[PATCH] perf metrics: Add literal for system TSC frequency

[PATCH] perf metrics: Add literal for system TSC frequency

[Ian Rogers]

Such a literal is useful to calculate things like the average frequency
[1]. The TSC frequency isn't exposed by sysfs although some experimental
drivers look to add it [2]. This change computes the value using the
frequency in /proc/cpuinfo which is accruate at least on Intel
processors.

[1] https://github.com/intel/perfmon-metrics/blob/5ad9ef7056f31075e8178b9f1fb732af183b2c8d/SKX/metrics/perf/skx_metric_perf.json#L11
[2] https://github.com/trailofbits/tsc_freq_khz

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/util/expr.c | 42 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 42 insertions(+)

diff --git a/tools/perf/util/expr.c b/tools/perf/util/expr.c
index 675f318ce7c1..7ad01ed57025 100644
--- a/tools/perf/util/expr.c
+++ b/tools/perf/util/expr.c
@@ -402,6 +402,43 @@ double expr_id_data__source_count(const struct expr_id_data *data)
 	return data->val.source_count;
 }
 
+/*
+ * Derive the TSC frequency in Hz from the /proc/cpuinfo, for example:
+ * ...
+ * model name      : Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz
+ * ...
+ * will return 3000000000.
+ */
+static double system_tsc_freq(void)
+{
+	double result = NAN;
+	FILE *cpuinfo;
+	char *line = NULL;
+	size_t len = 0;
+
+	cpuinfo = fopen("/proc/cpuinfo", "r");
+	if (!cpuinfo)
+		return NAN;
+
+	while (getline(&line, &len, cpuinfo) > 0) {
+		if (!strncmp(line, "model name", 10)) {
+			char *pos = line + 11;
+
+			while (*pos != '@')
+				pos++;
+			if (sscanf(pos, " %lfGHz", &result) == 1) {
+				result *= 1000000000;
+				goto out;
+			}
+		}
+	}
+
+out:
+	free(line);
+	fclose(cpuinfo);
+	return result;
+}
+
 double expr__get_literal(const char *literal)
 {
 	static struct cpu_topology *topology;
@@ -417,6 +454,11 @@ double expr__get_literal(const char *literal)
 		goto out;
 	}
 
+	if (!strcasecmp("#system_tsc_freq", literal)) {
+		result = system_tsc_freq();
+		goto out;
+	}
+
 	/*
 	 * Assume that topology strings are consistent, such as CPUs "0-1"
 	 * wouldn't be listed as "0,1", and so after deduplication the number of
-- 
2.35.0.rc2.247.g8bbb082509-goog

Re: [PATCH] perf metrics: Add literal for system TSC frequency

[Andi Kleen]

On 2/2/2022 5:07 PM, Ian Rogers wrote:
> Such a literal is useful to calculate things like the average frequency
> [1]. The TSC frequency isn't exposed by sysfs although some experimental
> drivers look to add it [2]. This change computes the value using the
> frequency in /proc/cpuinfo which is accruate at least on Intel
> processors.

FWIW the information in /proc/cpuinfo is going away.

You should really use CPUID for this, which has a new leaf for this. 
Unfortunately there

were a few systems were it was inaccurate, but I suppose you could still 
prefer /proc/cpuinfo

Also your method will not work when the metrics are used through perf 
stat report or perf script

and run on another system. For the perf script case there is support for 
putting the TSC into

perf.data (that is needed for PT), but the header is not dumped by 
default unfortunately.

So likely you would need some new mechanism for that.

-Andi

Re: [PATCH] perf metrics: Add literal for system TSC frequency

[Ian Rogers]

On Wed, Feb 2, 2022 at 7:59 PM Andi Kleen <ak@linux.intel.com> wrote:
>
>
> On 2/2/2022 5:07 PM, Ian Rogers wrote:
> > Such a literal is useful to calculate things like the average frequency
> > [1]. The TSC frequency isn't exposed by sysfs although some experimental
> > drivers look to add it [2]. This change computes the value using the
> > frequency in /proc/cpuinfo which is accruate at least on Intel
> > processors.
>
> FWIW the information in /proc/cpuinfo is going away.
>
> You should really use CPUID for this, which has a new leaf for this.
> Unfortunately there
>
> were a few systems were it was inaccurate, but I suppose you could still
> prefer /proc/cpuinfo
>
> Also your method will not work when the metrics are used through perf
> stat report or perf script
>
> and run on another system. For the perf script case there is support for
> putting the TSC into
>
> perf.data (that is needed for PT), but the header is not dumped by
> default unfortunately.
>
> So likely you would need some new mechanism for that.

Thanks Andi,

yep metrics with perf stat record are broken on a number of fronts and
this will be another that will need a fix. It looks like the cpuid
approach won't work pre-Skylake so sysfs could act as a fallback, but
I don't have an ETA on when that will be available. We could have
additional warnings when reading /proc/cpuinfo fails. Wdyt?

Thanks,
Ian

> -Andi
>