On 02/10/18 17:19, Jean-Philippe Brucker wrote:

On 02/10/2018 15:11, Jean-Philippe Brucker wrote:

quoted

quoted

+	cfgr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CFGR);

Something I missed previously: when SMMU_PMCG_CFGR.SID_FILTER_TYPE is 1,
filtering for all counters is configured by SMMU_PMCG_SMR0 and
SMMU_PMCG_EVTYPER0 (instead of having one separate filter per counter).

Oh, I hadn't even noticed it had that mode as well...

In that mode with your patch, if the user applies a filter to the first
event in the list passed to perf, it will be applied to all events.
Filter applied on any subsequent event will be ignored. Could we make
this more explicit? Maybe in the probe print that the PMCG is
global-filtering, and when attempting to apply a filter to something
else than EVCNTR0, return an error?

FWIW filtering is always per-counter-group on the SMMUv2 PMU, and it's 
actually pretty straightforward to cope with - pmu->add() just needs to 
reject the event if one with an incompatible configuration is already 
scheduled, so perf core handles it much like having more events than 
counters, by rotating the mutually-exclusive sets.

Robin.