System wide and NUMA domain wide counters support. PMC classes for ARM DMC-620 and CMN-600.

[tsoome]

Summary

Add support for system wide and NUMA domain wide counters support.
Add 3 new PMC classes for ARM DMC-620 and CMN-600 controllers PMU.

Sponsored By: ARM
Sponsored By: Ampere Computing

Original review: https://reviews.freebsd.org/D32319

I was asked by Allan to pick up this series of patches while ray is otherwise busy.

[pauamma_gundo.com]

lib/libpmc/pmc.3

204

Unclear to me. Do you mean something like "Counters can be tied/specific to each NUMA domain"?

224

Likewise, do you mean "Counters can have a single value shared among all NUMA domains" here?

[pauamma_gundo.com]

Tangentially: since MIPS CPUs are no longer supported per https://www.freebsd.org/platforms/, should they and the associated class still be listed in sys/sys/pmc.h?

[tsoome]

In D35342#801259, @pauamma_gundo.com wrote:

Tangentially: since MIPS CPUs are no longer supported per https://www.freebsd.org/platforms/, should they and the associated class still be listed in sys/sys/pmc.h?

I think,it would indeed need cleanup, but I do not want to do it here IMO it should be done in separate commit.

[pauamma_gundo.com]

In D35342#801269, @tsoome wrote:

In D35342#801259, @pauamma_gundo.com wrote:

Tangentially: since MIPS CPUs are no longer supported per https://www.freebsd.org/platforms/, should they and the associated class still be listed in sys/sys/pmc.h?

I think,it would indeed need cleanup, but I do not want to do it here IMO it should be done in separate commit.

Fair enough. (I can't do it myself, since I don't have the resources to test kernel or userland changes meaningfully.)

[allanjude]

lib/libpmc/pmc.3

204

It means that there will be a separate value/counter for each NUMA domain

224

One value for the entire system, rather than one per domain

[pauamma_gundo.com]

lib/libpmc/pmc.3

204

224

[tsoome]

lib/libpmc/pmc.3

204

thanks!

224

thanks!

[tsoome]

Updates to pmc.3 as suggested by pauamma_gundo.com.

[mhorne]

Seems to be the same as the original review with my two suggested tweaks? LGTM

[tsoome]

In D35342#807371, @mhorne wrote:

Seems to be the same as the original review with my two suggested tweaks? LGTM

Quite so, yes. Since Aleksandr can not complete this work now, we do not want it to get lost.

[gallatin]

usr.sbin/pmcstat/pmcstat.c

1190–1191

This utterly breaks system wide profiling for me on all arches that I've tried (amd64, arm64). The cause is that we don't add the pmc for CPU0, as the ID is not invalid. This causes samples from CPU0 to fail to be recognized, and to cause an assertion when analyzing logs:

Core was generated by `pmcstat -R out.pmclog_cpi01 -z 32 -G out.pmcstat_cpi01'.
Program terminated with signal SIGABRT, Aborted.
Sent by thr_kill() from pid 13680 and user 0.
#0 thr_kill () at thr_kill.S:4
4 thr_kill.S: No such file or directory.
(gdb) bt
#0 thr_kill () at thr_kill.S:4
#1 0x00000f8a5e2021a4 in __raise (s=s@entry=0x6) at /data/ocafirmware/FreeBSD/lib/libc/gen/raise.c:52
#2 0x00000f8a5e2b3039 in abort () at /data/ocafirmware/FreeBSD/lib/libc/stdlib/abort.c:67
#3 0x00000f8a5e1e5011 in __assert (func=<optimized out>, file=<optimized out>, line=<optimized out>,
failedexpr=<optimized out>) at /data/ocafirmware/FreeBSD/lib/libc/gen/assert.c:51
#4 0x00000f8236919e09 in pmcstat_analyze_log (args=0xf823691e858 <args>, plugins=0xf823691e120 <plugins>,
pmcstat_stats=0xf823691e980 <pmcstat_stats>, pmcstat_kernproc=0xf8a602a50f0, pmcstat_mergepmc=0x0,
pmcstat_npmcs=0xf823691e9b4 <pmcstat_npmcs>, ps_samples_period=0xf823691e9b8 <ps_samples_period>)
at /data/ocafirmware/FreeBSD/lib/libpmcstat/libpmcstat_logging.c:293
#5 0x00000f823691367a in pmcstat_process_log () at /data/ocafirmware/FreeBSD/usr.sbin/pmcstat/pmcstat_log.c:534
#6 0x00000f8236912fc8 in main (argc=<optimized out>, argv=<optimized out>)
at /data/ocafirmware/FreeBSD/usr.sbin/pmcstat/pmcstat.c:1411
(gdb) frame 4
#4 0x00000f8236919e09 in pmcstat_analyze_log (args=0xf823691e858 <args>, plugins=0xf823691e120 <plugins>,
pmcstat_stats=0xf823691e980 <pmcstat_stats>, pmcstat_kernproc=0xf8a602a50f0, pmcstat_mergepmc=0x0,
pmcstat_npmcs=0xf823691e9b4 <pmcstat_npmcs>, ps_samples_period=0xf823691e9b8 <ps_samples_period>)
at /data/ocafirmware/FreeBSD/lib/libpmcstat/libpmcstat_logging.c:293
293 /data/ocafirmware/FreeBSD/lib/libpmcstat/libpmcstat_logging.c: No such file or directory.

[kevans]

usr.sbin/pmcstat/pmcstat.c

738

@np hit an assertion similar to @gallatin's below in pmcstat -TS mode. It turns out that this is too early to allocate system events, the pmclog hasn't been configured yet so libpmcstat will not be made aware of any system-mode counters that are allocated. (Thus, logs recorded with -O are also not aware of these)

[mhorne]

usr.sbin/pmcstat/pmcstat.c

738

See D41978 for my attempt to remedy the situation.