Re: [PATCH v2 2/2] perf parse-events: Remove "not supported" hybrid cache events

From: Ian Rogers
Date: Fri Sep 23 2022 - 12:55:40 EST

Next message: Dmitry Torokhov: "Re: [PATCH v5 2/3] dt-bindings: input: touchscreen: elants_i2c: Add eth3915n touchscreen chip"
Previous message: Jason A. Donenfeld: "Re: [PATCH v2] minmax: clamp more efficiently by avoiding extra comparison"
In reply to: zhengjun . xing: "[PATCH v2 2/2] perf parse-events: Remove "not supported" hybrid cache events"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Thu, Sep 22, 2022 at 7:58 PM <zhengjun.xing@xxxxxxxxxxxxxxx> wrote:
>
> From: Zhengjun Xing <zhengjun.xing@xxxxxxxxxxxxxxx>
>
> By default, we create two hybrid cache events, one is for cpu_core, and
> another is for cpu_atom. But Some hybrid hardware cache events are only
> available on one CPU PMU. For example, the 'L1-dcache-load-misses' is only
> available on cpu_core, while the 'L1-icache-loads' is only available on
> cpu_atom. We need to remove "not supported" hybrid cache events. By
> extending is_event_supported() to global API and using it to check if the
> hybrid cache events are supported before being created, we can remove the
> "not supported" hybrid cache events.
>
> Before:
>
> # ./perf stat -e L1-dcache-load-misses,L1-icache-loads -a sleep 1
>
> Performance counter stats for 'system wide':
>
> 52,570 cpu_core/L1-dcache-load-misses/
> <not supported> cpu_atom/L1-dcache-load-misses/
> <not supported> cpu_core/L1-icache-loads/
> 1,471,817 cpu_atom/L1-icache-loads/
>
> 1.004915229 seconds time elapsed
>
> After:
>
> # ./perf stat -e L1-dcache-load-misses,L1-icache-loads -a sleep 1
>
> Performance counter stats for 'system wide':
>
> 54,510 cpu_core/L1-dcache-load-misses/
> 1,441,286 cpu_atom/L1-icache-loads/
>
> 1.005114281 seconds time elapsed
>
> Fixes: 30def61f64ba ("perf parse-events: Create two hybrid cache events")
> Signed-off-by: Zhengjun Xing <zhengjun.xing@xxxxxxxxxxxxxxx>
> Reviewed-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> Reported-by: Yi Ammy <ammy.yi@xxxxxxxxx>

Acked-by: Ian Rogers <irogers@xxxxxxxxxx>

Thanks,
Ian

> ---
> Change log:
> v2:
> * Adds a comment for removing "not supported" hybrid cache events.
> * Remove goto and add a strdup check
> * "is_event_supported" move to parse-events.c per Ian's suggestion.
> * Adds Reported-by from Yi Ammy <ammy.yi@xxxxxxxxx>
>
> tools/perf/util/parse-events-hybrid.c | 21 ++++++++++++---
> tools/perf/util/parse-events.c | 39 +++++++++++++++++++++++++++
> tools/perf/util/parse-events.h | 1 +
> tools/perf/util/print-events.c | 39 ---------------------------
> 4 files changed, 57 insertions(+), 43 deletions(-)
>
> diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c
> index 284f8eabd3b9..7c9f9150bad5 100644
> --- a/tools/perf/util/parse-events-hybrid.c
> +++ b/tools/perf/util/parse-events-hybrid.c
> @@ -33,7 +33,8 @@ static void config_hybrid_attr(struct perf_event_attr *attr,
> * If the PMU type ID is 0, the PERF_TYPE_RAW will be applied.
> */
> attr->type = type;
> - attr->config = attr->config | ((__u64)pmu_type << PERF_PMU_TYPE_SHIFT);
> + attr->config = (attr->config & PERF_HW_EVENT_MASK) |
> + ((__u64)pmu_type << PERF_PMU_TYPE_SHIFT);
> }
>
> static int create_event_hybrid(__u32 config_type, int *idx,
> @@ -48,13 +49,25 @@ static int create_event_hybrid(__u32 config_type, int *idx,
> __u64 config = attr->config;
>
> config_hybrid_attr(attr, config_type, pmu->type);
> +
> + /*
> + * Some hybrid hardware cache events are only available on one CPU
> + * PMU. For example, the 'L1-dcache-load-misses' is only available
> + * on cpu_core, while the 'L1-icache-loads' is only available on
> + * cpu_atom. We need to remove "not supported" hybrid cache events.
> + */
> + if (attr->type == PERF_TYPE_HW_CACHE
> + && !is_event_supported(attr->type, attr->config))
> + return 0;
> +
> evsel = parse_events__add_event_hybrid(list, idx, attr, name, metric_id,
> pmu, config_terms);
> - if (evsel)
> + if (evsel) {
> evsel->pmu_name = strdup(pmu->name);
> - else
> + if (!evsel->pmu_name)
> + return -ENOMEM;
> + } else
> return -ENOMEM;
> -
> attr->type = type;
> attr->config = config;
> return 0;
> diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> index f05e15acd33f..f3b2c2a87456 100644
> --- a/tools/perf/util/parse-events.c
> +++ b/tools/perf/util/parse-events.c
> @@ -28,6 +28,7 @@
> #include "util/parse-events-hybrid.h"
> #include "util/pmu-hybrid.h"
> #include "tracepoint.h"
> +#include "thread_map.h"
>
> #define MAX_NAME_LEN 100
>
> @@ -157,6 +158,44 @@ struct event_symbol event_symbols_sw[PERF_COUNT_SW_MAX] = {
> #define PERF_EVENT_TYPE(config) __PERF_EVENT_FIELD(config, TYPE)
> #define PERF_EVENT_ID(config) __PERF_EVENT_FIELD(config, EVENT)
>
> +bool is_event_supported(u8 type, u64 config)
> +{
> + bool ret = true;
> + int open_return;
> + struct evsel *evsel;
> + struct perf_event_attr attr = {
> + .type = type,
> + .config = config,
> + .disabled = 1,
> + };
> + struct perf_thread_map *tmap = thread_map__new_by_tid(0);
> +
> + if (tmap == NULL)
> + return false;
> +
> + evsel = evsel__new(&attr);
> + if (evsel) {
> + open_return = evsel__open(evsel, NULL, tmap);
> + ret = open_return >= 0;
> +
> + if (open_return == -EACCES) {
> + /*
> + * This happens if the paranoid value
> + * /proc/sys/kernel/perf_event_paranoid is set to 2
> + * Re-run with exclude_kernel set; we don't do that
> + * by default as some ARM machines do not support it.
> + *
> + */
> + evsel->core.attr.exclude_kernel = 1;
> + ret = evsel__open(evsel, NULL, tmap) >= 0;
> + }
> + evsel__delete(evsel);
> + }
> +
> + perf_thread_map__put(tmap);
> + return ret;
> +}
> +
> const char *event_type(int type)
> {
> switch (type) {
> diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h
> index 7e6a601d9cd0..07df7bb7b042 100644
> --- a/tools/perf/util/parse-events.h
> +++ b/tools/perf/util/parse-events.h
> @@ -19,6 +19,7 @@ struct option;
> struct perf_pmu;
>
> bool have_tracepoints(struct list_head *evlist);
> +bool is_event_supported(u8 type, u64 config);
>
> const char *event_type(int type);
>
> diff --git a/tools/perf/util/print-events.c b/tools/perf/util/print-events.c
> index 04050d4f6db8..c4d5d87fae2f 100644
> --- a/tools/perf/util/print-events.c
> +++ b/tools/perf/util/print-events.c
> @@ -22,7 +22,6 @@
> #include "probe-file.h"
> #include "string2.h"
> #include "strlist.h"
> -#include "thread_map.h"
> #include "tracepoint.h"
> #include "pfm.h"
> #include "pmu-hybrid.h"
> @@ -239,44 +238,6 @@ void print_sdt_events(const char *subsys_glob, const char *event_glob,
> strlist__delete(sdtlist);
> }
>
> -static bool is_event_supported(u8 type, u64 config)
> -{
> - bool ret = true;
> - int open_return;
> - struct evsel *evsel;
> - struct perf_event_attr attr = {
> - .type = type,
> - .config = config,
> - .disabled = 1,
> - };
> - struct perf_thread_map *tmap = thread_map__new_by_tid(0);
> -
> - if (tmap == NULL)
> - return false;
> -
> - evsel = evsel__new(&attr);
> - if (evsel) {
> - open_return = evsel__open(evsel, NULL, tmap);
> - ret = open_return >= 0;
> -
> - if (open_return == -EACCES) {
> - /*
> - * This happens if the paranoid value
> - * /proc/sys/kernel/perf_event_paranoid is set to 2
> - * Re-run with exclude_kernel set; we don't do that
> - * by default as some ARM machines do not support it.
> - *
> - */
> - evsel->core.attr.exclude_kernel = 1;
> - ret = evsel__open(evsel, NULL, tmap) >= 0;
> - }
> - evsel__delete(evsel);
> - }
> -
> - perf_thread_map__put(tmap);
> - return ret;
> -}
> -
> int print_hwcache_events(const char *event_glob, bool name_only)
> {
> unsigned int type, op, i, evt_i = 0, evt_num = 0, npmus = 0;
> --
> 2.25.1
>

Next message: Dmitry Torokhov: "Re: [PATCH v5 2/3] dt-bindings: input: touchscreen: elants_i2c: Add eth3915n touchscreen chip"
Previous message: Jason A. Donenfeld: "Re: [PATCH v2] minmax: clamp more efficiently by avoiding extra comparison"
In reply to: zhengjun . xing: "[PATCH v2 2/2] perf parse-events: Remove "not supported" hybrid cache events"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]