Re: [PATCH v6] posix-timers: add clock_compare system call

[Arnd Bergmann]

On Tue, Mar 12, 2024, at 13:15, Sagi Maimon wrote:
> On Tue, Mar 12, 2024 at 1:19 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
>> On Tue, Mar 12, 2024, at 10:50, Sagi Maimon wrote:
>> > +     kc_a = clockid_to_kclock(clock_a);
>> > +     if (!kc_a) {
>> > +             error = -EINVAL;
>> > +             return error;
>> > +     }
>> > +
>> > +     kc_b = clockid_to_kclock(clock_b);
>> > +     if (!kc_b) {
>> > +             error = -EINVAL;
>> > +             return error;
>> > +     }
>>
>> I'm not sure if we really need to have it generic enough to
>> support any combination of clocks here. It complicates the
>> implementation a bit but it also generalizes the user space
>> side of it.
>>
>> Can you think of cases where you want to compare against
>> something other than CLOCK_MONOTONIC_RAW or CLOCK_REALTIME,
>> or are these going to be the ones that you expect to
>> be used anyway?
>>
> sure, one example is syncing two different PHCs (which was originally
> why we needed this syscall)
> I hope that I have understand your note and that answers your question.

Right, that is clearly a sensible use case.

I'm still trying to understand the implementation for the case
where you have two different PHCs and both implement 
clock_get_crosstimespec(). Rather than averaging between
two snapshots here, I would expect this to result in
something like

      ktime_a1 += xtstamp_b.sys_monoraw - xtstamp_a1.sys_monoraw;

in order get two device timestamps ktime_a1 and ktime_b
that reflect the snapshots as if they were taken
simulatenously. Am I missing some finer detail here,
or is this something you should do?

>> > +     if (crosstime_support_a) {
>> > +             ktime_a1 = xtstamp_a1.device;
>> > +             ktime_a2 = xtstamp_a2.device;
>> > +     } else {
>> > +             ktime_a1 = timespec64_to_ktime(ts_a1);
>> > +             ktime_a2 = timespec64_to_ktime(ts_a2);
>> > +     }
>> > +
>> > +     ktime_a = ktime_add(ktime_a1, ktime_a2);
>> > +
>> > +     ts_offs = ktime_divns(ktime_a, 2);
>> > +
>> > +     ts_a1 = ns_to_timespec64(ts_offs);
>>
>> Converting nanoseconds to timespec64 is rather expensive,
>> so I wonder if this could be changed to something cheaper,
>> either by returning nanoseconds in the end and consistently
>> working on those, or by doing the calculation on the
>> timespec64 itself.
>>
> I prefer returning timespec64, so this system call aligns with other
> system calls like clock_gettime for example.
> As far as doing the calculation on timespec64 itself, that looks more
> expansive to me, but I might be wrong.

In the general case, dividing a 64-bit variable by some other
variable is really expensive and will take hundreds of cycles.
This one is a bit cheaper because the division is done using
a constant divider of NS_PER_SEC, which can get optimized fairly
well on many systems by turning it into an equivalent 128-bit
multiplication plus shift.

For the case where you start out with a timespec64, I would
expect it to be cheaper to calculate the nanosecond difference
between ts_a1 and ts_a2 to add half of that to the timespec
than to average two large 64-bit values and convert that back
to a timespec afterwards. This should be fairly easy to try
out if you can test a 32-bit kernel. We could decide that
there is no need to care about anything bug 64-bit kernels
here, in which case your current version should be just as
good for both the crosstime_support_a and !crosstime_support_a
cases.

     Arnd