Re: Consumers of libpmc/hwpmc

Hello Ali,

I've tested your patches on AMD Ryzen 5600X (Family 19h / Zen 3) running FreeBSD 16.0-CURRENT. We don't have PAPI or AMDuProf for performance analysis in our labs infrastructure.

Regarding the changes you mentioned:

PR #2076 (rdtsc instead of rdtscp for timestamps): Tested on Zen 3 — the patch consistently reduces discarded PMC events by ~11% across all sampling rates (1000 to 128000 samples/sec) with no panics or NMI errors. Full results posted at https://github.com/freebsd/freebsd-src/pull/2076

GET_CAPS (44a983d, PR #2058): Tested on the same hardware running 67b918e03475 — pmcstat works correctly with the new PMC_OP_GETCAPS ioctl path on Zen 3. No regressions observed.

I'm happy to continue testing your upcoming changes on Ryzen 5600X hardware as they land.

Regards,
Paulo
NLINK

Em 20/03/2026 17:00, Ali Mashtizadeh escreveu:
> Hello Adrian,
> 
> I don't see anything after a quick search.  If you find a link to the source, please send it my way.
> 
> At the moment the only breaking changes will affect users of the pmclog APIs.
>   - Andre has a diff up to fix the timestamp (this change's libpmc's pmclog_read but not the kernel interface).  This was a half baked feature that never worked as documented and it's unclear anyone is using it.
>   - I plan to change map-in/out pmclog events to report all map in/out events and extend the fields to report the length, file offset and flags.  We see some PCs that nobody seems to resolve correctly into a symbol, with IBS we now care about the address of the load or store, and this makes it easier for tools to lazily pull in elf info.
>   - I also plan to add a header with information that I need for analysis tools (e.g. cpuid, numa topology).
> 
> Existing tools also might want to make some changes for the following for new types of counters.
>   - GET_CAPS, domain and system-wide uncore counters
>   - Multipart payloads that allow extra information to be added to the callchain (AMD IBS) uses this.
> 
> I have a new framework/analysis passes for working with existing PMC data and the new AMD IBS bits I wrote.  Eventually I'd like to see these hooked up to Intel PEBS and ARM PT as well.
> 
> Best,
> Ali
> 
> 
> On Fri, Mar 20, 2026 at 10:27 AM Adrian Chadd <adrian@freebsd.org <mailto:adrian@freebsd.org>> wrote:
> 
>     there are some, i'm tinkering with something from sandvine (whose name
>     escapes me, even though I was tinkering with it two weeks ago!)
> 
>     What core changes are you thinking about?
> 
>     IMHO i think it'd be fine to break api a bit after such a long time
>     being stale..
> 
> 
> 
>     -adrian
> 
>     On Thu, 19 Mar 2026 at 14:46, Ali Mashtizadeh <mashtizadeh@gmail.com <mailto:mashtizadeh@gmail.com>> wrote:
>      >
>      > Hello everyone,
>      >
>      > I've been doing a lot of work on improving pmc/hwpmc on AMD systems but we need to make a few improvements to the pmc core.
>      >
>      > To be aware of compatibility issues, I was wondering if there are any 3rd party consumers of libpmc/hwpmc outside of PAPI and AMDuProf?
>      >
>      > Best,
>      > Ali
>