Re: [PATCH] CPU time clock support in clock_* syscalls

[Roland McGrath]

> Your approach means that only a part of the range may be used. 
> What happens if a pid of say 2^31-10 is used with your API?

That is certainly true.  On a 32-bit machine, a PID value above 2^29-1
cannot be used, by definition.  So there is no "what happens", it just can't.
Userland will need to reject huge PIDs before trying to encode them.
PID_MAX_LIMIT is 2^22, so no actually valid PID can be a problem in practice.

> Does this approach take into consideration that the TSC or cputimer may be
> different on each CPU of a SMP system? 

Yes.  If you consider the methodology I described, you'll see that it does.
The absolute sched_clock time is never relevant here, so it doesn't matter
that it differs between CPUs.  I take a sample when the thread gets
scheduled, and another when it gets descheduled (and perhaps others on the
timer interrupts in between).  So all I ever use is the difference between
two samples taken on the same CPU.

> I just reviewed the code and to my surprise the simple things like
> 
> clock_gettime(CLOCK_PROCESS_CPUTIME_ID) and
> clock_gettime(CLOCK_THREAD_CPUTIME_ID) are not supported. 

You seem to be confused.  A clockid_t for a CPU clock encodes a PID, which
can be zero to indicate the current thread or current process.  

> The thread specific time measurements have nothing to do with the posix
> standard and may best be kept separate.

Nonsense.  POSIX defines the notion of CPU clocks for these calls, and that
is what I have implemented.

Of course glibc is in charge of what the meaning of the POSIX APIs is.
That is true for every call.


Thanks,
Roland
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/