Re: [PATCH V3 4/8] perf metrics: Sort the Default metricgroup

From: Liang, Kan
Date: Thu Jun 15 2023 - 22:15:13 EST

Next message: Baolu Lu: "Re: [PATCH] iommu: export iommu_set_default_translated function"
Previous message: Jianmin Lv: "Re: [PATCH pci] PCI: don't skip probing entire device if first fn OF node has status = "disabled""
In reply to: Ian Rogers: "Re: [PATCH V3 4/8] perf metrics: Sort the Default metricgroup"
Next in thread: kan . liang: "[PATCH V3 6/8] pert tests: Update metric-value for perf stat JSON output"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On 2023-06-15 6:05 p.m., Ian Rogers wrote:
> On Thu, Jun 15, 2023 at 6:54 AM <kan.liang@xxxxxxxxxxxxxxx> wrote:
>>
>> From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>>
>> The new default mode will print the metrics as a metric group. The
>> metrics from the same metric group must be adjacent to each other in the
>> metric list. But the metric_list_cmp() sorts metrics by the number of
>> events.
>>
>> Add a new sort for the Default metricgroup, which sorts by
>> default_metricgroup_name and metric_name.
>>
>> Add is_default in the struct metric_event to indicate that it's from
>> the Default metricgroup.
>>
>> Store the displayed metricgroup name of the Default metricgroup into
>> the metric expr for output.
>>
>> Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>> ---
>> tools/perf/util/metricgroup.c | 37 +++++++++++++++++++++++++++++++++++
>> tools/perf/util/metricgroup.h | 3 +++
>> 2 files changed, 40 insertions(+)
>>
>> diff --git a/tools/perf/util/metricgroup.c b/tools/perf/util/metricgroup.c
>> index 8b19644ade7d..e20adbdd5b56 100644
>> --- a/tools/perf/util/metricgroup.c
>> +++ b/tools/perf/util/metricgroup.c
>> @@ -79,6 +79,7 @@ static struct rb_node *metric_event_new(struct rblist *rblist __maybe_unused,
>> return NULL;
>> memcpy(me, entry, sizeof(struct metric_event));
>> me->evsel = ((struct metric_event *)entry)->evsel;
>> + me->is_default = false;
>> INIT_LIST_HEAD(&me->head);
>> return &me->nd;
>> }
>> @@ -1160,6 +1161,25 @@ static int metric_list_cmp(void *priv __maybe_unused, const struct list_head *l,
>> return right_count - left_count;
>> }
>>
>> +/**
>> + * default_metricgroup_cmp - Implements complex key for the Default metricgroup
>> + * that first sorts by default_metricgroup_name, then
>> + * metric_name.
>> + */
>> +static int default_metricgroup_cmp(void *priv __maybe_unused,
>> + const struct list_head *l,
>> + const struct list_head *r)
>> +{
>> + const struct metric *left = container_of(l, struct metric, nd);
>> + const struct metric *right = container_of(r, struct metric, nd);
>> + int diff = strcmp(right->default_metricgroup_name, left->default_metricgroup_name);
>> +
>> + if (diff)
>> + return diff;
>> +
>> + return strcmp(right->metric_name, left->metric_name);
>> +}
>> +
>> struct metricgroup__add_metric_data {
>> struct list_head *list;
>> const char *pmu;
>> @@ -1515,6 +1535,7 @@ static int parse_groups(struct evlist *perf_evlist,
>> LIST_HEAD(metric_list);
>> struct metric *m;
>> bool tool_events[PERF_TOOL_MAX] = {false};
>> + bool is_default = !strcmp(str, "Default");
>> int ret;
>>
>> if (metric_events_list->nr_entries == 0)
>> @@ -1549,6 +1570,9 @@ static int parse_groups(struct evlist *perf_evlist,
>> goto out;
>> }
>>
>> + if (is_default)
>> + list_sort(NULL, &metric_list, default_metricgroup_cmp);
>> +
>> list_for_each_entry(m, &metric_list, nd) {
>> struct metric_event *me;
>> struct evsel **metric_events;
>> @@ -1637,6 +1661,19 @@ static int parse_groups(struct evlist *perf_evlist,
>> expr->metric_unit = m->metric_unit;
>> expr->metric_events = metric_events;
>> expr->runtime = m->pctx->sctx.runtime;
>> + if (m->pmu && strcmp(m->pmu, "cpu")) {
>
> This shouldn't compare with a PMU name like this. What happens for
> memory bandwidth which could be logically with a memory controller
> PMU?
>
>> + char *name;
>> +
>> + if (asprintf(&name, "%s (%s)", m->default_metricgroup_name, m->pmu) < 0)
>> + expr->default_metricgroup_name = m->default_metricgroup_name;
>
> Who owns the string in this case? Can't you end up with
> default_metricgroup_name pointing to a freed string? I think this
> feels a lot more like output code, so I'm unclear why we're setting it
> up in the metric.

Yes, we can decide when outputing the name. The below patch move it to
the output code. I will do the modification in V4.

diff --git a/tools/perf/util/metricgroup.c b/tools/perf/util/metricgroup.c
index acf86b15ee49..a6a5ed44a679 100644
--- a/tools/perf/util/metricgroup.c
+++ b/tools/perf/util/metricgroup.c
@@ -1661,17 +1661,8 @@ static int parse_groups(struct evlist *perf_evlist,
expr->metric_unit = m->metric_unit;
expr->metric_events = metric_events;
expr->runtime = m->pctx->sctx.runtime;
- if (m->pmu && strcmp(m->pmu, "cpu")) {
- char *name;
-
- if (asprintf(&name, "%s (%s)", m->default_metricgroup_name, m->pmu) < 0)
- expr->default_metricgroup_name = m->default_metricgroup_name;
- else
- expr->default_metricgroup_name = name;
- } else
- expr->default_metricgroup_name = m->default_metricgroup_name;
- if (is_default)
- me->is_default = true;
+ expr->default_metricgroup_name = m->default_metricgroup_name;
+ me->is_default = is_default;
list_add(&expr->nd, &me->head);
}

diff --git a/tools/perf/util/stat-shadow.c b/tools/perf/util/stat-shadow.c
index b25974670d30..1440b0fc7d00 100644
--- a/tools/perf/util/stat-shadow.c
+++ b/tools/perf/util/stat-shadow.c
@@ -539,6 +539,42 @@ double test_generic_metric(struct metric_expr
*mexp, int aggr_idx)
return ratio;
}

+static void perf_stat__print_metricgroup_header(struct perf_stat_config
*config,
+ struct evsel *evsel,
+ void *ctxp,
+ const char *name,
+ struct perf_stat_output_ctx *out)
+{
+ bool need_full_name = perf_pmus__num_core_pmus() > 1;
+ static const char *last_name;
+ static const char *last_pmu;
+ char full_name[64];
+
+ /*
+ * A metricgroup may have several metric events,
+ * e.g.,TopdownL1 on e-core of ADL.
+ * The name has been output by the first metric
+ * event. Only align with other metics from
+ * different metric events.
+ */
+ if (last_name && !strcmp(last_name, name)) {
+ if (!need_full_name || !strcmp(last_pmu, evsel->pmu_name)) {
+ out->print_metricgroup_header(config, ctxp, NULL);
+ return;
+ }
+ }
+
+ if (need_full_name)
+ scnprintf(full_name, sizeof(full_name), "%s (%s)", name,
evsel->pmu_name);
+ else
+ scnprintf(full_name, sizeof(full_name), "%s", name);
+
+ out->print_metricgroup_header(config, ctxp, full_name);
+
+ last_name = name;
+ last_pmu = evsel->pmu_name;
+}
+
/**
* perf_stat__print_shadow_stats_metricgroup - Print out metrics
associated with the evsel
* For the non-default, all metrics associated
@@ -563,7 +599,6 @@ void
*perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
void *ctxp = out->ctx;
bool header_printed = false;
const char *name = NULL;
- static const char *last_name;

me = metricgroup__lookup(metric_events, evsel, false);
if (me == NULL)
@@ -588,20 +623,8 @@ void
*perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
/* Only print the name of the metricgroup once */
if (!header_printed) {
header_printed = true;
- if (!last_name || strcmp(last_name, name)) {
- /* Print out the name for the new metricgroup. */
- out->print_metricgroup_header(config, ctxp, name);
- last_name = name;
- } else if (!strcmp(last_name, name)) {
- /*
- * A metricgroup may have several metric events,
- * e.g.,TopdownL1 on e-core of ADL.
- * The name has been output by the first metric
- * event. Only align with other metics from
- * different metric events.
- */
- out->print_metricgroup_header(config, ctxp, NULL);
- }
+ perf_stat__print_metricgroup_header(config, evsel, ctxp,
+ name, out);
}
}

>
>> + else {
>> + expr->default_metricgroup_name = strdup(name);
>
> But name was just allocated, why strdup?
>
> This is still leaking it is just the strdup now leaks and not the asprintf:
> ```
> ==2495793==ERROR: LeakSanitizer: detected memory leaks
>
> Direct leak of 6199 byte(s) in 340 object(s) allocated from:
> #0 0x7f175c27077b in __interceptor_strdup
> ../../../../src/libsanitizer/asan/asan_interceptors.cp
> p:439
> #1 0x5596323b2a28 in parse_groups util/metricgroup.c:1670
> #2 0x5596323b2f08 in metricgroup__parse_groups_test util/metricgroup.c:1721
> #3 0x5596322992bf in test__parsing_callback tests/pmu-events.c:837
> #4 0x559632658ac6 in pmu_metrics_table_for_each_metric
> /tmp/perf/pmu-events/pmu-events.c:61641
> #5 0x5596326590fc in pmu_for_each_core_metric
> /tmp/perf/pmu-events/pmu-events.c:61742
> #6 0x559632299b7c in test__parsing tests/pmu-events.c:898
> #7 0x5596322663b0 in run_test tests/builtin-test.c:236
> #8 0x559632266655 in test_and_print tests/builtin-test.c:265
> #9 0x55963226766f in __cmd_test tests/builtin-test.c:436
> #10 0x559632268953 in cmd_test tests/builtin-test.c:559
> #11 0x5596322f5916 in run_builtin
> /home/irogers/kernel.org/tools/perf/perf.c:323
> #12 0x5596322f5e87 in handle_internal_command
> /home/irogers/kernel.org/tools/perf/perf.c:377
> #13 0x5596322f624f in run_argv /home/irogers/kernel.org/tools/perf/perf.c:421
> #14 0x5596322f67b7 in main /home/irogers/kernel.org/tools/perf/perf.c:537
> #15 0x7f175bebf189 in __libc_start_call_main
> ../sysdeps/nptl/libc_start_call_main.h:58
>
> SUMMARY: AddressSanitizer: 6199 byte(s) leaked in 340 allocation(s).
> ```
>
>> + free(name);
>> + }
>> + } else
>> + expr->default_metricgroup_name = m->default_metricgroup_name;
>
> Who owns the string in this case? Can't you end up with
> default_metricgroup_name pointing to a freed string?
>
> I spent some time trying to rationalize this to add as a patch, but
> then the more I look at things like the strcmp with "cpu" my changes
> were going to modify behavior in a way that would need you to test and
> sign-off, so I'll hold back.

THe "cpu" check will be gone in v4.

>
> To test with leak/address sanitizer use the tmp.perf-tools-next
> branch, I build with:
> $ make -C tools/perf O=/tmp/perf DEBUG=1 EXTRA_CFLAGS="-O0 -g
> -fno-omit-frame-pointer -DREFCNT_CHECKING=1 -fsanitize=address"
> NO_LIBTRACEEVENT=1
> but the only bit you really need is "-fsanitize=address" which both
> clang and gcc support.
>
> Address sanitizer from apt-cache is:
> libasan6 - AddressSanitizer -- a fast memory error detector
>
> Fedora and other distro.s have it too.
>

I still cannot reproduce the memory leaks even with the above command.
I will remove the code in v4, so the memory leaks should be gone.

Thanks,
Kan

>
>> + if (is_default)
>> + me->is_default = true;
>> list_add(&expr->nd, &me->head);
>> }
>>
>> diff --git a/tools/perf/util/metricgroup.h b/tools/perf/util/metricgroup.h
>> index bf18274c15df..d5325c6ec8e1 100644
>> --- a/tools/perf/util/metricgroup.h
>> +++ b/tools/perf/util/metricgroup.h
>> @@ -22,6 +22,7 @@ struct cgroup;
>> struct metric_event {
>> struct rb_node nd;
>> struct evsel *evsel;
>> + bool is_default; /* the metric evsel from the Default metricgroup */
>> struct list_head head; /* list of metric_expr */
>> };
>>
>> @@ -55,6 +56,8 @@ struct metric_expr {
>> * more human intelligible) and then add "MiB" afterward when displayed.
>> */
>> const char *metric_unit;
>> + /** Displayed metricgroup name of the Default metricgroup */
>> + const char *default_metricgroup_name;
>> /** Null terminated array of events used by the metric. */
>> struct evsel **metric_events;
>> /** Null terminated array of referenced metrics. */
>> --
>> 2.35.1
>>

Next message: Baolu Lu: "Re: [PATCH] iommu: export iommu_set_default_translated function"
Previous message: Jianmin Lv: "Re: [PATCH pci] PCI: don't skip probing entire device if first fn OF node has status = "disabled""
In reply to: Ian Rogers: "Re: [PATCH V3 4/8] perf metrics: Sort the Default metricgroup"
Next in thread: kan . liang: "[PATCH V3 6/8] pert tests: Update metric-value for perf stat JSON output"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]