On Wed 05-12-18 12:45:06, Alexander Lochmann wrote:
>
> inode.i_flags might be altered without proper
> synchronisation when the inode belongs to devtmpfs.
> The following stacktrace shows how to get there:
> 13: entry_SYSENTER_32:460
> 12: do_fast_syscall_32:410
> 11: _static_cpu_has:146
> 10: do_syscall_32_irqs_on:322
> 09: SyS_pwrite64:636
> 08: SYSC_pwrite64:650
> 07: fdput:38
> 06: vfs_write:560
> 05: __vfs_write:512
> 04: new_sync_write:500
> 03: blkdev_write_iter:1977
> 02: __generic_file_write_iter:2897
> 01: file_remove_privs:1818
> 00: inode_has_no_xattr:3163
>
> Found by LockDoc (Alexander Lochmann, Horst Schirmeier and Olaf
> Spinczyk)
>
> Signed-off-by: Alexander Lochmann <alexander.lochmann@xxxxxxxxxxxxxx>
> Signed-off-by: Horst Schirmeier <horst.schirmeier@xxxxxxxxxxxxxx>
...
> +/*
> + * blkdev_write_iter() can call this without i_rwsem, need to be
> + * careful with i_flags update.
> + */
> static inline void inode_has_no_xattr(struct inode *inode)
> {
> if (!is_sxid(inode->i_mode) && (inode->i_sb->s_flags & SB_NOSEC))
> - inode->i_flags |= S_NOSEC;
> + inode_set_flags(inode, S_NOSEC, S_NOSEC);
> }
Thinking more about this I'm not sure if this is actually the right
solution. Because for example the write(2) can set S_NOSEC flag wrongly
when it would race with chmod adding SUID bit. So probably we rather need
to acquire i_rwsem in blkdev_write_iter() if file does not have S_NOSEC set
(we don't want to acquire it unconditionally as that would heavily impact
scalability of block device writes).
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
