Re: [PATCH v2] timekeeping: move multigrain timestamp floor handling into timekeeper

[John Stultz <jstultz@google.com>]

On Thu, Sep 12, 2024 at 1:33 PM John Stultz <jstultz@google.com> wrote:
>
> On Thu, Sep 12, 2024 at 1:18 PM Jeff Layton <jlayton@kernel.org> wrote:
> > On Thu, 2024-09-12 at 13:11 -0700, John Stultz wrote:
> > > On Thu, Sep 12, 2024 at 11:02 AM Jeff Layton <jlayton@kernel.org> wrote:
> > > But instead, it seems like if something has happened since the cookie
> > > value was saved (another cpu getting a fine grained timestamp), your
> > > ktime_get_real_ts64_mg() will fall back to returning the same coarse
> > > grained time saved to the cookie, as if no time had past?
> > >
> > > It seems like that could cause problems:
> > >
> > > cpu1                                     cpu2
> > > --------------------------------------------------------------------------
> > >                                          t2a = ktime_get_coarse_real_ts64_mg
> > > t1a = ktime_get_coarse_real_ts64_mg()
> > > t1b = ktime_get_real_ts64_mg(t1a)
> > >
> > >                                          t2b = ktime_get_real_ts64_mg(t2a)
> > >
> > > Where t2b will seem to be before t1b, even though it happened afterwards.
> > >
> >
> > Ahh no, the subtle thing about atomic64_try_cmpxchg is that it
> > overwrites "old" with the value that was currently there in the event
> > that the cmp fails.
>
> Ah, ok. Thank you for the explanation there!
>
> > So, the try_cmpxchg there will either swap the new value into place, or
> > if it was updated in the meantime, "old" will now refer to the value
> > that's currently in the floor word. Either is fine in this case, so we
> > don't need to retry anything.
>
>
> Though if cpu2 then made another call to
> ktime_get_coarse_real_ts64_mg(), the value returned there will be the
> same as t1b? and would be before t2b?

Oh, no. Apologies again, as I see  t2b would be the same as t1b as well. Ok.
-john