RE: [RFC PATCH v9 3/7] perf stat: Fork and launch perf record when perf stat needs to get retire latency value for a metric.
From: Wang, Weilin
Date: Fri May 24 2024 - 19:46:00 EST
> -----Original Message-----
> From: Namhyung Kim <namhyung@xxxxxxxxxx>
> Sent: Friday, May 24, 2024 4:02 PM
> To: Wang, Weilin <weilin.wang@xxxxxxxxx>
> Cc: Ian Rogers <irogers@xxxxxxxxxx>; Arnaldo Carvalho de Melo
> <acme@xxxxxxxxxx>; Peter Zijlstra <peterz@xxxxxxxxxxxxx>; Ingo Molnar
> <mingo@xxxxxxxxxx>; Alexander Shishkin
> <alexander.shishkin@xxxxxxxxxxxxxxx>; Jiri Olsa <jolsa@xxxxxxxxxx>; Hunter,
> Adrian <adrian.hunter@xxxxxxxxx>; Kan Liang <kan.liang@xxxxxxxxxxxxxxx>;
> linux-perf-users@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; Taylor, Perry
> <perry.taylor@xxxxxxxxx>; Alt, Samantha <samantha.alt@xxxxxxxxx>; Biggers,
> Caleb <caleb.biggers@xxxxxxxxx>
> Subject: Re: [RFC PATCH v9 3/7] perf stat: Fork and launch perf record when
> perf stat needs to get retire latency value for a metric.
>
> Hello,
>
> On Tue, May 21, 2024 at 10:40 AM <weilin.wang@xxxxxxxxx> wrote:
> >
> > From: Weilin Wang <weilin.wang@xxxxxxxxx>
> >
> > When retire_latency value is used in a metric formula, perf stat would fork a
> > perf record process with "-e" and "-W" options. Perf record will collect
> > required retire_latency values in parallel while perf stat is collecting
> > counting values.
> >
> > At the point of time that perf stat stops counting, it would send sigterm
> signal
> > to perf record process and receiving sampling data back from perf record
> from a
> > pipe. Perf stat will then process the received data to get retire latency data
> > and calculate metric result.
> >
> > Another thread is required to synchronize between perf stat and perf record
> > when we pass data through pipe.
> >
> > Signed-off-by: Weilin Wang <weilin.wang@xxxxxxxxx>
> > ---
> > tools/perf/builtin-stat.c | 17 ++
> > tools/perf/util/Build | 1 +
> > tools/perf/util/intel-tpebs.c | 301
> ++++++++++++++++++++++++++++++++++
> > tools/perf/util/intel-tpebs.h | 30 ++++
> > tools/perf/util/stat.h | 1 +
> > 5 files changed, 350 insertions(+)
> > create mode 100644 tools/perf/util/intel-tpebs.c
> > create mode 100644 tools/perf/util/intel-tpebs.h
> >
> > diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> > index 428e9721b908..c0e9dfa3b3c2 100644
> > --- a/tools/perf/builtin-stat.c
> > +++ b/tools/perf/builtin-stat.c
> > @@ -70,6 +70,7 @@
> > #include "util/bpf_counter.h"
> > #include "util/iostat.h"
> > #include "util/util.h"
> > +#include "util/intel-tpebs.h"
> > #include "asm/bug.h"
> >
> > #include <linux/time64.h>
> > @@ -162,6 +163,7 @@ static struct perf_stat_config stat_config = {
> > .ctl_fd = -1,
> > .ctl_fd_ack = -1,
> > .iostat_run = false,
> > + .tpebs_pid = -1,
>
> Where is this set?
>
> > };
> >
> > static void evlist__check_cpu_maps(struct evlist *evlist)
> > @@ -653,6 +655,8 @@ static enum counter_recovery
> stat_handle_error(struct evsel *counter)
> >
> > if (child_pid != -1)
> > kill(child_pid, SIGTERM);
> > + if (stat_config.tpebs_pid != -1)
> > + stop_tpebs();
> > return COUNTER_FATAL;
> > }
> >
> > @@ -673,6 +677,10 @@ static int __run_perf_stat(int argc, const char
> **argv, int run_idx)
> > int err;
> > bool second_pass = false;
> >
> > + err = start_tpebs(&stat_config, evsel_list);
> > + if (err < 0)
> > + return err;
> > +
> > if (forks) {
> > if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
> workload_exec_failed_signal) < 0) {
> > perror("failed to prepare workload");
> > @@ -878,6 +886,10 @@ static int __run_perf_stat(int argc, const char
> **argv, int run_idx)
> >
> > t1 = rdclock();
> >
> > + err = stop_tpebs();
> > + if (err < 0)
> > + return err;
> > +
> > if (stat_config.walltime_run_table)
> > stat_config.walltime_run[run_idx] = t1 - t0;
> >
> > @@ -985,6 +997,9 @@ static void sig_atexit(void)
> > if (child_pid != -1)
> > kill(child_pid, SIGTERM);
> >
> > + if (stat_config.tpebs_pid != -1)
> > + kill(stat_config.tpebs_pid, SIGTERM);
> > +
> > sigprocmask(SIG_SETMASK, &oset, NULL);
> >
> > if (signr == -1)
> > @@ -2918,5 +2933,7 @@ int cmd_stat(int argc, const char **argv)
> > metricgroup__rblist_exit(&stat_config.metric_events);
> > evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack,
> &stat_config.ctl_fd_close);
> >
> > + tpebs_data__delete();
> > +
> > return status;
> > }
> > diff --git a/tools/perf/util/Build b/tools/perf/util/Build
> > index 292170a99ab6..c9f1d0bb6bf8 100644
> > --- a/tools/perf/util/Build
> > +++ b/tools/perf/util/Build
> > @@ -153,6 +153,7 @@ perf-y += clockid.o
> > perf-y += list_sort.o
> > perf-y += mutex.o
> > perf-y += sharded_mutex.o
> > +perf-y += intel-tpebs.o
>
> Can we make this Intel (or x86) only?
>
> >
> > perf-$(CONFIG_LIBBPF) += bpf_map.o
> > perf-$(CONFIG_PERF_BPF_SKEL) += bpf_counter.o
> > diff --git a/tools/perf/util/intel-tpebs.c b/tools/perf/util/intel-tpebs.c
> > new file mode 100644
> > index 000000000000..13f65e39a845
> > --- /dev/null
> > +++ b/tools/perf/util/intel-tpebs.c
> > @@ -0,0 +1,301 @@
> > +// SPDX-License-Identifier: GPL-2.0-only
> > +/*
> > + * intel_tpebs.c: Intel TPEBS support
> > + */
> > +
> > +
> > +#include <sys/param.h>
> > +#include <subcmd/run-command.h>
> > +#include <thread.h>
> > +#include "intel-tpebs.h"
> > +#include <linux/list.h>
> > +#include <linux/zalloc.h>
> > +#include <linux/err.h>
> > +#include "sample.h"
> > +#include "debug.h"
> > +#include "evlist.h"
> > +#include "evsel.h"
> > +#include "session.h"
> > +#include "tool.h"
> > +#include "metricgroup.h"
> > +#include <sys/stat.h>
> > +#include <sys/file.h>
> > +#include <poll.h>
> > +
> > +#define PERF_DATA "-"
> > +
> > +struct list_head tpebs_results = LIST_HEAD_INIT(tpebs_results);
>
> static LIST_HEAD(tpebs_results);
>
>
> > +static pthread_t reader_thread;
> > +static struct child_process *cmd;
>
> Maybe better to have the 'tpebs_' prefix.
>
>
> > +static struct perf_stat_config *stat_config;
>
> Is this really needed? ...
>
>
> > +static size_t tpebs_event_size;
> > +
> > +static int get_perf_record_args(const char **record_argv, char buf[])
>
> ... I think you can just pass the stat_config argument
> from start_tpebs().
>
> > +{
> > + struct tpebs_retire_lat *e;
> > + int i = 0;
> > +
> > + pr_debug("Prepare perf record for retire_latency\n");
> > +
> > + record_argv[i++] = "perf";
> > + record_argv[i++] = "record";
> > + record_argv[i++] = "-W";
> > + record_argv[i++] = "--synth=no";
> > + record_argv[i++] = buf;
> > +
> > + if (stat_config->user_requested_cpu_list) {
> > + record_argv[i++] = "-C";
> > + record_argv[i++] = stat_config->user_requested_cpu_list;
> > + }
> > +
> > + if (stat_config->system_wide)
> > + record_argv[i++] = "-a";
> > +
> > + if (!stat_config->system_wide && !stat_config-
> >user_requested_cpu_list) {
> > + pr_err("Require -a or -C option to run sampling.\n");
> > + return -ECANCELED;
> > + }
> > +
> > + list_for_each_entry(e, &tpebs_results, nd) {
> > + record_argv[i++] = "-e";
> > + record_argv[i++] = e->name;
> > + }
> > +
> > + record_argv[i++] = "-o";
> > + record_argv[i++] = PERF_DATA;
> > +
> > + return 0;
> > +}
> > +
> > +static int prepare_run_command(const char **argv)
> > +{
> > + cmd = zalloc(sizeof(struct child_process));
> > + if (!cmd)
> > + return -ENOMEM;
> > + cmd->argv = argv;
> > + cmd->out = -1;
> > + return 0;
> > +}
> > +
> > +static int prepare_perf_record(int control_fd[], int ack_fd[])
> > +{
> > + const char **record_argv;
> > + int ret;
> > + char buf[32];
> > +
> > + scnprintf(buf, sizeof(buf), "--control=fd:%d,%d", control_fd[0],
> ack_fd[1]);
> > +
> > + record_argv = calloc(12 + 2 * tpebs_event_size, sizeof(char *));
> > + if (!record_argv)
> > + return -ENOMEM;
> > +
> > + ret = get_perf_record_args(record_argv, buf);
> > + if (ret)
> > + goto out;
> > +
> > + ret = prepare_run_command(record_argv);
> > + if (ret)
> > + goto out;
> > + ret = start_command(cmd);
> > +out:
> > + free(record_argv);
> > + return ret;
> > +}
>
> Please add a blank line.
>
>
> > +struct sample_data_reader {
> > + struct perf_tool tool;
> > + struct perf_session *session;
>
> You don't need this, right? Why not pass the 'tool' directly?
>
> > +};
> > +
> > +static void tpebs_retire_lat__delete(struct tpebs_retire_lat *r)
> > +{
> > + zfree(&r->name);
> > + zfree(&r->tpebs_name);
> > + free(r);
> > +}
> > +
> > +void tpebs_data__delete(void)
> > +{
> > + struct tpebs_retire_lat *r, *rtmp;
> > +
> > + list_for_each_entry_safe(r, rtmp, &tpebs_results, nd) {
> > + list_del_init(&r->nd);
> > + tpebs_retire_lat__delete(r);
> > + }
> > + free(cmd);
> > +}
> > +
> > +static int process_sample_event(struct perf_tool *tool __maybe_unused,
> > + union perf_event *event __maybe_unused,
> > + struct perf_sample *sample,
> > + struct evsel *evsel,
> > + struct machine *machine __maybe_unused)
> > +{
> > + int ret = 0;
> > + const char *evname;
> > + struct tpebs_retire_lat *t;
> > +
> > + evname = evsel__name(evsel);
> > +
> > + /*
> > + * Need to handle per core results? We are assuming average retire
>
> I don't know. It depends on the use case where you want
> different per-core retire-latency for the (per-core) metric.
>
>
> > + * latency value will be used. Save the number of samples and the sum
> of
> > + * retire latency value for each event.
> > + */
> > + list_for_each_entry(t, &tpebs_results, nd) {
> > + if (!strcmp(evname, t->name)) {
> > + t->count += 1;
> > + t->sum += sample->retire_lat;
> > + t->val = (double) t->sum / t->count;
> > + break;
> > + }
> > + }
> > +
> > + return ret;
> > +}
> > +
> > +static int process_feature_event(struct perf_session *session,
> > + union perf_event *event)
> > +{
> > + if (event->feat.feat_id < HEADER_LAST_FEATURE)
> > + return perf_event__process_feature(session, event);
> > + return 0;
> > +}
> > +
> > +static void *__sample_reader(void *arg)
> > +{
> > + struct child_process *child = arg;
> > + struct perf_session *session;
> > + struct perf_data data = {
> > + .mode = PERF_DATA_MODE_READ,
> > + .path = PERF_DATA,
> > + .file.fd = child->out,
> > + };
> > + struct sample_data_reader reader = {
> > + .tool = {
> > + .sample = process_sample_event,
> > + .feature = process_feature_event,
> > + .attr = perf_event__process_attr,
>
> Strange indentations.
Hi Namhyung,
The indentations here look correct on the webpage and in my environment.
I'm not sure why they are different here.
>
>
> > + },
> > + };
> > +
> > + session = perf_session__new(&data, &reader.tool);
> > + if (IS_ERR(session))
> > + return NULL;
> > + reader.session = session;
> > + perf_session__process_events(session);
> > + perf_session__delete(session);
> > +
> > + return NULL;
> > +}
> > +
> > +
> > +int start_tpebs(struct perf_stat_config *perf_stat_config, struct evlist
> *evsel_list)
> > +{
> > + int ret = 0;
> > + struct evsel *evsel;
> > +
> > + stat_config = perf_stat_config;
> > + /*
> > + * Prepare perf record for sampling event retire_latency before fork and
> > + * prepare workload
> > + */
> > + evlist__for_each_entry(evsel_list, evsel) {
> > + if (evsel->retire_lat) {
>
> To reduce the indentation, consider early continue like
>
> if (!evsel->retire_lat)
> continue;
>
>
> > + struct tpebs_retire_lat *new = malloc(sizeof(struct
> tpebs_retire_lat));
>
> I prefer sizeof(*new) instead but it doesn't matter. But you might use
> zalloc() and forget about the 0 initialization later.
>
> > + int i;
> > + char *name;
> > +
> > + pr_debug("perf stat retire latency %s required\n", evsel-
> >name);
> > + if (!new)
> > + return -1;
> > + for (i = strlen(evsel->name) - 1; i > 0; i--) {
> > + if (evsel->name[i] == 'R')
> > + break;
>
> I remember Ian checked '/' and ':' too. Otherwise it can detect
> R in the event name and treat it as a modifier.
>
> > + }
> > + if (i <= 0 || evsel->name[i] != 'R')
> > + return -1;
>
> You need to free the memory here and other places.
>
> > +
> > + name = strdup(evsel->name);
> > + if (!name)
> > + return -ENOMEM;
> > + name[i] = 'p';
> > + new->name = strdup(name);
> > + free(name);
>
> Seems like an unnecessary allocation. Why not use 'name'
> directly?
>
> > + new->tpebs_name = strdup(evsel->name);
> > + if (!new->tpebs_name)
> > + return -ENOMEM;
>
> Maybe orig_name? But I'm not sure if it's really needed..
> Can we just teach perf record to understand 'R' and to act
> like 'p'?
>
>
> > + new->count = 0;
> > + new->sum = 0;
> > + list_add_tail(&new->nd, &tpebs_results);
> > + tpebs_event_size += 1;
> > + }
> > + }
> > +
> > + if (tpebs_event_size > 0) {
> > + struct pollfd pollfd = { .events = POLLIN, };
> > + int control_fd[2], ack_fd[2], len;
> > + char ack_buf[8];
> > +
> > + /*Create control and ack fd for --control*/
> > + if (pipe(control_fd) < 0) {
> > + pr_err("Failed to create control fifo");
> > + return -1;
> > + }
> > + if (pipe(ack_fd) < 0) {
> > + pr_err("Failed to create control fifo");
> > + return -1;
>
> Need to close the pipes here and other places.
The pipes are closed at the end of the function.
>
> > + }
> > +
> > + ret = prepare_perf_record(control_fd, ack_fd);
> > + if (ret)
> > + return ret;
> > + if (pthread_create(&reader_thread, NULL, __sample_reader, cmd)) {
> > + kill(cmd->pid, SIGTERM);
> > + close(cmd->out);
> > + pr_err("Could not create thread to process sample data.\n");
> > + return -1;
> > + }
> > + /* Wait for perf record initialization.*/
> > + len = strlen("enable");
> > + ret = write(control_fd[1], "enable", len);
> > + if (ret != len) {
> > + pr_err("perf record control write control message failed\n");
> > + goto out;
> > + }
> > +
> > + pollfd.fd = ack_fd[0];
> > + if (!poll(&pollfd, 1, 2000))
> > + goto out;
>
> Do we need this? Why not just read?
I could remove this part if we don't need it.
Thanks,
Weilin
>
> > +
> > + ret = read(ack_fd[0], ack_buf, sizeof(ack_buf));
> > + if (ret > 0)
> > + ret = strcmp(ack_buf, "ack\n");
> > + else {
> > + pr_err("perf record control ack failed\n");
> > + goto out;
> > + }
> > + pr_debug("Received ack from perf record\n");
> > +out:
> > + close(control_fd[0]);
> > + close(control_fd[1]);
> > + close(ack_fd[0]);
> > + close(ack_fd[1]);
> > + }
> > + return ret;
> > +}
> > +
> > +
> > +int stop_tpebs(void)
> > +{
> > + int ret = 0;
> > +
> > + if (tpebs_event_size > 0) {
> > + kill(cmd->pid, SIGTERM);
> > + pthread_join(reader_thread, NULL);
> > + close(cmd->out);
> > + ret = finish_command(cmd);
> > + if (ret == -ERR_RUN_COMMAND_WAITPID_SIGNAL)
> > + ret = 0;
> > + }
> > + return ret;
> > +}
> > diff --git a/tools/perf/util/intel-tpebs.h b/tools/perf/util/intel-tpebs.h
> > new file mode 100644
> > index 000000000000..25e3e6729146
> > --- /dev/null
> > +++ b/tools/perf/util/intel-tpebs.h
> > @@ -0,0 +1,30 @@
> > +/* SPDX-License-Identifier: GPL-2.0-only */
> > +/*
> > + * intel_tpebs.h: Intel TEPBS support
> > + */
> > +#include "stat.h"
> > +
> > +#ifndef INCLUDE__PERF_INTEL_TPEBS_H__
> > +#define INCLUDE__PERF_INTEL_TPEBS_H__
> > +
> > +struct tpebs_retire_lat {
> > + struct list_head nd;
> > + /* Event name */
> > + const char *name;
> > + /* Event name with the TPEBS modifier R */
> > + const char *tpebs_name;
> > + /* Count of retire_latency values found in sample data */
> > + size_t count;
> > + /* Sum of all the retire_latency values in sample data */
> > + int sum;
> > + /* Average of retire_latency, val = sum / count */
> > + double val;
> > +};
> > +
> > +extern struct list_head tpebs_results;
> > +
> > +int start_tpebs(struct perf_stat_config *perf_stat_config, struct evlist
> *evsel_list);
> > +int stop_tpebs(void);
> > +void tpebs_data__delete(void);
>
> I think it's better to have the same prefix.
>
> tpebs_start()
> tpebs_stop()
> tpebs_delete()
>
> Thanks,
> Namhyung
>
> > +
> > +#endif
> > diff --git a/tools/perf/util/stat.h b/tools/perf/util/stat.h
> > index fd7a187551bd..6e86311bc75e 100644
> > --- a/tools/perf/util/stat.h
> > +++ b/tools/perf/util/stat.h
> > @@ -110,6 +110,7 @@ struct perf_stat_config {
> > struct cpu_aggr_map *cpus_aggr_map;
> > u64 *walltime_run;
> > struct rblist metric_events;
> > + pid_t tpebs_pid;
> > int ctl_fd;
> > int ctl_fd_ack;
> > bool ctl_fd_close;
> > --
> > 2.43.0
> >