Re: [PATCH v4 2/5] perf header: Parse non-cpu pmu capabilities

[Liang, Kan]

On 5/26/2022 11:08 AM, Ravi Bangoria wrote:
> Hi Kan,
>
> [...]
>
> > > +static int write_pmu_caps(struct feat_fd *ff, struct evlist *evlist __maybe_unused)
> > > +{
> > > +    struct perf_pmu_caps *caps = NULL;
> > > +    struct perf_pmu *pmu = NULL;
> > > +    u32 nr_pmus = 0;
> > > +    int ret;
> > > +
> > > +    while ((pmu = perf_pmu__scan(pmu))) {
> > > +        if (!pmu->name || !strncmp(pmu->name, "cpu", 3) ||
> > > +            perf_pmu__caps_parse(pmu) <= 0)
> > > +            continue;
> > > +        nr_pmus++;
> > > +    }
> > > +
> > > +    ret = do_write(ff, &nr_pmus, sizeof(nr_pmus));
> > > +    if (ret < 0)
> > > +        return ret;
> > > +
> > > +    if (!nr_pmus)
> > > +        return 0;
> > > +
> > > +    while ((pmu = perf_pmu__scan(pmu))) {
> > > +        if (!pmu->name || !strncmp(pmu->name, "cpu", 3) || !pmu->nr_caps)
> > > +            continue;
> > > +
> > > +        ret = do_write_string(ff, pmu->name);
> > > +        if (ret < 0)
> > > +            return ret;
> > > +
> > > +        ret = do_write(ff, &pmu->nr_caps, sizeof(pmu->nr_caps));
> > > +        if (ret < 0)
> > > +            return ret;
> > > +
> > > +        list_for_each_entry(caps, &pmu->caps, list) {
> > > +            ret = do_write_string(ff, caps->name);
> > > +            if (ret < 0)
> > > +                return ret;
> > > +
> > > +            ret = do_write_string(ff, caps->value);
> > > +            if (ret < 0)
> > > +                return ret;
> > > +        }
> > > +    }
> >
> > The write_per_cpu_pmu_caps() also does a similar thing. Can we factor out a generic write_pmu_caps() which works for both cpu and non-cpu pmu capabilities?
>
> I might be able to do this but..
>
> > It seems the print_pmu_caps()/process_pmu_caps() can also does similar factor out.
>
> not this, see below..
>
> > Actually, more aggressively, why not use the HEADER_PMU_CAPS to replace the HEADER_HYBRID_CPU_PMU_CAPS? The HEADER_HYBRID_CPU_PMU_CAPS is the last header feature. It seems doable. We can always write/print the "cpu_" kind of PMU first to be compatible with the old tools.
>
> There are some differences in how capabilities are stored in perf.data header
> as well as perf_env. In case of HEADER_CPU_PMU_CAPS or
> HEADER_HYBRID_CPU_PMU_CAPS, all capabilities are stored in a single string
> separated by NULL character. 

I think this is the format for the internal string, not the format of 
the perf.data header.

For the perf.data, here is the existing format for the 
HEADER_HYBRID_CPU_PMU_CAPS.

struct {
	u32 nr_pmu;
	struct {
		u32 nr_cpu_pmu_caps;
		{
			char	name[];
			char	value[];
		} [nr_cpu_pmu_caps];
		char pmu_name[];
	} [nr_pmu];
};

Here is your proposal.

+struct {
+	u32 nr_pmus;
+	struct {
+		char pmu_name[];
+		u32 nr_caps;
+		struct {
+			char name[];
+			char value[];
+		} [nr_caps];
+	} [nr_pmus];
+};

From my understanding, they are the same. (It doesn't matter where we 
put the char pmu_name[];)

That's also why I think we should merge the HEADER_HYBRID_CPU_PMU_CAPS 
and HEADER_PMU_CAPS. I don't think it make senses to basically handle 
the same thing with different codes.


> Whereas, in case of HEADER_PMU_CAPS, they are
> stored as an array of strings. The reason for this difference is, searching
> in an array is far easier compared to searching in a NULL separated string.

I think the hybrid_cpc_node can be replaced by the env_pmu_caps.
Then you don't need to modify the perf_env__find_pmu_cap().

Thanks,
Kan

> So, I don't think I can extend HEADER_HYBRID_CPU_PMU_CAPS as HEADER_PMU_CAPS
> without adding complexity in perf_env__find_pmu_cap().
>
> Thanks for the review,
> Ravi