@@ -578,9 +578,6 @@ int btrfs_validate_metadata_buffer(struct btrfs_io_bio *io_bio, u64 phy_offset,
goto err;
}
- 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)) {
@@ -4679,6 +4679,11 @@ 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 free'd 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);
@@ -5069,9 +5074,6 @@ static noinline int do_walk_down(struct btrfs_trans_handle *trans,
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);
@@ -5188,6 +5188,7 @@ 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++) {
@@ -6914,7 +6914,6 @@ 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
Both Filipe and Fedora QA recently hit the following lockdep splat WARNING: possible recursive locking detected 5.10.0-0.rc1.20201028gited8780e3f2ec.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.rc1.20201028gited8780e3f2ec.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 refcount'ed roots. However now all refcount'ed roots have the same class name, so we no longer need to worry about this. For non-refcount'ed tree's 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. Signed-off-by: Josef Bacik <josef@toxicpanda.com> --- fs/btrfs/disk-io.c | 3 --- fs/btrfs/extent-tree.c | 8 +++++--- fs/btrfs/extent_io.c | 1 + fs/btrfs/volumes.c | 1 - 4 files changed, 6 insertions(+), 7 deletions(-)