From [1]:

XFS is telling the allocator not to track this allocation with
lockdep, and that is getting passed down through the allocator which
has not passed it to lockdep (correct behaviour!), but then KASAN is
trying to track the allocation and that needs to do a memory
allocation.  __stack_depot_save() is passed the gfp mask from the
allocation context so it has __GFP_NOLOCKDEP right there, but it
does:

        if (unlikely(can_alloc && !smp_load_acquire(&next_slab_inited))) {
                /*
                 * Zero out zone modifiers, as we don't have specific zone
                 * requirements. Keep the flags related to allocation in atomic
                 * contexts and I/O.
                 */
                alloc_flags &= ~GFP_ZONEMASK;
>>>>>>>         alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
                alloc_flags |= __GFP_NOWARN;
                page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);

It masks masks out anything other than GFP_ATOMIC and GFP_KERNEL
related flags. This drops __GFP_NOLOCKDEP on the floor, hence
lockdep tracks an allocation in a context we've explicitly said not
to track. Hence lockdep (correctly!) explodes later when the
false positive "lock inode in reclaim context" situation triggers.

This is a KASAN bug. It should not be dropping __GFP_NOLOCKDEP from
the allocation context flags.

[1] https://lore.kernel.org/linux-xfs/20230119045253.GI360264@dread.disaster.area/