On Mon, Jan 30, 2006 at 06:54:29PM +0000, Philip Mucci wrote:
> In reference to the performance counting thread we had going earlier,
> I've noticed a 'feature' I need out of MIPS/Linux that isn't currently
> available. This has also recently come up on the oprofile list with one
> of the oprofile/mips tools not being able to grab cpu Mhz
> from /proc/cpuinfo because it's not there.
Surprise - the kernel doesn't actually have that information. The closest
available would be the counting speed of the c0 count register which - if
something like that is available at all - could be calibrated against an
external counter to figure out the clock speed. To complicate matters
not all processors have have a c0 counter, it's sometimes incrementing
every cycle but it might just as well counting at half the clock rate and
oh yes, more recently there is clock scaling.
> I have need to execute the mfc0 instruction on the config register and
> grok the results to find out things like cache size etc. In addition, it
> might be nice to also actually be able to find out the clock rate.
> (Currently I grab BogoMIPS and punt.)
Considering above complexities a BogoMIPS-based approach doesn't sound too
bad - but it has other problem to implement since it requires knowing the
exact execution timing. A think that could be done is meassuring two
loops containing a different number of nops and use that to compute the
execution time for the loop closing branch. Now knowing that based on
knowing the # of cycles per loop we can compute the clock speed - all it
takes is a timer running at a known speed.
A funny complexity is that some multiprocessor MIPS systems have processors
running at different speeds. So knowing the clock rate doesn't make life
too much easier ;-)
> On the intel and PPC systems, I believe you can execute similar
> instructions from user mode which makes things easy. However, of course
> an MFC0 is a privileged instruction...meaning that if the value or
> values aren't found in /proc/cpuinfo, I'm s.o.l.
We have a long tradition of emulating instructions, could use that to
permit access to a bunch of cp0 registers from userland ;)
> What does the list think about this? Making a mips /dev/cpuid is a bit
> gross but extending and grokking /proc/cpuinfo is perhaps grosser...and
> many tools do just this (like PAPI and oprofile's opreport...)
> Comments? I'm certainly willing to implement this, but I'd rather 'do it
> right the first time' rather than get rotten vegetables thrown my way.
/proc/cpuinfo is meant to have per processor information so I don't have
a problem with adding cache configuration information. I was considering
to do so to make such information available for bug reports.