btrfs: set the lockdep class for extent buffers on creation

Both Filipe and Fedora QA recently hit the following lockdep splat:

  WARNING: possible recursive locking detected
  5.10.0-0.rc1.20201028gited8780e3.57.fc34.x86_64 #1 Not tainted
  --------------------------------------------
  rsync/2610 is trying to acquire lock:
  ffff89617ed48f20 (&eb->lock){++++}-{2:2}, at: btrfs_tree_read_lock_atomic+0x34/0x140

  but task is already holding lock:
  ffff8961757b1130 (&eb->lock){++++}-{2:2}, at: btrfs_tree_read_lock_atomic+0x34/0x140

  other info that might help us debug this:
   Possible unsafe locking scenario:
	 CPU0
	 ----
    lock(&eb->lock);
    lock(&eb->lock);

   *** DEADLOCK ***
   May be due to missing lock nesting notation
  2 locks held by rsync/2610:
   #0: ffff896107212b90 (&type->i_mutex_dir_key#10){++++}-{3:3}, at: walk_component+0x10c/0x190
   #1: ffff8961757b1130 (&eb->lock){++++}-{2:2}, at: btrfs_tree_read_lock_atomic+0x34/0x140

  stack backtrace:
  CPU: 1 PID: 2610 Comm: rsync Not tainted 5.10.0-0.rc1.20201028gited8780e3.57.fc34.x86_64 #1
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 0.0.0 02/06/2015
  Call Trace:
   dump_stack+0x8b/0xb0
   __lock_acquire.cold+0x12d/0x2a4
   ? kvm_sched_clock_read+0x14/0x30
   ? sched_clock+0x5/0x10
   lock_acquire+0xc8/0x400
   ? btrfs_tree_read_lock_atomic+0x34/0x140
   ? read_block_for_search.isra.0+0xdd/0x320
   _raw_read_lock+0x3d/0xa0
   ? btrfs_tree_read_lock_atomic+0x34/0x140
   btrfs_tree_read_lock_atomic+0x34/0x140
   btrfs_search_slot+0x616/0x9a0
   btrfs_lookup_dir_item+0x6c/0xb0
   btrfs_lookup_dentry+0xa8/0x520
   ? lockdep_init_map_waits+0x4c/0x210
   btrfs_lookup+0xe/0x30
   __lookup_slow+0x10f/0x1e0
   walk_component+0x11b/0x190
   path_lookupat+0x72/0x1c0
   filename_lookup+0x97/0x180
   ? strncpy_from_user+0x96/0x1e0
   ? getname_flags.part.0+0x45/0x1a0
   vfs_statx+0x64/0x100
   ? lockdep_hardirqs_on_prepare+0xff/0x180
   ? _raw_spin_unlock_irqrestore+0x41/0x50
   __do_sys_newlstat+0x26/0x40
   ? lockdep_hardirqs_on_prepare+0xff/0x180
   ? syscall_enter_from_user_mode+0x27/0x80
   ? syscall_enter_from_user_mode+0x27/0x80
   do_syscall_64+0x33/0x40
   entry_SYSCALL_64_after_hwframe+0x44/0xa9

I have also seen a report of lockdep complaining about the lock class
that was looked up being the same as the lock class on the lock we were
using, but I can't find the report.

These are problems that occur because we do not have the lockdep class
set on the extent buffer until _after_ we read the eb in properly.  This
is problematic for concurrent readers, because we will create the extent
buffer, lock it, and then attempt to read the extent buffer.

If a second thread comes in and tries to do a search down the same path
they'll get the above lockdep splat because the class isn't set properly
on the extent buffer.

There was a good reason for this, we generally didn't know the real
owner of the eb until we read it, specifically in refcounted roots.

However now all refcounted roots have the same class name, so we no
longer need to worry about this.  For non-refcounted trees we know
which root we're on based on the parent.

Fix this by setting the lockdep class on the eb at creation time instead
of read time.  This will fix the splat and the weirdness where the class
changes in the middle of locking the block.
Reviewed-by: NFilipe Manana <fdmanana@suse.com>
Signed-off-by: NJosef Bacik <josef@toxicpanda.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

fs/btrfs/disk-io.c

@@ -554,9 +554,6 @@ static int validate_extent_buffer(struct extent_buffer *eb)
 		goto out;
 	}
 
-	btrfs_set_buffer_lockdep_class(btrfs_header_owner(eb),
-				       eb, found_level);
-
 	csum_tree_block(eb, result);
 
 	if (memcmp_extent_buffer(eb, result, 0, csum_size)) {

fs/btrfs/extent-tree.c

@@ -4634,6 +4634,11 @@ btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 		return ERR_PTR(-EUCLEAN);
 	}
 
+	/*
+	 * This needs to stay, because we could allocate a freed block from an
+	 * old tree into a new tree, so we need to make sure this new block is
+	 * set to the appropriate level and owner.
+	 */
 	btrfs_set_buffer_lockdep_class(owner, buf, level);
 	__btrfs_tree_lock(buf, nest);
 	btrfs_clean_tree_block(buf);
@@ -5022,9 +5027,6 @@ static noinline int do_walk_down(struct btrfs_trans_handle *trans,
 				root->root_key.objectid, level - 1);
 		if (IS_ERR(next))
 			return PTR_ERR(next);
-
-		btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
-					       level - 1);
 		reada = 1;
 	}
 	btrfs_tree_lock(next);

fs/btrfs/extent_io.c

@@ -5205,6 +5205,7 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
 	eb = __alloc_extent_buffer(fs_info, start, len);
 	if (!eb)
 		return ERR_PTR(-ENOMEM);
+	btrfs_set_buffer_lockdep_class(owner_root, eb, level);
 
 	num_pages = num_extent_pages(eb);
 	for (i = 0; i < num_pages; i++, index++) {

fs/btrfs/volumes.c

@@ -6908,7 +6908,6 @@ int btrfs_read_sys_array(struct btrfs_fs_info *fs_info)
 	if (IS_ERR(sb))
 		return PTR_ERR(sb);
 	set_extent_buffer_uptodate(sb);
-	btrfs_set_buffer_lockdep_class(root->root_key.objectid, sb, 0);
 	/*
 	 * The sb extent buffer is artificial and just used to read the system array.
 	 * set_extent_buffer_uptodate() call does not properly mark all it's