Re: [QUESTION] ARM DSU PMU: 32-bit counter assumption and handling of 64-bit event counters

[Robin Murphy]

On 2026-03-12 2:28 pm, Will Deacon wrote:
> [+Robin and Suzuki]
>
> On Fri, Feb 06, 2026 at 11:51:38AM +0800, feng.zhou wrote:
> > Hi Will and Mark,
> >
> > I have a question regarding the ARM DSU PMU driver implementation in
> > drivers/perf/arm_dsu_pmu.c, specifically about the counter width
> > assumptions.
> >
> > Current Implementation:
> > ======================
> > The driver currently assumes all event counters are 32-bit wide, with
> > only the cycle counter being 64-bit (lines 69-71):
> >
> > /* All event counters are 32bit, with a 64bit Cycle counter */
> > #define DSU_PMU_COUNTER_WIDTH(idx)	\
> > 	(((idx) == DSU_PMU_IDX_CYCLE_COUNTER) ? 64 : 32)
> >
> > This assumption is used throughout the code for handling counter
> > overflow and calculating deltas (e.g., line 344 in dsu_pmu_event_update).
> >
> > Issue Observed:
> > ===============
> > I've encountered overflow issues when using perf tools on newer ARM CPUs
> > where the event counters appear to be 64-bit rather than 32-bit. The
> > current implementation causes incorrect event counting and potential data
> > loss due to premature overflow handling based on the 32-bit mask.
>
> I suspect either the hardware or the architecture have evolved since the
> driver was initially written, possibly in a non-backwards-compatible way
> which is why you see problems.
>
> > Questions:
> > ==========
> > 1. Is the 32-bit event counter assumption based on the ARMv8 PMU
> >     specification or DSU-specific hardware documentation?
> >
> > 2. Are there known ARM CPU implementations where DSU event counters
> >     exceed 32 bits? If so, how should the driver handle this variation?
> >
> > 3. Would it be appropriate to detect the actual counter width at runtime
> >     (similar to how num_counters is probed in dsu_pmu_probe_pmu), or
> >     should this be handled through device tree / ACPI properties?
> >
> > 4. Should there be a mechanism to query the hardware for the actual
> >     counter width to ensure compatibility with future implementations?
> >
> > Thank you for your time and any insights you can provide.
>
> Suzuki and Robin should be able to help with these...

Yup, indeed newer DSUs subtly evolved, and the patches are already in 
7.0-rc ;)

Cheers,
Robin.