Re: [PATCH 1/7] cpu-timers: high-resolution CPU clocks for POSIX clock_* syscalls

[Andi Kleen <ak@suse.de>]

Roland McGrath <roland@redhat.com> writes:
>  
>  /*
> + * This is called on clock ticks and on context switches.
> + * Bank in p->sched_time the ns elapsed since the last tick or switch.
> + */
> +static void update_cpu_clock(task_t *p, runqueue_t *rq,
> +			     unsigned long long now)
> +{
> +	unsigned long long last = max(p->timestamp, rq->timestamp_last_tick);
> +	p->sched_time += now - last;
> +}

This will completely mess up the register allocation in schedule()
long long on i386 forces basically everything else out onto the stack
because it needs 4 aligned registers.

I suspect when you benchmark it it will become visibly slower.

In general it seems like a bad idea to polute the extremly critical
fast paths in schedule with support for such an obscure operation.
Is there really any real need for such a high resolution per process
timer anyways? I have my doubts about it, I would suspect most apps
are more interested in wall clock time.

I don't think this should be merged until a clear need from a useful
application is demonstrated for it.

-Andi