Linux Audio

Check our new training course

Linux kernel drivers training

May 6-19, 2025
Register
Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * builtin-stat.c
   4 *
   5 * Builtin stat command: Give a precise performance counters summary
   6 * overview about any workload, CPU or specific PID.
   7 *
   8 * Sample output:
   9
  10   $ perf stat ./hackbench 10
  11
  12  Time: 0.118
  13
  14  Performance counter stats for './hackbench 10':
  15
  16       1708.761321 task-clock                #   11.037 CPUs utilized
  17            41,190 context-switches          #    0.024 M/sec
  18             6,735 CPU-migrations            #    0.004 M/sec
  19            17,318 page-faults               #    0.010 M/sec
  20     5,205,202,243 cycles                    #    3.046 GHz
  21     3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  22     1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  23     2,603,501,247 instructions              #    0.50  insns per cycle
  24                                             #    1.48  stalled cycles per insn
  25       484,357,498 branches                  #  283.455 M/sec
  26         6,388,934 branch-misses             #    1.32% of all branches
  27
  28        0.154822978  seconds time elapsed
  29
  30 *
  31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  32 *
  33 * Improvements and fixes by:
  34 *
  35 *   Arjan van de Ven <arjan@linux.intel.com>
  36 *   Yanmin Zhang <yanmin.zhang@intel.com>
  37 *   Wu Fengguang <fengguang.wu@intel.com>
  38 *   Mike Galbraith <efault@gmx.de>
  39 *   Paul Mackerras <paulus@samba.org>
  40 *   Jaswinder Singh Rajput <jaswinder@kernel.org>
 
 
  41 */
  42
  43#include "builtin.h"
  44#include "perf.h"
  45#include "util/cgroup.h"
  46#include <subcmd/parse-options.h>
 
  47#include "util/parse-events.h"
  48#include "util/pmu.h"
  49#include "util/event.h"
  50#include "util/evlist.h"
  51#include "util/evlist-hybrid.h"
  52#include "util/evsel.h"
  53#include "util/debug.h"
  54#include "util/color.h"
  55#include "util/stat.h"
  56#include "util/header.h"
  57#include "util/cpumap.h"
 
  58#include "util/thread_map.h"
  59#include "util/counts.h"
  60#include "util/topdown.h"
  61#include "util/session.h"
  62#include "util/tool.h"
  63#include "util/string2.h"
  64#include "util/metricgroup.h"
  65#include "util/synthetic-events.h"
  66#include "util/target.h"
  67#include "util/time-utils.h"
  68#include "util/top.h"
  69#include "util/affinity.h"
  70#include "util/pfm.h"
  71#include "util/bpf_counter.h"
  72#include "util/iostat.h"
  73#include "util/pmu-hybrid.h"
  74#include "asm/bug.h"
  75
  76#include <linux/time64.h>
  77#include <linux/zalloc.h>
  78#include <api/fs/fs.h>
  79#include <errno.h>
  80#include <signal.h>
  81#include <stdlib.h>
  82#include <sys/prctl.h>
  83#include <inttypes.h>
  84#include <locale.h>
  85#include <math.h>
  86#include <sys/types.h>
  87#include <sys/stat.h>
  88#include <sys/wait.h>
  89#include <unistd.h>
  90#include <sys/time.h>
  91#include <sys/resource.h>
  92#include <linux/err.h>
  93
  94#include <linux/ctype.h>
  95#include <perf/evlist.h>
  96#include <internal/threadmap.h>
  97
  98#define DEFAULT_SEPARATOR	" "
  99#define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
 
 100
 101static void print_counters(struct timespec *ts, int argc, const char **argv);
 
 
 
 102
 103/* Default events used for perf stat -T */
 104static const char *transaction_attrs = {
 105	"task-clock,"
 106	"{"
 107	"instructions,"
 108	"cycles,"
 109	"cpu/cycles-t/,"
 110	"cpu/tx-start/,"
 111	"cpu/el-start/,"
 112	"cpu/cycles-ct/"
 113	"}"
 114};
 115
 116/* More limited version when the CPU does not have all events. */
 117static const char * transaction_limited_attrs = {
 118	"task-clock,"
 119	"{"
 120	"instructions,"
 121	"cycles,"
 122	"cpu/cycles-t/,"
 123	"cpu/tx-start/"
 124	"}"
 125};
 126
 127static const char * topdown_attrs[] = {
 128	"topdown-total-slots",
 129	"topdown-slots-retired",
 130	"topdown-recovery-bubbles",
 131	"topdown-fetch-bubbles",
 132	"topdown-slots-issued",
 133	NULL,
 134};
 135
 136static const char *topdown_metric_attrs[] = {
 137	"slots",
 138	"topdown-retiring",
 139	"topdown-bad-spec",
 140	"topdown-fe-bound",
 141	"topdown-be-bound",
 142	NULL,
 143};
 144
 145static const char *topdown_metric_L2_attrs[] = {
 146	"slots",
 147	"topdown-retiring",
 148	"topdown-bad-spec",
 149	"topdown-fe-bound",
 150	"topdown-be-bound",
 151	"topdown-heavy-ops",
 152	"topdown-br-mispredict",
 153	"topdown-fetch-lat",
 154	"topdown-mem-bound",
 155	NULL,
 156};
 157
 158#define TOPDOWN_MAX_LEVEL			2
 159
 160static const char *smi_cost_attrs = {
 161	"{"
 162	"msr/aperf/,"
 163	"msr/smi/,"
 164	"cycles"
 165	"}"
 166};
 167
 168static struct evlist	*evsel_list;
 169static bool all_counters_use_bpf = true;
 170
 171static struct target target = {
 172	.uid	= UINT_MAX,
 173};
 174
 175#define METRIC_ONLY_LEN 20
 
 
 
 
 
 176
 177static volatile sig_atomic_t	child_pid			= -1;
 
 
 
 
 
 178static int			detailed_run			=  0;
 179static bool			transaction_run;
 180static bool			topdown_run			= false;
 181static bool			smi_cost			= false;
 182static bool			smi_reset			= false;
 183static int			big_num_opt			=  -1;
 
 
 
 
 184static const char		*pre_cmd			= NULL;
 185static const char		*post_cmd			= NULL;
 186static bool			sync_run			= false;
 
 
 
 187static bool			forever				= false;
 188static bool			force_metric_only		= false;
 189static struct timespec		ref_time;
 190static bool			append_file;
 191static bool			interval_count;
 192static const char		*output_name;
 193static int			output_fd;
 194static char			*metrics;
 195
 196struct perf_stat {
 197	bool			 record;
 198	struct perf_data	 data;
 199	struct perf_session	*session;
 200	u64			 bytes_written;
 201	struct perf_tool	 tool;
 202	bool			 maps_allocated;
 203	struct perf_cpu_map	*cpus;
 204	struct perf_thread_map *threads;
 205	enum aggr_mode		 aggr_mode;
 206};
 207
 208static struct perf_stat		perf_stat;
 209#define STAT_RECORD		perf_stat.record
 210
 211static volatile sig_atomic_t done = 0;
 212
 213static struct perf_stat_config stat_config = {
 214	.aggr_mode		= AGGR_GLOBAL,
 215	.scale			= true,
 216	.unit_width		= 4, /* strlen("unit") */
 217	.run_count		= 1,
 218	.metric_only_len	= METRIC_ONLY_LEN,
 219	.walltime_nsecs_stats	= &walltime_nsecs_stats,
 220	.ru_stats		= &ru_stats,
 221	.big_num		= true,
 222	.ctl_fd			= -1,
 223	.ctl_fd_ack		= -1,
 224	.iostat_run		= false,
 225};
 226
 227static bool cpus_map_matched(struct evsel *a, struct evsel *b)
 228{
 229	if (!a->core.cpus && !b->core.cpus)
 230		return true;
 231
 232	if (!a->core.cpus || !b->core.cpus)
 233		return false;
 234
 235	if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
 236		return false;
 237
 238	for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
 239		if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
 240		    perf_cpu_map__cpu(b->core.cpus, i).cpu)
 241			return false;
 242	}
 243
 244	return true;
 245}
 246
 247static void evlist__check_cpu_maps(struct evlist *evlist)
 248{
 249	struct evsel *evsel, *pos, *leader;
 250	char buf[1024];
 251
 252	if (evlist__has_hybrid(evlist))
 253		evlist__warn_hybrid_group(evlist);
 254
 255	evlist__for_each_entry(evlist, evsel) {
 256		leader = evsel__leader(evsel);
 257
 258		/* Check that leader matches cpus with each member. */
 259		if (leader == evsel)
 260			continue;
 261		if (cpus_map_matched(leader, evsel))
 262			continue;
 263
 264		/* If there's mismatch disable the group and warn user. */
 265		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
 266		evsel__group_desc(leader, buf, sizeof(buf));
 267		pr_warning("  %s\n", buf);
 268
 269		if (verbose > 0) {
 270			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
 271			pr_warning("     %s: %s\n", leader->name, buf);
 272			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
 273			pr_warning("     %s: %s\n", evsel->name, buf);
 274		}
 275
 276		for_each_group_evsel(pos, leader)
 277			evsel__remove_from_group(pos, leader);
 278	}
 279}
 280
 281static inline void diff_timespec(struct timespec *r, struct timespec *a,
 282				 struct timespec *b)
 283{
 284	r->tv_sec = a->tv_sec - b->tv_sec;
 285	if (a->tv_nsec < b->tv_nsec) {
 286		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 287		r->tv_sec--;
 288	} else {
 289		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 290	}
 291}
 292
 293static void perf_stat__reset_stats(void)
 294{
 295	evlist__reset_stats(evsel_list);
 296	perf_stat__reset_shadow_stats();
 297}
 298
 299static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 300				     union perf_event *event,
 301				     struct perf_sample *sample __maybe_unused,
 302				     struct machine *machine __maybe_unused)
 303{
 304	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 305		pr_err("failed to write perf data, error: %m\n");
 306		return -1;
 307	}
 308
 309	perf_stat.bytes_written += event->header.size;
 310	return 0;
 311}
 312
 313static int write_stat_round_event(u64 tm, u64 type)
 314{
 315	return perf_event__synthesize_stat_round(NULL, tm, type,
 316						 process_synthesized_event,
 317						 NULL);
 318}
 319
 320#define WRITE_STAT_ROUND_EVENT(time, interval) \
 321	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 322
 323#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
 
 
 324
 325static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
 326				   struct perf_counts_values *count)
 327{
 328	struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
 329	struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
 330
 331	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 332					   process_synthesized_event, NULL);
 333}
 334
 335static int read_single_counter(struct evsel *counter, int cpu_map_idx,
 336			       int thread, struct timespec *rs)
 337{
 338	switch(counter->tool_event) {
 339		case PERF_TOOL_DURATION_TIME: {
 340			u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
 341			struct perf_counts_values *count =
 342				perf_counts(counter->counts, cpu_map_idx, thread);
 343			count->ena = count->run = val;
 344			count->val = val;
 345			return 0;
 346		}
 347		case PERF_TOOL_USER_TIME:
 348		case PERF_TOOL_SYSTEM_TIME: {
 349			u64 val;
 350			struct perf_counts_values *count =
 351				perf_counts(counter->counts, cpu_map_idx, thread);
 352			if (counter->tool_event == PERF_TOOL_USER_TIME)
 353				val = ru_stats.ru_utime_usec_stat.mean;
 354			else
 355				val = ru_stats.ru_stime_usec_stat.mean;
 356			count->ena = count->run = val;
 357			count->val = val;
 358			return 0;
 359		}
 360		default:
 361		case PERF_TOOL_NONE:
 362			return evsel__read_counter(counter, cpu_map_idx, thread);
 363		case PERF_TOOL_MAX:
 364			/* This should never be reached */
 365			return 0;
 366	}
 367}
 368
 369/*
 370 * Read out the results of a single counter:
 371 * do not aggregate counts across CPUs in system-wide mode
 372 */
 373static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
 374{
 375	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
 376	int thread;
 377
 378	if (!counter->supported)
 379		return -ENOENT;
 380
 381	for (thread = 0; thread < nthreads; thread++) {
 382		struct perf_counts_values *count;
 383
 384		count = perf_counts(counter->counts, cpu_map_idx, thread);
 385
 386		/*
 387		 * The leader's group read loads data into its group members
 388		 * (via evsel__read_counter()) and sets their count->loaded.
 389		 */
 390		if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
 391		    read_single_counter(counter, cpu_map_idx, thread, rs)) {
 392			counter->counts->scaled = -1;
 393			perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
 394			perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
 395			return -1;
 396		}
 397
 398		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
 399
 400		if (STAT_RECORD) {
 401			if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
 402				pr_err("failed to write stat event\n");
 403				return -1;
 404			}
 405		}
 406
 407		if (verbose > 1) {
 408			fprintf(stat_config.output,
 409				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 410					evsel__name(counter),
 411					perf_cpu_map__cpu(evsel__cpus(counter),
 412							  cpu_map_idx).cpu,
 413					count->val, count->ena, count->run);
 414		}
 415	}
 416
 417	return 0;
 418}
 419
 420static int read_affinity_counters(struct timespec *rs)
 421{
 422	struct evlist_cpu_iterator evlist_cpu_itr;
 423	struct affinity saved_affinity, *affinity;
 424
 425	if (all_counters_use_bpf)
 426		return 0;
 427
 428	if (!target__has_cpu(&target) || target__has_per_thread(&target))
 429		affinity = NULL;
 430	else if (affinity__setup(&saved_affinity) < 0)
 431		return -1;
 432	else
 433		affinity = &saved_affinity;
 434
 435	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 436		struct evsel *counter = evlist_cpu_itr.evsel;
 437
 438		if (evsel__is_bpf(counter))
 439			continue;
 440
 441		if (!counter->err) {
 442			counter->err = read_counter_cpu(counter, rs,
 443							evlist_cpu_itr.cpu_map_idx);
 444		}
 445	}
 446	if (affinity)
 447		affinity__cleanup(&saved_affinity);
 448
 449	return 0;
 450}
 451
 452static int read_bpf_map_counters(void)
 453{
 454	struct evsel *counter;
 455	int err;
 456
 457	evlist__for_each_entry(evsel_list, counter) {
 458		if (!evsel__is_bpf(counter))
 459			continue;
 460
 461		err = bpf_counter__read(counter);
 462		if (err)
 463			return err;
 
 464	}
 465	return 0;
 466}
 467
 468static int read_counters(struct timespec *rs)
 469{
 470	if (!stat_config.stop_read_counter) {
 471		if (read_bpf_map_counters() ||
 472		    read_affinity_counters(rs))
 473			return -1;
 
 
 
 474	}
 475	return 0;
 476}
 477
 478static void process_counters(void)
 479{
 480	struct evsel *counter;
 481
 482	evlist__for_each_entry(evsel_list, counter) {
 483		if (counter->err)
 484			pr_debug("failed to read counter %s\n", counter->name);
 485		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
 486			pr_warning("failed to process counter %s\n", counter->name);
 487		counter->err = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 488	}
 489
 490	perf_stat_merge_counters(&stat_config, evsel_list);
 491	perf_stat_process_percore(&stat_config, evsel_list);
 492	perf_stat_process_shadow_stats(&stat_config, evsel_list);
 493}
 494
 495static void process_interval(void)
 
 
 
 496{
 497	struct timespec ts, rs;
 498
 499	clock_gettime(CLOCK_MONOTONIC, &ts);
 500	diff_timespec(&rs, &ts, &ref_time);
 501
 502	perf_stat__reset_shadow_per_stat(&rt_stat);
 503	evlist__reset_aggr_stats(evsel_list);
 504
 505	if (read_counters(&rs) == 0)
 506		process_counters();
 507
 508	if (STAT_RECORD) {
 509		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 510			pr_err("failed to write stat round event\n");
 511	}
 512
 513	init_stats(&walltime_nsecs_stats);
 514	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
 515	print_counters(&rs, 0, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 516}
 517
 518static bool handle_interval(unsigned int interval, int *times)
 
 
 
 
 
 519{
 520	if (interval) {
 521		process_interval();
 522		if (interval_count && !(--(*times)))
 523			return true;
 524	}
 525	return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 526}
 527
 528static int enable_counters(void)
 
 
 
 
 529{
 530	struct evsel *evsel;
 531	int err;
 
 532
 533	evlist__for_each_entry(evsel_list, evsel) {
 534		if (!evsel__is_bpf(evsel))
 535			continue;
 536
 537		err = bpf_counter__enable(evsel);
 538		if (err)
 539			return err;
 540	}
 541
 542	/*
 543	 * We need to enable counters only if:
 544	 * - we don't have tracee (attaching to task or cpu)
 545	 * - we have initial delay configured
 546	 */
 547	if (!target__none(&target)) {
 548		if (!all_counters_use_bpf)
 549			evlist__enable(evsel_list);
 550	}
 551	return 0;
 552}
 553
 554static void disable_counters(void)
 555{
 556	struct evsel *counter;
 557
 558	/*
 559	 * If we don't have tracee (attaching to task or cpu), counters may
 560	 * still be running. To get accurate group ratios, we must stop groups
 561	 * from counting before reading their constituent counters.
 562	 */
 563	if (!target__none(&target)) {
 564		evlist__for_each_entry(evsel_list, counter)
 565			bpf_counter__disable(counter);
 566		if (!all_counters_use_bpf)
 567			evlist__disable(evsel_list);
 568	}
 569}
 570
 571static volatile sig_atomic_t workload_exec_errno;
 
 572
 573/*
 574 * evlist__prepare_workload will send a SIGUSR1
 575 * if the fork fails, since we asked by setting its
 576 * want_signal to true.
 577 */
 578static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 579					void *ucontext __maybe_unused)
 580{
 581	workload_exec_errno = info->si_value.sival_int;
 582}
 583
 584static bool evsel__should_store_id(struct evsel *counter)
 585{
 586	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
 587}
 588
 589static bool is_target_alive(struct target *_target,
 590			    struct perf_thread_map *threads)
 591{
 592	struct stat st;
 593	int i;
 594
 595	if (!target__has_task(_target))
 596		return true;
 597
 598	for (i = 0; i < threads->nr; i++) {
 599		char path[PATH_MAX];
 
 600
 601		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
 602			  threads->map[i].pid);
 603
 604		if (!stat(path, &st))
 605			return true;
 606	}
 607
 608	return false;
 609}
 610
 611static void process_evlist(struct evlist *evlist, unsigned int interval)
 612{
 613	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
 
 
 
 
 614
 615	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
 616		switch (cmd) {
 617		case EVLIST_CTL_CMD_ENABLE:
 618			__fallthrough;
 619		case EVLIST_CTL_CMD_DISABLE:
 620			if (interval)
 621				process_interval();
 622			break;
 623		case EVLIST_CTL_CMD_SNAPSHOT:
 624		case EVLIST_CTL_CMD_ACK:
 625		case EVLIST_CTL_CMD_UNSUPPORTED:
 626		case EVLIST_CTL_CMD_EVLIST:
 627		case EVLIST_CTL_CMD_STOP:
 628		case EVLIST_CTL_CMD_PING:
 629		default:
 630			break;
 631		}
 632	}
 633}
 634
 635static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
 636			int *time_to_sleep)
 637{
 638	int tts = *time_to_sleep;
 639	struct timespec time_diff;
 640
 641	diff_timespec(&time_diff, time_stop, time_start);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 642
 643	tts -= time_diff.tv_sec * MSEC_PER_SEC +
 644	       time_diff.tv_nsec / NSEC_PER_MSEC;
 645
 646	if (tts < 0)
 647		tts = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 648
 649	*time_to_sleep = tts;
 650}
 651
 652static int dispatch_events(bool forks, int timeout, int interval, int *times)
 653{
 654	int child_exited = 0, status = 0;
 655	int time_to_sleep, sleep_time;
 656	struct timespec time_start, time_stop;
 657
 658	if (interval)
 659		sleep_time = interval;
 660	else if (timeout)
 661		sleep_time = timeout;
 662	else
 663		sleep_time = 1000;
 664
 665	time_to_sleep = sleep_time;
 
 
 666
 667	while (!done) {
 668		if (forks)
 669			child_exited = waitpid(child_pid, &status, WNOHANG);
 670		else
 671			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
 672
 673		if (child_exited)
 674			break;
 675
 676		clock_gettime(CLOCK_MONOTONIC, &time_start);
 677		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
 678			if (timeout || handle_interval(interval, times))
 679				break;
 680			time_to_sleep = sleep_time;
 681		} else { /* fd revent */
 682			process_evlist(evsel_list, interval);
 683			clock_gettime(CLOCK_MONOTONIC, &time_stop);
 684			compute_tts(&time_start, &time_stop, &time_to_sleep);
 685		}
 686	}
 687
 688	return status;
 689}
 690
 691enum counter_recovery {
 692	COUNTER_SKIP,
 693	COUNTER_RETRY,
 694	COUNTER_FATAL,
 695};
 696
 697static enum counter_recovery stat_handle_error(struct evsel *counter)
 
 
 
 
 
 
 698{
 699	char msg[BUFSIZ];
 700	/*
 701	 * PPC returns ENXIO for HW counters until 2.6.37
 702	 * (behavior changed with commit b0a873e).
 703	 */
 704	if (errno == EINVAL || errno == ENOSYS ||
 705	    errno == ENOENT || errno == EOPNOTSUPP ||
 706	    errno == ENXIO) {
 707		if (verbose > 0)
 708			ui__warning("%s event is not supported by the kernel.\n",
 709				    evsel__name(counter));
 710		counter->supported = false;
 711		/*
 712		 * errored is a sticky flag that means one of the counter's
 713		 * cpu event had a problem and needs to be reexamined.
 714		 */
 715		counter->errored = true;
 716
 717		if ((evsel__leader(counter) != counter) ||
 718		    !(counter->core.leader->nr_members > 1))
 719			return COUNTER_SKIP;
 720	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
 721		if (verbose > 0)
 722			ui__warning("%s\n", msg);
 723		return COUNTER_RETRY;
 724	} else if (target__has_per_thread(&target) &&
 725		   evsel_list->core.threads &&
 726		   evsel_list->core.threads->err_thread != -1) {
 727		/*
 728		 * For global --per-thread case, skip current
 729		 * error thread.
 730		 */
 731		if (!thread_map__remove(evsel_list->core.threads,
 732					evsel_list->core.threads->err_thread)) {
 733			evsel_list->core.threads->err_thread = -1;
 734			return COUNTER_RETRY;
 735		}
 736	}
 737
 738	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
 739	ui__error("%s\n", msg);
 740
 741	if (child_pid != -1)
 742		kill(child_pid, SIGTERM);
 743	return COUNTER_FATAL;
 744}
 745
 746static int __run_perf_stat(int argc, const char **argv, int run_idx)
 747{
 748	int interval = stat_config.interval;
 749	int times = stat_config.times;
 750	int timeout = stat_config.timeout;
 751	char msg[BUFSIZ];
 752	unsigned long long t0, t1;
 753	struct evsel *counter;
 
 754	size_t l;
 755	int status = 0;
 756	const bool forks = (argc > 0);
 757	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 758	struct evlist_cpu_iterator evlist_cpu_itr;
 759	struct affinity saved_affinity, *affinity = NULL;
 760	int err;
 761	bool second_pass = false;
 
 
 
 762
 763	if (forks) {
 764		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
 
 765			perror("failed to prepare workload");
 766			return -1;
 767		}
 768		child_pid = evsel_list->workload.pid;
 769	}
 770
 771	if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
 772		if (affinity__setup(&saved_affinity) < 0)
 773			return -1;
 774		affinity = &saved_affinity;
 775	}
 776
 777	evlist__for_each_entry(evsel_list, counter) {
 778		counter->reset_group = false;
 779		if (bpf_counter__load(counter, &target))
 780			return -1;
 781		if (!evsel__is_bpf(counter))
 782			all_counters_use_bpf = false;
 783	}
 784
 785	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 786		counter = evlist_cpu_itr.evsel;
 787
 788		/*
 789		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
 790		 * no need to call it again here.
 791		 */
 792		if (target.use_bpf)
 793			break;
 794
 795		if (counter->reset_group || counter->errored)
 796			continue;
 797		if (evsel__is_bpf(counter))
 798			continue;
 799try_again:
 800		if (create_perf_stat_counter(counter, &stat_config, &target,
 801					     evlist_cpu_itr.cpu_map_idx) < 0) {
 802
 
 
 803			/*
 804			 * Weak group failed. We cannot just undo this here
 805			 * because earlier CPUs might be in group mode, and the kernel
 806			 * doesn't support mixing group and non group reads. Defer
 807			 * it to later.
 808			 * Don't close here because we're in the wrong affinity.
 809			 */
 810			if ((errno == EINVAL || errno == EBADF) &&
 811				evsel__leader(counter) != counter &&
 812				counter->weak_group) {
 813				evlist__reset_weak_group(evsel_list, counter, false);
 814				assert(counter->reset_group);
 815				second_pass = true;
 
 816				continue;
 817			}
 818
 819			switch (stat_handle_error(counter)) {
 820			case COUNTER_FATAL:
 821				return -1;
 822			case COUNTER_RETRY:
 823				goto try_again;
 824			case COUNTER_SKIP:
 825				continue;
 826			default:
 827				break;
 828			}
 829
 830		}
 831		counter->supported = true;
 832	}
 833
 834	if (second_pass) {
 835		/*
 836		 * Now redo all the weak group after closing them,
 837		 * and also close errored counters.
 838		 */
 839
 840		/* First close errored or weak retry */
 841		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 842			counter = evlist_cpu_itr.evsel;
 843
 844			if (!counter->reset_group && !counter->errored)
 845				continue;
 846
 847			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
 848		}
 849		/* Now reopen weak */
 850		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 851			counter = evlist_cpu_itr.evsel;
 852
 853			if (!counter->reset_group)
 854				continue;
 855try_again_reset:
 856			pr_debug2("reopening weak %s\n", evsel__name(counter));
 857			if (create_perf_stat_counter(counter, &stat_config, &target,
 858						     evlist_cpu_itr.cpu_map_idx) < 0) {
 859
 860				switch (stat_handle_error(counter)) {
 861				case COUNTER_FATAL:
 862					return -1;
 863				case COUNTER_RETRY:
 864					goto try_again_reset;
 865				case COUNTER_SKIP:
 866					continue;
 867				default:
 868					break;
 869				}
 870			}
 871			counter->supported = true;
 872		}
 873	}
 874	affinity__cleanup(affinity);
 875
 876	evlist__for_each_entry(evsel_list, counter) {
 877		if (!counter->supported) {
 878			perf_evsel__free_fd(&counter->core);
 879			continue;
 880		}
 
 881
 882		l = strlen(counter->unit);
 883		if (l > stat_config.unit_width)
 884			stat_config.unit_width = l;
 885
 886		if (evsel__should_store_id(counter) &&
 887		    evsel__store_ids(counter, evsel_list))
 888			return -1;
 889	}
 890
 891	if (evlist__apply_filters(evsel_list, &counter)) {
 892		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 893			counter->filter, evsel__name(counter), errno,
 894			str_error_r(errno, msg, sizeof(msg)));
 895		return -1;
 896	}
 897
 898	if (STAT_RECORD) {
 899		int fd = perf_data__fd(&perf_stat.data);
 900
 901		if (is_pipe) {
 902			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 903		} else {
 904			err = perf_session__write_header(perf_stat.session, evsel_list,
 905							 fd, false);
 906		}
 907
 908		if (err < 0)
 909			return err;
 910
 911		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
 912							 process_synthesized_event, is_pipe);
 913		if (err < 0)
 914			return err;
 915	}
 916
 917	if (stat_config.initial_delay) {
 918		pr_info(EVLIST_DISABLED_MSG);
 919	} else {
 920		err = enable_counters();
 921		if (err)
 922			return -1;
 923	}
 924
 925	/* Exec the command, if any */
 926	if (forks)
 927		evlist__start_workload(evsel_list);
 928
 929	if (stat_config.initial_delay > 0) {
 930		usleep(stat_config.initial_delay * USEC_PER_MSEC);
 931		err = enable_counters();
 932		if (err)
 933			return -1;
 934
 935		pr_info(EVLIST_ENABLED_MSG);
 936	}
 937
 938	t0 = rdclock();
 939	clock_gettime(CLOCK_MONOTONIC, &ref_time);
 940
 941	if (forks) {
 942		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
 943			status = dispatch_events(forks, timeout, interval, &times);
 944		if (child_pid != -1) {
 945			if (timeout)
 946				kill(child_pid, SIGTERM);
 947			wait4(child_pid, &status, 0, &stat_config.ru_data);
 
 
 948		}
 
 949
 950		if (workload_exec_errno) {
 951			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 952			pr_err("Workload failed: %s\n", emsg);
 953			return -1;
 954		}
 955
 956		if (WIFSIGNALED(status))
 957			psignal(WTERMSIG(status), argv[0]);
 958	} else {
 959		status = dispatch_events(forks, timeout, interval, &times);
 
 
 
 
 
 960	}
 961
 962	disable_counters();
 963
 964	t1 = rdclock();
 965
 966	if (stat_config.walltime_run_table)
 967		stat_config.walltime_run[run_idx] = t1 - t0;
 968
 969	if (interval && stat_config.summary) {
 970		stat_config.interval = 0;
 971		stat_config.stop_read_counter = true;
 972		init_stats(&walltime_nsecs_stats);
 973		update_stats(&walltime_nsecs_stats, t1 - t0);
 974
 975		evlist__copy_prev_raw_counts(evsel_list);
 976		evlist__reset_prev_raw_counts(evsel_list);
 977		evlist__reset_aggr_stats(evsel_list);
 978		perf_stat__reset_shadow_per_stat(&rt_stat);
 979	} else {
 980		update_stats(&walltime_nsecs_stats, t1 - t0);
 981		update_rusage_stats(&ru_stats, &stat_config.ru_data);
 
 
 982	}
 983
 984	/*
 985	 * Closing a group leader splits the group, and as we only disable
 986	 * group leaders, results in remaining events becoming enabled. To
 987	 * avoid arbitrary skew, we must read all counters before closing any
 988	 * group leaders.
 989	 */
 990	if (read_counters(&(struct timespec) { .tv_nsec = t1-t0 }) == 0)
 991		process_counters();
 992
 993	/*
 994	 * We need to keep evsel_list alive, because it's processed
 995	 * later the evsel_list will be closed after.
 996	 */
 997	if (!STAT_RECORD)
 998		evlist__close(evsel_list);
 999
1000	return WEXITSTATUS(status);
1001}
1002
1003static int run_perf_stat(int argc, const char **argv, int run_idx)
1004{
1005	int ret;
1006
1007	if (pre_cmd) {
1008		ret = system(pre_cmd);
1009		if (ret)
1010			return ret;
1011	}
1012
1013	if (sync_run)
1014		sync();
1015
1016	ret = __run_perf_stat(argc, argv, run_idx);
1017	if (ret)
1018		return ret;
1019
1020	if (post_cmd) {
1021		ret = system(post_cmd);
1022		if (ret)
1023			return ret;
1024	}
1025
1026	return ret;
1027}
1028
1029static void print_counters(struct timespec *ts, int argc, const char **argv)
1030{
1031	/* Do not print anything if we record to the pipe. */
1032	if (STAT_RECORD && perf_stat.data.is_pipe)
1033		return;
1034	if (quiet)
1035		return;
1036
1037	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
 
 
 
1038}
1039
1040static volatile sig_atomic_t signr = -1;
1041
1042static void skip_signal(int signo)
1043{
1044	if ((child_pid == -1) || stat_config.interval)
1045		done = 1;
1046
1047	signr = signo;
1048	/*
1049	 * render child_pid harmless
1050	 * won't send SIGTERM to a random
1051	 * process in case of race condition
1052	 * and fast PID recycling
1053	 */
1054	child_pid = -1;
1055}
1056
1057static void sig_atexit(void)
1058{
1059	sigset_t set, oset;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1060
1061	/*
1062	 * avoid race condition with SIGCHLD handler
1063	 * in skip_signal() which is modifying child_pid
1064	 * goal is to avoid send SIGTERM to a random
1065	 * process
1066	 */
1067	sigemptyset(&set);
1068	sigaddset(&set, SIGCHLD);
1069	sigprocmask(SIG_BLOCK, &set, &oset);
1070
1071	if (child_pid != -1)
1072		kill(child_pid, SIGTERM);
1073
1074	sigprocmask(SIG_SETMASK, &oset, NULL);
1075
1076	if (signr == -1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1077		return;
1078
1079	signal(signr, SIG_DFL);
1080	kill(getpid(), signr);
 
 
 
1081}
1082
1083void perf_stat__set_big_num(int set)
1084{
1085	stat_config.big_num = (set != 0);
1086}
 
 
 
1087
1088void perf_stat__set_no_csv_summary(int set)
1089{
1090	stat_config.no_csv_summary = (set != 0);
1091}
 
 
 
 
1092
1093static int stat__set_big_num(const struct option *opt __maybe_unused,
1094			     const char *s __maybe_unused, int unset)
1095{
1096	big_num_opt = unset ? 0 : 1;
1097	perf_stat__set_big_num(!unset);
1098	return 0;
 
 
1099}
1100
1101static int enable_metric_only(const struct option *opt __maybe_unused,
1102			      const char *s __maybe_unused, int unset)
 
1103{
1104	force_metric_only = true;
1105	stat_config.metric_only = !unset;
1106	return 0;
1107}
1108
1109static int append_metric_groups(const struct option *opt __maybe_unused,
1110			       const char *str,
1111			       int unset __maybe_unused)
1112{
1113	if (metrics) {
1114		char *tmp;
1115
1116		if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1117			return -ENOMEM;
1118		free(metrics);
1119		metrics = tmp;
1120	} else {
1121		metrics = strdup(str);
1122		if (!metrics)
1123			return -ENOMEM;
1124	}
1125	return 0;
1126}
1127
1128static int parse_control_option(const struct option *opt,
1129				const char *str,
1130				int unset __maybe_unused)
1131{
1132	struct perf_stat_config *config = opt->value;
1133
1134	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
 
 
 
 
1135}
1136
1137static int parse_stat_cgroups(const struct option *opt,
1138			      const char *str, int unset)
 
1139{
1140	if (stat_config.cgroup_list) {
1141		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1142		return -1;
1143	}
 
 
 
 
 
1144
1145	return parse_cgroups(opt, str, unset);
 
 
1146}
1147
1148static int parse_hybrid_type(const struct option *opt,
1149			     const char *str,
1150			     int unset __maybe_unused)
1151{
1152	struct evlist *evlist = *(struct evlist **)opt->value;
 
1153
1154	if (!list_empty(&evlist->core.entries)) {
1155		fprintf(stderr, "Must define cputype before events/metrics\n");
1156		return -1;
1157	}
1158
1159	evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1160	if (!evlist->hybrid_pmu_name) {
1161		fprintf(stderr, "--cputype %s is not supported!\n", str);
1162		return -1;
1163	}
1164
1165	return 0;
 
 
 
 
1166}
1167
1168static struct option stat_options[] = {
1169	OPT_BOOLEAN('T', "transaction", &transaction_run,
1170		    "hardware transaction statistics"),
1171	OPT_CALLBACK('e', "event", &evsel_list, "event",
1172		     "event selector. use 'perf list' to list available events",
1173		     parse_events_option),
1174	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1175		     "event filter", parse_filter),
1176	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1177		    "child tasks do not inherit counters"),
1178	OPT_STRING('p', "pid", &target.pid, "pid",
1179		   "stat events on existing process id"),
1180	OPT_STRING('t', "tid", &target.tid, "tid",
1181		   "stat events on existing thread id"),
1182#ifdef HAVE_BPF_SKEL
1183	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1184		   "stat events on existing bpf program id"),
1185	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1186		    "use bpf program to count events"),
1187	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1188		   "path to perf_event_attr map"),
1189#endif
1190	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1191		    "system-wide collection from all CPUs"),
1192	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1193		    "Use --no-scale to disable counter scaling for multiplexing"),
1194	OPT_INCR('v', "verbose", &verbose,
1195		    "be more verbose (show counter open errors, etc)"),
1196	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1197		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1198	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1199		    "display details about each run (only with -r option)"),
1200	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1201		    "null run - dont start any counters"),
1202	OPT_INCR('d', "detailed", &detailed_run,
1203		    "detailed run - start a lot of events"),
1204	OPT_BOOLEAN('S', "sync", &sync_run,
1205		    "call sync() before starting a run"),
1206	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1207			   "print large numbers with thousands\' separators",
1208			   stat__set_big_num),
1209	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1210		    "list of cpus to monitor in system-wide"),
1211	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1212		    "disable CPU count aggregation", AGGR_NONE),
1213	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1214	OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1215		    "Merge identical named hybrid events"),
1216	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1217		   "print counts with custom separator"),
1218	OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1219		   "print counts in JSON format"),
1220	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1221		     "monitor event in cgroup name only", parse_stat_cgroups),
1222	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1223		    "expand events for each cgroup"),
1224	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1225	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1226	OPT_INTEGER(0, "log-fd", &output_fd,
1227		    "log output to fd, instead of stderr"),
1228	OPT_STRING(0, "pre", &pre_cmd, "command",
1229			"command to run prior to the measured command"),
1230	OPT_STRING(0, "post", &post_cmd, "command",
1231			"command to run after to the measured command"),
1232	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1233		    "print counts at regular interval in ms "
1234		    "(overhead is possible for values <= 100ms)"),
1235	OPT_INTEGER(0, "interval-count", &stat_config.times,
1236		    "print counts for fixed number of times"),
1237	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1238		    "clear screen in between new interval"),
1239	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1240		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1241	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1242		     "aggregate counts per processor socket", AGGR_SOCKET),
1243	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1244		     "aggregate counts per processor die", AGGR_DIE),
1245	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1246		     "aggregate counts per physical processor core", AGGR_CORE),
1247	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1248		     "aggregate counts per thread", AGGR_THREAD),
1249	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1250		     "aggregate counts per numa node", AGGR_NODE),
1251	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1252		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1253	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1254			"Only print computed metrics. No raw values", enable_metric_only),
1255	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1256		       "don't group metric events, impacts multiplexing"),
1257	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1258		       "don't try to share events between metrics in a group"),
1259	OPT_BOOLEAN(0, "topdown", &topdown_run,
1260			"measure top-down statistics"),
1261	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1262			"Set the metrics level for the top-down statistics (0: max level)"),
1263	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1264			"measure SMI cost"),
1265	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1266		     "monitor specified metrics or metric groups (separated by ,)",
1267		     append_metric_groups),
1268	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1269			 "Configure all used events to run in kernel space.",
1270			 PARSE_OPT_EXCLUSIVE),
1271	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1272			 "Configure all used events to run in user space.",
1273			 PARSE_OPT_EXCLUSIVE),
1274	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1275		    "Use with 'percore' event qualifier to show the event "
1276		    "counts of one hardware thread by sum up total hardware "
1277		    "threads of same physical core"),
1278	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1279		       "print summary for interval mode"),
1280	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1281		       "don't print 'summary' for CSV summary output"),
1282	OPT_BOOLEAN(0, "quiet", &quiet,
1283			"don't print any output, messages or warnings (useful with record)"),
1284	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1285		     "Only enable events on applying cpu with this type "
1286		     "for hybrid platform (e.g. core or atom)",
1287		     parse_hybrid_type),
1288#ifdef HAVE_LIBPFM
1289	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1290		"libpfm4 event selector. use 'perf list' to list available events",
1291		parse_libpfm_events_option),
1292#endif
1293	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1294		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1295		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1296		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1297		      parse_control_option),
1298	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1299			    "measure I/O performance metrics provided by arch/platform",
1300			    iostat_parse),
1301	OPT_END()
1302};
1303
1304static const char *const aggr_mode__string[] = {
1305	[AGGR_CORE] = "core",
1306	[AGGR_DIE] = "die",
1307	[AGGR_GLOBAL] = "global",
1308	[AGGR_NODE] = "node",
1309	[AGGR_NONE] = "none",
1310	[AGGR_SOCKET] = "socket",
1311	[AGGR_THREAD] = "thread",
1312	[AGGR_UNSET] = "unset",
1313};
1314
1315static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1316						struct perf_cpu cpu)
1317{
1318	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1319}
1320
1321static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1322					     struct perf_cpu cpu)
1323{
1324	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1325}
1326
1327static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1328					      struct perf_cpu cpu)
1329{
1330	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1331}
1332
1333static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1334					      struct perf_cpu cpu)
 
1335{
1336	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1337}
1338
1339static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1340						struct perf_cpu cpu)
1341{
1342	return aggr_cpu_id__global(cpu, /*data=*/NULL);
1343}
1344
1345static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1346					     struct perf_cpu cpu)
1347{
1348	return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
 
 
 
 
1349}
1350
1351static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1352					      aggr_get_id_t get_id, struct perf_cpu cpu)
 
1353{
1354	struct aggr_cpu_id id;
 
1355
1356	/* per-process mode - should use global aggr mode */
1357	if (cpu.cpu == -1)
1358		return get_id(config, cpu);
1359
1360	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1361		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1362
1363	id = config->cpus_aggr_map->map[cpu.cpu];
1364	return id;
1365}
1366
1367static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1368						       struct perf_cpu cpu)
1369{
1370	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1371}
1372
1373static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1374						    struct perf_cpu cpu)
 
1375{
1376	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1377}
1378
1379static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1380						     struct perf_cpu cpu)
1381{
1382	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1383}
1384
1385static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1386						     struct perf_cpu cpu)
1387{
1388	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1389}
1390
1391static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1392						       struct perf_cpu cpu)
1393{
1394	return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
 
1395}
1396
1397static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1398						    struct perf_cpu cpu)
 
1399{
1400	return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
 
 
 
 
 
 
 
 
 
 
 
 
1401}
1402
1403static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1404{
1405	switch (aggr_mode) {
1406	case AGGR_SOCKET:
1407		return aggr_cpu_id__socket;
1408	case AGGR_DIE:
1409		return aggr_cpu_id__die;
1410	case AGGR_CORE:
1411		return aggr_cpu_id__core;
1412	case AGGR_NODE:
1413		return aggr_cpu_id__node;
1414	case AGGR_NONE:
1415		return aggr_cpu_id__cpu;
1416	case AGGR_GLOBAL:
1417		return aggr_cpu_id__global;
1418	case AGGR_THREAD:
1419	case AGGR_UNSET:
1420	case AGGR_MAX:
1421	default:
1422		return NULL;
1423	}
1424}
1425
1426static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1427{
1428	switch (aggr_mode) {
1429	case AGGR_SOCKET:
1430		return perf_stat__get_socket_cached;
1431	case AGGR_DIE:
1432		return perf_stat__get_die_cached;
1433	case AGGR_CORE:
1434		return perf_stat__get_core_cached;
1435	case AGGR_NODE:
1436		return perf_stat__get_node_cached;
1437	case AGGR_NONE:
1438		return perf_stat__get_cpu_cached;
1439	case AGGR_GLOBAL:
1440		return perf_stat__get_global_cached;
1441	case AGGR_THREAD:
1442	case AGGR_UNSET:
1443	case AGGR_MAX:
1444	default:
1445		return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1446	}
1447}
1448
1449static int perf_stat_init_aggr_mode(void)
1450{
1451	int nr;
1452	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
 
 
1453
1454	if (get_id) {
1455		bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1456		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1457							 get_id, /*data=*/NULL, needs_sort);
1458		if (!stat_config.aggr_map) {
1459			pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1460			return -1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1461		}
1462		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1463	}
 
1464
1465	if (stat_config.aggr_mode == AGGR_THREAD) {
1466		nr = perf_thread_map__nr(evsel_list->core.threads);
1467		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1468		if (stat_config.aggr_map == NULL)
1469			return -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1470
1471		for (int s = 0; s < nr; s++) {
1472			struct aggr_cpu_id id = aggr_cpu_id__empty();
1473
1474			id.thread_idx = s;
1475			stat_config.aggr_map->map[s] = id;
1476		}
1477		return 0;
1478	}
1479
1480	/*
1481	 * The evsel_list->cpus is the base we operate on,
1482	 * taking the highest cpu number to be the size of
1483	 * the aggregation translate cpumap.
1484	 */
1485	if (evsel_list->core.user_requested_cpus)
1486		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1487	else
1488		nr = 0;
1489	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1490	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1491}
1492
1493static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1494{
1495	if (map) {
1496		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1497			  "cpu_aggr_map refcnt unbalanced\n");
1498		free(map);
1499	}
1500}
1501
1502static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1503{
1504	if (map && refcount_dec_and_test(&map->refcnt))
1505		cpu_aggr_map__delete(map);
1506}
1507
1508static void perf_stat__exit_aggr_mode(void)
1509{
1510	cpu_aggr_map__put(stat_config.aggr_map);
1511	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1512	stat_config.aggr_map = NULL;
1513	stat_config.cpus_aggr_map = NULL;
1514}
1515
1516static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
 
 
 
 
1517{
1518	struct perf_env *env = data;
1519	struct aggr_cpu_id id = aggr_cpu_id__empty();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1520
1521	if (cpu.cpu != -1)
1522		id.socket = env->cpu[cpu.cpu].socket_id;
 
1523
1524	return id;
1525}
1526
1527static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1528{
1529	struct perf_env *env = data;
1530	struct aggr_cpu_id id = aggr_cpu_id__empty();
1531
1532	if (cpu.cpu != -1) {
1533		/*
1534		 * die_id is relative to socket, so start
1535		 * with the socket ID and then add die to
1536		 * make a unique ID.
1537		 */
1538		id.socket = env->cpu[cpu.cpu].socket_id;
1539		id.die = env->cpu[cpu.cpu].die_id;
1540	}
1541
1542	return id;
 
 
 
 
 
1543}
1544
1545static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1546{
1547	struct perf_env *env = data;
1548	struct aggr_cpu_id id = aggr_cpu_id__empty();
1549
1550	if (cpu.cpu != -1) {
1551		/*
1552		 * core_id is relative to socket and die,
1553		 * we need a global id. So we set
1554		 * socket, die id and core id
1555		 */
1556		id.socket = env->cpu[cpu.cpu].socket_id;
1557		id.die = env->cpu[cpu.cpu].die_id;
1558		id.core = env->cpu[cpu.cpu].core_id;
1559	}
1560
1561	return id;
1562}
 
 
 
 
 
 
 
 
 
 
 
 
 
1563
1564static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1565{
1566	struct perf_env *env = data;
1567	struct aggr_cpu_id id = aggr_cpu_id__empty();
 
1568
1569	if (cpu.cpu != -1) {
1570		/*
1571		 * core_id is relative to socket and die,
1572		 * we need a global id. So we set
1573		 * socket, die id and core id
1574		 */
1575		id.socket = env->cpu[cpu.cpu].socket_id;
1576		id.die = env->cpu[cpu.cpu].die_id;
1577		id.core = env->cpu[cpu.cpu].core_id;
1578		id.cpu = cpu;
 
 
 
 
 
1579	}
1580
1581	return id;
 
 
 
 
 
 
 
 
 
 
 
1582}
1583
1584static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
 
 
1585{
1586	struct aggr_cpu_id id = aggr_cpu_id__empty();
 
1587
1588	id.node = perf_env__numa_node(data, cpu);
1589	return id;
 
 
 
 
 
 
1590}
1591
1592static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1593							   void *data __maybe_unused)
1594{
1595	struct aggr_cpu_id id = aggr_cpu_id__empty();
1596
1597	/* it always aggregates to the cpu 0 */
1598	id.cpu = (struct perf_cpu){ .cpu = 0 };
1599	return id;
1600}
 
 
 
 
 
1601
1602static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1603						     struct perf_cpu cpu)
1604{
1605	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1606}
1607static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1608						  struct perf_cpu cpu)
1609{
1610	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1611}
1612
1613static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1614						   struct perf_cpu cpu)
1615{
1616	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1617}
1618
1619static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1620						  struct perf_cpu cpu)
1621{
1622	return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1623}
1624
1625static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1626						   struct perf_cpu cpu)
1627{
1628	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1629}
1630
1631static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1632						     struct perf_cpu cpu)
1633{
1634	return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
 
1635}
1636
1637static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1638{
1639	switch (aggr_mode) {
1640	case AGGR_SOCKET:
1641		return perf_env__get_socket_aggr_by_cpu;
1642	case AGGR_DIE:
1643		return perf_env__get_die_aggr_by_cpu;
 
 
 
1644	case AGGR_CORE:
1645		return perf_env__get_core_aggr_by_cpu;
1646	case AGGR_NODE:
1647		return perf_env__get_node_aggr_by_cpu;
1648	case AGGR_GLOBAL:
1649		return perf_env__get_global_aggr_by_cpu;
 
1650	case AGGR_NONE:
1651		return perf_env__get_cpu_aggr_by_cpu;
1652	case AGGR_THREAD:
1653	case AGGR_UNSET:
1654	case AGGR_MAX:
1655	default:
1656		return NULL;
1657	}
1658}
1659
1660static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1661{
1662	switch (aggr_mode) {
1663	case AGGR_SOCKET:
1664		return perf_stat__get_socket_file;
1665	case AGGR_DIE:
1666		return perf_stat__get_die_file;
1667	case AGGR_CORE:
1668		return perf_stat__get_core_file;
1669	case AGGR_NODE:
1670		return perf_stat__get_node_file;
1671	case AGGR_GLOBAL:
1672		return perf_stat__get_global_file;
1673	case AGGR_NONE:
1674		return perf_stat__get_cpu_file;
1675	case AGGR_THREAD:
1676	case AGGR_UNSET:
1677	case AGGR_MAX:
1678	default:
1679		return NULL;
1680	}
 
1681}
1682
1683static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1684{
1685	struct perf_env *env = &st->session->header.env;
1686	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1687	bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1688
1689	if (stat_config.aggr_mode == AGGR_THREAD) {
1690		int nr = perf_thread_map__nr(evsel_list->core.threads);
1691
1692		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1693		if (stat_config.aggr_map == NULL)
1694			return -ENOMEM;
1695
1696		for (int s = 0; s < nr; s++) {
1697			struct aggr_cpu_id id = aggr_cpu_id__empty();
1698
1699			id.thread_idx = s;
1700			stat_config.aggr_map->map[s] = id;
1701		}
1702		return 0;
1703	}
1704
1705	if (!get_id)
1706		return 0;
1707
1708	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1709						 get_id, env, needs_sort);
1710	if (!stat_config.aggr_map) {
1711		pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1712		return -1;
1713	}
1714	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1715	return 0;
1716}
1717
1718/*
1719 * Add default attributes, if there were no attributes specified or
1720 * if -d/--detailed, -d -d or -d -d -d is used:
1721 */
1722static int add_default_attributes(void)
1723{
1724	int err;
1725	struct perf_event_attr default_attrs0[] = {
1726
1727  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1728  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1729  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1730  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1731
1732  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1733};
1734	struct perf_event_attr frontend_attrs[] = {
1735  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1736};
1737	struct perf_event_attr backend_attrs[] = {
1738  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1739};
1740	struct perf_event_attr default_attrs1[] = {
1741  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1742  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1743  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1744
1745};
1746
1747/*
1748 * Detailed stats (-d), covering the L1 and last level data caches:
1749 */
1750	struct perf_event_attr detailed_attrs[] = {
1751
1752  { .type = PERF_TYPE_HW_CACHE,
1753    .config =
1754	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1755	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1756	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1757
1758  { .type = PERF_TYPE_HW_CACHE,
1759    .config =
1760	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1761	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1762	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1763
1764  { .type = PERF_TYPE_HW_CACHE,
1765    .config =
1766	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1767	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1768	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1769
1770  { .type = PERF_TYPE_HW_CACHE,
1771    .config =
1772	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1773	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1774	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1775};
1776
1777/*
1778 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1779 */
1780	struct perf_event_attr very_detailed_attrs[] = {
1781
1782  { .type = PERF_TYPE_HW_CACHE,
1783    .config =
1784	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1785	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1786	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1787
1788  { .type = PERF_TYPE_HW_CACHE,
1789    .config =
1790	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1791	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1792	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1793
1794  { .type = PERF_TYPE_HW_CACHE,
1795    .config =
1796	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1797	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1798	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1799
1800  { .type = PERF_TYPE_HW_CACHE,
1801    .config =
1802	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1803	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1804	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1805
1806  { .type = PERF_TYPE_HW_CACHE,
1807    .config =
1808	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1809	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1810	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1811
1812  { .type = PERF_TYPE_HW_CACHE,
1813    .config =
1814	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1815	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1816	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1817
1818};
1819
1820/*
1821 * Very, very detailed stats (-d -d -d), adding prefetch events:
1822 */
1823	struct perf_event_attr very_very_detailed_attrs[] = {
1824
1825  { .type = PERF_TYPE_HW_CACHE,
1826    .config =
1827	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1828	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1829	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1830
1831  { .type = PERF_TYPE_HW_CACHE,
1832    .config =
1833	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1834	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1835	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1836};
1837
1838	struct perf_event_attr default_null_attrs[] = {};
1839
1840	/* Set attrs if no event is selected and !null_run: */
1841	if (stat_config.null_run)
1842		return 0;
1843
1844	if (transaction_run) {
1845		struct parse_events_error errinfo;
1846		/* Handle -T as -M transaction. Once platform specific metrics
1847		 * support has been added to the json files, all architectures
1848		 * will use this approach. To determine transaction support
1849		 * on an architecture test for such a metric name.
1850		 */
1851		if (metricgroup__has_metric("transaction")) {
1852			return metricgroup__parse_groups(evsel_list, "transaction",
1853							 stat_config.metric_no_group,
1854							 stat_config.metric_no_merge,
1855							 stat_config.user_requested_cpu_list,
1856							 stat_config.system_wide,
1857							 &stat_config.metric_events);
1858		}
1859
1860		parse_events_error__init(&errinfo);
1861		if (pmu_have_event("cpu", "cycles-ct") &&
1862		    pmu_have_event("cpu", "el-start"))
1863			err = parse_events(evsel_list, transaction_attrs,
1864					   &errinfo);
1865		else
1866			err = parse_events(evsel_list,
1867					   transaction_limited_attrs,
1868					   &errinfo);
1869		if (err) {
1870			fprintf(stderr, "Cannot set up transaction events\n");
1871			parse_events_error__print(&errinfo, transaction_attrs);
1872		}
1873		parse_events_error__exit(&errinfo);
1874		return err ? -1 : 0;
1875	}
1876
1877	if (smi_cost) {
1878		struct parse_events_error errinfo;
1879		int smi;
1880
1881		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1882			fprintf(stderr, "freeze_on_smi is not supported.\n");
1883			return -1;
1884		}
1885
1886		if (!smi) {
1887			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1888				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1889				return -1;
1890			}
1891			smi_reset = true;
1892		}
1893
1894		if (!pmu_have_event("msr", "aperf") ||
1895		    !pmu_have_event("msr", "smi")) {
1896			fprintf(stderr, "To measure SMI cost, it needs "
1897				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1898			return -1;
1899		}
1900		if (!force_metric_only)
1901			stat_config.metric_only = true;
1902
1903		parse_events_error__init(&errinfo);
1904		err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1905		if (err) {
1906			parse_events_error__print(&errinfo, smi_cost_attrs);
1907			fprintf(stderr, "Cannot set up SMI cost events\n");
1908		}
1909		parse_events_error__exit(&errinfo);
1910		return err ? -1 : 0;
1911	}
1912
1913	if (topdown_run) {
1914		const char **metric_attrs = topdown_metric_attrs;
1915		unsigned int max_level = 1;
1916		char *str = NULL;
1917		bool warn = false;
1918		const char *pmu_name = arch_get_topdown_pmu_name(evsel_list, true);
1919
1920		if (!force_metric_only)
1921			stat_config.metric_only = true;
1922
1923		if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1924			metric_attrs = topdown_metric_L2_attrs;
1925			max_level = 2;
1926		}
1927
1928		if (stat_config.topdown_level > max_level) {
1929			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1930			return -1;
1931		} else if (!stat_config.topdown_level)
1932			stat_config.topdown_level = max_level;
1933
1934		if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1935			pr_err("Out of memory\n");
1936			return -1;
1937		}
1938
1939		if (metric_attrs[0] && str) {
1940			if (!stat_config.interval && !stat_config.metric_only) {
1941				fprintf(stat_config.output,
1942					"Topdown accuracy may decrease when measuring long periods.\n"
1943					"Please print the result regularly, e.g. -I1000\n");
1944			}
1945			goto setup_metrics;
1946		}
1947
1948		zfree(&str);
1949
1950		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1951		    stat_config.aggr_mode != AGGR_CORE) {
1952			pr_err("top down event configuration requires --per-core mode\n");
1953			return -1;
1954		}
1955		stat_config.aggr_mode = AGGR_CORE;
1956		if (nr_cgroups || !target__has_cpu(&target)) {
1957			pr_err("top down event configuration requires system-wide mode (-a)\n");
1958			return -1;
1959		}
1960
1961		if (topdown_filter_events(topdown_attrs, &str,
1962				arch_topdown_check_group(&warn),
1963				pmu_name) < 0) {
1964			pr_err("Out of memory\n");
1965			return -1;
1966		}
1967
1968		if (topdown_attrs[0] && str) {
1969			struct parse_events_error errinfo;
1970			if (warn)
1971				arch_topdown_group_warn();
1972setup_metrics:
1973			parse_events_error__init(&errinfo);
1974			err = parse_events(evsel_list, str, &errinfo);
1975			if (err) {
1976				fprintf(stderr,
1977					"Cannot set up top down events %s: %d\n",
1978					str, err);
1979				parse_events_error__print(&errinfo, str);
1980				parse_events_error__exit(&errinfo);
1981				free(str);
1982				return -1;
1983			}
1984			parse_events_error__exit(&errinfo);
1985		} else {
1986			fprintf(stderr, "System does not support topdown\n");
1987			return -1;
1988		}
1989		free(str);
1990	}
1991
1992	if (!stat_config.topdown_level)
1993		stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1994
1995	if (!evsel_list->core.nr_entries) {
1996		if (target__has_cpu(&target))
1997			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1998
1999		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2000			return -1;
2001		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2002			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
2003				return -1;
2004		}
2005		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2006			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2007				return -1;
2008		}
2009		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2010			return -1;
2011		/* Platform specific attrs */
2012		if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2013			return -1;
2014	}
2015
2016	/* Detailed events get appended to the event list: */
2017
2018	if (detailed_run <  1)
2019		return 0;
2020
2021	/* Append detailed run extra attributes: */
2022	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2023		return -1;
2024
2025	if (detailed_run < 2)
2026		return 0;
2027
2028	/* Append very detailed run extra attributes: */
2029	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2030		return -1;
2031
2032	if (detailed_run < 3)
2033		return 0;
2034
2035	/* Append very, very detailed run extra attributes: */
2036	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2037}
2038
2039static const char * const stat_record_usage[] = {
2040	"perf stat record [<options>]",
2041	NULL,
2042};
2043
2044static void init_features(struct perf_session *session)
2045{
2046	int feat;
2047
2048	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2049		perf_header__set_feat(&session->header, feat);
2050
2051	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2052	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2053	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2054	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2055	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2056}
2057
2058static int __cmd_record(int argc, const char **argv)
2059{
2060	struct perf_session *session;
2061	struct perf_data *data = &perf_stat.data;
2062
2063	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2064			     PARSE_OPT_STOP_AT_NON_OPTION);
2065
2066	if (output_name)
2067		data->path = output_name;
2068
2069	if (stat_config.run_count != 1 || forever) {
2070		pr_err("Cannot use -r option with perf stat record.\n");
2071		return -1;
2072	}
2073
2074	session = perf_session__new(data, NULL);
2075	if (IS_ERR(session)) {
2076		pr_err("Perf session creation failed\n");
2077		return PTR_ERR(session);
2078	}
2079
2080	init_features(session);
2081
2082	session->evlist   = evsel_list;
2083	perf_stat.session = session;
2084	perf_stat.record  = true;
2085	return argc;
2086}
2087
2088static int process_stat_round_event(struct perf_session *session,
2089				    union perf_event *event)
2090{
2091	struct perf_record_stat_round *stat_round = &event->stat_round;
2092	struct timespec tsh, *ts = NULL;
2093	const char **argv = session->header.env.cmdline_argv;
2094	int argc = session->header.env.nr_cmdline;
2095
2096	process_counters();
2097
2098	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2099		update_stats(&walltime_nsecs_stats, stat_round->time);
2100
2101	if (stat_config.interval && stat_round->time) {
2102		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2103		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2104		ts = &tsh;
2105	}
2106
2107	print_counters(ts, argc, argv);
2108	return 0;
2109}
2110
2111static
2112int process_stat_config_event(struct perf_session *session,
2113			      union perf_event *event)
2114{
2115	struct perf_tool *tool = session->tool;
2116	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2117
2118	perf_event__read_stat_config(&stat_config, &event->stat_config);
2119
2120	if (perf_cpu_map__empty(st->cpus)) {
2121		if (st->aggr_mode != AGGR_UNSET)
2122			pr_warning("warning: processing task data, aggregation mode not set\n");
2123	} else if (st->aggr_mode != AGGR_UNSET) {
2124		stat_config.aggr_mode = st->aggr_mode;
2125	}
2126
2127	if (perf_stat.data.is_pipe)
2128		perf_stat_init_aggr_mode();
2129	else
2130		perf_stat_init_aggr_mode_file(st);
2131
2132	if (stat_config.aggr_map) {
2133		int nr_aggr = stat_config.aggr_map->nr;
2134
2135		if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2136			pr_err("cannot allocate aggr counts\n");
2137			return -1;
2138		}
2139	}
2140	return 0;
2141}
2142
2143static int set_maps(struct perf_stat *st)
2144{
2145	if (!st->cpus || !st->threads)
2146		return 0;
2147
2148	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2149		return -EINVAL;
2150
2151	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2152
2153	if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2154		return -ENOMEM;
2155
2156	st->maps_allocated = true;
2157	return 0;
2158}
2159
2160static
2161int process_thread_map_event(struct perf_session *session,
2162			     union perf_event *event)
2163{
2164	struct perf_tool *tool = session->tool;
2165	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2166
2167	if (st->threads) {
2168		pr_warning("Extra thread map event, ignoring.\n");
2169		return 0;
2170	}
2171
2172	st->threads = thread_map__new_event(&event->thread_map);
2173	if (!st->threads)
2174		return -ENOMEM;
2175
2176	return set_maps(st);
2177}
2178
2179static
2180int process_cpu_map_event(struct perf_session *session,
2181			  union perf_event *event)
2182{
2183	struct perf_tool *tool = session->tool;
2184	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2185	struct perf_cpu_map *cpus;
2186
2187	if (st->cpus) {
2188		pr_warning("Extra cpu map event, ignoring.\n");
2189		return 0;
2190	}
2191
2192	cpus = cpu_map__new_data(&event->cpu_map.data);
2193	if (!cpus)
2194		return -ENOMEM;
2195
2196	st->cpus = cpus;
2197	return set_maps(st);
2198}
2199
2200static const char * const stat_report_usage[] = {
2201	"perf stat report [<options>]",
2202	NULL,
2203};
2204
2205static struct perf_stat perf_stat = {
2206	.tool = {
2207		.attr		= perf_event__process_attr,
2208		.event_update	= perf_event__process_event_update,
2209		.thread_map	= process_thread_map_event,
2210		.cpu_map	= process_cpu_map_event,
2211		.stat_config	= process_stat_config_event,
2212		.stat		= perf_event__process_stat_event,
2213		.stat_round	= process_stat_round_event,
2214	},
2215	.aggr_mode = AGGR_UNSET,
2216};
2217
2218static int __cmd_report(int argc, const char **argv)
2219{
2220	struct perf_session *session;
2221	const struct option options[] = {
2222	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2223	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2224		     "aggregate counts per processor socket", AGGR_SOCKET),
2225	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2226		     "aggregate counts per processor die", AGGR_DIE),
2227	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2228		     "aggregate counts per physical processor core", AGGR_CORE),
2229	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2230		     "aggregate counts per numa node", AGGR_NODE),
2231	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2232		     "disable CPU count aggregation", AGGR_NONE),
2233	OPT_END()
2234	};
2235	struct stat st;
2236	int ret;
2237
2238	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2239
2240	if (!input_name || !strlen(input_name)) {
2241		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2242			input_name = "-";
2243		else
2244			input_name = "perf.data";
2245	}
2246
2247	perf_stat__init_shadow_stats();
2248
2249	perf_stat.data.path = input_name;
2250	perf_stat.data.mode = PERF_DATA_MODE_READ;
2251
2252	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2253	if (IS_ERR(session))
2254		return PTR_ERR(session);
2255
2256	perf_stat.session  = session;
2257	stat_config.output = stderr;
2258	evsel_list         = session->evlist;
2259
2260	ret = perf_session__process_events(session);
2261	if (ret)
2262		return ret;
2263
2264	perf_session__delete(session);
2265	return 0;
2266}
2267
2268static void setup_system_wide(int forks)
2269{
2270	/*
2271	 * Make system wide (-a) the default target if
2272	 * no target was specified and one of following
2273	 * conditions is met:
2274	 *
2275	 *   - there's no workload specified
2276	 *   - there is workload specified but all requested
2277	 *     events are system wide events
2278	 */
2279	if (!target__none(&target))
2280		return;
2281
2282	if (!forks)
2283		target.system_wide = true;
2284	else {
2285		struct evsel *counter;
2286
2287		evlist__for_each_entry(evsel_list, counter) {
2288			if (!counter->core.requires_cpu &&
2289			    strcmp(counter->name, "duration_time")) {
2290				return;
2291			}
2292		}
2293
2294		if (evsel_list->core.nr_entries)
2295			target.system_wide = true;
2296	}
2297}
2298
2299int cmd_stat(int argc, const char **argv)
2300{
2301	const char * const stat_usage[] = {
2302		"perf stat [<options>] [<command>]",
2303		NULL
2304	};
2305	int status = -EINVAL, run_idx, err;
2306	const char *mode;
2307	FILE *output = stderr;
2308	unsigned int interval, timeout;
2309	const char * const stat_subcommands[] = { "record", "report" };
2310	char errbuf[BUFSIZ];
2311
2312	setlocale(LC_ALL, "");
2313
2314	evsel_list = evlist__new();
2315	if (evsel_list == NULL)
2316		return -ENOMEM;
2317
2318	parse_events__shrink_config_terms();
2319
2320	/* String-parsing callback-based options would segfault when negated */
2321	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2322	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2323	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2324
2325	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2326					(const char **) stat_usage,
2327					PARSE_OPT_STOP_AT_NON_OPTION);
2328
2329	if (stat_config.csv_sep) {
2330		stat_config.csv_output = true;
2331		if (!strcmp(stat_config.csv_sep, "\\t"))
2332			stat_config.csv_sep = "\t";
2333	} else
2334		stat_config.csv_sep = DEFAULT_SEPARATOR;
2335
2336	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2337		argc = __cmd_record(argc, argv);
2338		if (argc < 0)
2339			return -1;
2340	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2341		return __cmd_report(argc, argv);
2342
2343	interval = stat_config.interval;
2344	timeout = stat_config.timeout;
2345
2346	/*
2347	 * For record command the -o is already taken care of.
2348	 */
2349	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2350		output = NULL;
2351
2352	if (output_name && output_fd) {
2353		fprintf(stderr, "cannot use both --output and --log-fd\n");
2354		parse_options_usage(stat_usage, stat_options, "o", 1);
2355		parse_options_usage(NULL, stat_options, "log-fd", 0);
2356		goto out;
2357	}
2358
2359	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2360		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2361		goto out;
2362	}
2363
2364	if (stat_config.metric_only && stat_config.run_count > 1) {
2365		fprintf(stderr, "--metric-only is not supported with -r\n");
2366		goto out;
2367	}
2368
2369	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2370		fprintf(stderr, "--table is only supported with -r\n");
2371		parse_options_usage(stat_usage, stat_options, "r", 1);
2372		parse_options_usage(NULL, stat_options, "table", 0);
2373		goto out;
2374	}
2375
2376	if (output_fd < 0) {
2377		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2378		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2379		goto out;
2380	}
2381
2382	if (!output && !quiet) {
2383		struct timespec tm;
2384		mode = append_file ? "a" : "w";
2385
2386		output = fopen(output_name, mode);
2387		if (!output) {
2388			perror("failed to create output file");
2389			return -1;
2390		}
2391		clock_gettime(CLOCK_REALTIME, &tm);
2392		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2393	} else if (output_fd > 0) {
2394		mode = append_file ? "a" : "w";
2395		output = fdopen(output_fd, mode);
2396		if (!output) {
2397			perror("Failed opening logfd");
2398			return -errno;
2399		}
2400	}
2401
2402	if (stat_config.interval_clear && !isatty(fileno(output))) {
2403		fprintf(stderr, "--interval-clear does not work with output\n");
2404		parse_options_usage(stat_usage, stat_options, "o", 1);
2405		parse_options_usage(NULL, stat_options, "log-fd", 0);
2406		parse_options_usage(NULL, stat_options, "interval-clear", 0);
2407		return -1;
2408	}
2409
2410	stat_config.output = output;
2411
2412	/*
2413	 * let the spreadsheet do the pretty-printing
2414	 */
2415	if (stat_config.csv_output) {
2416		/* User explicitly passed -B? */
2417		if (big_num_opt == 1) {
2418			fprintf(stderr, "-B option not supported with -x\n");
2419			parse_options_usage(stat_usage, stat_options, "B", 1);
2420			parse_options_usage(NULL, stat_options, "x", 1);
2421			goto out;
2422		} else /* Nope, so disable big number formatting */
2423			stat_config.big_num = false;
2424	} else if (big_num_opt == 0) /* User passed --no-big-num */
2425		stat_config.big_num = false;
2426
2427	err = target__validate(&target);
2428	if (err) {
2429		target__strerror(&target, err, errbuf, BUFSIZ);
2430		pr_warning("%s\n", errbuf);
2431	}
2432
2433	setup_system_wide(argc);
2434
2435	/*
2436	 * Display user/system times only for single
2437	 * run and when there's specified tracee.
2438	 */
2439	if ((stat_config.run_count == 1) && target__none(&target))
2440		stat_config.ru_display = true;
2441
2442	if (stat_config.run_count < 0) {
2443		pr_err("Run count must be a positive number\n");
2444		parse_options_usage(stat_usage, stat_options, "r", 1);
2445		goto out;
2446	} else if (stat_config.run_count == 0) {
2447		forever = true;
2448		stat_config.run_count = 1;
2449	}
2450
2451	if (stat_config.walltime_run_table) {
2452		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2453		if (!stat_config.walltime_run) {
2454			pr_err("failed to setup -r option");
2455			goto out;
2456		}
2457	}
2458
2459	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2460		!target__has_task(&target)) {
2461		if (!target.system_wide || target.cpu_list) {
2462			fprintf(stderr, "The --per-thread option is only "
2463				"available when monitoring via -p -t -a "
2464				"options or only --per-thread.\n");
2465			parse_options_usage(NULL, stat_options, "p", 1);
2466			parse_options_usage(NULL, stat_options, "t", 1);
2467			goto out;
2468		}
2469	}
2470
2471	/*
2472	 * no_aggr, cgroup are for system-wide only
2473	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2474	 */
2475	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2476	      stat_config.aggr_mode != AGGR_THREAD) ||
2477	     (nr_cgroups || stat_config.cgroup_list)) &&
2478	    !target__has_cpu(&target)) {
2479		fprintf(stderr, "both cgroup and no-aggregation "
2480			"modes only available in system-wide mode\n");
2481
2482		parse_options_usage(stat_usage, stat_options, "G", 1);
2483		parse_options_usage(NULL, stat_options, "A", 1);
2484		parse_options_usage(NULL, stat_options, "a", 1);
2485		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2486		goto out;
2487	}
2488
2489	if (stat_config.iostat_run) {
2490		status = iostat_prepare(evsel_list, &stat_config);
2491		if (status)
2492			goto out;
2493		if (iostat_mode == IOSTAT_LIST) {
2494			iostat_list(evsel_list, &stat_config);
2495			goto out;
2496		} else if (verbose > 0)
2497			iostat_list(evsel_list, &stat_config);
2498		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2499			target.system_wide = true;
2500	}
2501
2502	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2503		target.per_thread = true;
2504
2505	stat_config.system_wide = target.system_wide;
2506	if (target.cpu_list) {
2507		stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2508		if (!stat_config.user_requested_cpu_list) {
2509			status = -ENOMEM;
2510			goto out;
2511		}
2512	}
2513
2514	/*
2515	 * Metric parsing needs to be delayed as metrics may optimize events
2516	 * knowing the target is system-wide.
2517	 */
2518	if (metrics) {
2519		metricgroup__parse_groups(evsel_list, metrics,
2520					stat_config.metric_no_group,
2521					stat_config.metric_no_merge,
2522					stat_config.user_requested_cpu_list,
2523					stat_config.system_wide,
2524					&stat_config.metric_events);
2525		zfree(&metrics);
2526	}
2527	perf_stat__collect_metric_expr(evsel_list);
2528	perf_stat__init_shadow_stats();
2529
2530	if (add_default_attributes())
2531		goto out;
2532
2533	if (stat_config.cgroup_list) {
2534		if (nr_cgroups > 0) {
2535			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2536			parse_options_usage(stat_usage, stat_options, "G", 1);
2537			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2538			goto out;
2539		}
2540
2541		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2542					  &stat_config.metric_events, true) < 0) {
2543			parse_options_usage(stat_usage, stat_options,
2544					    "for-each-cgroup", 0);
2545			goto out;
2546		}
2547	}
2548
2549	if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2550		pr_err("failed to use cpu list %s\n", target.cpu_list);
2551		goto out;
2552	}
2553
2554	target.hybrid = perf_pmu__has_hybrid();
2555	if (evlist__create_maps(evsel_list, &target) < 0) {
2556		if (target__has_task(&target)) {
2557			pr_err("Problems finding threads of monitor\n");
2558			parse_options_usage(stat_usage, stat_options, "p", 1);
2559			parse_options_usage(NULL, stat_options, "t", 1);
2560		} else if (target__has_cpu(&target)) {
2561			perror("failed to parse CPUs map");
2562			parse_options_usage(stat_usage, stat_options, "C", 1);
2563			parse_options_usage(NULL, stat_options, "a", 1);
2564		}
2565		goto out;
2566	}
2567
2568	evlist__check_cpu_maps(evsel_list);
2569
2570	/*
2571	 * Initialize thread_map with comm names,
2572	 * so we could print it out on output.
2573	 */
2574	if (stat_config.aggr_mode == AGGR_THREAD) {
2575		thread_map__read_comms(evsel_list->core.threads);
2576	}
2577
2578	if (stat_config.aggr_mode == AGGR_NODE)
2579		cpu__setup_cpunode_map();
2580
2581	if (stat_config.times && interval)
2582		interval_count = true;
2583	else if (stat_config.times && !interval) {
2584		pr_err("interval-count option should be used together with "
2585				"interval-print.\n");
2586		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2587		parse_options_usage(stat_usage, stat_options, "I", 1);
2588		goto out;
2589	}
2590
2591	if (timeout && timeout < 100) {
2592		if (timeout < 10) {
2593			pr_err("timeout must be >= 10ms.\n");
2594			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2595			goto out;
2596		} else
2597			pr_warning("timeout < 100ms. "
2598				   "The overhead percentage could be high in some cases. "
2599				   "Please proceed with caution.\n");
2600	}
2601	if (timeout && interval) {
2602		pr_err("timeout option is not supported with interval-print.\n");
2603		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2604		parse_options_usage(stat_usage, stat_options, "I", 1);
2605		goto out;
2606	}
2607
2608	if (perf_stat_init_aggr_mode())
2609		goto out;
2610
2611	if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2612		goto out;
2613
2614	/*
2615	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2616	 * while avoiding that older tools show confusing messages.
2617	 *
2618	 * However for pipe sessions we need to keep it zero,
2619	 * because script's perf_evsel__check_attr is triggered
2620	 * by attr->sample_type != 0, and we can't run it on
2621	 * stat sessions.
2622	 */
2623	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2624
2625	/*
2626	 * We dont want to block the signals - that would cause
2627	 * child tasks to inherit that and Ctrl-C would not work.
2628	 * What we want is for Ctrl-C to work in the exec()-ed
2629	 * task, but being ignored by perf stat itself:
2630	 */
2631	atexit(sig_atexit);
2632	if (!forever)
2633		signal(SIGINT,  skip_signal);
2634	signal(SIGCHLD, skip_signal);
2635	signal(SIGALRM, skip_signal);
2636	signal(SIGABRT, skip_signal);
2637
2638	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2639		goto out;
2640
2641	/* Enable ignoring missing threads when -p option is defined. */
2642	evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2643	status = 0;
2644	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2645		if (stat_config.run_count != 1 && verbose > 0)
2646			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2647				run_idx + 1);
2648
2649		if (run_idx != 0)
2650			evlist__reset_prev_raw_counts(evsel_list);
2651
2652		status = run_perf_stat(argc, argv, run_idx);
2653		if (forever && status != -1 && !interval) {
2654			print_counters(NULL, argc, argv);
2655			perf_stat__reset_stats();
2656		}
2657	}
2658
2659	if (!forever && status != -1 && (!interval || stat_config.summary))
2660		print_counters(NULL, argc, argv);
2661
2662	evlist__finalize_ctlfd(evsel_list);
2663
2664	if (STAT_RECORD) {
2665		/*
2666		 * We synthesize the kernel mmap record just so that older tools
2667		 * don't emit warnings about not being able to resolve symbols
2668		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2669		 * a saner message about no samples being in the perf.data file.
2670		 *
2671		 * This also serves to suppress a warning about f_header.data.size == 0
2672		 * in header.c at the moment 'perf stat record' gets introduced, which
2673		 * is not really needed once we start adding the stat specific PERF_RECORD_
2674		 * records, but the need to suppress the kptr_restrict messages in older
2675		 * tools remain  -acme
2676		 */
2677		int fd = perf_data__fd(&perf_stat.data);
2678
2679		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2680							 process_synthesized_event,
2681							 &perf_stat.session->machines.host);
2682		if (err) {
2683			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2684				   "older tools may produce warnings about this file\n.");
2685		}
2686
2687		if (!interval) {
2688			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2689				pr_err("failed to write stat round event\n");
2690		}
2691
2692		if (!perf_stat.data.is_pipe) {
2693			perf_stat.session->header.data_size += perf_stat.bytes_written;
2694			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2695		}
2696
2697		evlist__close(evsel_list);
2698		perf_session__delete(perf_stat.session);
2699	}
2700
2701	perf_stat__exit_aggr_mode();
2702	evlist__free_stats(evsel_list);
2703out:
2704	if (stat_config.iostat_run)
2705		iostat_release(evsel_list);
2706
2707	zfree(&stat_config.walltime_run);
2708	zfree(&stat_config.user_requested_cpu_list);
2709
2710	if (smi_cost && smi_reset)
2711		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2712
2713	evlist__delete(evsel_list);
2714
2715	metricgroup__rblist_exit(&stat_config.metric_events);
2716	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2717
2718	return status;
2719}
v3.15
 
   1/*
   2 * builtin-stat.c
   3 *
   4 * Builtin stat command: Give a precise performance counters summary
   5 * overview about any workload, CPU or specific PID.
   6 *
   7 * Sample output:
   8
   9   $ perf stat ./hackbench 10
  10
  11  Time: 0.118
  12
  13  Performance counter stats for './hackbench 10':
  14
  15       1708.761321 task-clock                #   11.037 CPUs utilized
  16            41,190 context-switches          #    0.024 M/sec
  17             6,735 CPU-migrations            #    0.004 M/sec
  18            17,318 page-faults               #    0.010 M/sec
  19     5,205,202,243 cycles                    #    3.046 GHz
  20     3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  21     1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  22     2,603,501,247 instructions              #    0.50  insns per cycle
  23                                             #    1.48  stalled cycles per insn
  24       484,357,498 branches                  #  283.455 M/sec
  25         6,388,934 branch-misses             #    1.32% of all branches
  26
  27        0.154822978  seconds time elapsed
  28
  29 *
  30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  31 *
  32 * Improvements and fixes by:
  33 *
  34 *   Arjan van de Ven <arjan@linux.intel.com>
  35 *   Yanmin Zhang <yanmin.zhang@intel.com>
  36 *   Wu Fengguang <fengguang.wu@intel.com>
  37 *   Mike Galbraith <efault@gmx.de>
  38 *   Paul Mackerras <paulus@samba.org>
  39 *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  40 *
  41 * Released under the GPL v2. (and only v2, not any later version)
  42 */
  43
 
  44#include "perf.h"
  45#include "builtin.h"
  46#include "util/util.h"
  47#include "util/parse-options.h"
  48#include "util/parse-events.h"
  49#include "util/pmu.h"
  50#include "util/event.h"
  51#include "util/evlist.h"
 
  52#include "util/evsel.h"
  53#include "util/debug.h"
  54#include "util/color.h"
  55#include "util/stat.h"
  56#include "util/header.h"
  57#include "util/cpumap.h"
  58#include "util/thread.h"
  59#include "util/thread_map.h"
  60
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  61#include <stdlib.h>
  62#include <sys/prctl.h>
 
  63#include <locale.h>
 
 
 
 
 
 
 
 
 
 
 
 
  64
  65#define DEFAULT_SEPARATOR	" "
  66#define CNTR_NOT_SUPPORTED	"<not supported>"
  67#define CNTR_NOT_COUNTED	"<not counted>"
  68
  69static void print_stat(int argc, const char **argv);
  70static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
  71static void print_counter(struct perf_evsel *counter, char *prefix);
  72static void print_aggr(char *prefix);
  73
  74/* Default events used for perf stat -T */
  75static const char * const transaction_attrs[] = {
  76	"task-clock",
  77	"{"
  78	"instructions,"
  79	"cycles,"
  80	"cpu/cycles-t/,"
  81	"cpu/tx-start/,"
  82	"cpu/el-start/,"
  83	"cpu/cycles-ct/"
  84	"}"
  85};
  86
  87/* More limited version when the CPU does not have all events. */
  88static const char * const transaction_limited_attrs[] = {
  89	"task-clock",
  90	"{"
  91	"instructions,"
  92	"cycles,"
  93	"cpu/cycles-t/,"
  94	"cpu/tx-start/"
  95	"}"
  96};
  97
  98/* must match transaction_attrs and the beginning limited_attrs */
  99enum {
 100	T_TASK_CLOCK,
 101	T_INSTRUCTIONS,
 102	T_CYCLES,
 103	T_CYCLES_IN_TX,
 104	T_TRANSACTION_START,
 105	T_ELISION_START,
 106	T_CYCLES_IN_TX_CP,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 107};
 108
 109static struct perf_evlist	*evsel_list;
 
 
 
 
 
 
 
 
 
 
 
 110
 111static struct target target = {
 112	.uid	= UINT_MAX,
 113};
 114
 115enum aggr_mode {
 116	AGGR_NONE,
 117	AGGR_GLOBAL,
 118	AGGR_SOCKET,
 119	AGGR_CORE,
 120};
 121
 122static int			run_count			=  1;
 123static bool			no_inherit			= false;
 124static bool			scale				=  true;
 125static enum aggr_mode		aggr_mode			= AGGR_GLOBAL;
 126static volatile pid_t		child_pid			= -1;
 127static bool			null_run			=  false;
 128static int			detailed_run			=  0;
 129static bool			transaction_run;
 130static bool			big_num				=  true;
 
 
 131static int			big_num_opt			=  -1;
 132static const char		*csv_sep			= NULL;
 133static bool			csv_output			= false;
 134static bool			group				= false;
 135static FILE			*output				= NULL;
 136static const char		*pre_cmd			= NULL;
 137static const char		*post_cmd			= NULL;
 138static bool			sync_run			= false;
 139static unsigned int		interval			= 0;
 140static unsigned int		initial_delay			= 0;
 141static unsigned int		unit_width			= 4; /* strlen("unit") */
 142static bool			forever				= false;
 
 143static struct timespec		ref_time;
 144static struct cpu_map		*aggr_map;
 145static int			(*aggr_get_id)(struct cpu_map *m, int cpu);
 
 
 
 146
 147static volatile int done = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 148
 149struct perf_stat {
 150	struct stats	  res_stats[3];
 
 
 
 
 
 
 
 
 
 
 151};
 152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 153static inline void diff_timespec(struct timespec *r, struct timespec *a,
 154				 struct timespec *b)
 155{
 156	r->tv_sec = a->tv_sec - b->tv_sec;
 157	if (a->tv_nsec < b->tv_nsec) {
 158		r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 159		r->tv_sec--;
 160	} else {
 161		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 162	}
 163}
 164
 165static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
 166{
 167	return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
 
 168}
 169
 170static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
 
 
 
 171{
 172	return perf_evsel__cpus(evsel)->nr;
 
 
 
 
 
 
 173}
 174
 175static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 176{
 177	int i;
 178	struct perf_stat *ps = evsel->priv;
 
 
 
 
 
 179
 180	for (i = 0; i < 3; i++)
 181		init_stats(&ps->res_stats[i]);
 182}
 183
 184static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 
 185{
 186	evsel->priv = zalloc(sizeof(struct perf_stat));
 187	if (evsel == NULL)
 188		return -ENOMEM;
 189	perf_evsel__reset_stat_priv(evsel);
 190	return 0;
 191}
 192
 193static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 
 194{
 195	zfree(&evsel->priv);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 196}
 197
 198static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
 
 
 
 
 199{
 200	void *addr;
 201	size_t sz;
 202
 203	sz = sizeof(*evsel->counts) +
 204	     (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
 205
 206	addr = zalloc(sz);
 207	if (!addr)
 208		return -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 209
 210	evsel->prev_raw_counts =  addr;
 
 
 
 
 
 
 
 
 211
 212	return 0;
 213}
 214
 215static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
 216{
 217	zfree(&evsel->prev_raw_counts);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 218}
 219
 220static void perf_evlist__free_stats(struct perf_evlist *evlist)
 221{
 222	struct perf_evsel *evsel;
 
 
 
 
 
 223
 224	evlist__for_each(evlist, evsel) {
 225		perf_evsel__free_stat_priv(evsel);
 226		perf_evsel__free_counts(evsel);
 227		perf_evsel__free_prev_raw_counts(evsel);
 228	}
 
 229}
 230
 231static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
 232{
 233	struct perf_evsel *evsel;
 234
 235	evlist__for_each(evlist, evsel) {
 236		if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 237		    perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
 238		    (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
 239			goto out_free;
 240	}
 
 
 241
 242	return 0;
 
 
 243
 244out_free:
 245	perf_evlist__free_stats(evlist);
 246	return -1;
 247}
 248
 249static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
 250static struct stats runtime_cycles_stats[MAX_NR_CPUS];
 251static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
 252static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
 253static struct stats runtime_branches_stats[MAX_NR_CPUS];
 254static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
 255static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
 256static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
 257static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
 258static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
 259static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
 260static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
 261static struct stats walltime_nsecs_stats;
 262static struct stats runtime_transaction_stats[MAX_NR_CPUS];
 263static struct stats runtime_elision_stats[MAX_NR_CPUS];
 264
 265static void perf_stat__reset_stats(struct perf_evlist *evlist)
 266{
 267	struct perf_evsel *evsel;
 268
 269	evlist__for_each(evlist, evsel) {
 270		perf_evsel__reset_stat_priv(evsel);
 271		perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
 272	}
 273
 274	memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
 275	memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
 276	memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
 277	memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
 278	memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
 279	memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
 280	memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
 281	memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
 282	memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
 283	memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
 284	memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
 285	memset(runtime_cycles_in_tx_stats, 0,
 286			sizeof(runtime_cycles_in_tx_stats));
 287	memset(runtime_transaction_stats, 0,
 288		sizeof(runtime_transaction_stats));
 289	memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
 290	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
 291}
 292
 293static int create_perf_stat_counter(struct perf_evsel *evsel)
 294{
 295	struct perf_event_attr *attr = &evsel->attr;
 296
 297	if (scale)
 298		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 299				    PERF_FORMAT_TOTAL_TIME_RUNNING;
 300
 301	attr->inherit = !no_inherit;
 302
 303	if (target__has_cpu(&target))
 304		return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 305
 306	if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
 307		attr->disabled = 1;
 308		if (!initial_delay)
 309			attr->enable_on_exec = 1;
 310	}
 311
 312	return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 
 
 313}
 314
 315/*
 316 * Does the counter have nsecs as a unit?
 317 */
 318static inline int nsec_counter(struct perf_evsel *evsel)
 319{
 320	if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 321	    perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 322		return 1;
 
 323
 324	return 0;
 325}
 
 
 
 
 
 
 
 
 326
 327static struct perf_evsel *nth_evsel(int n)
 328{
 329	static struct perf_evsel **array;
 330	static int array_len;
 331	struct perf_evsel *ev;
 332	int j;
 333
 334	/* Assumes this only called when evsel_list does not change anymore. */
 335	if (!array) {
 336		evlist__for_each(evsel_list, ev)
 337			array_len++;
 338		array = malloc(array_len * sizeof(void *));
 339		if (!array)
 340			exit(ENOMEM);
 341		j = 0;
 342		evlist__for_each(evsel_list, ev)
 343			array[j++] = ev;
 344	}
 345	if (n < array_len)
 346		return array[n];
 347	return NULL;
 348}
 349
 350/*
 351 * Update various tracking values we maintain to print
 352 * more semantic information such as miss/hit ratios,
 353 * instruction rates, etc:
 354 */
 355static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
 356{
 357	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 358		update_stats(&runtime_nsecs_stats[0], count[0]);
 359	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 360		update_stats(&runtime_cycles_stats[0], count[0]);
 361	else if (transaction_run &&
 362		 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
 363		update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
 364	else if (transaction_run &&
 365		 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
 366		update_stats(&runtime_transaction_stats[0], count[0]);
 367	else if (transaction_run &&
 368		 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
 369		update_stats(&runtime_elision_stats[0], count[0]);
 370	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
 371		update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
 372	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 373		update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
 374	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 375		update_stats(&runtime_branches_stats[0], count[0]);
 376	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 377		update_stats(&runtime_cacherefs_stats[0], count[0]);
 378	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
 379		update_stats(&runtime_l1_dcache_stats[0], count[0]);
 380	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
 381		update_stats(&runtime_l1_icache_stats[0], count[0]);
 382	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
 383		update_stats(&runtime_ll_cache_stats[0], count[0]);
 384	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
 385		update_stats(&runtime_dtlb_cache_stats[0], count[0]);
 386	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
 387		update_stats(&runtime_itlb_cache_stats[0], count[0]);
 388}
 389
 390/*
 391 * Read out the results of a single counter:
 392 * aggregate counts across CPUs in system-wide mode
 393 */
 394static int read_counter_aggr(struct perf_evsel *counter)
 395{
 396	struct perf_stat *ps = counter->priv;
 397	u64 *count = counter->counts->aggr.values;
 398	int i;
 399
 400	if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
 401			       thread_map__nr(evsel_list->threads), scale) < 0)
 402		return -1;
 403
 404	for (i = 0; i < 3; i++)
 405		update_stats(&ps->res_stats[i], count[i]);
 
 
 406
 407	if (verbose) {
 408		fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 409			perf_evsel__name(counter), count[0], count[1], count[2]);
 
 
 
 
 
 410	}
 
 
 
 
 
 
 411
 412	/*
 413	 * Save the full runtime - to allow normalization during printout:
 
 
 414	 */
 415	update_shadow_stats(counter, count);
 
 
 
 
 
 
 416
 417	return 0;
 418}
 419
 420/*
 421 * Read out the results of a single counter:
 422 * do not aggregate counts across CPUs in system-wide mode
 
 423 */
 424static int read_counter(struct perf_evsel *counter)
 
 
 
 
 
 
 425{
 426	u64 *count;
 427	int cpu;
 
 
 
 
 
 
 
 
 
 428
 429	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 430		if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 431			return -1;
 432
 433		count = counter->counts->cpu[cpu].values;
 
 434
 435		update_shadow_stats(counter, count);
 
 436	}
 437
 438	return 0;
 439}
 440
 441static void print_interval(void)
 442{
 443	static int num_print_interval;
 444	struct perf_evsel *counter;
 445	struct perf_stat *ps;
 446	struct timespec ts, rs;
 447	char prefix[64];
 448
 449	if (aggr_mode == AGGR_GLOBAL) {
 450		evlist__for_each(evsel_list, counter) {
 451			ps = counter->priv;
 452			memset(ps->res_stats, 0, sizeof(ps->res_stats));
 453			read_counter_aggr(counter);
 454		}
 455	} else	{
 456		evlist__for_each(evsel_list, counter) {
 457			ps = counter->priv;
 458			memset(ps->res_stats, 0, sizeof(ps->res_stats));
 459			read_counter(counter);
 
 
 
 
 
 460		}
 461	}
 
 462
 463	clock_gettime(CLOCK_MONOTONIC, &ts);
 464	diff_timespec(&rs, &ts, &ref_time);
 465	sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
 
 
 466
 467	if (num_print_interval == 0 && !csv_output) {
 468		switch (aggr_mode) {
 469		case AGGR_SOCKET:
 470			fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
 471			break;
 472		case AGGR_CORE:
 473			fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
 474			break;
 475		case AGGR_NONE:
 476			fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
 477			break;
 478		case AGGR_GLOBAL:
 479		default:
 480			fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
 481		}
 482	}
 483
 484	if (++num_print_interval == 25)
 485		num_print_interval = 0;
 486
 487	switch (aggr_mode) {
 488	case AGGR_CORE:
 489	case AGGR_SOCKET:
 490		print_aggr(prefix);
 491		break;
 492	case AGGR_NONE:
 493		evlist__for_each(evsel_list, counter)
 494			print_counter(counter, prefix);
 495		break;
 496	case AGGR_GLOBAL:
 497	default:
 498		evlist__for_each(evsel_list, counter)
 499			print_counter_aggr(counter, prefix);
 500	}
 501
 502	fflush(output);
 503}
 504
 505static void handle_initial_delay(void)
 506{
 507	struct perf_evsel *counter;
 
 
 
 
 
 
 
 
 
 508
 509	if (initial_delay) {
 510		const int ncpus = cpu_map__nr(evsel_list->cpus),
 511			nthreads = thread_map__nr(evsel_list->threads);
 512
 513		usleep(initial_delay * 1000);
 514		evlist__for_each(evsel_list, counter)
 515			perf_evsel__enable(counter, ncpus, nthreads);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 516	}
 
 
 517}
 518
 519static volatile int workload_exec_errno;
 
 
 
 
 520
 521/*
 522 * perf_evlist__prepare_workload will send a SIGUSR1
 523 * if the fork fails, since we asked by setting its
 524 * want_signal to true.
 525 */
 526static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 527					void *ucontext __maybe_unused)
 528{
 529	workload_exec_errno = info->si_value.sival_int;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 530}
 531
 532static int __run_perf_stat(int argc, const char **argv)
 533{
 534	char msg[512];
 
 
 
 535	unsigned long long t0, t1;
 536	struct perf_evsel *counter;
 537	struct timespec ts;
 538	size_t l;
 539	int status = 0;
 540	const bool forks = (argc > 0);
 541
 542	if (interval) {
 543		ts.tv_sec  = interval / 1000;
 544		ts.tv_nsec = (interval % 1000) * 1000000;
 545	} else {
 546		ts.tv_sec  = 1;
 547		ts.tv_nsec = 0;
 548	}
 549
 550	if (forks) {
 551		if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
 552						  workload_exec_failed_signal) < 0) {
 553			perror("failed to prepare workload");
 554			return -1;
 555		}
 556		child_pid = evsel_list->workload.pid;
 557	}
 558
 559	if (group)
 560		perf_evlist__set_leader(evsel_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 561
 562	evlist__for_each(evsel_list, counter) {
 563		if (create_perf_stat_counter(counter) < 0) {
 564			/*
 565			 * PPC returns ENXIO for HW counters until 2.6.37
 566			 * (behavior changed with commit b0a873e).
 
 
 
 567			 */
 568			if (errno == EINVAL || errno == ENOSYS ||
 569			    errno == ENOENT || errno == EOPNOTSUPP ||
 570			    errno == ENXIO) {
 571				if (verbose)
 572					ui__warning("%s event is not supported by the kernel.\n",
 573						    perf_evsel__name(counter));
 574				counter->supported = false;
 575				continue;
 576			}
 577
 578			perf_evsel__open_strerror(counter, &target,
 579						  errno, msg, sizeof(msg));
 580			ui__error("%s\n", msg);
 
 
 
 
 
 
 
 581
 582			if (child_pid != -1)
 583				kill(child_pid, SIGTERM);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 584
 585			return -1;
 
 
 
 586		}
 587		counter->supported = true;
 588
 589		l = strlen(counter->unit);
 590		if (l > unit_width)
 591			unit_width = l;
 
 
 
 
 592	}
 593
 594	if (perf_evlist__apply_filters(evsel_list)) {
 595		error("failed to set filter with %d (%s)\n", errno,
 596			strerror(errno));
 
 597		return -1;
 598	}
 599
 600	/*
 601	 * Enable counters and exec the command:
 602	 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 603	t0 = rdclock();
 604	clock_gettime(CLOCK_MONOTONIC, &ref_time);
 605
 606	if (forks) {
 607		perf_evlist__start_workload(evsel_list);
 608		handle_initial_delay();
 609
 610		if (interval) {
 611			while (!waitpid(child_pid, &status, WNOHANG)) {
 612				nanosleep(&ts, NULL);
 613				print_interval();
 614			}
 615		}
 616		wait(&status);
 617
 618		if (workload_exec_errno) {
 619			const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 620			pr_err("Workload failed: %s\n", emsg);
 621			return -1;
 622		}
 623
 624		if (WIFSIGNALED(status))
 625			psignal(WTERMSIG(status), argv[0]);
 626	} else {
 627		handle_initial_delay();
 628		while (!done) {
 629			nanosleep(&ts, NULL);
 630			if (interval)
 631				print_interval();
 632		}
 633	}
 634
 
 
 635	t1 = rdclock();
 636
 637	update_stats(&walltime_nsecs_stats, t1 - t0);
 
 638
 639	if (aggr_mode == AGGR_GLOBAL) {
 640		evlist__for_each(evsel_list, counter) {
 641			read_counter_aggr(counter);
 642			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 643					     thread_map__nr(evsel_list->threads));
 644		}
 
 
 
 
 645	} else {
 646		evlist__for_each(evsel_list, counter) {
 647			read_counter(counter);
 648			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
 649		}
 650	}
 651
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 652	return WEXITSTATUS(status);
 653}
 654
 655static int run_perf_stat(int argc, const char **argv)
 656{
 657	int ret;
 658
 659	if (pre_cmd) {
 660		ret = system(pre_cmd);
 661		if (ret)
 662			return ret;
 663	}
 664
 665	if (sync_run)
 666		sync();
 667
 668	ret = __run_perf_stat(argc, argv);
 669	if (ret)
 670		return ret;
 671
 672	if (post_cmd) {
 673		ret = system(post_cmd);
 674		if (ret)
 675			return ret;
 676	}
 677
 678	return ret;
 679}
 680
 681static void print_noise_pct(double total, double avg)
 682{
 683	double pct = rel_stddev_stats(total, avg);
 
 
 
 
 684
 685	if (csv_output)
 686		fprintf(output, "%s%.2f%%", csv_sep, pct);
 687	else if (pct)
 688		fprintf(output, "  ( +-%6.2f%% )", pct);
 689}
 690
 691static void print_noise(struct perf_evsel *evsel, double avg)
 
 
 692{
 693	struct perf_stat *ps;
 
 694
 695	if (run_count == 1)
 696		return;
 697
 698	ps = evsel->priv;
 699	print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 
 
 
 700}
 701
 702static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 703{
 704	switch (aggr_mode) {
 705	case AGGR_CORE:
 706		fprintf(output, "S%d-C%*d%s%*d%s",
 707			cpu_map__id_to_socket(id),
 708			csv_output ? 0 : -8,
 709			cpu_map__id_to_cpu(id),
 710			csv_sep,
 711			csv_output ? 0 : 4,
 712			nr,
 713			csv_sep);
 714		break;
 715	case AGGR_SOCKET:
 716		fprintf(output, "S%*d%s%*d%s",
 717			csv_output ? 0 : -5,
 718			id,
 719			csv_sep,
 720			csv_output ? 0 : 4,
 721			nr,
 722			csv_sep);
 723			break;
 724	case AGGR_NONE:
 725		fprintf(output, "CPU%*d%s",
 726			csv_output ? 0 : -4,
 727			perf_evsel__cpus(evsel)->map[id], csv_sep);
 728		break;
 729	case AGGR_GLOBAL:
 730	default:
 731		break;
 732	}
 733}
 734
 735static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 736{
 737	double msecs = avg / 1e6;
 738	const char *fmt_v, *fmt_n;
 739	char name[25];
 
 
 
 
 740
 741	fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
 742	fmt_n = csv_output ? "%s" : "%-25s";
 743
 744	aggr_printout(evsel, cpu, nr);
 745
 746	scnprintf(name, sizeof(name), "%s%s",
 747		  perf_evsel__name(evsel), csv_output ? "" : " (msec)");
 748
 749	fprintf(output, fmt_v, msecs, csv_sep);
 750
 751	if (csv_output)
 752		fprintf(output, "%s%s", evsel->unit, csv_sep);
 753	else
 754		fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
 755
 756	fprintf(output, fmt_n, name);
 757
 758	if (evsel->cgrp)
 759		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 760
 761	if (csv_output || interval)
 762		return;
 763
 764	if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 765		fprintf(output, " # %8.3f CPUs utilized          ",
 766			avg / avg_stats(&walltime_nsecs_stats));
 767	else
 768		fprintf(output, "                                   ");
 769}
 770
 771/* used for get_ratio_color() */
 772enum grc_type {
 773	GRC_STALLED_CYCLES_FE,
 774	GRC_STALLED_CYCLES_BE,
 775	GRC_CACHE_MISSES,
 776	GRC_MAX_NR
 777};
 778
 779static const char *get_ratio_color(enum grc_type type, double ratio)
 780{
 781	static const double grc_table[GRC_MAX_NR][3] = {
 782		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
 783		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
 784		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
 785	};
 786	const char *color = PERF_COLOR_NORMAL;
 787
 788	if (ratio > grc_table[type][0])
 789		color = PERF_COLOR_RED;
 790	else if (ratio > grc_table[type][1])
 791		color = PERF_COLOR_MAGENTA;
 792	else if (ratio > grc_table[type][2])
 793		color = PERF_COLOR_YELLOW;
 794
 795	return color;
 796}
 797
 798static void print_stalled_cycles_frontend(int cpu,
 799					  struct perf_evsel *evsel
 800					  __maybe_unused, double avg)
 801{
 802	double total, ratio = 0.0;
 803	const char *color;
 
 
 804
 805	total = avg_stats(&runtime_cycles_stats[cpu]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 806
 807	if (total)
 808		ratio = avg / total * 100.0;
 
 
 
 809
 810	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
 811
 812	fprintf(output, " #  ");
 813	color_fprintf(output, color, "%6.2f%%", ratio);
 814	fprintf(output, " frontend cycles idle   ");
 815}
 816
 817static void print_stalled_cycles_backend(int cpu,
 818					 struct perf_evsel *evsel
 819					 __maybe_unused, double avg)
 820{
 821	double total, ratio = 0.0;
 822	const char *color;
 823
 824	total = avg_stats(&runtime_cycles_stats[cpu]);
 825
 826	if (total)
 827		ratio = avg / total * 100.0;
 828
 829	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
 830
 831	fprintf(output, " #  ");
 832	color_fprintf(output, color, "%6.2f%%", ratio);
 833	fprintf(output, " backend  cycles idle   ");
 834}
 835
 836static void print_branch_misses(int cpu,
 837				struct perf_evsel *evsel __maybe_unused,
 838				double avg)
 839{
 840	double total, ratio = 0.0;
 841	const char *color;
 842
 843	total = avg_stats(&runtime_branches_stats[cpu]);
 
 
 
 844
 845	if (total)
 846		ratio = avg / total * 100.0;
 
 
 
 847
 848	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 849
 850	fprintf(output, " #  ");
 851	color_fprintf(output, color, "%6.2f%%", ratio);
 852	fprintf(output, " of all branches        ");
 853}
 854
 855static void print_l1_dcache_misses(int cpu,
 856				   struct perf_evsel *evsel __maybe_unused,
 857				   double avg)
 858{
 859	double total, ratio = 0.0;
 860	const char *color;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 861
 862	total = avg_stats(&runtime_l1_dcache_stats[cpu]);
 
 
 
 
 
 
 
 
 
 863
 864	if (total)
 865		ratio = avg / total * 100.0;
 
 
 
 866
 867	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 
 
 
 
 868
 869	fprintf(output, " #  ");
 870	color_fprintf(output, color, "%6.2f%%", ratio);
 871	fprintf(output, " of all L1-dcache hits  ");
 
 872}
 873
 874static void print_l1_icache_misses(int cpu,
 875				   struct perf_evsel *evsel __maybe_unused,
 876				   double avg)
 877{
 878	double total, ratio = 0.0;
 879	const char *color;
 880
 881	total = avg_stats(&runtime_l1_icache_stats[cpu]);
 
 
 
 
 882
 883	if (total)
 884		ratio = avg / total * 100.0;
 885
 886	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 887
 888	fprintf(output, " #  ");
 889	color_fprintf(output, color, "%6.2f%%", ratio);
 890	fprintf(output, " of all L1-icache hits  ");
 891}
 892
 893static void print_dtlb_cache_misses(int cpu,
 894				    struct perf_evsel *evsel __maybe_unused,
 895				    double avg)
 896{
 897	double total, ratio = 0.0;
 898	const char *color;
 899
 900	total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
 
 
 901
 902	if (total)
 903		ratio = avg / total * 100.0;
 904
 905	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 
 
 906
 907	fprintf(output, " #  ");
 908	color_fprintf(output, color, "%6.2f%%", ratio);
 909	fprintf(output, " of all dTLB cache hits ");
 
 910}
 911
 912static void print_itlb_cache_misses(int cpu,
 913				    struct perf_evsel *evsel __maybe_unused,
 914				    double avg)
 915{
 916	double total, ratio = 0.0;
 917	const char *color;
 918
 919	total = avg_stats(&runtime_itlb_cache_stats[cpu]);
 
 
 
 
 920
 921	if (total)
 922		ratio = avg / total * 100.0;
 
 
 
 923
 924	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 925
 926	fprintf(output, " #  ");
 927	color_fprintf(output, color, "%6.2f%%", ratio);
 928	fprintf(output, " of all iTLB cache hits ");
 929}
 930
 931static void print_ll_cache_misses(int cpu,
 932				  struct perf_evsel *evsel __maybe_unused,
 933				  double avg)
 934{
 935	double total, ratio = 0.0;
 936	const char *color;
 937
 938	total = avg_stats(&runtime_ll_cache_stats[cpu]);
 939
 940	if (total)
 941		ratio = avg / total * 100.0;
 942
 943	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 944
 945	fprintf(output, " #  ");
 946	color_fprintf(output, color, "%6.2f%%", ratio);
 947	fprintf(output, " of all LL-cache hits   ");
 948}
 949
 950static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 951{
 952	double total, ratio = 0.0, total2;
 953	double sc =  evsel->scale;
 954	const char *fmt;
 955
 956	if (csv_output) {
 957		fmt = sc != 1.0 ?  "%.2f%s" : "%.0f%s";
 958	} else {
 959		if (big_num)
 960			fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
 961		else
 962			fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
 
 
 
 
 
 
 
 963	}
 
 964
 965	aggr_printout(evsel, cpu, nr);
 966
 967	if (aggr_mode == AGGR_GLOBAL)
 968		cpu = 0;
 969
 970	fprintf(output, fmt, avg, csv_sep);
 971
 972	if (evsel->unit)
 973		fprintf(output, "%-*s%s",
 974			csv_output ? 0 : unit_width,
 975			evsel->unit, csv_sep);
 976
 977	fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
 978
 979	if (evsel->cgrp)
 980		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 981
 982	if (csv_output || interval)
 983		return;
 984
 985	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 986		total = avg_stats(&runtime_cycles_stats[cpu]);
 987		if (total) {
 988			ratio = avg / total;
 989			fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
 990		}
 991		total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
 992		total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
 993
 994		if (total && avg) {
 995			ratio = total / avg;
 996			fprintf(output, "\n");
 997			if (aggr_mode == AGGR_NONE)
 998				fprintf(output, "        ");
 999			fprintf(output, "                                                  #   %5.2f  stalled cycles per insn", ratio);
1000		}
1001
1002	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1003			runtime_branches_stats[cpu].n != 0) {
1004		print_branch_misses(cpu, evsel, avg);
1005	} else if (
1006		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1007		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
1008					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1009					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1010			runtime_l1_dcache_stats[cpu].n != 0) {
1011		print_l1_dcache_misses(cpu, evsel, avg);
1012	} else if (
1013		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1014		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
1015					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1016					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1017			runtime_l1_icache_stats[cpu].n != 0) {
1018		print_l1_icache_misses(cpu, evsel, avg);
1019	} else if (
1020		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1021		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
1022					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1023					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1024			runtime_dtlb_cache_stats[cpu].n != 0) {
1025		print_dtlb_cache_misses(cpu, evsel, avg);
1026	} else if (
1027		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1028		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
1029					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1030					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1031			runtime_itlb_cache_stats[cpu].n != 0) {
1032		print_itlb_cache_misses(cpu, evsel, avg);
1033	} else if (
1034		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1035		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
1036					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1037					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1038			runtime_ll_cache_stats[cpu].n != 0) {
1039		print_ll_cache_misses(cpu, evsel, avg);
1040	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1041			runtime_cacherefs_stats[cpu].n != 0) {
1042		total = avg_stats(&runtime_cacherefs_stats[cpu]);
1043
1044		if (total)
1045			ratio = avg * 100 / total;
1046
1047		fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
1048
1049	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1050		print_stalled_cycles_frontend(cpu, evsel, avg);
1051	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1052		print_stalled_cycles_backend(cpu, evsel, avg);
1053	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1054		total = avg_stats(&runtime_nsecs_stats[cpu]);
1055
1056		if (total) {
1057			ratio = avg / total;
1058			fprintf(output, " # %8.3f GHz                    ", ratio);
1059		}
1060	} else if (transaction_run &&
1061		   perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1062		total = avg_stats(&runtime_cycles_stats[cpu]);
1063		if (total)
1064			fprintf(output,
1065				" #   %5.2f%% transactional cycles   ",
1066				100.0 * (avg / total));
1067	} else if (transaction_run &&
1068		   perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1069		total = avg_stats(&runtime_cycles_stats[cpu]);
1070		total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1071		if (total2 < avg)
1072			total2 = avg;
1073		if (total)
1074			fprintf(output,
1075				" #   %5.2f%% aborted cycles         ",
1076				100.0 * ((total2-avg) / total));
1077	} else if (transaction_run &&
1078		   perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1079		   avg > 0 &&
1080		   runtime_cycles_in_tx_stats[cpu].n != 0) {
1081		total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1082
1083		if (total)
1084			ratio = total / avg;
1085
1086		fprintf(output, " # %8.0f cycles / transaction   ", ratio);
1087	} else if (transaction_run &&
1088		   perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1089		   avg > 0 &&
1090		   runtime_cycles_in_tx_stats[cpu].n != 0) {
1091		total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1092
1093		if (total)
1094			ratio = total / avg;
1095
1096		fprintf(output, " # %8.0f cycles / elision       ", ratio);
1097	} else if (runtime_nsecs_stats[cpu].n != 0) {
1098		char unit = 'M';
1099
1100		total = avg_stats(&runtime_nsecs_stats[cpu]);
1101
1102		if (total)
1103			ratio = 1000.0 * avg / total;
1104		if (ratio < 0.001) {
1105			ratio *= 1000;
1106			unit = 'K';
1107		}
1108
1109		fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
1110	} else {
1111		fprintf(output, "                                   ");
1112	}
1113}
1114
1115static void print_aggr(char *prefix)
1116{
1117	struct perf_evsel *counter;
1118	int cpu, cpu2, s, s2, id, nr;
1119	double uval;
1120	u64 ena, run, val;
1121
1122	if (!(aggr_map || aggr_get_id))
1123		return;
1124
1125	for (s = 0; s < aggr_map->nr; s++) {
1126		id = aggr_map->map[s];
1127		evlist__for_each(evsel_list, counter) {
1128			val = ena = run = 0;
1129			nr = 0;
1130			for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1131				cpu2 = perf_evsel__cpus(counter)->map[cpu];
1132				s2 = aggr_get_id(evsel_list->cpus, cpu2);
1133				if (s2 != id)
1134					continue;
1135				val += counter->counts->cpu[cpu].val;
1136				ena += counter->counts->cpu[cpu].ena;
1137				run += counter->counts->cpu[cpu].run;
1138				nr++;
1139			}
1140			if (prefix)
1141				fprintf(output, "%s", prefix);
1142
1143			if (run == 0 || ena == 0) {
1144				aggr_printout(counter, id, nr);
1145
1146				fprintf(output, "%*s%s",
1147					csv_output ? 0 : 18,
1148					counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1149					csv_sep);
1150
1151				fprintf(output, "%-*s%s",
1152					csv_output ? 0 : unit_width,
1153					counter->unit, csv_sep);
1154
1155				fprintf(output, "%*s",
1156					csv_output ? 0 : -25,
1157					perf_evsel__name(counter));
1158
1159				if (counter->cgrp)
1160					fprintf(output, "%s%s",
1161						csv_sep, counter->cgrp->name);
1162
1163				fputc('\n', output);
1164				continue;
1165			}
1166			uval = val * counter->scale;
1167
1168			if (nsec_counter(counter))
1169				nsec_printout(id, nr, counter, uval);
1170			else
1171				abs_printout(id, nr, counter, uval);
1172
1173			if (!csv_output) {
1174				print_noise(counter, 1.0);
1175
1176				if (run != ena)
1177					fprintf(output, "  (%.2f%%)",
1178						100.0 * run / ena);
1179			}
1180			fputc('\n', output);
1181		}
 
1182	}
1183}
1184
1185/*
1186 * Print out the results of a single counter:
1187 * aggregated counts in system-wide mode
1188 */
1189static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1190{
1191	struct perf_stat *ps = counter->priv;
1192	double avg = avg_stats(&ps->res_stats[0]);
1193	int scaled = counter->counts->scaled;
1194	double uval;
1195
1196	if (prefix)
1197		fprintf(output, "%s", prefix);
1198
1199	if (scaled == -1) {
1200		fprintf(output, "%*s%s",
1201			csv_output ? 0 : 18,
1202			counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1203			csv_sep);
1204		fprintf(output, "%-*s%s",
1205			csv_output ? 0 : unit_width,
1206			counter->unit, csv_sep);
1207		fprintf(output, "%*s",
1208			csv_output ? 0 : -25,
1209			perf_evsel__name(counter));
1210
1211		if (counter->cgrp)
1212			fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1213
1214		fputc('\n', output);
1215		return;
 
 
1216	}
1217
1218	uval = avg * counter->scale;
1219
1220	if (nsec_counter(counter))
1221		nsec_printout(-1, 0, counter, uval);
 
 
 
1222	else
1223		abs_printout(-1, 0, counter, uval);
 
 
 
1224
1225	print_noise(counter, avg);
1226
1227	if (csv_output) {
1228		fputc('\n', output);
1229		return;
 
1230	}
 
1231
1232	if (scaled) {
1233		double avg_enabled, avg_running;
 
 
 
1234
1235		avg_enabled = avg_stats(&ps->res_stats[1]);
1236		avg_running = avg_stats(&ps->res_stats[2]);
1237
1238		fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1239	}
1240	fprintf(output, "\n");
1241}
1242
1243/*
1244 * Print out the results of a single counter:
1245 * does not use aggregated count in system-wide
1246 */
1247static void print_counter(struct perf_evsel *counter, char *prefix)
1248{
1249	u64 ena, run, val;
1250	double uval;
1251	int cpu;
1252
1253	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1254		val = counter->counts->cpu[cpu].val;
1255		ena = counter->counts->cpu[cpu].ena;
1256		run = counter->counts->cpu[cpu].run;
1257
1258		if (prefix)
1259			fprintf(output, "%s", prefix);
1260
1261		if (run == 0 || ena == 0) {
1262			fprintf(output, "CPU%*d%s%*s%s",
1263				csv_output ? 0 : -4,
1264				perf_evsel__cpus(counter)->map[cpu], csv_sep,
1265				csv_output ? 0 : 18,
1266				counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1267				csv_sep);
1268
1269				fprintf(output, "%-*s%s",
1270					csv_output ? 0 : unit_width,
1271					counter->unit, csv_sep);
1272
1273				fprintf(output, "%*s",
1274					csv_output ? 0 : -25,
1275					perf_evsel__name(counter));
1276
1277			if (counter->cgrp)
1278				fprintf(output, "%s%s",
1279					csv_sep, counter->cgrp->name);
1280
1281			fputc('\n', output);
1282			continue;
1283		}
1284
1285		uval = val * counter->scale;
 
1286
1287		if (nsec_counter(counter))
1288			nsec_printout(cpu, 0, counter, uval);
1289		else
1290			abs_printout(cpu, 0, counter, uval);
1291
1292		if (!csv_output) {
1293			print_noise(counter, 1.0);
 
 
 
 
 
 
 
1294
1295			if (run != ena)
1296				fprintf(output, "  (%.2f%%)",
1297					100.0 * run / ena);
1298		}
1299		fputc('\n', output);
1300	}
1301}
1302
1303static void print_stat(int argc, const char **argv)
1304{
1305	struct perf_evsel *counter;
1306	int i;
1307
1308	fflush(stdout);
 
 
 
 
 
 
 
 
 
1309
1310	if (!csv_output) {
1311		fprintf(output, "\n");
1312		fprintf(output, " Performance counter stats for ");
1313		if (target.system_wide)
1314			fprintf(output, "\'system wide");
1315		else if (target.cpu_list)
1316			fprintf(output, "\'CPU(s) %s", target.cpu_list);
1317		else if (!target__has_task(&target)) {
1318			fprintf(output, "\'%s", argv[0]);
1319			for (i = 1; i < argc; i++)
1320				fprintf(output, " %s", argv[i]);
1321		} else if (target.pid)
1322			fprintf(output, "process id \'%s", target.pid);
1323		else
1324			fprintf(output, "thread id \'%s", target.tid);
1325
1326		fprintf(output, "\'");
1327		if (run_count > 1)
1328			fprintf(output, " (%d runs)", run_count);
1329		fprintf(output, ":\n\n");
1330	}
1331
1332	switch (aggr_mode) {
1333	case AGGR_CORE:
1334	case AGGR_SOCKET:
1335		print_aggr(NULL);
1336		break;
1337	case AGGR_GLOBAL:
1338		evlist__for_each(evsel_list, counter)
1339			print_counter_aggr(counter, NULL);
1340		break;
1341	case AGGR_NONE:
1342		evlist__for_each(evsel_list, counter)
1343			print_counter(counter, NULL);
1344		break;
1345	default:
1346		break;
1347	}
1348
1349	if (!csv_output) {
1350		if (!null_run)
1351			fprintf(output, "\n");
1352		fprintf(output, " %17.9f seconds time elapsed",
1353				avg_stats(&walltime_nsecs_stats)/1e9);
1354		if (run_count > 1) {
1355			fprintf(output, "                                        ");
1356			print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1357					avg_stats(&walltime_nsecs_stats));
1358		}
1359		fprintf(output, "\n\n");
1360	}
1361}
1362
1363static volatile int signr = -1;
1364
1365static void skip_signal(int signo)
1366{
1367	if ((child_pid == -1) || interval)
1368		done = 1;
1369
1370	signr = signo;
1371	/*
1372	 * render child_pid harmless
1373	 * won't send SIGTERM to a random
1374	 * process in case of race condition
1375	 * and fast PID recycling
1376	 */
1377	child_pid = -1;
1378}
1379
1380static void sig_atexit(void)
 
1381{
1382	sigset_t set, oset;
1383
1384	/*
1385	 * avoid race condition with SIGCHLD handler
1386	 * in skip_signal() which is modifying child_pid
1387	 * goal is to avoid send SIGTERM to a random
1388	 * process
1389	 */
1390	sigemptyset(&set);
1391	sigaddset(&set, SIGCHLD);
1392	sigprocmask(SIG_BLOCK, &set, &oset);
1393
1394	if (child_pid != -1)
1395		kill(child_pid, SIGTERM);
 
 
 
 
 
 
 
 
1396
1397	sigprocmask(SIG_SETMASK, &oset, NULL);
 
 
 
 
1398
1399	if (signr == -1)
1400		return;
 
 
 
1401
1402	signal(signr, SIG_DFL);
1403	kill(getpid(), signr);
 
 
1404}
1405
1406static int stat__set_big_num(const struct option *opt __maybe_unused,
1407			     const char *s __maybe_unused, int unset)
1408{
1409	big_num_opt = unset ? 0 : 1;
1410	return 0;
1411}
1412
1413static int perf_stat_init_aggr_mode(void)
1414{
1415	switch (aggr_mode) {
1416	case AGGR_SOCKET:
1417		if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1418			perror("cannot build socket map");
1419			return -1;
1420		}
1421		aggr_get_id = cpu_map__get_socket;
1422		break;
1423	case AGGR_CORE:
1424		if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1425			perror("cannot build core map");
1426			return -1;
1427		}
1428		aggr_get_id = cpu_map__get_core;
1429		break;
1430	case AGGR_NONE:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1431	case AGGR_GLOBAL:
 
 
 
 
 
 
1432	default:
1433		break;
1434	}
1435	return 0;
1436}
1437
1438static int setup_events(const char * const *attrs, unsigned len)
1439{
1440	unsigned i;
 
 
 
 
 
 
 
 
 
1441
1442	for (i = 0; i < len; i++) {
1443		if (parse_events(evsel_list, attrs[i]))
1444			return -1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1445	}
 
1446	return 0;
1447}
1448
1449/*
1450 * Add default attributes, if there were no attributes specified or
1451 * if -d/--detailed, -d -d or -d -d -d is used:
1452 */
1453static int add_default_attributes(void)
1454{
1455	struct perf_event_attr default_attrs[] = {
 
1456
1457  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1458  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1459  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1460  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1461
1462  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
 
 
1463  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
 
 
1464  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
 
 
1465  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1466  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1467  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1468
1469};
1470
1471/*
1472 * Detailed stats (-d), covering the L1 and last level data caches:
1473 */
1474	struct perf_event_attr detailed_attrs[] = {
1475
1476  { .type = PERF_TYPE_HW_CACHE,
1477    .config =
1478	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1479	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1480	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1481
1482  { .type = PERF_TYPE_HW_CACHE,
1483    .config =
1484	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1485	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1486	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1487
1488  { .type = PERF_TYPE_HW_CACHE,
1489    .config =
1490	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1491	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1492	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1493
1494  { .type = PERF_TYPE_HW_CACHE,
1495    .config =
1496	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1497	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1498	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1499};
1500
1501/*
1502 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1503 */
1504	struct perf_event_attr very_detailed_attrs[] = {
1505
1506  { .type = PERF_TYPE_HW_CACHE,
1507    .config =
1508	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1509	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1510	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1511
1512  { .type = PERF_TYPE_HW_CACHE,
1513    .config =
1514	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1515	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1516	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1517
1518  { .type = PERF_TYPE_HW_CACHE,
1519    .config =
1520	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1521	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1522	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1523
1524  { .type = PERF_TYPE_HW_CACHE,
1525    .config =
1526	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1527	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1528	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1529
1530  { .type = PERF_TYPE_HW_CACHE,
1531    .config =
1532	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1533	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1534	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1535
1536  { .type = PERF_TYPE_HW_CACHE,
1537    .config =
1538	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1539	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1540	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1541
1542};
1543
1544/*
1545 * Very, very detailed stats (-d -d -d), adding prefetch events:
1546 */
1547	struct perf_event_attr very_very_detailed_attrs[] = {
1548
1549  { .type = PERF_TYPE_HW_CACHE,
1550    .config =
1551	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1552	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1553	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1554
1555  { .type = PERF_TYPE_HW_CACHE,
1556    .config =
1557	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1558	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1559	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1560};
1561
 
 
1562	/* Set attrs if no event is selected and !null_run: */
1563	if (null_run)
1564		return 0;
1565
1566	if (transaction_run) {
1567		int err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1568		if (pmu_have_event("cpu", "cycles-ct") &&
1569		    pmu_have_event("cpu", "el-start"))
1570			err = setup_events(transaction_attrs,
1571					ARRAY_SIZE(transaction_attrs));
1572		else
1573			err = setup_events(transaction_limited_attrs,
1574				 ARRAY_SIZE(transaction_limited_attrs));
1575		if (err < 0) {
 
1576			fprintf(stderr, "Cannot set up transaction events\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1577			return -1;
1578		}
1579		return 0;
1580	}
1581
1582	if (!evsel_list->nr_entries) {
1583		if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1584			return -1;
1585	}
1586
1587	/* Detailed events get appended to the event list: */
1588
1589	if (detailed_run <  1)
1590		return 0;
1591
1592	/* Append detailed run extra attributes: */
1593	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1594		return -1;
1595
1596	if (detailed_run < 2)
1597		return 0;
1598
1599	/* Append very detailed run extra attributes: */
1600	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1601		return -1;
1602
1603	if (detailed_run < 3)
1604		return 0;
1605
1606	/* Append very, very detailed run extra attributes: */
1607	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1608}
1609
1610int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 
1611{
1612	bool append_file = false;
1613	int output_fd = 0;
1614	const char *output_name	= NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1615	const struct option options[] = {
1616	OPT_BOOLEAN('T', "transaction", &transaction_run,
1617		    "hardware transaction statistics"),
1618	OPT_CALLBACK('e', "event", &evsel_list, "event",
1619		     "event selector. use 'perf list' to list available events",
1620		     parse_events_option),
1621	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1622		     "event filter", parse_filter),
1623	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1624		    "child tasks do not inherit counters"),
1625	OPT_STRING('p', "pid", &target.pid, "pid",
1626		   "stat events on existing process id"),
1627	OPT_STRING('t', "tid", &target.tid, "tid",
1628		   "stat events on existing thread id"),
1629	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1630		    "system-wide collection from all CPUs"),
1631	OPT_BOOLEAN('g', "group", &group,
1632		    "put the counters into a counter group"),
1633	OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1634	OPT_INCR('v', "verbose", &verbose,
1635		    "be more verbose (show counter open errors, etc)"),
1636	OPT_INTEGER('r', "repeat", &run_count,
1637		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1638	OPT_BOOLEAN('n', "null", &null_run,
1639		    "null run - dont start any counters"),
1640	OPT_INCR('d', "detailed", &detailed_run,
1641		    "detailed run - start a lot of events"),
1642	OPT_BOOLEAN('S', "sync", &sync_run,
1643		    "call sync() before starting a run"),
1644	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 
1645			   "print large numbers with thousands\' separators",
1646			   stat__set_big_num),
1647	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1648		    "list of cpus to monitor in system-wide"),
1649	OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1650		    "disable CPU count aggregation", AGGR_NONE),
1651	OPT_STRING('x', "field-separator", &csv_sep, "separator",
1652		   "print counts with custom separator"),
1653	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1654		     "monitor event in cgroup name only", parse_cgroups),
1655	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1656	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1657	OPT_INTEGER(0, "log-fd", &output_fd,
1658		    "log output to fd, instead of stderr"),
1659	OPT_STRING(0, "pre", &pre_cmd, "command",
1660			"command to run prior to the measured command"),
1661	OPT_STRING(0, "post", &post_cmd, "command",
1662			"command to run after to the measured command"),
1663	OPT_UINTEGER('I', "interval-print", &interval,
1664		    "print counts at regular interval in ms (>= 100)"),
1665	OPT_SET_UINT(0, "per-socket", &aggr_mode,
1666		     "aggregate counts per processor socket", AGGR_SOCKET),
1667	OPT_SET_UINT(0, "per-core", &aggr_mode,
 
 
1668		     "aggregate counts per physical processor core", AGGR_CORE),
1669	OPT_UINTEGER('D', "delay", &initial_delay,
1670		     "ms to wait before starting measurement after program start"),
 
 
1671	OPT_END()
1672	};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1673	const char * const stat_usage[] = {
1674		"perf stat [<options>] [<command>]",
1675		NULL
1676	};
1677	int status = -EINVAL, run_idx;
1678	const char *mode;
 
 
 
 
1679
1680	setlocale(LC_ALL, "");
1681
1682	evsel_list = perf_evlist__new();
1683	if (evsel_list == NULL)
1684		return -ENOMEM;
1685
1686	argc = parse_options(argc, argv, options, stat_usage,
1687		PARSE_OPT_STOP_AT_NON_OPTION);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1688
1689	output = stderr;
1690	if (output_name && strcmp(output_name, "-"))
 
 
 
 
 
1691		output = NULL;
1692
1693	if (output_name && output_fd) {
1694		fprintf(stderr, "cannot use both --output and --log-fd\n");
1695		parse_options_usage(stat_usage, options, "o", 1);
1696		parse_options_usage(NULL, options, "log-fd", 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1697		goto out;
1698	}
1699
1700	if (output_fd < 0) {
1701		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1702		parse_options_usage(stat_usage, options, "log-fd", 0);
1703		goto out;
1704	}
1705
1706	if (!output) {
1707		struct timespec tm;
1708		mode = append_file ? "a" : "w";
1709
1710		output = fopen(output_name, mode);
1711		if (!output) {
1712			perror("failed to create output file");
1713			return -1;
1714		}
1715		clock_gettime(CLOCK_REALTIME, &tm);
1716		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1717	} else if (output_fd > 0) {
1718		mode = append_file ? "a" : "w";
1719		output = fdopen(output_fd, mode);
1720		if (!output) {
1721			perror("Failed opening logfd");
1722			return -errno;
1723		}
1724	}
1725
1726	if (csv_sep) {
1727		csv_output = true;
1728		if (!strcmp(csv_sep, "\\t"))
1729			csv_sep = "\t";
1730	} else
1731		csv_sep = DEFAULT_SEPARATOR;
 
 
 
1732
1733	/*
1734	 * let the spreadsheet do the pretty-printing
1735	 */
1736	if (csv_output) {
1737		/* User explicitly passed -B? */
1738		if (big_num_opt == 1) {
1739			fprintf(stderr, "-B option not supported with -x\n");
1740			parse_options_usage(stat_usage, options, "B", 1);
1741			parse_options_usage(NULL, options, "x", 1);
1742			goto out;
1743		} else /* Nope, so disable big number formatting */
1744			big_num = false;
1745	} else if (big_num_opt == 0) /* User passed --no-big-num */
1746		big_num = false;
 
 
 
 
 
 
 
 
1747
1748	if (!argc && target__none(&target))
1749		usage_with_options(stat_usage, options);
 
 
 
 
1750
1751	if (run_count < 0) {
1752		pr_err("Run count must be a positive number\n");
1753		parse_options_usage(stat_usage, options, "r", 1);
1754		goto out;
1755	} else if (run_count == 0) {
1756		forever = true;
1757		run_count = 1;
1758	}
1759
1760	/* no_aggr, cgroup are for system-wide only */
1761	if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1762	    !target__has_cpu(&target)) {
1763		fprintf(stderr, "both cgroup and no-aggregation "
1764			"modes only available in system-wide mode\n");
1765
1766		parse_options_usage(stat_usage, options, "G", 1);
1767		parse_options_usage(NULL, options, "A", 1);
1768		parse_options_usage(NULL, options, "a", 1);
 
1769		goto out;
1770	}
1771
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1772	if (add_default_attributes())
1773		goto out;
1774
1775	target__validate(&target);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1776
1777	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
 
1778		if (target__has_task(&target)) {
1779			pr_err("Problems finding threads of monitor\n");
1780			parse_options_usage(stat_usage, options, "p", 1);
1781			parse_options_usage(NULL, options, "t", 1);
1782		} else if (target__has_cpu(&target)) {
1783			perror("failed to parse CPUs map");
1784			parse_options_usage(stat_usage, options, "C", 1);
1785			parse_options_usage(NULL, options, "a", 1);
1786		}
1787		goto out;
1788	}
1789	if (interval && interval < 100) {
1790		pr_err("print interval must be >= 100ms\n");
1791		parse_options_usage(stat_usage, options, "I", 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1792		goto out;
1793	}
1794
1795	if (perf_evlist__alloc_stats(evsel_list, interval))
 
 
 
 
 
 
 
 
 
 
 
 
 
1796		goto out;
 
1797
1798	if (perf_stat_init_aggr_mode())
1799		goto out;
1800
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1801	/*
1802	 * We dont want to block the signals - that would cause
1803	 * child tasks to inherit that and Ctrl-C would not work.
1804	 * What we want is for Ctrl-C to work in the exec()-ed
1805	 * task, but being ignored by perf stat itself:
1806	 */
1807	atexit(sig_atexit);
1808	if (!forever)
1809		signal(SIGINT,  skip_signal);
1810	signal(SIGCHLD, skip_signal);
1811	signal(SIGALRM, skip_signal);
1812	signal(SIGABRT, skip_signal);
1813
 
 
 
 
 
1814	status = 0;
1815	for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1816		if (run_count != 1 && verbose)
1817			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1818				run_idx + 1);
1819
1820		status = run_perf_stat(argc, argv);
1821		if (forever && status != -1) {
1822			print_stat(argc, argv);
1823			perf_stat__reset_stats(evsel_list);
 
 
 
1824		}
1825	}
1826
1827	if (!forever && status != -1 && !interval)
1828		print_stat(argc, argv);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1829
1830	perf_evlist__free_stats(evsel_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1831out:
1832	perf_evlist__delete(evsel_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
1833	return status;
1834}