lkml.org 
[lkml]   [2012]   [Aug]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [RFC PATCH] perf/x86: Add cpumask for uncore PMU.
From
On Mon, Aug 6, 2012 at 9:35 AM, Yan, Zheng <zheng.z.yan@intel.com> wrote:
>
> From: "Yan, Zheng" <zheng.z.yan@intel.com>
>
> This RFC patch adds a cpumask file to the uncore pmu sysfs directory.
> If user doesn't explicitly specify CPU list, perf-stat only collects
> uncore events on CPUs listed in the cpumask file.
>
> Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
> ---
> arch/x86/kernel/cpu/perf_event_intel_uncore.c | 28 ++++++++++++++++++++++---
> arch/x86/kernel/cpu/perf_event_intel_uncore.h | 6 ++++--
> tools/perf/builtin-stat.c | 30 ++++++++++++++++++---------
> tools/perf/util/cpumap.c | 22 +++++++++++++-------
> tools/perf/util/cpumap.h | 2 +-
> tools/perf/util/evsel.h | 1 +
> tools/perf/util/parse-events.c | 14 +++++++------
> tools/perf/util/pmu.c | 30 +++++++++++++++++++++++++++
> tools/perf/util/pmu.h | 1 +
> 9 files changed, 105 insertions(+), 29 deletions(-)
>
> diff --git a/arch/x86/kernel/cpu/perf_event_intel_uncore.c b/arch/x86/kernel/cpu/perf_event_intel_uncore.c
> index 0a55710..62ec3e6 100644
> --- a/arch/x86/kernel/cpu/perf_event_intel_uncore.c
> +++ b/arch/x86/kernel/cpu/perf_event_intel_uncore.c
> @@ -2341,6 +2341,27 @@ int uncore_pmu_event_init(struct perf_event *event)
> return ret;
> }
>
> +static ssize_t uncore_get_attr_cpumask(struct device *dev,
> + struct device_attribute *attr, char *buf)
> +{
> + int n = cpulist_scnprintf(buf, PAGE_SIZE - 2, &uncore_cpu_mask);
> +
> + buf[n++] = '\n';

Why the \n. I thought there were no \n in sysfs files.

> + buf[n] = '\0';
> + return n;
> +}
> +
> +static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL);
> +
> +static struct attribute *uncore_pmu_attrs[] = {
> + &dev_attr_cpumask.attr,
> + NULL,
> +};
> +
> +static struct attribute_group uncore_pmu_attr_group = {
> + .attrs = uncore_pmu_attrs,
> +};
> +
> static int __init uncore_pmu_register(struct intel_uncore_pmu *pmu)
> {
> int ret;
> @@ -2378,8 +2399,8 @@ static void __init uncore_type_exit(struct intel_uncore_type *type)
> free_percpu(type->pmus[i].box);
> kfree(type->pmus);
> type->pmus = NULL;
> - kfree(type->attr_groups[1]);
> - type->attr_groups[1] = NULL;
> + kfree(type->events_group);
> + type->events_group = NULL;
> }
>
> static void __init uncore_types_exit(struct intel_uncore_type **types)
> @@ -2431,9 +2452,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type)
> for (j = 0; j < i; j++)
> attrs[j] = &type->event_descs[j].attr.attr;
>
> - type->attr_groups[1] = events_group;
> + type->events_group = events_group;
> }
>
> + type->pmu_group = &uncore_pmu_attr_group;
> type->pmus = pmus;
> return 0;
> fail:
> diff --git a/arch/x86/kernel/cpu/perf_event_intel_uncore.h b/arch/x86/kernel/cpu/perf_event_intel_uncore.h
> index 5b81c18..e68a455 100644
> --- a/arch/x86/kernel/cpu/perf_event_intel_uncore.h
> +++ b/arch/x86/kernel/cpu/perf_event_intel_uncore.h
> @@ -369,10 +369,12 @@ struct intel_uncore_type {
> struct intel_uncore_pmu *pmus;
> struct intel_uncore_ops *ops;
> struct uncore_event_desc *event_descs;
> - const struct attribute_group *attr_groups[3];
> + const struct attribute_group *attr_groups[4];
> };
>
> -#define format_group attr_groups[0]
> +#define pmu_group attr_groups[0]
> +#define format_group attr_groups[1]
> +#define events_group attr_groups[2]
>
> struct intel_uncore_ops {
> void (*init_box)(struct intel_uncore_box *);
> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> index 861f0ae..8b4275c 100644
> --- a/tools/perf/builtin-stat.c
> +++ b/tools/perf/builtin-stat.c
> @@ -220,6 +220,16 @@ static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
> evsel->priv = NULL;
> }
>
> +static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
> +{
> + return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
> +}
> +
> +static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
> +{
> + return perf_evsel__cpus(evsel)->nr;
> +}
> +
> static void update_stats(struct stats *stats, u64 val)
> {
> double delta;
> @@ -299,7 +309,7 @@ retry:
> evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
>
> if (perf_target__has_cpu(&target)) {
> - ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
> + ret = perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel),
> group, group_fd);
> if (ret)
> goto check_ret;
> @@ -382,7 +392,7 @@ static int read_counter_aggr(struct perf_evsel *counter)
> u64 *count = counter->counts->aggr.values;
> int i;
>
> - if (__perf_evsel__read(counter, evsel_list->cpus->nr,
> + if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
> evsel_list->threads->nr, scale) < 0)
> return -1;
>
> @@ -411,7 +421,7 @@ static int read_counter(struct perf_evsel *counter)
> u64 *count;
> int cpu;
>
> - for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
> + for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
> if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
> return -1;
>
> @@ -546,12 +556,12 @@ static int run_perf_stat(int argc __used, const char **argv)
> if (no_aggr) {
> list_for_each_entry(counter, &evsel_list->entries, node) {
> read_counter(counter);
> - perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
> + perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
> }
> } else {
> list_for_each_entry(counter, &evsel_list->entries, node) {
> read_counter_aggr(counter);
> - perf_evsel__close_fd(counter, evsel_list->cpus->nr,
> + perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
> evsel_list->threads->nr);
> }
> }
> @@ -592,7 +602,7 @@ static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
> if (no_aggr)
> sprintf(cpustr, "CPU%*d%s",
> csv_output ? 0 : -4,
> - evsel_list->cpus->map[cpu], csv_sep);
> + perf_evsel__cpus(evsel)->map[cpu], csv_sep);
>
> fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel));
>
> @@ -788,7 +798,7 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
> if (no_aggr)
> sprintf(cpustr, "CPU%*d%s",
> csv_output ? 0 : -4,
> - evsel_list->cpus->map[cpu], csv_sep);
> + perf_evsel__cpus(evsel)->map[cpu], csv_sep);
> else
> cpu = 0;
>
> @@ -949,14 +959,14 @@ static void print_counter(struct perf_evsel *counter)
> u64 ena, run, val;
> int cpu;
>
> - for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
> + for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
> val = counter->counts->cpu[cpu].val;
> ena = counter->counts->cpu[cpu].ena;
> run = counter->counts->cpu[cpu].run;
> if (run == 0 || ena == 0) {
> fprintf(output, "CPU%*d%s%*s%s%*s",
> csv_output ? 0 : -4,
> - evsel_list->cpus->map[cpu], csv_sep,
> + perf_evsel__cpus(counter)->map[cpu], csv_sep,
> csv_output ? 0 : 18,
> counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
> csv_sep,
> @@ -1255,7 +1265,7 @@ int cmd_stat(int argc, const char **argv, const char *prefix __used)
>
> list_for_each_entry(pos, &evsel_list->entries, node) {
> if (perf_evsel__alloc_stat_priv(pos) < 0 ||
> - perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
> + perf_evsel__alloc_counts(pos, perf_evsel__nr_cpus(pos)) < 0)
> goto out_free_fd;
> }
>
> diff --git a/tools/perf/util/cpumap.c b/tools/perf/util/cpumap.c
> index adc72f0..2b32ffa 100644
> --- a/tools/perf/util/cpumap.c
> +++ b/tools/perf/util/cpumap.c
> @@ -38,24 +38,19 @@ static struct cpu_map *cpu_map__trim_new(int nr_cpus, int *tmp_cpus)
> return cpus;
> }
>
> -static struct cpu_map *cpu_map__read_all_cpu_map(void)
> +struct cpu_map *cpu_map__read(FILE *file)
> {
> struct cpu_map *cpus = NULL;
> - FILE *onlnf;
> int nr_cpus = 0;
> int *tmp_cpus = NULL, *tmp;
> int max_entries = 0;
> int n, cpu, prev;
> char sep;
>
> - onlnf = fopen("/sys/devices/system/cpu/online", "r");
> - if (!onlnf)
> - return cpu_map__default_new();
> -
> sep = 0;
> prev = -1;
> for (;;) {
> - n = fscanf(onlnf, "%u%c", &cpu, &sep);
> + n = fscanf(file, "%u%c", &cpu, &sep);
> if (n <= 0)
> break;
> if (prev >= 0) {
> @@ -95,6 +90,19 @@ static struct cpu_map *cpu_map__read_all_cpu_map(void)
> cpus = cpu_map__default_new();
> out_free_tmp:
> free(tmp_cpus);
> + return cpus;
> +}
> +
> +static struct cpu_map *cpu_map__read_all_cpu_map(void)
> +{
> + struct cpu_map *cpus = NULL;
> + FILE *onlnf;
> +
> + onlnf = fopen("/sys/devices/system/cpu/online", "r");
> + if (!onlnf)
> + return cpu_map__default_new();
> +
> + cpus = cpu_map__read(onlnf);
> fclose(onlnf);
> return cpus;
> }
> diff --git a/tools/perf/util/cpumap.h b/tools/perf/util/cpumap.h
> index c415185..17b5264 100644
> --- a/tools/perf/util/cpumap.h
> +++ b/tools/perf/util/cpumap.h
> @@ -11,7 +11,7 @@ struct cpu_map {
> struct cpu_map *cpu_map__new(const char *cpu_list);
> struct cpu_map *cpu_map__dummy_new(void);
> void cpu_map__delete(struct cpu_map *map);
> -
> +struct cpu_map *cpu_map__read(FILE *file);
> size_t cpu_map__fprintf(struct cpu_map *map, FILE *fp);
>
> #endif /* __PERF_CPUMAP_H */
> diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
> index b559929..26c0095 100644
> --- a/tools/perf/util/evsel.h
> +++ b/tools/perf/util/evsel.h
> @@ -65,6 +65,7 @@ struct perf_evsel {
> void *func;
> void *data;
> } handler;
> + struct cpu_map *cpus;
> unsigned int sample_size;
> bool supported;
> };
> diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> index 74a5af4..af07c4f 100644
> --- a/tools/perf/util/parse-events.c
> +++ b/tools/perf/util/parse-events.c
> @@ -240,7 +240,8 @@ const char *event_type(int type)
> }
>
> static int add_event(struct list_head **_list, int *idx,
> - struct perf_event_attr *attr, char *name)
> + struct perf_event_attr *attr,
> + struct cpu_map *cpus, char *name)
> {
> struct perf_evsel *evsel;
> struct list_head *list = *_list;
> @@ -260,6 +261,7 @@ static int add_event(struct list_head **_list, int *idx,
> return -ENOMEM;
> }
>
> + evsel->cpus = cpus;
> if (name)
> evsel->name = strdup(name);
> list_add_tail(&evsel->node, list);
> @@ -343,7 +345,7 @@ int parse_events_add_cache(struct list_head **list, int *idx,
> memset(&attr, 0, sizeof(attr));
> attr.config = cache_type | (cache_op << 8) | (cache_result << 16);
> attr.type = PERF_TYPE_HW_CACHE;
> - return add_event(list, idx, &attr, name);
> + return add_event(list, idx, &attr, NULL, name);
> }
>
> static int add_tracepoint(struct list_head **list, int *idx,
> @@ -381,7 +383,7 @@ static int add_tracepoint(struct list_head **list, int *idx,
> attr.sample_period = 1;
>
> snprintf(name, MAX_NAME_LEN, "%s:%s", sys_name, evt_name);
> - return add_event(list, idx, &attr, name);
> + return add_event(list, idx, &attr, NULL, name);
> }
>
> static int add_tracepoint_multi(struct list_head **list, int *idx,
> @@ -492,7 +494,7 @@ int parse_events_add_breakpoint(struct list_head **list, int *idx,
> attr.type = PERF_TYPE_BREAKPOINT;
> attr.sample_period = 1;
>
> - return add_event(list, idx, &attr, NULL);
> + return add_event(list, idx, &attr, NULL, NULL);
> }
>
> static int config_term(struct perf_event_attr *attr,
> @@ -564,7 +566,7 @@ int parse_events_add_numeric(struct list_head **list, int *idx,
> config_attr(&attr, head_config, 1))
> return -EINVAL;
>
> - return add_event(list, idx, &attr, NULL);
> + return add_event(list, idx, &attr, NULL, NULL);
> }
>
> static int parse_events__is_name_term(struct parse_events__term *term)
> @@ -607,7 +609,7 @@ int parse_events_add_pmu(struct list_head **list, int *idx,
> if (perf_pmu__config(pmu, &attr, head_config))
> return -EINVAL;
>
> - return add_event(list, idx, &attr,
> + return add_event(list, idx, &attr, pmu->cpus,
> pmu_event_name(head_config));
> }
>
> diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
> index 67715a4..0b9aca6 100644
> --- a/tools/perf/util/pmu.c
> +++ b/tools/perf/util/pmu.c
> @@ -9,6 +9,7 @@
> #include "util.h"
> #include "pmu.h"
> #include "parse-events.h"
> +#include "cpumap.h"
>
> int perf_pmu_parse(struct list_head *list, char *name);
> extern FILE *perf_pmu_in;
> @@ -222,6 +223,33 @@ static int pmu_type(char *name, __u32 *type)
> return ret;
> }
>
> +static struct cpu_map *pmu_cpumask(char *name)
> +{
> + struct stat st;
> + char path[PATH_MAX];
> + const char *sysfs;
> + FILE *file;
> + struct cpu_map *cpus;
> +
> + sysfs = sysfs_find_mountpoint();
> + if (!sysfs)
> + return NULL;
> +
> + snprintf(path, PATH_MAX,
> + "%s/bus/event_source/devices/%s/cpumask", sysfs, name);
> +
> + if (stat(path, &st) < 0)
> + return NULL;
> +
> + file = fopen(path, "r");
> + if (!file)
> + return NULL;
> +
> + cpus = cpu_map__read(file);
> + fclose(file);
> + return cpus;
> +}
> +
> static struct perf_pmu *pmu_lookup(char *name)
> {
> struct perf_pmu *pmu;
> @@ -244,6 +272,8 @@ static struct perf_pmu *pmu_lookup(char *name)
> if (!pmu)
> return NULL;
>
> + pmu->cpus = pmu_cpumask(name);
> +
> pmu_aliases(name, &aliases);
>
> INIT_LIST_HEAD(&pmu->format);
> diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h
> index 535f2c5..277c874 100644
> --- a/tools/perf/util/pmu.h
> +++ b/tools/perf/util/pmu.h
> @@ -28,6 +28,7 @@ struct perf_pmu__alias {
> struct perf_pmu {
> char *name;
> __u32 type;
> + struct cpu_map *cpus;
> struct list_head format;
> struct list_head aliases;
> struct list_head list;
> --
> 1.7.11.2
>

The key here is to ensure that perf stat -a picks one CPU on each socket
and that from that CPU we can measure events on ALL the uncore PMU
boxes on the socket. I think this is the case today. For instance, it is does
not really make sense to measure one C-Box at a time. For any C-box-level
metric, you need to cover all 8 of them.

How does this play with the -C option of perf stat? I assume -C overrides.

In case perf stat picks up the CPUs from cpumask, then I'd like to keep
a trace of which CPUs were actually used. In other words, I think it would
be useful to have this printed somewhere in the output. Unless this shows
up only when you use the -A option.


\
 
 \ /
  Last update: 2012-08-17 10:41    [W:0.262 / U:0.028 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site