Hi Manoje,

So thanks for running those experiments for me. What I seem to have
found is that there is some instability in the counts that is very
infrequent. It doesn't seem to matter if I use one counter or two or
whether the system is loaded or not. Another modification I made to self
was that I removed the '3 passes' and just did '1 pass' but I moved the
load context up before the write_pmc/write_pmd pair. That also caused
intermittent failures...and they were VERY intermittent...sometimes 1
out of 50 runs. If I put a loop around the
writepmd/start/loop/stop/read/print section, I could never get the
failures to appear. I am putting more debugging into the kernel and will
be testing it after I get back from Sri Lanka here in a few weeks. I was
thinking about maybe why you didn't see the error on your system...I
think it could be because of the clock rate...We're on a malta which is
a 24K system running at 640Mhz with no L2 cache.

Another problem I had spotted was with self notification, but that
didn't show up in the libpfm test suite, In PAPI I use the sigpoll.si_fd
value from the signal handler to find out which FD I should read() the
data from. If you look in notify_self, you'll see that Stefane cheats
and stores it in a global variable. ;-) (BTW, Stefane this should be
changed to read it from the context...it's more useful to users that way
and tests more of the system) Now, I'm using a 64 bit 2.6.15 kernel with
64 bit user land and the si_fd was zero according to the sig handler.
Inspecting the signal context indicated that there was an wrong offset
somewhere...because the correct FD was in the structure, just at an
offset I didn't expect.

I would recommend that on your system, you try digging out the si_fd
from the signal context and see if that value is 0 or if it is a valid
context. That would be a linux bug (maybe in the libc headers) and needs
to be fixed.

Well, that's the latest.

Regards,

Phil




_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/

Reply via email to