Philip J. Mucci wrote:
Hi Will,
Thanks very much for the report, it helps me greatly. I think I know
what the problem is, it is those pesky umasks. One of the problems being
a virtual employee of no less than 3 institutions is my lack of control
over real hardware. I'm beginning to think about soliciting donations
for old boxes to test perfmon on. HP would rather kill me than give
outsiders access to their machines and I assume Redhat is nearly as
careful. We of course have dozens of machines at UT, but as I said,
their perfctr bound due to contractual support of production systems.
Sound familiar? Ugh.
Hi Phil,
Virtual employee at three institutions?!! Only three institutions. :)
Yes, I have had similar issues testing. You end up collecting machines for this
kind of work, because the development requires root access and there are a
variety of performance monitoring hardware. Getting root access inside a
corporation firewall is a difficult thing. Pretty much have to be an employee
there. Maybe if lucky there are tests systems setup outside the firewall e.g.
Fedora.
http://fedoraproject.org/wiki/FedoraTesting
I am willing to try things out as needed to help track down these problems. I
would like to see perfmon really get execised and in the upstream kernel. Having
PAPI use it would be one of the motivations to get it into the kernel.
I looked at the code in perfmon_rw.c to see what triggers the error message. It
test to see that the pmc is in range and that the pmc is available. Looking at
/sys/kernel/perfmon/pmu_desc
$ more /sys/kernel/perfmon/pmu_desc/
counter_width pmc1/ pmc3/ pmd2/
model pmc2/ pmd1/ pmd3/
Notice there is no pmc0 listed. I assume it is being reserved for a watchdog
timer. Doe the PAPI code check to see whether the register is available?
In regards to your lockup, Stefane has informed me that numerous systems
have issue with printk...I have seen hard lockups with this as well when
I enabled debugging, both on my p3 laptop and various MIPS beasties
I thought that the prink went to a circular buffer and there should be a problem
with lockup.
BTW, 2 totally unrelated things. a) Did you ever get that message I sent
from the broadcom guys about fixes to the event maps for pfm? (Which I
generated from oprofile's files, did you see that script? I've attached
it for fun. It does not handle umasks.) b) I have updated
perfmon/oprofile kernel support in my copy of the kernel tree for MIPS.
I haven't tested it, but will soon. I know there are still a few things
broken in user land (clockrate for example), but I want to see if at
least the oprofile+perfmon kernel component is doing the right thing.
Will keep you posted.
Again, thanks for you help.
phil
_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/
