RE: [PATCH 6/6] trace: add PMU

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Tomasz Duszynski [mailto:tduszynski@marvell.com]
> Sent: Monday, 16 June 2025 11.49
> 
> >16/06/2025 08:53, Tomasz Duszynski:
> >> @@ -86,6 +86,7 @@ always_enable = [
> >>          'ring',
> >>          'stack',
> >>          'telemetry',
> >> +        'pmu',
> >>  ]
> >
> >This list is alphabetically sorted.
> 
> Right, missed that.
> 
> >
> >What is the justification for always enabling PMU lib?
> >Is it a good idea to always enable an experimental library?
> 
> Well, since on Linux eal depends on pmu, disabling pmu ends up disabling
> eal, which breaks minimal build.
> 
> As for the second question - no. I think ideally lib should remain
> optional. However, since trace is part of eal, that indirectly forces
> pmu to be built.

Trace is optional, controlled by RTE_TRACE in /config/rte_config.h.

> 
> Alternatively, I guess we could make the library explicitly optional by
> extra meson option. But given that there are already options for
> controlling libraries that approach is not perfect either.

PMU should depend on EAL, not the other way around.

Please refer to this fix, where pmu.c is updated to call rte_mem_page_size(), which includes error handling, instead of directly calling sysconf(_SC_PAGE_SIZE) without handling errors from it:
https://patchwork.dpdk.org/project/dpdk/patch/20250612140657.313785-1-mb@smartsharesystems.com/

If there are cross dependency issues due to integration between PMU and Trace, adding the PMU related trace directly in the Trace library is a much better solution than making EAL depend on PMU.

Also, I agree with Thomas that PMU should not be considered a mandatory core library (always_enable).