Re: [RFC PATCH Mach] Add a mach host operation which returns elapsed time since bootup

[Samuel Thibault]

Hello,

Diego Nieto Cid, le mar. 17 déc. 2024 23:02:01 -0300, a ecrit:
> El mar, 17 dic 2024 a las 22:21, Zhaoming Luo (<zhmingluo@163.com>) escribió:
> >
> > The precision of this implmentation is 10ms. Not sure how to do with the
> > possible data race.
> >
> > Signed-off-by: Zhaoming Luo <zhmingluo@163.com>
> > ---
> >  include/mach/mach_host.defs |  7 +++++++
> >  kern/mach_clock.c           | 20 ++++++++++++++++++++
> >  2 files changed, 27 insertions(+)
> >
> > diff --git a/include/mach/mach_host.defs b/include/mach/mach_host.defs
> > index 8fd9d6b3..67f72cda 100644
> > --- a/include/mach/mach_host.defs
> > +++ b/include/mach/mach_host.defs
> > @@ -386,3 +386,10 @@ routine    host_adjust_time64(
> >  routine        host_get_kernel_version(
> >                 host            : host_t;
> >         out     kernel_version  : new_kernel_version_t);
> > +
> > +/*
> > + *     Get the elapsed time on this host.
> > + */
> > +routine        host_get_elapsed_time(
> > +               host            : host_t;
> > +       out     elapsed_time    : time_value_t);
> 
> If I'm not mistaken, this is most commonly called `uptime` (like in
> `host_get_uptime` or similar)
> 
> > +
> > +       read_elapsed_ticks = elapsed_ticks;
> > +       elapsed_usec = read_elapsed_ticks * tick;
> > +       elapsed_time->seconds = elapsed_usec / MICROSECONDS_IN_ONE_SECOND;
> > +       elapsed_time->microseconds = elapsed_usec % 
> > MICROSECONDS_IN_ONE_SECOND;
> 
> `read_elapsed_ticks` is only used to calculate elapsed_usec and thus
> may be inlined, removing the otherwise unused variable
> 
> As for the data races, I believe you can use atomic builtins[1], like below:
> 
>     elapsed_usec = __sync_fetch_and_add(&elapsed_tikcs, 0) * tick
> 
> Or you could also use the lock used to increment `elapsed_ticks`:
> 
>      s = simple_lock_irq(&timer_lock);
>     elapsed_usec = elapsed_tikcs * tick;
>     simple_unlock_irq(s, &timer_lock);
> 
> I'm not sure which works best.

The former will be way better: it'll just introduce memory barriers to
make sure to get a coherent value, and read it only once (otherwise the
compiler would be allowed to read it several times, for filling
elapsed_time->seconds and elapsed_time->microseconds).

Samuel