On Tue, Jan 26, 2021 at 09:30:45AM -0500, Josef Bacik wrote:

On 1/26/21 4:02 AM, Nikolay Borisov wrote:

quoted

On 25.01.21 г. 23:42 ч., Josef Bacik wrote:

quoted

In __btrfs_return_cluster_to_free_space we will bail doing the cleanup
of the cluster if the block group we passed in doesn't match the block
group on the cluster.  However we drop a reference to block_group, as
the cluster holds a reference to the block group while it's attached to
the cluster.  If cluster->block_group != block_group however then this
is an extra put, which means we'll go negative and free this block group
down the line, leading to a UAF.

Was this found by code inspection or did you hit in production. Also why
in btrfs_remove_free_space_cache just before
__btrfs_return_cluster_to_free_space there is:

It was found in production sort of halfway.  I was doing something for WhatsApp 
and had to convert our block group reference counting to the refcount stuff so I 
could find where I made a mistake.  Turns out this was where the problem was, my 
stuff had just made it way more likely to happen.  I don't have the stack trace 
because this was like 6 months ago, I'm going through all my WhatsApp magic and 
getting them actually usable for upstream.

quoted

WARN_ON(cluster->block_group != block_group);

IMO this patch should also remove the WARN_ON if it's a valid condition
to have the passed bg be different than the one in the cluster. Also
that WARN_ON is likely racy since it's done outside of cluster->lock.

Yup that's in a follow up thing, I wanted to get the actual fix out before I got 
distracted by my mountain of meetings this week.  Thanks,

Removing the WARN_ON in a separate patch sounds ok to me, this patch
clearly fixes the refcounting bug, the warning condition is the same but
would need a different reasoning.

Nikolay, if you're ok with current patch version let me know if you want
a rev-by added.