Re: [PATCH] btrfs: fix encoded write i_size corruption

[Josef Bacik <josef@toxicpanda.com>]

On Fri, Apr 28, 2023 at 02:02:11PM -0700, Boris Burkov wrote:
> We have observed a btrfs filesystem corruption on workloads using
> NOHOLES and encoded writes via sendstream v2. The symptom is that a file
> appears to be truncated to the end of its last aligned extent, even
> though the final unaligned extent and even the file extent and otherwise
> correctly updated inode item have been written.
> 
> So if we were writing out a 1MiB+X file via 8 128K extents and one
> extent of length X, isize would be set to 1MiB, but the ninth extent,
> nbyte, etc.. would all appear correct otherwise.
> 
> The source of the race is a narrow (one code line..) window in which a
> noholes fs has read in an updated isize, but has not yet set a shared
> disk_i_size variable to write. Therefore, if two ordered extents run in
> parallel (par for the course for receive workloads), the following
> sequence can play out: (following "threads" a bit loosely, since there
> are callbacks involved for endio but extra threads aren't needed to
> cause the issue)
> 
> ENC-WR1 (second to last)                                         ENC-WR2 (last)
> -------                                                          -------
> btrfs_do_encoded_write
>   set isize = 1M
>   submit bio B1 ending at 1M
> endio B1
> btrfs_inode_safe_disk_i_size_write
>   local isize = 1M
>   falls off a cliff for some reason
>                                                             btrfs_do_encoded_write
>                                                               set isize = 1M+X
>                                                               submit bio B2 ending at 1M+X
>                                                             endio B2
> 							    btrfs_inode_safe_disk_i_size_write
>                                                               local isize = 1M+X
>                                                               disk_i_size = 1M+X
>   disk_i_size = 1M
> 							    btrfs_delayed_update_inode
>   btrfs_delayed_update_inode
> 
> And the delayed inode ends up filled with nbytes=1M+X and isize=1M, and
> writes respect isize and present a corruted file missing its last
> extents.
> 
> Fix this by holding the inode lock in the noholes case so that a thread
> can't sneak in a write to disk_i_size that gets overwritten with an out
> of date isize.
> 
> Fixes: 41a2ee75aab0290 btrfs: introduce per-inode file extent tree
> Signed-off-by: Boris Burkov <boris@bur.io>

Reviewed-by: Josef Bacik <josef@toxicpanda.com>

Thanks,

Josef