Re: [jlayton:ctime] [ext4] ff9aaf58e8: ltp.statx06.fail

[Jeff Layton <jlayton@kernel.org>]

On Tue, 2023-05-02 at 10:39 +1000, Dave Chinner wrote:
> On Mon, May 01, 2023 at 12:05:17PM -0400, Jeff Layton wrote:
> > On Mon, 2023-05-01 at 22:09 +0800, kernel test robot wrote:
> > The test does this:
> > 
> >         SAFE_CLOCK_GETTIME(CLOCK_REALTIME_COARSE, &before_time);
> >         clock_wait_tick();
> >         tc->operation();
> >         clock_wait_tick();
> >         SAFE_CLOCK_GETTIME(CLOCK_REALTIME_COARSE, &after_time);
> > 
> > ...and with that, I usually end up with before/after_times that are 1ns
> > apart, since my machine is reporting a 1ns granularity.
> > 
> > The first problem is that the coarse grained timestamps represent the
> > lower bound of what time could end up in the inode. With multigrain
> > ctimes, we can end up grabbing a fine-grained timestamp to store in the
> > inode that will be later than either coarse grained time that was
> > fetched.
> > 
> > That's easy enough to fix -- grab a coarse time for "before" and a fine-
> > grained time for "after".
> > 
> > The clock_getres function though returns that it has a 1ns granularity
> > (since it does). With multigrain ctimes, we no longer have that at the
> > filesystem level. It's a 2ns granularity now (as we need the lowest bit
> > for the flag).
> 
> Why are you even using the low bit for this? Nanosecond resolution
> only uses 30 bits, leaving the upper two bits of a 32 bit tv_nsec
> field available for internal status bits. As long as we mask out the
> internal bits when reading the VFS timestamp tv_nsec field, then
> we don't need to change the timestamp resolution, right?
> 

Yeah, that should work. Let me give that a shot on the next pass.

Thanks,
-- 
Jeff Layton <jlayton@kernel.org>