Re: [PATCH 09/10] fs: ceph: Replace CURRENT_TIME by ktime_get_real_ts()

[Arnd Bergmann]

On Thursday 04 February 2016 10:01:31 Ilya Dryomov wrote:
> On Thu, Feb 4, 2016 at 9:30 AM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > On Thursday 04 February 2016 10:00:19 Yan, Zheng wrote:
> >> > On Feb 4, 2016, at 05:27, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> >
> > static inline void ceph_decode_timespec(struct timespec *ts,
> >                                         const struct ceph_timespec *tv)
> > {
> >         ts->tv_sec = (__kernel_time_t)le32_to_cpu(tv->tv_sec);
> >         ts->tv_nsec = (long)le32_to_cpu(tv->tv_nsec);
> > }
> >
> > Is that intentional and documented? If yes, what is your plan to deal
> > with y2038 support?
> 
> tv_sec is used as a time_t, so signed.  The problem is that ceph_timespec is
> not only passed over the wire, but is also stored on disk, part of quite a few
> other data structures. 

That is only part of the issue though:

Most file systems that store a timespec on disk define the function
differently:

static inline void ceph_decode_timespec(struct timespec *ts,
                                        const struct ceph_timespec *tv)
{
        ts->tv_sec = (time_t)(u32)le32_to_cpu(tv->tv_sec);
        ts->tv_nsec = (long)le32_to_cpu(tv->tv_nsec);
}

On systems that have a 64-bit time_t, the 1902..1970 interval
(0xffffffff80000000..0xffffffffffffffff) and the 2038..2106
interval (0x0000000080000000..0x00000000ffffffff) are written
as the same 32-bit numbers, so when reading back you have to
decide which interpretation you want, and your cast to
__kernel_time_t means that you get the first representation on
both 32-bit and 64-bit systems.

On systems with a 32-bit time_t, this is the only option you
have anyway, and some other file systems (ext2/3/4, xfs, ...)
made the same decision in order to behave in a consistent way
independent of what kernel (32-bit or 64-bit) you use. This
is generally a reasonable goal, but it means that you get the
overflow in 2038 rather than 2106.

Alex Elder changed the cephs behavior in 2013 to be the same
way, but from the changelog c3f56102f28d ("libceph: validate
timespec conversions"), I guess this was not intentional, as
he was also adding a comparison against U32_MAX, which should
have been S32_MAX.

A lot of other file systems (jfs, jffs2, hpfs, minix) apparently
prefer the 1970..2106 interpretation of time values.

> The plan is to eventually switch to a 64-bit tv_sec and
> tv_nsec, bump the version on all the structures that contain it and add
> a cluster-wide feature bit to deal with older clients.  We've recently had
> a discussion about this, so it may even happen in a not so distant future, but
> no promises 

Ok. We have a (rough) plan to deal with file systems that don't support
extended time stamps in the meantime, so depending on user preferences
we would either allow them to be used as before with times clamped
to the 2038 overflow date, or only mounted readonly for users that want
to ensure their systems can survive without regressions in 2038.

	Arnd