make lockdep happy.
+        */
+       spin_lock(&fs_info->trans_lock);
+       if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
+               committing = true;
+       spin_unlock(&fs_info->trans_lock);
+
+       if (!committing)
+               mutex_lock(&fs_info->qgroup_ioctl_lock);
        if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
                goto out;
 
@@ -2785,7 +2804,8 @@ int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
 unlock:
        spin_unlock(&fs_info->qgroup_lock);
 out:
-       mutex_unlock(&fs_info->qgroup_ioctl_lock);
+       if (!committing)
+               mutex_unlock(&fs_info->qgroup_ioctl_lock);
        return ret;
 }