Re: [BUG] fs: btrfs: several possible ABBA deadlocks

[Josef Bacik]

On Sat, Nov 27, 2021 at 04:23:37PM +0800, Jia-Ju Bai wrote:
> Hello,
> 
> My static analysis tool reports several possible ABBA deadlocks in the btrfs
> module in Linux 5.10:
> 
> # DEADLOCK 1:
> __clear_extent_bit()
>   spin_lock(&tree->lock); --> Line 733 (Lock A)
>   split_state()
>     btrfs_split_delalloc_extent()
>       spin_lock(&BTRFS_I(inode)->lock); --> Line 1870 (Lock B)
> 
> btrfs_inode_safe_disk_i_size_write()
>   spin_lock(&BTRFS_I(inode)->lock); --> Line 53 (Lock B)
>   find_contiguous_extent_bit()
>     spin_lock(&tree->lock); --> Line 1620 (Lock A)
> 
> When __clear_extent_bit() and btrfs_inode_safe_disk_i_size_write() are
> concurrently executed, the deadlock can occur.
> 
> # DEADLOCK 2:
> __set_extent_bit()
>   spin_lock(&tree->lock); --> Line 995 (Lock A)
>   set_state_bits()
>     btrfs_set_delalloc_extent()
>       spin_lock(&BTRFS_I(inode)->lock); --> Line 2007 or 2017 or 2029 (Lock
> B)
> 
> btrfs_inode_safe_disk_i_size_write()
>   spin_lock(&BTRFS_I(inode)->lock); --> Line 53 (Lock B)
>   find_contiguous_extent_bit()
>     spin_lock(&tree->lock); --> Line 1620 (Lock A)
> 
> When __set_extent_bit() and btrfs_inode_safe_disk_i_size_write() are
> concurrently executed, the deadlock can occur.
> 
> # DEADLOCK 3:
> convert_extent_bit()
>   spin_lock(&tree->lock); --> Line 1241 (Lock A)
>   set_state_bits()
>     btrfs_set_delalloc_extent()
>       spin_lock(&BTRFS_I(inode)->lock); --> Line 2007 or 2017 or 2029 (Lock
> B)
> 
> btrfs_inode_safe_disk_i_size_write()
>   spin_lock(&BTRFS_I(inode)->lock); --> Line 53 (Lock B)
>   find_contiguous_extent_bit()
>     spin_lock(&tree->lock); --> Line 1620 (Lock A)
> 
> When convert_extent_bit() and btrfs_inode_safe_disk_i_size_write() are
> concurrently executed, the deadlock can occur.
> 
> I am not quite sure whether these possible deadlocks are real and how to fix
> them if they are real.
> Any feedback would be appreciated, thanks :)
> 

Hey Jia-Ju,

This is pretty good work, unfortunately it's wrong but it's in a subtle way that
a tool wouldn't be able to catch.  The btrfs_inode_safe_disk_i_size_write()
helper only messes with BTRFS_I(inode)->file_extent_tree, which is separate from
the BTRFS_I(inode)->io_tree.  io_tree gets the btrfs_set_delalloc_extent() stuff
called on it, but the file_extent_tree does not.  The file_extent_tree has
inode->lock -> tree->lock as the locking order, whereas the file_extent_tree has
inode->lock -> tree->lock as the locking order.  Thanks,

Josef