Re: [PATCH v2 01/10] btrfs: use an xarray to track open inodes in a root

[David Sterba <dsterba@suse.cz>]

On Fri, May 10, 2024 at 06:32:49PM +0100, fdmanana@kernel.org wrote:
> From: Filipe Manana <fdmanana@suse.com>

The xarray pattern in this patch seems to solve problem with
preallocation and search, I was stuck on that with the extent buffer
conversion from radix -> xarray, so I'm only clarifying things loud for
myself:

> +	ret = xa_reserve(&root->inodes, ino, GFP_NOFS);
> +	if (ret)
> +		return ret;

Unlocked reservation with NOFS

> +
>  	spin_lock(&root->inode_lock);
> -	p = &root->inode_tree.rb_node;
> -	while (*p) {
> -		parent = *p;
> -		entry = rb_entry(parent, struct btrfs_inode, rb_node);
> +	existing = xa_store(&root->inodes, ino, inode, GFP_ATOMIC);

Locked insertion with preallocation, thus no ATOMIC needed

>  	spin_lock(&root->inode_lock);

...

> +		inode = xa_find(&root->inodes, &from, ULONG_MAX, XA_PRESENT);

Locked search that does not need to handle the preallocated but not
inserted element.

What I did with the delayed inodes tree was xa_reserve + xa_load, this
needs the special handling of potentially finding NULL by xa_load. This
worked for the delayed notes but not for the extent buffers. We'd have
to do special cases for that, loop and wait, but xa_find(..., XA_PRESENT)
seems to be doing exactly that.