Re: [PATCH] fs: fix a data race in i_size_write/i_size_read

[David Sterba]

On Tue, Feb 18, 2020 at 11:04:26PM -0500, Qian Cai wrote:
> inode::i_size could be accessed concurently as noticed by KCSAN,
> 
>  BUG: KCSAN: data-race in iomap_do_writepage / iomap_write_end
> 
> The write is protected by exclusive inode::i_rwsem (in
> xfs_file_buffered_aio_write()) but the read is not. A shattered value
> could introduce a logic bug. Fixed it using a pair of WRITE/READ_ONCE().

We had a different problem with lack of READ_ONCE/WRITE_ONCE for i_size,
the fix was the same though. There was i_size_read(inode) used in max()
macro and compiled code two reads (unlocked), and this led to a race
when where different value was checked and then used.

The thread:
https://lore.kernel.org/linux-fsdevel/20191011202050.8656-1-josef@xxxxxxxxxxxxxx/

We had to apply a workaround to btrfs code because the generic fix was
not merged, even with several reviews and fixing a real bug. The report
from KCSAN seems to require some sort of splitting the values. What we
saw happened on 64bit platform without that effect so I'd call that a
more likely to happen because the pattern max(i_size_read(inode), ...) is
not something we'd instinctively consider unsafe.