]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
perf tool_pmu: More accurately set the cpus for tool events
authorIan Rogers <irogers@google.com>
Thu, 13 Nov 2025 18:05:13 +0000 (10:05 -0800)
committerNamhyung Kim <namhyung@kernel.org>
Tue, 18 Nov 2025 02:43:09 +0000 (18:43 -0800)
The user and system time events can record on different CPUs, but for
all other events a single CPU map of just CPU 0 makes sense. In
parse-events detect a tool PMU and then pass the perf_event_attr so
that the tool_pmu can return CPUs specific for the event. This avoids
a CPU map of all online CPUs being used for events like
duration_time. Avoiding this avoids the evlist CPUs containing CPUs
for which duration_time just gives 0. Minimizing the evlist CPUs can
remove unnecessary sched_setaffinity syscalls that delay metric
calculations.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
tools/perf/util/parse-events.c
tools/perf/util/tool_pmu.c
tools/perf/util/tool_pmu.h

index e522082c5a214038a50d1bcbf2b9ff6edd16ce08..5ffeb7b416d93958fa8044394cb32aeffb2744ad 100644 (file)
@@ -30,6 +30,7 @@
 #include "util/event.h"
 #include "util/bpf-filter.h"
 #include "util/stat.h"
+#include "util/tool_pmu.h"
 #include "util/util.h"
 #include "tracepoint.h"
 #include <api/fs/tracing_path.h>
@@ -227,8 +228,12 @@ __add_event(struct list_head *list, int *idx,
        if (pmu) {
                is_pmu_core = pmu->is_core;
                pmu_cpus = perf_cpu_map__get(pmu->cpus);
-               if (perf_cpu_map__is_empty(pmu_cpus))
-                       pmu_cpus = cpu_map__online();
+               if (perf_cpu_map__is_empty(pmu_cpus)) {
+                       if (perf_pmu__is_tool(pmu))
+                               pmu_cpus = tool_pmu__cpus(attr);
+                       else
+                               pmu_cpus = cpu_map__online();
+               }
        } else {
                is_pmu_core = (attr->type == PERF_TYPE_HARDWARE ||
                               attr->type == PERF_TYPE_HW_CACHE);
index 6a9df3dc0e076ee87f4595e638c021c9dce87cd4..37c4eae0bef1bc6a60a75fd4db2296f1b0fd1603 100644 (file)
@@ -2,6 +2,7 @@
 #include "cgroup.h"
 #include "counts.h"
 #include "cputopo.h"
+#include "debug.h"
 #include "evsel.h"
 #include "pmu.h"
 #include "print-events.h"
@@ -13,6 +14,7 @@
 #include <api/fs/fs.h>
 #include <api/io.h>
 #include <internal/threadmap.h>
+#include <perf/cpumap.h>
 #include <perf/threadmap.h>
 #include <fcntl.h>
 #include <strings.h>
@@ -109,6 +111,23 @@ const char *evsel__tool_pmu_event_name(const struct evsel *evsel)
        return tool_pmu__event_to_str(evsel->core.attr.config);
 }
 
+struct perf_cpu_map *tool_pmu__cpus(struct perf_event_attr *attr)
+{
+       static struct perf_cpu_map *cpu0_map;
+       enum tool_pmu_event event = (enum tool_pmu_event)attr->config;
+
+       if (event <= TOOL_PMU__EVENT_NONE || event >= TOOL_PMU__EVENT_MAX) {
+               pr_err("Invalid tool PMU event config %llx\n", attr->config);
+               return NULL;
+       }
+       if (event == TOOL_PMU__EVENT_USER_TIME || event == TOOL_PMU__EVENT_SYSTEM_TIME)
+               return cpu_map__online();
+
+       if (!cpu0_map)
+               cpu0_map = perf_cpu_map__new_int(0);
+       return perf_cpu_map__get(cpu0_map);
+}
+
 static bool read_until_char(struct io *io, char e)
 {
        int c;
index f1714001bc1d92b30cdea3aefd6a69ad2c7e3818..ea343d1983d3a83e5d1cf196d2739bee892b8916 100644 (file)
@@ -46,6 +46,7 @@ bool tool_pmu__read_event(enum tool_pmu_event ev,
 u64 tool_pmu__cpu_slots_per_cycle(void);
 
 bool perf_pmu__is_tool(const struct perf_pmu *pmu);
+struct perf_cpu_map *tool_pmu__cpus(struct perf_event_attr *attr);
 
 bool evsel__is_tool(const struct evsel *evsel);
 enum tool_pmu_event evsel__tool_event(const struct evsel *evsel);