On Sat, May 31, 2014 at 05:37:52PM +0200, Arnd Bergmann wrote:
> On Saturday 31 May 2014 11:14:50 Dave Chinner wrote:
> > On Fri, May 30, 2014 at 05:41:14PM -0700, H. Peter Anvin wrote:
> > > On 05/30/2014 05:37 PM, Dave Chinner wrote:
> > > >
> > > > IOWs, the filesystem has to be able to reject any attempt to set a
> > > > timestamp that is can't represent on disk otherwise Bad Stuff will
> > > > happen,
> > >
> > > Actually it is questionable if it is worse to reject a timestamp or just
> > > let it wrap. Rejecting a valid timestamp is a bit like "You don't
> > > exist, go away."
> > I think having the new systems calls being able to
> > return EINVAL if the value cannot be stored permanently on disk
> > correctly is the right thing to do. Having it silently mangled
> > by the filesystem and returning "everything is just fine, trust me"
> > is close to the worst solution I can think of. That's exactly what
> > leads to overflow bugs occurring....
> While going through the file systems, I was wondering whether
> we should have the times stop at the end of each file systems
> epoch rather than wrap around.
> > > > and filesystems have to be able to specify in their on
> > > > disk format what timestamp encoding is being used. The solution will
> > > > be different for every filesystem that needs to support time beyond
> > > > 2038.
> > >
> > > Actually the cutoff can be really different for each filesystem, not
> > > necessarily 2038. However, I maintain the above still holds.
> > Sure, but all filesystems are supposed to handle at least the
> > current unix epoch.
> In my list at http://kernelnewbies.org/y2038, I found that almost
> all file systems at least times until 2106, because they treat
> the on-disk value as unsigned on 64-bit systems, or they use
> a completely different representation. My guess is that somebody
> earlier spent a lot of work on making that happen.
> The exceptions are:
> * exofs uses signed values, which can probably be changed to be
> consistent with the others.
> * isofs has a bug that limits it until 2027 on architectures with
> a signed 'char' type (otherwise it's 2155).
> * udf can represent times for many thousands of years through a
> 16-bit year representation, but the code to convert to epoch
> uses a const array that ends at 2038.
> * afs uses signed seconds and can probably be fixed
> * coda relies on user space time representation getting passed
> through an ioctl.
> * I miscategorized xfs/ext2/ext3 as having unsigned 32-bit seconds,
> where they really use signed.
> I was confused about XFS since I didn't noticed that there are
> separate xfs_ictimestamp_t and xfs_timestamp_t types, so I expected
> XFS to also use the 1970-2106 time range on 64-bit systems today.
You've missed an awful lot more than just the implications for the
core kernel code.
There's a good chance such changes propagate to APIs elsewhere in
the filesystems, because something you haven't realised is that XFS
effectively exposes the on-disk timestamp format directly to
userspace via the bulkstat interface (see struct xfs_bstat). It also
affects the XFS open-by-handle ioctl and the swap extent ioctl used
by the online defragmenter.
IOWs, if we are changing the on-disk timestamp format then this
affects several ioctl()s and hence quite a few of the XFS userspace
utilities. The hardest to fix will be xfsdump which would need a new
dump format to store the extended timestamp ranges, and then
xfs_restore will need to be able to handle restoring such timestamps
on filesystems that don't have extended timestamp support...
Put simply, changing the structure of system time isn't as straight
forward as changing the kernel structures. System time gets stored
permanently, and that has a cascade effect through the kernel all
to all of the filesystem utilities that know about that permanent
storage in some way....
So yes, you can change the kernel definition, but until the
permanent storage of system time can be extended to support the same
range as the kernel the *system* will still have nasty, silent epoch
overflow, truncation or corruption issues.
> If we are using the variant of my patch that extends
> indode_time->tv_sec to s64, nothing should change for XFS
> at all, the main difference is that we if it gets extended
> to wider on-disk timestamps, they will work the same way on
> 32-bit and 64-bit kernels.
"nothing should change" except for the fact that a 64 bit timestamp
gets silently truncated to 32 bits and the timestamp is not what the
user expects it to be. The user does not find out until the inode
passes out of cache and is re-read from disk, and then it's wrong.
To put it politely: that is broken, obnoxious behaviour and we don't
design new interfaces with such ugly warts anymore. Define an
EOVERFLOW, EINVAL or ERANGE error in the new syscalls to handle this
case and *hard fail* if the storage cannot support the extended
timestamp being passed in. There is no excuse for silently mangling
out-of-range data, especially as we have plenty of time to add
support to the filesystems so that such errors don't occur. It might
take us a year to implement, but it will be done long before the
And, FWIW, this patchset needs a set of regression tests that ensure
timestamps beyond 2038 and 2106 don't change across unmount/mount.
Written for xfstests, preferably, so that it's run as part of every
filesystem developer's daily workflow. This is the only way we are
going to ensure that the filesystem and VFS code works correctly and
continues to work correctly up to the end of the current epoch....