I did some testing on special-file copyup in unionfs, whereby unionfs calls
vfs_mknod on a lower f/s to create (copyup) a device. The copyup
functionality worked fine when unionfs was stacked on top of all file
systems other than xfs. I found out that when I create a device with
<maj,min> on xfs, then I stat(1) through the union, I get the sysv encoding
of the <maj,min>.
The problem appears to be in the xfs_vn_mknod() function in xfs_iops.c: the
code changes the rdev value using rdev = sysv_encode_dev(rdev); so it can
fall through that case of the first switch statement in that function, and
onto xfs_create(). Later in that function, however, it copies the
sysv_encode'd rdev value into the inode that is going to be d_instantiate'd,
thus exposing the sysv encoding to the VFS -- and to any stackable file
system that may be using fsstack_copy_attr_all() to copy lower inode
attributes to an upper inode.
The following small patch seems to fix the problem. Bug and fix were
verified on v2.6.24-rc5-245-gc63a119. Someone with better understanding of
XFS's internals should verify this patch.
Andrew, a related question: kdev_t.h defines several device <maj,min>
encoding formats, which different file systems seems to use. Assuming that
a file system can choose whatever internal encoding it wants, is there a
uniform standard for what the VFS-level inode->i_rdev format should be?
(vfs.txt is silent about the issue.)
Thanks,
Erez.
XFS: don't expose internal sysv encoding to VFS inode->i_rdev
Signed:-off-by: Erez Zadok <[email protected]>
diff --git a/fs/xfs/linux-2.6/xfs_iops.c b/fs/xfs/linux-2.6/xfs_iops.c
index 37e1167..daaa060 100644
--- a/fs/xfs/linux-2.6/xfs_iops.c
+++ b/fs/xfs/linux-2.6/xfs_iops.c
@@ -333,7 +333,8 @@ xfs_vn_mknod(
ip = vn_to_inode(vp);
if (S_ISCHR(mode) || S_ISBLK(mode))
- ip->i_rdev = rdev;
+ ip->i_rdev = MKDEV(sysv_major(rdev) & 0x1ff,
+ sysv_minor(rdev));
else if (S_ISDIR(mode))
xfs_validate_fields(ip);
d_instantiate(dentry, ip);
I've sent Dave a patch yesterday that fixes the problem in an even nicer
way by just removing the assignment in xfs_vn_mknod because we're doing
the proper translation deeper down in the stack. The patch should
already be on it's way to Linus.