On Thu, Oct 04, 2018 at 05:44:47PM -0700, Darrick J. Wong wrote:
> From: Darrick J. Wong <darrick.wong@xxxxxxxxxx>
>
> Refactor all the reflink preparation steps into a separate helper that
> we'll use to land all the upcoming fixes for insufficient input checks.
>
> Signed-off-by: Darrick J. Wong <darrick.wong@xxxxxxxxxx>
.....
> +xfs_reflink_remap_range(
> + struct file *file_in,
> + loff_t pos_in,
> + struct file *file_out,
> + loff_t pos_out,
> + u64 len,
> + bool is_dedupe)
> +{
> + struct inode *inode_in = file_inode(file_in);
> + struct xfs_inode *src = XFS_I(inode_in);
> + struct inode *inode_out = file_inode(file_out);
> + struct xfs_inode *dest = XFS_I(inode_out);
> + struct xfs_mount *mp = src->i_mount;
> + xfs_fileoff_t sfsbno, dfsbno;
> + xfs_filblks_t fsblen;
> + xfs_extlen_t cowextsize;
> + ssize_t ret;
> +
> + if (!xfs_sb_version_hasreflink(&mp->m_sb))
> + return -EOPNOTSUPP;
> +
> + if (XFS_FORCED_SHUTDOWN(mp))
> + return -EIO;
> +
> + /* Prepare and then clone file data. */
> + ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
> + len, is_dedupe);
> + if (ret)
> + return ret;
generic/013 indicates there's a double unlock bug here.
vfs_clone_file_prep_inodes() can return zero (do nothing, but don't
fail!), and when that happens xfs_reflink_remap_prep() unlocks the
inodes and returns 0. This new code doesn't catch it, we do the
remap on unlocked inodes, and then trip lock debugging bugs
> @@ -1300,12 +1351,7 @@ xfs_reflink_remap_range(
> is_dedupe);
>
> out_unlock:
> - xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
> - if (!same_inode)
> - xfs_iunlock(src, XFS_MMAPLOCK_SHARED);
> - inode_unlock(inode_out);
> - if (!same_inode)
> - inode_unlock_shared(inode_in);
> + xfs_reflink_remap_unlock(file_in, file_out);
here:
DEBUG_LOCKS_WARN_ON(sem->owner != get_current())
WARNING: CPU: 3 PID: 4766 at kernel/locking/rwsem.c:133 up_write+0x66/0x70
CPU: 3 PID: 4766 Comm: fsstress Not tainted 4.19.0-rc6-dgc+ #671
....
Call Trace:
xfs_iunlock+0x152/0x220
xfs_reflink_remap_unlock+0x22/0x70
xfs_reflink_remap_range+0x129/0x2a0
do_clone_file_range+0x119/0x200
vfs_clone_file_range+0x35/0xa0
ioctl_file_clone+0x8a/0xa0
do_vfs_ioctl+0x2e1/0x6c0
ksys_ioctl+0x70/0x80
__x64_sys_ioctl+0x16/0x20
do_syscall_64+0x5a/0x180
entry_SYSCALL_64_after_hwframe+0x49/0xbe
I'll fix it for the moment by making xfs_reflink_remap_prep() behave
like vfs_clone_file_prep_inodes() - it will return 1 on success,
0 for nothing to do and < 0 for an error and catch it in this code.
I note that later patches in the series change the
vfs_clone_file_prep_inodes() behaviour so this behaviour is probably
masked by those later changes. It's still a nasty bisect landmine,
though, so I'll fix it here.
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
