On Mon, Feb 08, 2021 at 10:26:54AM +0200, Nikolay Borisov wrote:
> During allocation the allocator will try to allocate an extent using
> cluster policy. Once the current cluster is exhausted it will remove the
> its entry under btrfs_free_cluster::lock and subsequently acquire
> btrfs_free_space_ctl::tree_lock to dispose of the already-deleted
> entry and adjust btrfs_free_space_ctl::total_bitmap. This poses a
> problem because there exists a race condition between removing the
> entry under one lock and doing the necessary accounting holding a
> different lock since extent freeing only uses the 2nd lock. This can
> result in the following situation:
>
> T1: T2:
> btrfs_alloc_from_cluster insert_into_bitmap <holds tree_lock>
> if (entry->bytes == 0) if (block_group &&
> !list_empty(&block_group->cluster_list)) {
> rb_erase(entry)
>
> spin_unlock(&cluster->lock);
> (total_bitmaps is still 4) spin_lock(&cluster->lock);
> <doesn't find entry in cluster->root>
> spin_lock(&ctl->tree_lock); <goes to new_bitmap label, adds
> <blocked since T2 holds tree_lock> <a new entry and calls
> add_new_bitmap>
> recalculate_thresholds <crashes,
> due to total_bitmaps
> becoming 5 and triggering
> an ASSERT>
>
> To fix this ensure that once depleted, the cluster entry is deleted when
> both cluster lock and tree locks are held in the allocator (T1), this
> ensures that even if there is a race with a concurrent
> insert_into_bitmap call it will correctly find the entry in the cluster
> and add the new space to it.
>
> Signed-off-by: Nikolay Borisov <nbori...@suse.com>
> Cc: <sta...@vger.kernel.org>
Added to for-next, targeting 5.12-rc, thanks.
