On Wed, Nov 10, 2021 at 4:56 AM Jiri Olsa [off-list ref] wrote:

On Fri, Nov 05, 2021 at 10:09:42AM -0700, Ian Rogers wrote:

quoted

Allow the number of cpus, cores, dies and packages to be queried by a
metric expression.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/tests/expr.c | 12 +++++++++++-
 tools/perf/util/expr.c  | 27 +++++++++++++++++++++++++++
 2 files changed, 38 insertions(+), 1 deletion(-)

diff --git a/tools/perf/tests/expr.c b/tools/perf/tests/expr.c
index 9ee2dc91c27b..0c09ccc76665 100644
--- a/tools/perf/tests/expr.c
+++ b/tools/perf/tests/expr.c

@@ -66,7 +66,7 @@ int test__expr(struct test *t __maybe_unused, int subtest __maybe_unused)
 {
      struct expr_id_data *val_ptr;
      const char *p;
-     double val;
+     double val, num_cpus, num_cores, num_dies, num_packages;
      int ret;
      struct expr_parse_ctx *ctx;

@@ -161,6 +161,16 @@ int test__expr(struct test *t __maybe_unused, int subtest __maybe_unused)
                      NULL, ctx) == 0);
      TEST_ASSERT_VAL("find ids", hashmap__size(ctx->ids) == 0);

+     /* Test toplogy constants appear well ordered. */
+     expr__ctx_clear(ctx);
+     TEST_ASSERT_VAL("#num_cpus", expr__parse(&num_cpus, ctx, "#num_cpus") == 0);
+     TEST_ASSERT_VAL("#num_cores", expr__parse(&num_cores, ctx, "#num_cores") == 0);
+     TEST_ASSERT_VAL("#num_cpus >= #num_cores", num_cpus >= num_cores);
+     TEST_ASSERT_VAL("#num_dies", expr__parse(&num_dies, ctx, "#num_dies") == 0);
+     TEST_ASSERT_VAL("#num_cores >= #num_dies", num_cores >= num_dies);
+     TEST_ASSERT_VAL("#num_packages", expr__parse(&num_packages, ctx, "#num_packages") == 0);
+     TEST_ASSERT_VAL("#num_dies >= #num_packages", num_dies >= num_packages);
+
      expr__ctx_free(ctx);

      return 0;

diff --git a/tools/perf/util/expr.c b/tools/perf/util/expr.c
index 7464739c2890..15af8b8ef5e7 100644
--- a/tools/perf/util/expr.c
+++ b/tools/perf/util/expr.c

@@ -5,6 +5,8 @@
 #include <stdlib.h>
 #include <string.h>
 #include "metricgroup.h"
+#include "cpumap.h"
+#include "cputopo.h"
 #include "debug.h"
 #include "expr.h"
 #include "expr-bison.h"

@@ -375,9 +377,34 @@ double expr_id_data__value(const struct expr_id_data *data)

 double expr__get_literal(const char *literal)
 {
+     static struct cpu_topology *topology;
+
      if (!strcmp("#smt_on", literal))
              return smt_on() > 0 ? 1.0 : 0.0;

+     if (!strcmp("#num_cpus", literal))
+             return cpu__max_present_cpu();
+
+     /*
+      * 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
+      * these strings gives an indication of the number of packages, dies,
+      * etc.
+      */
+     if (!topology) {
+             topology = cpu_topology__new();

any chance we could propagate expr_scanner_ctx in here and store topology
to it and release it at the end? I think we have several places like this,
so it'd be nice not to make more if it's possible ;-)

The topology here is static and so will only get computed once per
execution rather than once pre expression parse. I was worried about
the cost of recomputing the topology for something like 'perf stat -I
1000 -M ...' in which case the static will do less recomputation.

Thanks,
Ian

thanks,
jirka

quoted

+             if (!topology) {
+                     pr_err("Error creating CPU topology");
+                     return NAN;
+             }
+     }
+     if (!strcmp("#num_packages", literal))
+             return topology->package_cpus_lists;
+     if (!strcmp("#num_dies", literal))
+             return topology->die_cpus_lists;
+     if (!strcmp("#num_cores", literal))
+             return topology->core_cpus_lists;
+
      pr_err("Unrecognized literal '%s'", literal);
      return NAN;
 }
--
2.34.0.rc0.344.g81b53c2807-goog