On Tue, Jul 20, 2021 at 04:05:23PM +0100, fdmanana@kernel.org wrote:

From: Filipe Manana <redacted>

In commit 351cbf6e4410e7 ("btrfs: use nofs allocations for running delayed
items") we wrapped all btree updates when running delayed items with
memalloc_nofs_save() and memalloc_nofs_restore(), due to a lock inversion
detected by lockdep involving reclaim and the mutex of delayed nodes.

The problem is because the ref verify tool does some memory allocations
with GFP_KERNEL, which can trigger reclaim and reclaim can trigger inode
eviction, which requires locking the mutex of an inode's delayed node.
On the other hand the ref verify tool is called when allocating metadata
extents as part of operations that modify a btree, which is a problem when
running delayed nodes, where we do btree updates while holding the mutex
of a delayed node. This is what caused the lockdep warning.

Instead of wrapping every btree update when running delayed nodes, change
the ref verify tool to never do GFP_KERNEL allocations, because:

1) We get less repeated code, which at the moment does not even have a
   comment mentioning why we need to setup the NOFS context, which is a
   recommended good practice as mentioned at
   Documentation/core-api/gfp_mask-from-fs-io.rst

2) The ref verify tool is something meant only for debugging and not
   something that should be enabled on non-debug / non-development
   kernels;

3) We may have yet more places outside delayed-inode.c where we have
   similar problem: doing btree updates while holding some lock and
   then having the GFP_KERNEL memory allocations, from the ref verify
   tool, trigger reclaim and trying again to acquire the same lock
   through the reclaim path.
   Or we could get more such cases in the future, therefore this change
   prevents getting into similar cases when using the ref verify tool.

That all sounds reasonable to me regarding the GFP flags use.