Re: [PATCH 1/2] btrfs: convert block group refcount to refcount_t

[David Sterba <dsterba@suse.cz>]

On Wed, Jul 01, 2020 at 04:22:18PM -0400, Josef Bacik wrote:
> I was experimenting with some new allocator changes and I noticed that I
> was getting a UAF with the block groups.  In order to help figure this
> out I converted the block group to use the refcount_t infrastructure.
> This is a generally good idea anyway, so kill the atomic and use
> refcount_t so we can get the benefit of loud complaints when refcounting
> goes wrong.

I don't mind adding some background or motivation of the patch but the
technical part and explanation should be still there and after reading
this paragraph I'm still mssing it.

> Signed-off-by: Josef Bacik <josef@toxicpanda.com>
> ---
>  fs/btrfs/block-group.c | 8 ++++----
>  fs/btrfs/block-group.h | 2 +-
>  2 files changed, 5 insertions(+), 5 deletions(-)
> 
> diff --git a/fs/btrfs/block-group.c b/fs/btrfs/block-group.c
> index 09b796a081dd..7c0075413b08 100644
> --- a/fs/btrfs/block-group.c
> +++ b/fs/btrfs/block-group.c
> @@ -118,12 +118,12 @@ u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
>  
>  void btrfs_get_block_group(struct btrfs_block_group *cache)
>  {
> -	atomic_inc(&cache->count);
> +	refcount_inc(&cache->count);
>  }
>  
>  void btrfs_put_block_group(struct btrfs_block_group *cache)
>  {
> -	if (atomic_dec_and_test(&cache->count)) {
> +	if (refcount_dec_and_test(&cache->count)) {
>  		WARN_ON(cache->pinned > 0);
>  		WARN_ON(cache->reserved > 0);
>  
> @@ -1805,7 +1805,7 @@ static struct btrfs_block_group *btrfs_create_block_group_cache(
>  
>  	cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
>  
> -	atomic_set(&cache->count, 1);
> +	refcount_set(&cache->count, 1);
>  	spin_lock_init(&cache->lock);
>  	init_rwsem(&cache->data_rwsem);
>  	INIT_LIST_HEAD(&cache->list);
> @@ -3427,7 +3427,7 @@ int btrfs_free_block_groups(struct btrfs_fs_info *info)
>  		ASSERT(list_empty(&block_group->dirty_list));
>  		ASSERT(list_empty(&block_group->io_list));
>  		ASSERT(list_empty(&block_group->bg_list));
> -		ASSERT(atomic_read(&block_group->count) == 1);
> +		ASSERT(refcount_read(&block_group->count) == 1);
>  		btrfs_put_block_group(block_group);
>  
>  		spin_lock(&info->block_group_cache_lock);
> diff --git a/fs/btrfs/block-group.h b/fs/btrfs/block-group.h
> index b6ee70a039c7..705e48050163 100644
> --- a/fs/btrfs/block-group.h
> +++ b/fs/btrfs/block-group.h
> @@ -115,7 +115,7 @@ struct btrfs_block_group {
>  	struct list_head list;
>  
>  	/* Usage count */
> -	atomic_t count;
> +	refcount_t count;

Originally the comment says 'usage', which is a bit different from
refcounts. The atomics will allow 1->0 and 0->1, which might be valid
pattern and reflects the 'usage counter' semantics. The refcounts catch
the 0->1 increment as bug and complain. This is what we want and that's
why you switch that.

So I suggest to drop the comment and rename it to 'refs' that we use
for refcounts elsewhere.