Hello,
FYI: don't use the oprofile tools tarball - use the latest from the CVS
site on sourceforge. The last tarball on the website is more than a bit
out of date.
I'll follow up on your other questions in a separate email (in a little
while....).
/Mark
>
> #1: I looked at oprofile-0.9.1 and it lists this event for SB1 in the
> events/mips/sb1 directory
> --------------------------------------------------------------
> ----------
> ---
> event:10 counters:1,2,3 um:zero minimum:500
> name:DCACHE_FILLED_SHD_NONC_EXC :Dcache is filled (shared, nonc,
> exclusive)
> ----------------------------------------------------------------
>
> However this doesn't have an equivalent performance source
> Listed in the sibyte manual (table 33 system performance
> counter sources).
Are you looking in Sb1-UM100-RDS.pdf? This is in table 95 in my copy,
on page 96/97 (section 11). Also note that the table isn't in order --
event #10 appears on the top of the 2nd page. It's logical grouping,
not numeric.
> #2: how does the mapping from event numbers to performance
> sources work for sibyte ?
>
> I looked at the op_model_mipsxx.c file in the 2.6 CVS tree
> and the macro it uses doesn't seem to match the format
> specified for perf_cnt_cfg register in sibyte.
Are you using the kernel from CVS instead of git? The SB1 oprofile
support didn't turn up in the kernel until sometime mid-January, and is
only available through the git version (the CVS version of the kernel is
long out of date).
HTH,
Mark
|