Re: small perfctr bug or misunderstanding

[bert hubert <ahu@ds9a.nl>]

On Sat, Jul 03, 2004 at 04:58:10PM +0200, Mikael Pettersson wrote:
> Currently no; I removed them while we've been debating the
> API to the (IMO more important) per-process counters.
> I intend to add them back once the current stuff has been
> Linus-approved.

Ok - I'd love the ability to diagnose an entire system. Furthermore, it'd be
very cool if it were possible to profile another process, like strace -p
pid.

I think this means looking at 'virtual counters' for arbitrary processes.
Would this be possible?

I currently have a client using a 2.6.7 kernel and they have performance
problems and applications I can't recompile. It'd be very good if I could
spot which of their many application is thrashing the cache.

> The driver sees ENABLE set in EVNTSEL1 on your P-M,
> and properly returns an error.

Ahhhh, I see. With this line things work as intended:
d_control.cpu_control.evntsel[count] = v | (1 << 16) | (!count << 22) | (unit << 8); 

> handle any quirks. For P6 vs K7 the differences are
> minor, but to program the P4 you _really_ need helper
> procedures.

Indeed. Thanks. I'll make a P6PerfCtr and an AMDPerfCtr and a P4PerfCtr. The
pentium 1/2 people can work it out for themselves :-)

Regards,

bert

-- 
http://www.PowerDNS.com      Open source, database driven DNS Software 
http://lartc.org           Linux Advanced Routing & Traffic Control HOWTO