Linux Audio

Check our new training course

Loading...
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * builtin-record.c
   4 *
   5 * Builtin record command: Record the profile of a workload
   6 * (or a CPU, or a PID) into the perf.data output file - for
   7 * later analysis via perf report.
   8 */
   9#include "builtin.h"
  10
  11#include "util/build-id.h"
  12#include <subcmd/parse-options.h>
  13#include <internal/xyarray.h>
  14#include "util/parse-events.h"
  15#include "util/config.h"
  16
  17#include "util/callchain.h"
  18#include "util/cgroup.h"
  19#include "util/header.h"
  20#include "util/event.h"
  21#include "util/evlist.h"
  22#include "util/evsel.h"
  23#include "util/debug.h"
  24#include "util/mmap.h"
  25#include "util/mutex.h"
  26#include "util/target.h"
  27#include "util/session.h"
  28#include "util/tool.h"
  29#include "util/symbol.h"
  30#include "util/record.h"
  31#include "util/cpumap.h"
  32#include "util/thread_map.h"
  33#include "util/data.h"
  34#include "util/perf_regs.h"
  35#include "util/auxtrace.h"
  36#include "util/tsc.h"
  37#include "util/parse-branch-options.h"
  38#include "util/parse-regs-options.h"
  39#include "util/perf_api_probe.h"
  40#include "util/trigger.h"
  41#include "util/perf-hooks.h"
  42#include "util/cpu-set-sched.h"
  43#include "util/synthetic-events.h"
  44#include "util/time-utils.h"
  45#include "util/units.h"
  46#include "util/bpf-event.h"
  47#include "util/util.h"
  48#include "util/pfm.h"
  49#include "util/pmu.h"
  50#include "util/pmus.h"
  51#include "util/clockid.h"
  52#include "util/off_cpu.h"
  53#include "util/bpf-filter.h"
  54#include "asm/bug.h"
  55#include "perf.h"
  56#include "cputopo.h"
  57
  58#include <errno.h>
  59#include <inttypes.h>
  60#include <locale.h>
  61#include <poll.h>
  62#include <pthread.h>
  63#include <unistd.h>
  64#ifndef HAVE_GETTID
  65#include <syscall.h>
  66#endif
  67#include <sched.h>
  68#include <signal.h>
  69#ifdef HAVE_EVENTFD_SUPPORT
  70#include <sys/eventfd.h>
  71#endif
  72#include <sys/mman.h>
  73#include <sys/wait.h>
  74#include <sys/types.h>
  75#include <sys/stat.h>
  76#include <fcntl.h>
  77#include <linux/err.h>
  78#include <linux/string.h>
  79#include <linux/time64.h>
  80#include <linux/zalloc.h>
  81#include <linux/bitmap.h>
  82#include <sys/time.h>
  83
  84struct switch_output {
  85	bool		 enabled;
  86	bool		 signal;
  87	unsigned long	 size;
  88	unsigned long	 time;
  89	const char	*str;
  90	bool		 set;
  91	char		 **filenames;
  92	int		 num_files;
  93	int		 cur_file;
  94};
  95
  96struct thread_mask {
  97	struct mmap_cpu_mask	maps;
  98	struct mmap_cpu_mask	affinity;
  99};
 100
 101struct record_thread {
 102	pid_t			tid;
 103	struct thread_mask	*mask;
 104	struct {
 105		int		msg[2];
 106		int		ack[2];
 107	} pipes;
 108	struct fdarray		pollfd;
 109	int			ctlfd_pos;
 110	int			nr_mmaps;
 111	struct mmap		**maps;
 112	struct mmap		**overwrite_maps;
 113	struct record		*rec;
 114	unsigned long long	samples;
 115	unsigned long		waking;
 116	u64			bytes_written;
 117	u64			bytes_transferred;
 118	u64			bytes_compressed;
 119};
 120
 121static __thread struct record_thread *thread;
 122
 123enum thread_msg {
 124	THREAD_MSG__UNDEFINED = 0,
 125	THREAD_MSG__READY,
 126	THREAD_MSG__MAX,
 127};
 128
 129static const char *thread_msg_tags[THREAD_MSG__MAX] = {
 130	"UNDEFINED", "READY"
 131};
 132
 133enum thread_spec {
 134	THREAD_SPEC__UNDEFINED = 0,
 135	THREAD_SPEC__CPU,
 136	THREAD_SPEC__CORE,
 137	THREAD_SPEC__PACKAGE,
 138	THREAD_SPEC__NUMA,
 139	THREAD_SPEC__USER,
 140	THREAD_SPEC__MAX,
 141};
 142
 143static const char *thread_spec_tags[THREAD_SPEC__MAX] = {
 144	"undefined", "cpu", "core", "package", "numa", "user"
 145};
 146
 147struct pollfd_index_map {
 148	int evlist_pollfd_index;
 149	int thread_pollfd_index;
 150};
 151
 152struct record {
 153	struct perf_tool	tool;
 154	struct record_opts	opts;
 155	u64			bytes_written;
 156	u64			thread_bytes_written;
 157	struct perf_data	data;
 158	struct auxtrace_record	*itr;
 159	struct evlist	*evlist;
 160	struct perf_session	*session;
 161	struct evlist		*sb_evlist;
 162	pthread_t		thread_id;
 163	int			realtime_prio;
 164	bool			switch_output_event_set;
 165	bool			no_buildid;
 166	bool			no_buildid_set;
 167	bool			no_buildid_cache;
 168	bool			no_buildid_cache_set;
 169	bool			buildid_all;
 170	bool			buildid_mmap;
 171	bool			timestamp_filename;
 172	bool			timestamp_boundary;
 173	bool			off_cpu;
 
 174	struct switch_output	switch_output;
 175	unsigned long long	samples;
 176	unsigned long		output_max_size;	/* = 0: unlimited */
 177	struct perf_debuginfod	debuginfod;
 178	int			nr_threads;
 179	struct thread_mask	*thread_masks;
 180	struct record_thread	*thread_data;
 181	struct pollfd_index_map	*index_map;
 182	size_t			index_map_sz;
 183	size_t			index_map_cnt;
 184};
 185
 186static volatile int done;
 187
 188static volatile int auxtrace_record__snapshot_started;
 189static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
 190static DEFINE_TRIGGER(switch_output_trigger);
 191
 192static const char *affinity_tags[PERF_AFFINITY_MAX] = {
 193	"SYS", "NODE", "CPU"
 194};
 195
 
 
 
 
 
 
 
 
 
 196#ifndef HAVE_GETTID
 197static inline pid_t gettid(void)
 198{
 199	return (pid_t)syscall(__NR_gettid);
 200}
 201#endif
 202
 203static int record__threads_enabled(struct record *rec)
 204{
 205	return rec->opts.threads_spec;
 206}
 207
 208static bool switch_output_signal(struct record *rec)
 209{
 210	return rec->switch_output.signal &&
 211	       trigger_is_ready(&switch_output_trigger);
 212}
 213
 214static bool switch_output_size(struct record *rec)
 215{
 216	return rec->switch_output.size &&
 217	       trigger_is_ready(&switch_output_trigger) &&
 218	       (rec->bytes_written >= rec->switch_output.size);
 219}
 220
 221static bool switch_output_time(struct record *rec)
 222{
 223	return rec->switch_output.time &&
 224	       trigger_is_ready(&switch_output_trigger);
 225}
 226
 227static u64 record__bytes_written(struct record *rec)
 228{
 229	return rec->bytes_written + rec->thread_bytes_written;
 230}
 231
 232static bool record__output_max_size_exceeded(struct record *rec)
 233{
 234	return rec->output_max_size &&
 235	       (record__bytes_written(rec) >= rec->output_max_size);
 236}
 237
 238static int record__write(struct record *rec, struct mmap *map __maybe_unused,
 239			 void *bf, size_t size)
 240{
 241	struct perf_data_file *file = &rec->session->data->file;
 242
 243	if (map && map->file)
 244		file = map->file;
 245
 246	if (perf_data_file__write(file, bf, size) < 0) {
 247		pr_err("failed to write perf data, error: %m\n");
 248		return -1;
 249	}
 250
 251	if (map && map->file) {
 252		thread->bytes_written += size;
 253		rec->thread_bytes_written += size;
 254	} else {
 255		rec->bytes_written += size;
 256	}
 257
 258	if (record__output_max_size_exceeded(rec) && !done) {
 259		fprintf(stderr, "[ perf record: perf size limit reached (%" PRIu64 " KB),"
 260				" stopping session ]\n",
 261				record__bytes_written(rec) >> 10);
 262		done = 1;
 263	}
 264
 265	if (switch_output_size(rec))
 266		trigger_hit(&switch_output_trigger);
 267
 268	return 0;
 269}
 270
 271static int record__aio_enabled(struct record *rec);
 272static int record__comp_enabled(struct record *rec);
 273static ssize_t zstd_compress(struct perf_session *session, struct mmap *map,
 274			    void *dst, size_t dst_size, void *src, size_t src_size);
 275
 276#ifdef HAVE_AIO_SUPPORT
 277static int record__aio_write(struct aiocb *cblock, int trace_fd,
 278		void *buf, size_t size, off_t off)
 279{
 280	int rc;
 281
 282	cblock->aio_fildes = trace_fd;
 283	cblock->aio_buf    = buf;
 284	cblock->aio_nbytes = size;
 285	cblock->aio_offset = off;
 286	cblock->aio_sigevent.sigev_notify = SIGEV_NONE;
 287
 288	do {
 289		rc = aio_write(cblock);
 290		if (rc == 0) {
 291			break;
 292		} else if (errno != EAGAIN) {
 293			cblock->aio_fildes = -1;
 294			pr_err("failed to queue perf data, error: %m\n");
 295			break;
 296		}
 297	} while (1);
 298
 299	return rc;
 300}
 301
 302static int record__aio_complete(struct mmap *md, struct aiocb *cblock)
 303{
 304	void *rem_buf;
 305	off_t rem_off;
 306	size_t rem_size;
 307	int rc, aio_errno;
 308	ssize_t aio_ret, written;
 309
 310	aio_errno = aio_error(cblock);
 311	if (aio_errno == EINPROGRESS)
 312		return 0;
 313
 314	written = aio_ret = aio_return(cblock);
 315	if (aio_ret < 0) {
 316		if (aio_errno != EINTR)
 317			pr_err("failed to write perf data, error: %m\n");
 318		written = 0;
 319	}
 320
 321	rem_size = cblock->aio_nbytes - written;
 322
 323	if (rem_size == 0) {
 324		cblock->aio_fildes = -1;
 325		/*
 326		 * md->refcount is incremented in record__aio_pushfn() for
 327		 * every aio write request started in record__aio_push() so
 328		 * decrement it because the request is now complete.
 329		 */
 330		perf_mmap__put(&md->core);
 331		rc = 1;
 332	} else {
 333		/*
 334		 * aio write request may require restart with the
 335		 * reminder if the kernel didn't write whole
 336		 * chunk at once.
 337		 */
 338		rem_off = cblock->aio_offset + written;
 339		rem_buf = (void *)(cblock->aio_buf + written);
 340		record__aio_write(cblock, cblock->aio_fildes,
 341				rem_buf, rem_size, rem_off);
 342		rc = 0;
 343	}
 344
 345	return rc;
 346}
 347
 348static int record__aio_sync(struct mmap *md, bool sync_all)
 349{
 350	struct aiocb **aiocb = md->aio.aiocb;
 351	struct aiocb *cblocks = md->aio.cblocks;
 352	struct timespec timeout = { 0, 1000 * 1000  * 1 }; /* 1ms */
 353	int i, do_suspend;
 354
 355	do {
 356		do_suspend = 0;
 357		for (i = 0; i < md->aio.nr_cblocks; ++i) {
 358			if (cblocks[i].aio_fildes == -1 || record__aio_complete(md, &cblocks[i])) {
 359				if (sync_all)
 360					aiocb[i] = NULL;
 361				else
 362					return i;
 363			} else {
 364				/*
 365				 * Started aio write is not complete yet
 366				 * so it has to be waited before the
 367				 * next allocation.
 368				 */
 369				aiocb[i] = &cblocks[i];
 370				do_suspend = 1;
 371			}
 372		}
 373		if (!do_suspend)
 374			return -1;
 375
 376		while (aio_suspend((const struct aiocb **)aiocb, md->aio.nr_cblocks, &timeout)) {
 377			if (!(errno == EAGAIN || errno == EINTR))
 378				pr_err("failed to sync perf data, error: %m\n");
 379		}
 380	} while (1);
 381}
 382
 383struct record_aio {
 384	struct record	*rec;
 385	void		*data;
 386	size_t		size;
 387};
 388
 389static int record__aio_pushfn(struct mmap *map, void *to, void *buf, size_t size)
 390{
 391	struct record_aio *aio = to;
 392
 393	/*
 394	 * map->core.base data pointed by buf is copied into free map->aio.data[] buffer
 395	 * to release space in the kernel buffer as fast as possible, calling
 396	 * perf_mmap__consume() from perf_mmap__push() function.
 397	 *
 398	 * That lets the kernel to proceed with storing more profiling data into
 399	 * the kernel buffer earlier than other per-cpu kernel buffers are handled.
 400	 *
 401	 * Coping can be done in two steps in case the chunk of profiling data
 402	 * crosses the upper bound of the kernel buffer. In this case we first move
 403	 * part of data from map->start till the upper bound and then the reminder
 404	 * from the beginning of the kernel buffer till the end of the data chunk.
 405	 */
 406
 407	if (record__comp_enabled(aio->rec)) {
 408		ssize_t compressed = zstd_compress(aio->rec->session, NULL, aio->data + aio->size,
 409						   mmap__mmap_len(map) - aio->size,
 410						   buf, size);
 411		if (compressed < 0)
 412			return (int)compressed;
 413
 414		size = compressed;
 415	} else {
 416		memcpy(aio->data + aio->size, buf, size);
 417	}
 418
 419	if (!aio->size) {
 420		/*
 421		 * Increment map->refcount to guard map->aio.data[] buffer
 422		 * from premature deallocation because map object can be
 423		 * released earlier than aio write request started on
 424		 * map->aio.data[] buffer is complete.
 425		 *
 426		 * perf_mmap__put() is done at record__aio_complete()
 427		 * after started aio request completion or at record__aio_push()
 428		 * if the request failed to start.
 429		 */
 430		perf_mmap__get(&map->core);
 431	}
 432
 433	aio->size += size;
 434
 435	return size;
 436}
 437
 438static int record__aio_push(struct record *rec, struct mmap *map, off_t *off)
 439{
 440	int ret, idx;
 441	int trace_fd = rec->session->data->file.fd;
 442	struct record_aio aio = { .rec = rec, .size = 0 };
 443
 444	/*
 445	 * Call record__aio_sync() to wait till map->aio.data[] buffer
 446	 * becomes available after previous aio write operation.
 447	 */
 448
 449	idx = record__aio_sync(map, false);
 450	aio.data = map->aio.data[idx];
 451	ret = perf_mmap__push(map, &aio, record__aio_pushfn);
 452	if (ret != 0) /* ret > 0 - no data, ret < 0 - error */
 453		return ret;
 454
 455	rec->samples++;
 456	ret = record__aio_write(&(map->aio.cblocks[idx]), trace_fd, aio.data, aio.size, *off);
 457	if (!ret) {
 458		*off += aio.size;
 459		rec->bytes_written += aio.size;
 460		if (switch_output_size(rec))
 461			trigger_hit(&switch_output_trigger);
 462	} else {
 463		/*
 464		 * Decrement map->refcount incremented in record__aio_pushfn()
 465		 * back if record__aio_write() operation failed to start, otherwise
 466		 * map->refcount is decremented in record__aio_complete() after
 467		 * aio write operation finishes successfully.
 468		 */
 469		perf_mmap__put(&map->core);
 470	}
 471
 472	return ret;
 473}
 474
 475static off_t record__aio_get_pos(int trace_fd)
 476{
 477	return lseek(trace_fd, 0, SEEK_CUR);
 478}
 479
 480static void record__aio_set_pos(int trace_fd, off_t pos)
 481{
 482	lseek(trace_fd, pos, SEEK_SET);
 483}
 484
 485static void record__aio_mmap_read_sync(struct record *rec)
 486{
 487	int i;
 488	struct evlist *evlist = rec->evlist;
 489	struct mmap *maps = evlist->mmap;
 490
 491	if (!record__aio_enabled(rec))
 492		return;
 493
 494	for (i = 0; i < evlist->core.nr_mmaps; i++) {
 495		struct mmap *map = &maps[i];
 496
 497		if (map->core.base)
 498			record__aio_sync(map, true);
 499	}
 500}
 501
 502static int nr_cblocks_default = 1;
 503static int nr_cblocks_max = 4;
 504
 505static int record__aio_parse(const struct option *opt,
 506			     const char *str,
 507			     int unset)
 508{
 509	struct record_opts *opts = (struct record_opts *)opt->value;
 510
 511	if (unset) {
 512		opts->nr_cblocks = 0;
 513	} else {
 514		if (str)
 515			opts->nr_cblocks = strtol(str, NULL, 0);
 516		if (!opts->nr_cblocks)
 517			opts->nr_cblocks = nr_cblocks_default;
 518	}
 519
 520	return 0;
 521}
 522#else /* HAVE_AIO_SUPPORT */
 523static int nr_cblocks_max = 0;
 524
 525static int record__aio_push(struct record *rec __maybe_unused, struct mmap *map __maybe_unused,
 526			    off_t *off __maybe_unused)
 527{
 528	return -1;
 529}
 530
 531static off_t record__aio_get_pos(int trace_fd __maybe_unused)
 532{
 533	return -1;
 534}
 535
 536static void record__aio_set_pos(int trace_fd __maybe_unused, off_t pos __maybe_unused)
 537{
 538}
 539
 540static void record__aio_mmap_read_sync(struct record *rec __maybe_unused)
 541{
 542}
 543#endif
 544
 545static int record__aio_enabled(struct record *rec)
 546{
 547	return rec->opts.nr_cblocks > 0;
 548}
 549
 550#define MMAP_FLUSH_DEFAULT 1
 551static int record__mmap_flush_parse(const struct option *opt,
 552				    const char *str,
 553				    int unset)
 554{
 555	int flush_max;
 556	struct record_opts *opts = (struct record_opts *)opt->value;
 557	static struct parse_tag tags[] = {
 558			{ .tag  = 'B', .mult = 1       },
 559			{ .tag  = 'K', .mult = 1 << 10 },
 560			{ .tag  = 'M', .mult = 1 << 20 },
 561			{ .tag  = 'G', .mult = 1 << 30 },
 562			{ .tag  = 0 },
 563	};
 564
 565	if (unset)
 566		return 0;
 567
 568	if (str) {
 569		opts->mmap_flush = parse_tag_value(str, tags);
 570		if (opts->mmap_flush == (int)-1)
 571			opts->mmap_flush = strtol(str, NULL, 0);
 572	}
 573
 574	if (!opts->mmap_flush)
 575		opts->mmap_flush = MMAP_FLUSH_DEFAULT;
 576
 577	flush_max = evlist__mmap_size(opts->mmap_pages);
 578	flush_max /= 4;
 579	if (opts->mmap_flush > flush_max)
 580		opts->mmap_flush = flush_max;
 581
 582	return 0;
 583}
 584
 585#ifdef HAVE_ZSTD_SUPPORT
 586static unsigned int comp_level_default = 1;
 587
 588static int record__parse_comp_level(const struct option *opt, const char *str, int unset)
 589{
 590	struct record_opts *opts = opt->value;
 591
 592	if (unset) {
 593		opts->comp_level = 0;
 594	} else {
 595		if (str)
 596			opts->comp_level = strtol(str, NULL, 0);
 597		if (!opts->comp_level)
 598			opts->comp_level = comp_level_default;
 599	}
 600
 601	return 0;
 602}
 603#endif
 604static unsigned int comp_level_max = 22;
 605
 606static int record__comp_enabled(struct record *rec)
 607{
 608	return rec->opts.comp_level > 0;
 609}
 610
 611static int process_synthesized_event(struct perf_tool *tool,
 612				     union perf_event *event,
 613				     struct perf_sample *sample __maybe_unused,
 614				     struct machine *machine __maybe_unused)
 615{
 616	struct record *rec = container_of(tool, struct record, tool);
 617	return record__write(rec, NULL, event, event->header.size);
 618}
 619
 620static struct mutex synth_lock;
 621
 622static int process_locked_synthesized_event(struct perf_tool *tool,
 623				     union perf_event *event,
 624				     struct perf_sample *sample __maybe_unused,
 625				     struct machine *machine __maybe_unused)
 626{
 627	int ret;
 628
 629	mutex_lock(&synth_lock);
 630	ret = process_synthesized_event(tool, event, sample, machine);
 631	mutex_unlock(&synth_lock);
 632	return ret;
 633}
 634
 635static int record__pushfn(struct mmap *map, void *to, void *bf, size_t size)
 636{
 637	struct record *rec = to;
 638
 639	if (record__comp_enabled(rec)) {
 640		ssize_t compressed = zstd_compress(rec->session, map, map->data,
 641						   mmap__mmap_len(map), bf, size);
 642
 643		if (compressed < 0)
 644			return (int)compressed;
 645
 646		size = compressed;
 647		bf   = map->data;
 648	}
 649
 650	thread->samples++;
 651	return record__write(rec, map, bf, size);
 652}
 653
 654static volatile sig_atomic_t signr = -1;
 655static volatile sig_atomic_t child_finished;
 656#ifdef HAVE_EVENTFD_SUPPORT
 657static volatile sig_atomic_t done_fd = -1;
 658#endif
 659
 660static void sig_handler(int sig)
 661{
 662	if (sig == SIGCHLD)
 663		child_finished = 1;
 664	else
 665		signr = sig;
 666
 667	done = 1;
 668#ifdef HAVE_EVENTFD_SUPPORT
 669	if (done_fd >= 0) {
 670		u64 tmp = 1;
 671		int orig_errno = errno;
 672
 673		/*
 674		 * It is possible for this signal handler to run after done is
 675		 * checked in the main loop, but before the perf counter fds are
 676		 * polled. If this happens, the poll() will continue to wait
 677		 * even though done is set, and will only break out if either
 678		 * another signal is received, or the counters are ready for
 679		 * read. To ensure the poll() doesn't sleep when done is set,
 680		 * use an eventfd (done_fd) to wake up the poll().
 681		 */
 682		if (write(done_fd, &tmp, sizeof(tmp)) < 0)
 683			pr_err("failed to signal wakeup fd, error: %m\n");
 684
 685		errno = orig_errno;
 686	}
 687#endif // HAVE_EVENTFD_SUPPORT
 688}
 689
 690static void sigsegv_handler(int sig)
 691{
 692	perf_hooks__recover();
 693	sighandler_dump_stack(sig);
 694}
 695
 696static void record__sig_exit(void)
 697{
 698	if (signr == -1)
 699		return;
 700
 701	signal(signr, SIG_DFL);
 702	raise(signr);
 703}
 704
 705#ifdef HAVE_AUXTRACE_SUPPORT
 706
 707static int record__process_auxtrace(struct perf_tool *tool,
 708				    struct mmap *map,
 709				    union perf_event *event, void *data1,
 710				    size_t len1, void *data2, size_t len2)
 711{
 712	struct record *rec = container_of(tool, struct record, tool);
 713	struct perf_data *data = &rec->data;
 714	size_t padding;
 715	u8 pad[8] = {0};
 716
 717	if (!perf_data__is_pipe(data) && perf_data__is_single_file(data)) {
 718		off_t file_offset;
 719		int fd = perf_data__fd(data);
 720		int err;
 721
 722		file_offset = lseek(fd, 0, SEEK_CUR);
 723		if (file_offset == -1)
 724			return -1;
 725		err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 726						     event, file_offset);
 727		if (err)
 728			return err;
 729	}
 730
 731	/* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 732	padding = (len1 + len2) & 7;
 733	if (padding)
 734		padding = 8 - padding;
 735
 736	record__write(rec, map, event, event->header.size);
 737	record__write(rec, map, data1, len1);
 738	if (len2)
 739		record__write(rec, map, data2, len2);
 740	record__write(rec, map, &pad, padding);
 741
 742	return 0;
 743}
 744
 745static int record__auxtrace_mmap_read(struct record *rec,
 746				      struct mmap *map)
 747{
 748	int ret;
 749
 750	ret = auxtrace_mmap__read(map, rec->itr, &rec->tool,
 751				  record__process_auxtrace);
 752	if (ret < 0)
 753		return ret;
 754
 755	if (ret)
 756		rec->samples++;
 757
 758	return 0;
 759}
 760
 761static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 762					       struct mmap *map)
 763{
 764	int ret;
 765
 766	ret = auxtrace_mmap__read_snapshot(map, rec->itr, &rec->tool,
 767					   record__process_auxtrace,
 768					   rec->opts.auxtrace_snapshot_size);
 769	if (ret < 0)
 770		return ret;
 771
 772	if (ret)
 773		rec->samples++;
 774
 775	return 0;
 776}
 777
 778static int record__auxtrace_read_snapshot_all(struct record *rec)
 779{
 780	int i;
 781	int rc = 0;
 782
 783	for (i = 0; i < rec->evlist->core.nr_mmaps; i++) {
 784		struct mmap *map = &rec->evlist->mmap[i];
 785
 786		if (!map->auxtrace_mmap.base)
 787			continue;
 788
 789		if (record__auxtrace_mmap_read_snapshot(rec, map) != 0) {
 790			rc = -1;
 791			goto out;
 792		}
 793	}
 794out:
 795	return rc;
 796}
 797
 798static void record__read_auxtrace_snapshot(struct record *rec, bool on_exit)
 799{
 800	pr_debug("Recording AUX area tracing snapshot\n");
 801	if (record__auxtrace_read_snapshot_all(rec) < 0) {
 802		trigger_error(&auxtrace_snapshot_trigger);
 803	} else {
 804		if (auxtrace_record__snapshot_finish(rec->itr, on_exit))
 805			trigger_error(&auxtrace_snapshot_trigger);
 806		else
 807			trigger_ready(&auxtrace_snapshot_trigger);
 808	}
 809}
 810
 811static int record__auxtrace_snapshot_exit(struct record *rec)
 812{
 813	if (trigger_is_error(&auxtrace_snapshot_trigger))
 814		return 0;
 815
 816	if (!auxtrace_record__snapshot_started &&
 817	    auxtrace_record__snapshot_start(rec->itr))
 818		return -1;
 819
 820	record__read_auxtrace_snapshot(rec, true);
 821	if (trigger_is_error(&auxtrace_snapshot_trigger))
 822		return -1;
 823
 824	return 0;
 825}
 826
 827static int record__auxtrace_init(struct record *rec)
 828{
 829	int err;
 830
 831	if ((rec->opts.auxtrace_snapshot_opts || rec->opts.auxtrace_sample_opts)
 832	    && record__threads_enabled(rec)) {
 833		pr_err("AUX area tracing options are not available in parallel streaming mode.\n");
 834		return -EINVAL;
 835	}
 836
 837	if (!rec->itr) {
 838		rec->itr = auxtrace_record__init(rec->evlist, &err);
 839		if (err)
 840			return err;
 841	}
 842
 843	err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
 844					      rec->opts.auxtrace_snapshot_opts);
 845	if (err)
 846		return err;
 847
 848	err = auxtrace_parse_sample_options(rec->itr, rec->evlist, &rec->opts,
 849					    rec->opts.auxtrace_sample_opts);
 850	if (err)
 851		return err;
 852
 853	auxtrace_regroup_aux_output(rec->evlist);
 854
 855	return auxtrace_parse_filters(rec->evlist);
 856}
 857
 858#else
 859
 860static inline
 861int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 862			       struct mmap *map __maybe_unused)
 863{
 864	return 0;
 865}
 866
 867static inline
 868void record__read_auxtrace_snapshot(struct record *rec __maybe_unused,
 869				    bool on_exit __maybe_unused)
 870{
 871}
 872
 873static inline
 874int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 875{
 876	return 0;
 877}
 878
 879static inline
 880int record__auxtrace_snapshot_exit(struct record *rec __maybe_unused)
 881{
 882	return 0;
 883}
 884
 885static int record__auxtrace_init(struct record *rec __maybe_unused)
 886{
 887	return 0;
 888}
 889
 890#endif
 891
 892static int record__config_text_poke(struct evlist *evlist)
 893{
 894	struct evsel *evsel;
 895
 896	/* Nothing to do if text poke is already configured */
 897	evlist__for_each_entry(evlist, evsel) {
 898		if (evsel->core.attr.text_poke)
 899			return 0;
 900	}
 901
 902	evsel = evlist__add_dummy_on_all_cpus(evlist);
 903	if (!evsel)
 904		return -ENOMEM;
 905
 906	evsel->core.attr.text_poke = 1;
 907	evsel->core.attr.ksymbol = 1;
 908	evsel->immediate = true;
 909	evsel__set_sample_bit(evsel, TIME);
 910
 911	return 0;
 912}
 913
 914static int record__config_off_cpu(struct record *rec)
 915{
 916	return off_cpu_prepare(rec->evlist, &rec->opts.target, &rec->opts);
 917}
 918
 919static bool record__tracking_system_wide(struct record *rec)
 920{
 921	struct evlist *evlist = rec->evlist;
 922	struct evsel *evsel;
 923
 924	/*
 925	 * If non-dummy evsel exists, system_wide sideband is need to
 926	 * help parse sample information.
 927	 * For example, PERF_EVENT_MMAP event to help parse symbol,
 928	 * and PERF_EVENT_COMM event to help parse task executable name.
 929	 */
 930	evlist__for_each_entry(evlist, evsel) {
 931		if (!evsel__is_dummy_event(evsel))
 932			return true;
 933	}
 934
 935	return false;
 936}
 937
 938static int record__config_tracking_events(struct record *rec)
 939{
 940	struct record_opts *opts = &rec->opts;
 941	struct evlist *evlist = rec->evlist;
 942	bool system_wide = false;
 943	struct evsel *evsel;
 944
 945	/*
 946	 * For initial_delay, system wide or a hybrid system, we need to add
 947	 * tracking event so that we can track PERF_RECORD_MMAP to cover the
 948	 * delay of waiting or event synthesis.
 949	 */
 950	if (opts->target.initial_delay || target__has_cpu(&opts->target) ||
 951	    perf_pmus__num_core_pmus() > 1) {
 952
 953		/*
 954		 * User space tasks can migrate between CPUs, so when tracing
 955		 * selected CPUs, sideband for all CPUs is still needed.
 956		 */
 957		if (!!opts->target.cpu_list && record__tracking_system_wide(rec))
 958			system_wide = true;
 959
 960		evsel = evlist__findnew_tracking_event(evlist, system_wide);
 961		if (!evsel)
 962			return -ENOMEM;
 963
 964		/*
 965		 * Enable the tracking event when the process is forked for
 966		 * initial_delay, immediately for system wide.
 967		 */
 968		if (opts->target.initial_delay && !evsel->immediate &&
 969		    !target__has_cpu(&opts->target))
 970			evsel->core.attr.enable_on_exec = 1;
 971		else
 972			evsel->immediate = 1;
 973	}
 974
 975	return 0;
 976}
 977
 978static bool record__kcore_readable(struct machine *machine)
 979{
 980	char kcore[PATH_MAX];
 981	int fd;
 982
 983	scnprintf(kcore, sizeof(kcore), "%s/proc/kcore", machine->root_dir);
 984
 985	fd = open(kcore, O_RDONLY);
 986	if (fd < 0)
 987		return false;
 988
 989	close(fd);
 990
 991	return true;
 992}
 993
 994static int record__kcore_copy(struct machine *machine, struct perf_data *data)
 995{
 996	char from_dir[PATH_MAX];
 997	char kcore_dir[PATH_MAX];
 998	int ret;
 999
1000	snprintf(from_dir, sizeof(from_dir), "%s/proc", machine->root_dir);
1001
1002	ret = perf_data__make_kcore_dir(data, kcore_dir, sizeof(kcore_dir));
1003	if (ret)
1004		return ret;
1005
1006	return kcore_copy(from_dir, kcore_dir);
1007}
1008
1009static void record__thread_data_init_pipes(struct record_thread *thread_data)
1010{
1011	thread_data->pipes.msg[0] = -1;
1012	thread_data->pipes.msg[1] = -1;
1013	thread_data->pipes.ack[0] = -1;
1014	thread_data->pipes.ack[1] = -1;
1015}
1016
1017static int record__thread_data_open_pipes(struct record_thread *thread_data)
1018{
1019	if (pipe(thread_data->pipes.msg))
1020		return -EINVAL;
1021
1022	if (pipe(thread_data->pipes.ack)) {
1023		close(thread_data->pipes.msg[0]);
1024		thread_data->pipes.msg[0] = -1;
1025		close(thread_data->pipes.msg[1]);
1026		thread_data->pipes.msg[1] = -1;
1027		return -EINVAL;
1028	}
1029
1030	pr_debug2("thread_data[%p]: msg=[%d,%d], ack=[%d,%d]\n", thread_data,
1031		 thread_data->pipes.msg[0], thread_data->pipes.msg[1],
1032		 thread_data->pipes.ack[0], thread_data->pipes.ack[1]);
1033
1034	return 0;
1035}
1036
1037static void record__thread_data_close_pipes(struct record_thread *thread_data)
1038{
1039	if (thread_data->pipes.msg[0] != -1) {
1040		close(thread_data->pipes.msg[0]);
1041		thread_data->pipes.msg[0] = -1;
1042	}
1043	if (thread_data->pipes.msg[1] != -1) {
1044		close(thread_data->pipes.msg[1]);
1045		thread_data->pipes.msg[1] = -1;
1046	}
1047	if (thread_data->pipes.ack[0] != -1) {
1048		close(thread_data->pipes.ack[0]);
1049		thread_data->pipes.ack[0] = -1;
1050	}
1051	if (thread_data->pipes.ack[1] != -1) {
1052		close(thread_data->pipes.ack[1]);
1053		thread_data->pipes.ack[1] = -1;
1054	}
1055}
1056
1057static bool evlist__per_thread(struct evlist *evlist)
1058{
1059	return cpu_map__is_dummy(evlist->core.user_requested_cpus);
1060}
1061
1062static int record__thread_data_init_maps(struct record_thread *thread_data, struct evlist *evlist)
1063{
1064	int m, tm, nr_mmaps = evlist->core.nr_mmaps;
1065	struct mmap *mmap = evlist->mmap;
1066	struct mmap *overwrite_mmap = evlist->overwrite_mmap;
1067	struct perf_cpu_map *cpus = evlist->core.all_cpus;
1068	bool per_thread = evlist__per_thread(evlist);
1069
1070	if (per_thread)
1071		thread_data->nr_mmaps = nr_mmaps;
1072	else
1073		thread_data->nr_mmaps = bitmap_weight(thread_data->mask->maps.bits,
1074						      thread_data->mask->maps.nbits);
1075	if (mmap) {
1076		thread_data->maps = zalloc(thread_data->nr_mmaps * sizeof(struct mmap *));
1077		if (!thread_data->maps)
1078			return -ENOMEM;
1079	}
1080	if (overwrite_mmap) {
1081		thread_data->overwrite_maps = zalloc(thread_data->nr_mmaps * sizeof(struct mmap *));
1082		if (!thread_data->overwrite_maps) {
1083			zfree(&thread_data->maps);
1084			return -ENOMEM;
1085		}
1086	}
1087	pr_debug2("thread_data[%p]: nr_mmaps=%d, maps=%p, ow_maps=%p\n", thread_data,
1088		 thread_data->nr_mmaps, thread_data->maps, thread_data->overwrite_maps);
1089
1090	for (m = 0, tm = 0; m < nr_mmaps && tm < thread_data->nr_mmaps; m++) {
1091		if (per_thread ||
1092		    test_bit(perf_cpu_map__cpu(cpus, m).cpu, thread_data->mask->maps.bits)) {
1093			if (thread_data->maps) {
1094				thread_data->maps[tm] = &mmap[m];
1095				pr_debug2("thread_data[%p]: cpu%d: maps[%d] -> mmap[%d]\n",
1096					  thread_data, perf_cpu_map__cpu(cpus, m).cpu, tm, m);
1097			}
1098			if (thread_data->overwrite_maps) {
1099				thread_data->overwrite_maps[tm] = &overwrite_mmap[m];
1100				pr_debug2("thread_data[%p]: cpu%d: ow_maps[%d] -> ow_mmap[%d]\n",
1101					  thread_data, perf_cpu_map__cpu(cpus, m).cpu, tm, m);
1102			}
1103			tm++;
1104		}
1105	}
1106
1107	return 0;
1108}
1109
1110static int record__thread_data_init_pollfd(struct record_thread *thread_data, struct evlist *evlist)
1111{
1112	int f, tm, pos;
1113	struct mmap *map, *overwrite_map;
1114
1115	fdarray__init(&thread_data->pollfd, 64);
1116
1117	for (tm = 0; tm < thread_data->nr_mmaps; tm++) {
1118		map = thread_data->maps ? thread_data->maps[tm] : NULL;
1119		overwrite_map = thread_data->overwrite_maps ?
1120				thread_data->overwrite_maps[tm] : NULL;
1121
1122		for (f = 0; f < evlist->core.pollfd.nr; f++) {
1123			void *ptr = evlist->core.pollfd.priv[f].ptr;
1124
1125			if ((map && ptr == map) || (overwrite_map && ptr == overwrite_map)) {
1126				pos = fdarray__dup_entry_from(&thread_data->pollfd, f,
1127							      &evlist->core.pollfd);
1128				if (pos < 0)
1129					return pos;
1130				pr_debug2("thread_data[%p]: pollfd[%d] <- event_fd=%d\n",
1131					 thread_data, pos, evlist->core.pollfd.entries[f].fd);
1132			}
1133		}
1134	}
1135
1136	return 0;
1137}
1138
1139static void record__free_thread_data(struct record *rec)
1140{
1141	int t;
1142	struct record_thread *thread_data = rec->thread_data;
1143
1144	if (thread_data == NULL)
1145		return;
1146
1147	for (t = 0; t < rec->nr_threads; t++) {
1148		record__thread_data_close_pipes(&thread_data[t]);
1149		zfree(&thread_data[t].maps);
1150		zfree(&thread_data[t].overwrite_maps);
1151		fdarray__exit(&thread_data[t].pollfd);
1152	}
1153
1154	zfree(&rec->thread_data);
1155}
1156
1157static int record__map_thread_evlist_pollfd_indexes(struct record *rec,
1158						    int evlist_pollfd_index,
1159						    int thread_pollfd_index)
1160{
1161	size_t x = rec->index_map_cnt;
1162
1163	if (realloc_array_as_needed(rec->index_map, rec->index_map_sz, x, NULL))
1164		return -ENOMEM;
1165	rec->index_map[x].evlist_pollfd_index = evlist_pollfd_index;
1166	rec->index_map[x].thread_pollfd_index = thread_pollfd_index;
1167	rec->index_map_cnt += 1;
1168	return 0;
1169}
1170
1171static int record__update_evlist_pollfd_from_thread(struct record *rec,
1172						    struct evlist *evlist,
1173						    struct record_thread *thread_data)
1174{
1175	struct pollfd *e_entries = evlist->core.pollfd.entries;
1176	struct pollfd *t_entries = thread_data->pollfd.entries;
1177	int err = 0;
1178	size_t i;
1179
1180	for (i = 0; i < rec->index_map_cnt; i++) {
1181		int e_pos = rec->index_map[i].evlist_pollfd_index;
1182		int t_pos = rec->index_map[i].thread_pollfd_index;
1183
1184		if (e_entries[e_pos].fd != t_entries[t_pos].fd ||
1185		    e_entries[e_pos].events != t_entries[t_pos].events) {
1186			pr_err("Thread and evlist pollfd index mismatch\n");
1187			err = -EINVAL;
1188			continue;
1189		}
1190		e_entries[e_pos].revents = t_entries[t_pos].revents;
1191	}
1192	return err;
1193}
1194
1195static int record__dup_non_perf_events(struct record *rec,
1196				       struct evlist *evlist,
1197				       struct record_thread *thread_data)
1198{
1199	struct fdarray *fda = &evlist->core.pollfd;
1200	int i, ret;
1201
1202	for (i = 0; i < fda->nr; i++) {
1203		if (!(fda->priv[i].flags & fdarray_flag__non_perf_event))
1204			continue;
1205		ret = fdarray__dup_entry_from(&thread_data->pollfd, i, fda);
1206		if (ret < 0) {
1207			pr_err("Failed to duplicate descriptor in main thread pollfd\n");
1208			return ret;
1209		}
1210		pr_debug2("thread_data[%p]: pollfd[%d] <- non_perf_event fd=%d\n",
1211			  thread_data, ret, fda->entries[i].fd);
1212		ret = record__map_thread_evlist_pollfd_indexes(rec, i, ret);
1213		if (ret < 0) {
1214			pr_err("Failed to map thread and evlist pollfd indexes\n");
1215			return ret;
1216		}
1217	}
1218	return 0;
1219}
1220
1221static int record__alloc_thread_data(struct record *rec, struct evlist *evlist)
1222{
1223	int t, ret;
1224	struct record_thread *thread_data;
1225
1226	rec->thread_data = zalloc(rec->nr_threads * sizeof(*(rec->thread_data)));
1227	if (!rec->thread_data) {
1228		pr_err("Failed to allocate thread data\n");
1229		return -ENOMEM;
1230	}
1231	thread_data = rec->thread_data;
1232
1233	for (t = 0; t < rec->nr_threads; t++)
1234		record__thread_data_init_pipes(&thread_data[t]);
1235
1236	for (t = 0; t < rec->nr_threads; t++) {
1237		thread_data[t].rec = rec;
1238		thread_data[t].mask = &rec->thread_masks[t];
1239		ret = record__thread_data_init_maps(&thread_data[t], evlist);
1240		if (ret) {
1241			pr_err("Failed to initialize thread[%d] maps\n", t);
1242			goto out_free;
1243		}
1244		ret = record__thread_data_init_pollfd(&thread_data[t], evlist);
1245		if (ret) {
1246			pr_err("Failed to initialize thread[%d] pollfd\n", t);
1247			goto out_free;
1248		}
1249		if (t) {
1250			thread_data[t].tid = -1;
1251			ret = record__thread_data_open_pipes(&thread_data[t]);
1252			if (ret) {
1253				pr_err("Failed to open thread[%d] communication pipes\n", t);
1254				goto out_free;
1255			}
1256			ret = fdarray__add(&thread_data[t].pollfd, thread_data[t].pipes.msg[0],
1257					   POLLIN | POLLERR | POLLHUP, fdarray_flag__nonfilterable);
1258			if (ret < 0) {
1259				pr_err("Failed to add descriptor to thread[%d] pollfd\n", t);
1260				goto out_free;
1261			}
1262			thread_data[t].ctlfd_pos = ret;
1263			pr_debug2("thread_data[%p]: pollfd[%d] <- ctl_fd=%d\n",
1264				 thread_data, thread_data[t].ctlfd_pos,
1265				 thread_data[t].pipes.msg[0]);
1266		} else {
1267			thread_data[t].tid = gettid();
1268
1269			ret = record__dup_non_perf_events(rec, evlist, &thread_data[t]);
1270			if (ret < 0)
1271				goto out_free;
1272
1273			thread_data[t].ctlfd_pos = -1; /* Not used */
1274		}
1275	}
1276
1277	return 0;
1278
1279out_free:
1280	record__free_thread_data(rec);
1281
1282	return ret;
1283}
1284
1285static int record__mmap_evlist(struct record *rec,
1286			       struct evlist *evlist)
1287{
1288	int i, ret;
1289	struct record_opts *opts = &rec->opts;
1290	bool auxtrace_overwrite = opts->auxtrace_snapshot_mode ||
1291				  opts->auxtrace_sample_mode;
1292	char msg[512];
1293
1294	if (opts->affinity != PERF_AFFINITY_SYS)
1295		cpu__setup_cpunode_map();
1296
1297	if (evlist__mmap_ex(evlist, opts->mmap_pages,
1298				 opts->auxtrace_mmap_pages,
1299				 auxtrace_overwrite,
1300				 opts->nr_cblocks, opts->affinity,
1301				 opts->mmap_flush, opts->comp_level) < 0) {
1302		if (errno == EPERM) {
1303			pr_err("Permission error mapping pages.\n"
1304			       "Consider increasing "
1305			       "/proc/sys/kernel/perf_event_mlock_kb,\n"
1306			       "or try again with a smaller value of -m/--mmap_pages.\n"
1307			       "(current value: %u,%u)\n",
1308			       opts->mmap_pages, opts->auxtrace_mmap_pages);
1309			return -errno;
1310		} else {
1311			pr_err("failed to mmap with %d (%s)\n", errno,
1312				str_error_r(errno, msg, sizeof(msg)));
1313			if (errno)
1314				return -errno;
1315			else
1316				return -EINVAL;
1317		}
1318	}
1319
1320	if (evlist__initialize_ctlfd(evlist, opts->ctl_fd, opts->ctl_fd_ack))
1321		return -1;
1322
1323	ret = record__alloc_thread_data(rec, evlist);
1324	if (ret)
1325		return ret;
1326
1327	if (record__threads_enabled(rec)) {
1328		ret = perf_data__create_dir(&rec->data, evlist->core.nr_mmaps);
1329		if (ret) {
1330			pr_err("Failed to create data directory: %s\n", strerror(-ret));
1331			return ret;
1332		}
1333		for (i = 0; i < evlist->core.nr_mmaps; i++) {
1334			if (evlist->mmap)
1335				evlist->mmap[i].file = &rec->data.dir.files[i];
1336			if (evlist->overwrite_mmap)
1337				evlist->overwrite_mmap[i].file = &rec->data.dir.files[i];
1338		}
1339	}
1340
1341	return 0;
1342}
1343
1344static int record__mmap(struct record *rec)
1345{
1346	return record__mmap_evlist(rec, rec->evlist);
1347}
1348
1349static int record__open(struct record *rec)
1350{
1351	char msg[BUFSIZ];
1352	struct evsel *pos;
1353	struct evlist *evlist = rec->evlist;
1354	struct perf_session *session = rec->session;
1355	struct record_opts *opts = &rec->opts;
1356	int rc = 0;
1357
1358	evlist__for_each_entry(evlist, pos) {
1359try_again:
1360		if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
1361			if (evsel__fallback(pos, &opts->target, errno, msg, sizeof(msg))) {
1362				if (verbose > 0)
1363					ui__warning("%s\n", msg);
1364				goto try_again;
1365			}
1366			if ((errno == EINVAL || errno == EBADF) &&
1367			    pos->core.leader != &pos->core &&
1368			    pos->weak_group) {
1369			        pos = evlist__reset_weak_group(evlist, pos, true);
1370				goto try_again;
1371			}
1372			rc = -errno;
1373			evsel__open_strerror(pos, &opts->target, errno, msg, sizeof(msg));
1374			ui__error("%s\n", msg);
1375			goto out;
1376		}
1377
1378		pos->supported = true;
1379	}
1380
1381	if (symbol_conf.kptr_restrict && !evlist__exclude_kernel(evlist)) {
1382		pr_warning(
1383"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1384"check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1385"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1386"file is not found in the buildid cache or in the vmlinux path.\n\n"
1387"Samples in kernel modules won't be resolved at all.\n\n"
1388"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1389"even with a suitable vmlinux or kallsyms file.\n\n");
1390	}
1391
1392	if (evlist__apply_filters(evlist, &pos)) {
1393		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
1394			pos->filter ?: "BPF", evsel__name(pos), errno,
1395			str_error_r(errno, msg, sizeof(msg)));
1396		rc = -1;
1397		goto out;
1398	}
1399
1400	rc = record__mmap(rec);
1401	if (rc)
1402		goto out;
1403
1404	session->evlist = evlist;
1405	perf_session__set_id_hdr_size(session);
1406out:
1407	return rc;
1408}
1409
1410static void set_timestamp_boundary(struct record *rec, u64 sample_time)
1411{
1412	if (rec->evlist->first_sample_time == 0)
1413		rec->evlist->first_sample_time = sample_time;
1414
1415	if (sample_time)
1416		rec->evlist->last_sample_time = sample_time;
1417}
1418
1419static int process_sample_event(struct perf_tool *tool,
1420				union perf_event *event,
1421				struct perf_sample *sample,
1422				struct evsel *evsel,
1423				struct machine *machine)
1424{
1425	struct record *rec = container_of(tool, struct record, tool);
1426
1427	set_timestamp_boundary(rec, sample->time);
1428
1429	if (rec->buildid_all)
1430		return 0;
1431
1432	rec->samples++;
1433	return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
1434}
1435
1436static int process_buildids(struct record *rec)
1437{
1438	struct perf_session *session = rec->session;
1439
1440	if (perf_data__size(&rec->data) == 0)
1441		return 0;
1442
1443	/*
1444	 * During this process, it'll load kernel map and replace the
1445	 * dso->long_name to a real pathname it found.  In this case
1446	 * we prefer the vmlinux path like
1447	 *   /lib/modules/3.16.4/build/vmlinux
1448	 *
1449	 * rather than build-id path (in debug directory).
1450	 *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
1451	 */
1452	symbol_conf.ignore_vmlinux_buildid = true;
1453
1454	/*
1455	 * If --buildid-all is given, it marks all DSO regardless of hits,
1456	 * so no need to process samples. But if timestamp_boundary is enabled,
1457	 * it still needs to walk on all samples to get the timestamps of
1458	 * first/last samples.
1459	 */
1460	if (rec->buildid_all && !rec->timestamp_boundary)
1461		rec->tool.sample = NULL;
1462
1463	return perf_session__process_events(session);
1464}
1465
1466static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
1467{
1468	int err;
1469	struct perf_tool *tool = data;
1470	/*
1471	 *As for guest kernel when processing subcommand record&report,
1472	 *we arrange module mmap prior to guest kernel mmap and trigger
1473	 *a preload dso because default guest module symbols are loaded
1474	 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
1475	 *method is used to avoid symbol missing when the first addr is
1476	 *in module instead of in guest kernel.
1477	 */
1478	err = perf_event__synthesize_modules(tool, process_synthesized_event,
1479					     machine);
1480	if (err < 0)
1481		pr_err("Couldn't record guest kernel [%d]'s reference"
1482		       " relocation symbol.\n", machine->pid);
1483
1484	/*
1485	 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
1486	 * have no _text sometimes.
1487	 */
1488	err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1489						 machine);
1490	if (err < 0)
1491		pr_err("Couldn't record guest kernel [%d]'s reference"
1492		       " relocation symbol.\n", machine->pid);
1493}
1494
1495static struct perf_event_header finished_round_event = {
1496	.size = sizeof(struct perf_event_header),
1497	.type = PERF_RECORD_FINISHED_ROUND,
1498};
1499
1500static struct perf_event_header finished_init_event = {
1501	.size = sizeof(struct perf_event_header),
1502	.type = PERF_RECORD_FINISHED_INIT,
1503};
1504
1505static void record__adjust_affinity(struct record *rec, struct mmap *map)
1506{
1507	if (rec->opts.affinity != PERF_AFFINITY_SYS &&
1508	    !bitmap_equal(thread->mask->affinity.bits, map->affinity_mask.bits,
1509			  thread->mask->affinity.nbits)) {
1510		bitmap_zero(thread->mask->affinity.bits, thread->mask->affinity.nbits);
1511		bitmap_or(thread->mask->affinity.bits, thread->mask->affinity.bits,
1512			  map->affinity_mask.bits, thread->mask->affinity.nbits);
1513		sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&thread->mask->affinity),
1514					(cpu_set_t *)thread->mask->affinity.bits);
1515		if (verbose == 2) {
1516			pr_debug("threads[%d]: running on cpu%d: ", thread->tid, sched_getcpu());
1517			mmap_cpu_mask__scnprintf(&thread->mask->affinity, "affinity");
1518		}
1519	}
1520}
1521
1522static size_t process_comp_header(void *record, size_t increment)
1523{
1524	struct perf_record_compressed *event = record;
1525	size_t size = sizeof(*event);
1526
1527	if (increment) {
1528		event->header.size += increment;
1529		return increment;
1530	}
1531
1532	event->header.type = PERF_RECORD_COMPRESSED;
1533	event->header.size = size;
1534
1535	return size;
1536}
1537
1538static ssize_t zstd_compress(struct perf_session *session, struct mmap *map,
1539			    void *dst, size_t dst_size, void *src, size_t src_size)
1540{
1541	ssize_t compressed;
1542	size_t max_record_size = PERF_SAMPLE_MAX_SIZE - sizeof(struct perf_record_compressed) - 1;
1543	struct zstd_data *zstd_data = &session->zstd_data;
1544
1545	if (map && map->file)
1546		zstd_data = &map->zstd_data;
1547
1548	compressed = zstd_compress_stream_to_records(zstd_data, dst, dst_size, src, src_size,
1549						     max_record_size, process_comp_header);
1550	if (compressed < 0)
1551		return compressed;
1552
1553	if (map && map->file) {
1554		thread->bytes_transferred += src_size;
1555		thread->bytes_compressed  += compressed;
1556	} else {
1557		session->bytes_transferred += src_size;
1558		session->bytes_compressed  += compressed;
1559	}
1560
1561	return compressed;
1562}
1563
1564static int record__mmap_read_evlist(struct record *rec, struct evlist *evlist,
1565				    bool overwrite, bool synch)
1566{
1567	u64 bytes_written = rec->bytes_written;
1568	int i;
1569	int rc = 0;
1570	int nr_mmaps;
1571	struct mmap **maps;
1572	int trace_fd = rec->data.file.fd;
1573	off_t off = 0;
1574
1575	if (!evlist)
1576		return 0;
1577
1578	nr_mmaps = thread->nr_mmaps;
1579	maps = overwrite ? thread->overwrite_maps : thread->maps;
1580
1581	if (!maps)
1582		return 0;
1583
1584	if (overwrite && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
1585		return 0;
1586
1587	if (record__aio_enabled(rec))
1588		off = record__aio_get_pos(trace_fd);
1589
1590	for (i = 0; i < nr_mmaps; i++) {
1591		u64 flush = 0;
1592		struct mmap *map = maps[i];
1593
1594		if (map->core.base) {
1595			record__adjust_affinity(rec, map);
1596			if (synch) {
1597				flush = map->core.flush;
1598				map->core.flush = 1;
1599			}
1600			if (!record__aio_enabled(rec)) {
1601				if (perf_mmap__push(map, rec, record__pushfn) < 0) {
1602					if (synch)
1603						map->core.flush = flush;
1604					rc = -1;
1605					goto out;
1606				}
1607			} else {
1608				if (record__aio_push(rec, map, &off) < 0) {
1609					record__aio_set_pos(trace_fd, off);
1610					if (synch)
1611						map->core.flush = flush;
1612					rc = -1;
1613					goto out;
1614				}
1615			}
1616			if (synch)
1617				map->core.flush = flush;
1618		}
1619
1620		if (map->auxtrace_mmap.base && !rec->opts.auxtrace_snapshot_mode &&
1621		    !rec->opts.auxtrace_sample_mode &&
1622		    record__auxtrace_mmap_read(rec, map) != 0) {
1623			rc = -1;
1624			goto out;
1625		}
1626	}
1627
1628	if (record__aio_enabled(rec))
1629		record__aio_set_pos(trace_fd, off);
1630
1631	/*
1632	 * Mark the round finished in case we wrote
1633	 * at least one event.
1634	 *
1635	 * No need for round events in directory mode,
1636	 * because per-cpu maps and files have data
1637	 * sorted by kernel.
1638	 */
1639	if (!record__threads_enabled(rec) && bytes_written != rec->bytes_written)
1640		rc = record__write(rec, NULL, &finished_round_event, sizeof(finished_round_event));
1641
1642	if (overwrite)
1643		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
1644out:
1645	return rc;
1646}
1647
1648static int record__mmap_read_all(struct record *rec, bool synch)
1649{
1650	int err;
1651
1652	err = record__mmap_read_evlist(rec, rec->evlist, false, synch);
1653	if (err)
1654		return err;
1655
1656	return record__mmap_read_evlist(rec, rec->evlist, true, synch);
1657}
1658
1659static void record__thread_munmap_filtered(struct fdarray *fda, int fd,
1660					   void *arg __maybe_unused)
1661{
1662	struct perf_mmap *map = fda->priv[fd].ptr;
1663
1664	if (map)
1665		perf_mmap__put(map);
1666}
1667
1668static void *record__thread(void *arg)
1669{
1670	enum thread_msg msg = THREAD_MSG__READY;
1671	bool terminate = false;
1672	struct fdarray *pollfd;
1673	int err, ctlfd_pos;
1674
1675	thread = arg;
1676	thread->tid = gettid();
1677
1678	err = write(thread->pipes.ack[1], &msg, sizeof(msg));
1679	if (err == -1)
1680		pr_warning("threads[%d]: failed to notify on start: %s\n",
1681			   thread->tid, strerror(errno));
1682
1683	pr_debug("threads[%d]: started on cpu%d\n", thread->tid, sched_getcpu());
1684
1685	pollfd = &thread->pollfd;
1686	ctlfd_pos = thread->ctlfd_pos;
1687
1688	for (;;) {
1689		unsigned long long hits = thread->samples;
1690
1691		if (record__mmap_read_all(thread->rec, false) < 0 || terminate)
1692			break;
1693
1694		if (hits == thread->samples) {
1695
1696			err = fdarray__poll(pollfd, -1);
1697			/*
1698			 * Propagate error, only if there's any. Ignore positive
1699			 * number of returned events and interrupt error.
1700			 */
1701			if (err > 0 || (err < 0 && errno == EINTR))
1702				err = 0;
1703			thread->waking++;
1704
1705			if (fdarray__filter(pollfd, POLLERR | POLLHUP,
1706					    record__thread_munmap_filtered, NULL) == 0)
1707				break;
1708		}
1709
1710		if (pollfd->entries[ctlfd_pos].revents & POLLHUP) {
1711			terminate = true;
1712			close(thread->pipes.msg[0]);
1713			thread->pipes.msg[0] = -1;
1714			pollfd->entries[ctlfd_pos].fd = -1;
1715			pollfd->entries[ctlfd_pos].events = 0;
1716		}
1717
1718		pollfd->entries[ctlfd_pos].revents = 0;
1719	}
1720	record__mmap_read_all(thread->rec, true);
1721
1722	err = write(thread->pipes.ack[1], &msg, sizeof(msg));
1723	if (err == -1)
1724		pr_warning("threads[%d]: failed to notify on termination: %s\n",
1725			   thread->tid, strerror(errno));
1726
1727	return NULL;
1728}
1729
1730static void record__init_features(struct record *rec)
1731{
1732	struct perf_session *session = rec->session;
1733	int feat;
1734
1735	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1736		perf_header__set_feat(&session->header, feat);
1737
1738	if (rec->no_buildid)
1739		perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1740
1741#ifdef HAVE_LIBTRACEEVENT
1742	if (!have_tracepoints(&rec->evlist->core.entries))
1743		perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1744#endif
1745
1746	if (!rec->opts.branch_stack)
1747		perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1748
1749	if (!rec->opts.full_auxtrace)
1750		perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1751
1752	if (!(rec->opts.use_clockid && rec->opts.clockid_res_ns))
1753		perf_header__clear_feat(&session->header, HEADER_CLOCKID);
1754
1755	if (!rec->opts.use_clockid)
1756		perf_header__clear_feat(&session->header, HEADER_CLOCK_DATA);
1757
1758	if (!record__threads_enabled(rec))
1759		perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1760
1761	if (!record__comp_enabled(rec))
1762		perf_header__clear_feat(&session->header, HEADER_COMPRESSED);
1763
1764	perf_header__clear_feat(&session->header, HEADER_STAT);
1765}
1766
1767static void
1768record__finish_output(struct record *rec)
1769{
1770	int i;
1771	struct perf_data *data = &rec->data;
1772	int fd = perf_data__fd(data);
1773
1774	if (data->is_pipe) {
1775		/* Just to display approx. size */
1776		data->file.size = rec->bytes_written;
1777		return;
1778	}
1779
1780	rec->session->header.data_size += rec->bytes_written;
1781	data->file.size = lseek(perf_data__fd(data), 0, SEEK_CUR);
1782	if (record__threads_enabled(rec)) {
1783		for (i = 0; i < data->dir.nr; i++)
1784			data->dir.files[i].size = lseek(data->dir.files[i].fd, 0, SEEK_CUR);
1785	}
1786
1787	if (!rec->no_buildid) {
1788		process_buildids(rec);
1789
1790		if (rec->buildid_all)
1791			dsos__hit_all(rec->session);
1792	}
1793	perf_session__write_header(rec->session, rec->evlist, fd, true);
1794
1795	return;
1796}
1797
1798static int record__synthesize_workload(struct record *rec, bool tail)
1799{
1800	int err;
1801	struct perf_thread_map *thread_map;
1802	bool needs_mmap = rec->opts.synth & PERF_SYNTH_MMAP;
1803
1804	if (rec->opts.tail_synthesize != tail)
1805		return 0;
1806
1807	thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
1808	if (thread_map == NULL)
1809		return -1;
1810
1811	err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
1812						 process_synthesized_event,
1813						 &rec->session->machines.host,
1814						 needs_mmap,
1815						 rec->opts.sample_address);
1816	perf_thread_map__put(thread_map);
1817	return err;
1818}
1819
1820static int write_finished_init(struct record *rec, bool tail)
1821{
1822	if (rec->opts.tail_synthesize != tail)
1823		return 0;
1824
1825	return record__write(rec, NULL, &finished_init_event, sizeof(finished_init_event));
1826}
1827
1828static int record__synthesize(struct record *rec, bool tail);
1829
1830static int
1831record__switch_output(struct record *rec, bool at_exit)
1832{
1833	struct perf_data *data = &rec->data;
1834	char *new_filename = NULL;
1835	int fd, err;
1836
1837	/* Same Size:      "2015122520103046"*/
1838	char timestamp[] = "InvalidTimestamp";
1839
1840	record__aio_mmap_read_sync(rec);
1841
1842	write_finished_init(rec, true);
1843
1844	record__synthesize(rec, true);
1845	if (target__none(&rec->opts.target))
1846		record__synthesize_workload(rec, true);
1847
1848	rec->samples = 0;
1849	record__finish_output(rec);
1850	err = fetch_current_timestamp(timestamp, sizeof(timestamp));
1851	if (err) {
1852		pr_err("Failed to get current timestamp\n");
1853		return -EINVAL;
1854	}
1855
1856	fd = perf_data__switch(data, timestamp,
1857			       rec->session->header.data_offset,
1858			       at_exit, &new_filename);
1859	if (fd >= 0 && !at_exit) {
1860		rec->bytes_written = 0;
1861		rec->session->header.data_size = 0;
1862	}
1863
1864	if (!quiet) {
1865		fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
1866			data->path, timestamp);
1867	}
1868
1869	if (rec->switch_output.num_files) {
1870		int n = rec->switch_output.cur_file + 1;
1871
1872		if (n >= rec->switch_output.num_files)
1873			n = 0;
1874		rec->switch_output.cur_file = n;
1875		if (rec->switch_output.filenames[n]) {
1876			remove(rec->switch_output.filenames[n]);
1877			zfree(&rec->switch_output.filenames[n]);
1878		}
1879		rec->switch_output.filenames[n] = new_filename;
1880	} else {
1881		free(new_filename);
1882	}
1883
1884	/* Output tracking events */
1885	if (!at_exit) {
1886		record__synthesize(rec, false);
1887
1888		/*
1889		 * In 'perf record --switch-output' without -a,
1890		 * record__synthesize() in record__switch_output() won't
1891		 * generate tracking events because there's no thread_map
1892		 * in evlist. Which causes newly created perf.data doesn't
1893		 * contain map and comm information.
1894		 * Create a fake thread_map and directly call
1895		 * perf_event__synthesize_thread_map() for those events.
1896		 */
1897		if (target__none(&rec->opts.target))
1898			record__synthesize_workload(rec, false);
1899		write_finished_init(rec, false);
1900	}
1901	return fd;
1902}
1903
1904static void __record__save_lost_samples(struct record *rec, struct evsel *evsel,
1905					struct perf_record_lost_samples *lost,
1906					int cpu_idx, int thread_idx, u64 lost_count,
1907					u16 misc_flag)
1908{
1909	struct perf_sample_id *sid;
1910	struct perf_sample sample = {};
1911	int id_hdr_size;
1912
1913	lost->lost = lost_count;
1914	if (evsel->core.ids) {
1915		sid = xyarray__entry(evsel->core.sample_id, cpu_idx, thread_idx);
1916		sample.id = sid->id;
1917	}
1918
1919	id_hdr_size = perf_event__synthesize_id_sample((void *)(lost + 1),
1920						       evsel->core.attr.sample_type, &sample);
1921	lost->header.size = sizeof(*lost) + id_hdr_size;
1922	lost->header.misc = misc_flag;
1923	record__write(rec, NULL, lost, lost->header.size);
1924}
1925
1926static void record__read_lost_samples(struct record *rec)
1927{
1928	struct perf_session *session = rec->session;
1929	struct perf_record_lost_samples *lost = NULL;
1930	struct evsel *evsel;
1931
1932	/* there was an error during record__open */
1933	if (session->evlist == NULL)
1934		return;
1935
1936	evlist__for_each_entry(session->evlist, evsel) {
1937		struct xyarray *xy = evsel->core.sample_id;
1938		u64 lost_count;
1939
1940		if (xy == NULL || evsel->core.fd == NULL)
1941			continue;
1942		if (xyarray__max_x(evsel->core.fd) != xyarray__max_x(xy) ||
1943		    xyarray__max_y(evsel->core.fd) != xyarray__max_y(xy)) {
1944			pr_debug("Unmatched FD vs. sample ID: skip reading LOST count\n");
1945			continue;
1946		}
1947
1948		for (int x = 0; x < xyarray__max_x(xy); x++) {
1949			for (int y = 0; y < xyarray__max_y(xy); y++) {
1950				struct perf_counts_values count;
1951
1952				if (perf_evsel__read(&evsel->core, x, y, &count) < 0) {
1953					pr_debug("read LOST count failed\n");
1954					goto out;
1955				}
1956
1957				if (count.lost) {
1958					if (!lost) {
1959						lost = zalloc(sizeof(*lost) +
1960							      session->machines.host.id_hdr_size);
1961						if (!lost) {
1962							pr_debug("Memory allocation failed\n");
1963							return;
1964						}
1965						lost->header.type = PERF_RECORD_LOST_SAMPLES;
1966					}
1967					__record__save_lost_samples(rec, evsel, lost,
1968								    x, y, count.lost, 0);
1969				}
1970			}
1971		}
1972
1973		lost_count = perf_bpf_filter__lost_count(evsel);
1974		if (lost_count) {
1975			if (!lost) {
1976				lost = zalloc(sizeof(*lost) +
1977					      session->machines.host.id_hdr_size);
1978				if (!lost) {
1979					pr_debug("Memory allocation failed\n");
1980					return;
1981				}
1982				lost->header.type = PERF_RECORD_LOST_SAMPLES;
1983			}
1984			__record__save_lost_samples(rec, evsel, lost, 0, 0, lost_count,
1985						    PERF_RECORD_MISC_LOST_SAMPLES_BPF);
1986		}
1987	}
1988out:
1989	free(lost);
1990}
1991
1992static volatile sig_atomic_t workload_exec_errno;
1993
1994/*
1995 * evlist__prepare_workload will send a SIGUSR1
1996 * if the fork fails, since we asked by setting its
1997 * want_signal to true.
1998 */
1999static void workload_exec_failed_signal(int signo __maybe_unused,
2000					siginfo_t *info,
2001					void *ucontext __maybe_unused)
2002{
2003	workload_exec_errno = info->si_value.sival_int;
2004	done = 1;
2005	child_finished = 1;
2006}
2007
2008static void snapshot_sig_handler(int sig);
2009static void alarm_sig_handler(int sig);
2010
2011static const struct perf_event_mmap_page *evlist__pick_pc(struct evlist *evlist)
2012{
2013	if (evlist) {
2014		if (evlist->mmap && evlist->mmap[0].core.base)
2015			return evlist->mmap[0].core.base;
2016		if (evlist->overwrite_mmap && evlist->overwrite_mmap[0].core.base)
2017			return evlist->overwrite_mmap[0].core.base;
2018	}
2019	return NULL;
2020}
2021
2022static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
2023{
2024	const struct perf_event_mmap_page *pc = evlist__pick_pc(rec->evlist);
2025	if (pc)
2026		return pc;
2027	return NULL;
2028}
2029
2030static int record__synthesize(struct record *rec, bool tail)
2031{
2032	struct perf_session *session = rec->session;
2033	struct machine *machine = &session->machines.host;
2034	struct perf_data *data = &rec->data;
2035	struct record_opts *opts = &rec->opts;
2036	struct perf_tool *tool = &rec->tool;
2037	int err = 0;
2038	event_op f = process_synthesized_event;
2039
2040	if (rec->opts.tail_synthesize != tail)
2041		return 0;
2042
2043	if (data->is_pipe) {
2044		err = perf_event__synthesize_for_pipe(tool, session, data,
2045						      process_synthesized_event);
2046		if (err < 0)
2047			goto out;
2048
2049		rec->bytes_written += err;
2050	}
2051
2052	err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
2053					  process_synthesized_event, machine);
2054	if (err)
2055		goto out;
2056
2057	/* Synthesize id_index before auxtrace_info */
2058	err = perf_event__synthesize_id_index(tool,
2059					      process_synthesized_event,
2060					      session->evlist, machine);
2061	if (err)
2062		goto out;
2063
2064	if (rec->opts.full_auxtrace) {
2065		err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
2066					session, process_synthesized_event);
2067		if (err)
2068			goto out;
2069	}
2070
2071	if (!evlist__exclude_kernel(rec->evlist)) {
2072		err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
2073							 machine);
2074		WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
2075				   "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
2076				   "Check /proc/kallsyms permission or run as root.\n");
2077
2078		err = perf_event__synthesize_modules(tool, process_synthesized_event,
2079						     machine);
2080		WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
2081				   "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
2082				   "Check /proc/modules permission or run as root.\n");
2083	}
2084
2085	if (perf_guest) {
2086		machines__process_guests(&session->machines,
2087					 perf_event__synthesize_guest_os, tool);
2088	}
2089
2090	err = perf_event__synthesize_extra_attr(&rec->tool,
2091						rec->evlist,
2092						process_synthesized_event,
2093						data->is_pipe);
2094	if (err)
2095		goto out;
2096
2097	err = perf_event__synthesize_thread_map2(&rec->tool, rec->evlist->core.threads,
2098						 process_synthesized_event,
2099						NULL);
2100	if (err < 0) {
2101		pr_err("Couldn't synthesize thread map.\n");
2102		return err;
2103	}
2104
2105	err = perf_event__synthesize_cpu_map(&rec->tool, rec->evlist->core.all_cpus,
2106					     process_synthesized_event, NULL);
2107	if (err < 0) {
2108		pr_err("Couldn't synthesize cpu map.\n");
2109		return err;
2110	}
2111
2112	err = perf_event__synthesize_bpf_events(session, process_synthesized_event,
2113						machine, opts);
2114	if (err < 0) {
2115		pr_warning("Couldn't synthesize bpf events.\n");
2116		err = 0;
2117	}
2118
2119	if (rec->opts.synth & PERF_SYNTH_CGROUP) {
2120		err = perf_event__synthesize_cgroups(tool, process_synthesized_event,
2121						     machine);
2122		if (err < 0) {
2123			pr_warning("Couldn't synthesize cgroup events.\n");
2124			err = 0;
2125		}
2126	}
2127
2128	if (rec->opts.nr_threads_synthesize > 1) {
2129		mutex_init(&synth_lock);
2130		perf_set_multithreaded();
2131		f = process_locked_synthesized_event;
2132	}
2133
2134	if (rec->opts.synth & PERF_SYNTH_TASK) {
2135		bool needs_mmap = rec->opts.synth & PERF_SYNTH_MMAP;
2136
2137		err = __machine__synthesize_threads(machine, tool, &opts->target,
2138						    rec->evlist->core.threads,
2139						    f, needs_mmap, opts->sample_address,
2140						    rec->opts.nr_threads_synthesize);
2141	}
2142
2143	if (rec->opts.nr_threads_synthesize > 1) {
2144		perf_set_singlethreaded();
2145		mutex_destroy(&synth_lock);
2146	}
2147
2148out:
2149	return err;
2150}
2151
2152static int record__process_signal_event(union perf_event *event __maybe_unused, void *data)
2153{
2154	struct record *rec = data;
2155	pthread_kill(rec->thread_id, SIGUSR2);
2156	return 0;
2157}
2158
2159static int record__setup_sb_evlist(struct record *rec)
2160{
2161	struct record_opts *opts = &rec->opts;
2162
2163	if (rec->sb_evlist != NULL) {
2164		/*
2165		 * We get here if --switch-output-event populated the
2166		 * sb_evlist, so associate a callback that will send a SIGUSR2
2167		 * to the main thread.
2168		 */
2169		evlist__set_cb(rec->sb_evlist, record__process_signal_event, rec);
2170		rec->thread_id = pthread_self();
2171	}
2172#ifdef HAVE_LIBBPF_SUPPORT
2173	if (!opts->no_bpf_event) {
2174		if (rec->sb_evlist == NULL) {
2175			rec->sb_evlist = evlist__new();
2176
2177			if (rec->sb_evlist == NULL) {
2178				pr_err("Couldn't create side band evlist.\n.");
2179				return -1;
2180			}
2181		}
2182
2183		if (evlist__add_bpf_sb_event(rec->sb_evlist, &rec->session->header.env)) {
2184			pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
2185			return -1;
2186		}
2187	}
2188#endif
2189	if (evlist__start_sb_thread(rec->sb_evlist, &rec->opts.target)) {
2190		pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
2191		opts->no_bpf_event = true;
2192	}
2193
2194	return 0;
2195}
2196
2197static int record__init_clock(struct record *rec)
2198{
2199	struct perf_session *session = rec->session;
2200	struct timespec ref_clockid;
2201	struct timeval ref_tod;
2202	u64 ref;
2203
2204	if (!rec->opts.use_clockid)
2205		return 0;
2206
2207	if (rec->opts.use_clockid && rec->opts.clockid_res_ns)
2208		session->header.env.clock.clockid_res_ns = rec->opts.clockid_res_ns;
2209
2210	session->header.env.clock.clockid = rec->opts.clockid;
2211
2212	if (gettimeofday(&ref_tod, NULL) != 0) {
2213		pr_err("gettimeofday failed, cannot set reference time.\n");
2214		return -1;
2215	}
2216
2217	if (clock_gettime(rec->opts.clockid, &ref_clockid)) {
2218		pr_err("clock_gettime failed, cannot set reference time.\n");
2219		return -1;
2220	}
2221
2222	ref = (u64) ref_tod.tv_sec * NSEC_PER_SEC +
2223	      (u64) ref_tod.tv_usec * NSEC_PER_USEC;
2224
2225	session->header.env.clock.tod_ns = ref;
2226
2227	ref = (u64) ref_clockid.tv_sec * NSEC_PER_SEC +
2228	      (u64) ref_clockid.tv_nsec;
2229
2230	session->header.env.clock.clockid_ns = ref;
2231	return 0;
2232}
2233
2234static void hit_auxtrace_snapshot_trigger(struct record *rec)
2235{
2236	if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
2237		trigger_hit(&auxtrace_snapshot_trigger);
2238		auxtrace_record__snapshot_started = 1;
2239		if (auxtrace_record__snapshot_start(rec->itr))
2240			trigger_error(&auxtrace_snapshot_trigger);
2241	}
2242}
2243
2244static int record__terminate_thread(struct record_thread *thread_data)
2245{
2246	int err;
2247	enum thread_msg ack = THREAD_MSG__UNDEFINED;
2248	pid_t tid = thread_data->tid;
2249
2250	close(thread_data->pipes.msg[1]);
2251	thread_data->pipes.msg[1] = -1;
2252	err = read(thread_data->pipes.ack[0], &ack, sizeof(ack));
2253	if (err > 0)
2254		pr_debug2("threads[%d]: sent %s\n", tid, thread_msg_tags[ack]);
2255	else
2256		pr_warning("threads[%d]: failed to receive termination notification from %d\n",
2257			   thread->tid, tid);
2258
2259	return 0;
2260}
2261
2262static int record__start_threads(struct record *rec)
2263{
2264	int t, tt, err, ret = 0, nr_threads = rec->nr_threads;
2265	struct record_thread *thread_data = rec->thread_data;
2266	sigset_t full, mask;
2267	pthread_t handle;
2268	pthread_attr_t attrs;
2269
2270	thread = &thread_data[0];
2271
2272	if (!record__threads_enabled(rec))
2273		return 0;
2274
2275	sigfillset(&full);
2276	if (sigprocmask(SIG_SETMASK, &full, &mask)) {
2277		pr_err("Failed to block signals on threads start: %s\n", strerror(errno));
2278		return -1;
2279	}
2280
2281	pthread_attr_init(&attrs);
2282	pthread_attr_setdetachstate(&attrs, PTHREAD_CREATE_DETACHED);
2283
2284	for (t = 1; t < nr_threads; t++) {
2285		enum thread_msg msg = THREAD_MSG__UNDEFINED;
2286
2287#ifdef HAVE_PTHREAD_ATTR_SETAFFINITY_NP
2288		pthread_attr_setaffinity_np(&attrs,
2289					    MMAP_CPU_MASK_BYTES(&(thread_data[t].mask->affinity)),
2290					    (cpu_set_t *)(thread_data[t].mask->affinity.bits));
2291#endif
2292		if (pthread_create(&handle, &attrs, record__thread, &thread_data[t])) {
2293			for (tt = 1; tt < t; tt++)
2294				record__terminate_thread(&thread_data[t]);
2295			pr_err("Failed to start threads: %s\n", strerror(errno));
2296			ret = -1;
2297			goto out_err;
2298		}
2299
2300		err = read(thread_data[t].pipes.ack[0], &msg, sizeof(msg));
2301		if (err > 0)
2302			pr_debug2("threads[%d]: sent %s\n", rec->thread_data[t].tid,
2303				  thread_msg_tags[msg]);
2304		else
2305			pr_warning("threads[%d]: failed to receive start notification from %d\n",
2306				   thread->tid, rec->thread_data[t].tid);
2307	}
2308
2309	sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&thread->mask->affinity),
2310			(cpu_set_t *)thread->mask->affinity.bits);
2311
2312	pr_debug("threads[%d]: started on cpu%d\n", thread->tid, sched_getcpu());
2313
2314out_err:
2315	pthread_attr_destroy(&attrs);
2316
2317	if (sigprocmask(SIG_SETMASK, &mask, NULL)) {
2318		pr_err("Failed to unblock signals on threads start: %s\n", strerror(errno));
2319		ret = -1;
2320	}
2321
2322	return ret;
2323}
2324
2325static int record__stop_threads(struct record *rec)
2326{
2327	int t;
2328	struct record_thread *thread_data = rec->thread_data;
2329
2330	for (t = 1; t < rec->nr_threads; t++)
2331		record__terminate_thread(&thread_data[t]);
2332
2333	for (t = 0; t < rec->nr_threads; t++) {
2334		rec->samples += thread_data[t].samples;
2335		if (!record__threads_enabled(rec))
2336			continue;
2337		rec->session->bytes_transferred += thread_data[t].bytes_transferred;
2338		rec->session->bytes_compressed += thread_data[t].bytes_compressed;
2339		pr_debug("threads[%d]: samples=%lld, wakes=%ld, ", thread_data[t].tid,
2340			 thread_data[t].samples, thread_data[t].waking);
2341		if (thread_data[t].bytes_transferred && thread_data[t].bytes_compressed)
2342			pr_debug("transferred=%" PRIu64 ", compressed=%" PRIu64 "\n",
2343				 thread_data[t].bytes_transferred, thread_data[t].bytes_compressed);
2344		else
2345			pr_debug("written=%" PRIu64 "\n", thread_data[t].bytes_written);
2346	}
2347
2348	return 0;
2349}
2350
2351static unsigned long record__waking(struct record *rec)
2352{
2353	int t;
2354	unsigned long waking = 0;
2355	struct record_thread *thread_data = rec->thread_data;
2356
2357	for (t = 0; t < rec->nr_threads; t++)
2358		waking += thread_data[t].waking;
2359
2360	return waking;
2361}
2362
2363static int __cmd_record(struct record *rec, int argc, const char **argv)
2364{
2365	int err;
2366	int status = 0;
2367	const bool forks = argc > 0;
2368	struct perf_tool *tool = &rec->tool;
2369	struct record_opts *opts = &rec->opts;
2370	struct perf_data *data = &rec->data;
2371	struct perf_session *session;
2372	bool disabled = false, draining = false;
2373	int fd;
2374	float ratio = 0;
2375	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
2376
2377	atexit(record__sig_exit);
2378	signal(SIGCHLD, sig_handler);
2379	signal(SIGINT, sig_handler);
2380	signal(SIGTERM, sig_handler);
2381	signal(SIGSEGV, sigsegv_handler);
2382
2383	if (rec->opts.record_namespaces)
2384		tool->namespace_events = true;
2385
2386	if (rec->opts.record_cgroup) {
2387#ifdef HAVE_FILE_HANDLE
2388		tool->cgroup_events = true;
2389#else
2390		pr_err("cgroup tracking is not supported\n");
2391		return -1;
2392#endif
2393	}
2394
2395	if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
2396		signal(SIGUSR2, snapshot_sig_handler);
2397		if (rec->opts.auxtrace_snapshot_mode)
2398			trigger_on(&auxtrace_snapshot_trigger);
2399		if (rec->switch_output.enabled)
2400			trigger_on(&switch_output_trigger);
2401	} else {
2402		signal(SIGUSR2, SIG_IGN);
2403	}
2404
 
 
 
 
 
 
 
 
 
 
 
 
2405	session = perf_session__new(data, tool);
2406	if (IS_ERR(session)) {
2407		pr_err("Perf session creation failed.\n");
2408		return PTR_ERR(session);
2409	}
2410
2411	if (record__threads_enabled(rec)) {
2412		if (perf_data__is_pipe(&rec->data)) {
2413			pr_err("Parallel trace streaming is not available in pipe mode.\n");
2414			return -1;
2415		}
2416		if (rec->opts.full_auxtrace) {
2417			pr_err("Parallel trace streaming is not available in AUX area tracing mode.\n");
2418			return -1;
2419		}
2420	}
2421
2422	fd = perf_data__fd(data);
2423	rec->session = session;
2424
2425	if (zstd_init(&session->zstd_data, rec->opts.comp_level) < 0) {
2426		pr_err("Compression initialization failed.\n");
2427		return -1;
2428	}
2429#ifdef HAVE_EVENTFD_SUPPORT
2430	done_fd = eventfd(0, EFD_NONBLOCK);
2431	if (done_fd < 0) {
2432		pr_err("Failed to create wakeup eventfd, error: %m\n");
2433		status = -1;
2434		goto out_delete_session;
2435	}
2436	err = evlist__add_wakeup_eventfd(rec->evlist, done_fd);
2437	if (err < 0) {
2438		pr_err("Failed to add wakeup eventfd to poll list\n");
2439		status = err;
2440		goto out_delete_session;
2441	}
2442#endif // HAVE_EVENTFD_SUPPORT
2443
2444	session->header.env.comp_type  = PERF_COMP_ZSTD;
2445	session->header.env.comp_level = rec->opts.comp_level;
2446
2447	if (rec->opts.kcore &&
2448	    !record__kcore_readable(&session->machines.host)) {
2449		pr_err("ERROR: kcore is not readable.\n");
2450		return -1;
2451	}
2452
2453	if (record__init_clock(rec))
2454		return -1;
2455
2456	record__init_features(rec);
2457
2458	if (forks) {
2459		err = evlist__prepare_workload(rec->evlist, &opts->target, argv, data->is_pipe,
2460					       workload_exec_failed_signal);
2461		if (err < 0) {
2462			pr_err("Couldn't run the workload!\n");
2463			status = err;
2464			goto out_delete_session;
2465		}
2466	}
2467
2468	/*
2469	 * If we have just single event and are sending data
2470	 * through pipe, we need to force the ids allocation,
2471	 * because we synthesize event name through the pipe
2472	 * and need the id for that.
2473	 */
2474	if (data->is_pipe && rec->evlist->core.nr_entries == 1)
2475		rec->opts.sample_id = true;
2476
2477	if (rec->timestamp_filename && perf_data__is_pipe(data)) {
2478		rec->timestamp_filename = false;
2479		pr_warning("WARNING: --timestamp-filename option is not available in pipe mode.\n");
2480	}
2481
2482	evlist__uniquify_name(rec->evlist);
2483
2484	evlist__config(rec->evlist, opts, &callchain_param);
2485
2486	/* Debug message used by test scripts */
2487	pr_debug3("perf record opening and mmapping events\n");
2488	if (record__open(rec) != 0) {
2489		err = -1;
2490		goto out_free_threads;
2491	}
2492	/* Debug message used by test scripts */
2493	pr_debug3("perf record done opening and mmapping events\n");
2494	session->header.env.comp_mmap_len = session->evlist->core.mmap_len;
2495
2496	if (rec->opts.kcore) {
2497		err = record__kcore_copy(&session->machines.host, data);
2498		if (err) {
2499			pr_err("ERROR: Failed to copy kcore\n");
2500			goto out_free_threads;
2501		}
2502	}
2503
2504	/*
2505	 * Normally perf_session__new would do this, but it doesn't have the
2506	 * evlist.
2507	 */
2508	if (rec->tool.ordered_events && !evlist__sample_id_all(rec->evlist)) {
2509		pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
2510		rec->tool.ordered_events = false;
2511	}
2512
2513	if (evlist__nr_groups(rec->evlist) == 0)
2514		perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
2515
2516	if (data->is_pipe) {
2517		err = perf_header__write_pipe(fd);
2518		if (err < 0)
2519			goto out_free_threads;
2520	} else {
2521		err = perf_session__write_header(session, rec->evlist, fd, false);
2522		if (err < 0)
2523			goto out_free_threads;
2524	}
2525
2526	err = -1;
2527	if (!rec->no_buildid
2528	    && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
2529		pr_err("Couldn't generate buildids. "
2530		       "Use --no-buildid to profile anyway.\n");
2531		goto out_free_threads;
2532	}
2533
2534	err = record__setup_sb_evlist(rec);
2535	if (err)
2536		goto out_free_threads;
2537
2538	err = record__synthesize(rec, false);
2539	if (err < 0)
2540		goto out_free_threads;
2541
2542	if (rec->realtime_prio) {
2543		struct sched_param param;
2544
2545		param.sched_priority = rec->realtime_prio;
2546		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
2547			pr_err("Could not set realtime priority.\n");
2548			err = -1;
2549			goto out_free_threads;
2550		}
2551	}
2552
2553	if (record__start_threads(rec))
2554		goto out_free_threads;
2555
2556	/*
2557	 * When perf is starting the traced process, all the events
2558	 * (apart from group members) have enable_on_exec=1 set,
2559	 * so don't spoil it by prematurely enabling them.
2560	 */
2561	if (!target__none(&opts->target) && !opts->target.initial_delay)
2562		evlist__enable(rec->evlist);
2563
2564	/*
2565	 * Let the child rip
2566	 */
2567	if (forks) {
2568		struct machine *machine = &session->machines.host;
2569		union perf_event *event;
2570		pid_t tgid;
2571
2572		event = malloc(sizeof(event->comm) + machine->id_hdr_size);
2573		if (event == NULL) {
2574			err = -ENOMEM;
2575			goto out_child;
2576		}
2577
2578		/*
2579		 * Some H/W events are generated before COMM event
2580		 * which is emitted during exec(), so perf script
2581		 * cannot see a correct process name for those events.
2582		 * Synthesize COMM event to prevent it.
2583		 */
2584		tgid = perf_event__synthesize_comm(tool, event,
2585						   rec->evlist->workload.pid,
2586						   process_synthesized_event,
2587						   machine);
2588		free(event);
2589
2590		if (tgid == -1)
2591			goto out_child;
2592
2593		event = malloc(sizeof(event->namespaces) +
2594			       (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
2595			       machine->id_hdr_size);
2596		if (event == NULL) {
2597			err = -ENOMEM;
2598			goto out_child;
2599		}
2600
2601		/*
2602		 * Synthesize NAMESPACES event for the command specified.
2603		 */
2604		perf_event__synthesize_namespaces(tool, event,
2605						  rec->evlist->workload.pid,
2606						  tgid, process_synthesized_event,
2607						  machine);
2608		free(event);
2609
2610		evlist__start_workload(rec->evlist);
2611	}
2612
2613	if (opts->target.initial_delay) {
2614		pr_info(EVLIST_DISABLED_MSG);
2615		if (opts->target.initial_delay > 0) {
2616			usleep(opts->target.initial_delay * USEC_PER_MSEC);
2617			evlist__enable(rec->evlist);
2618			pr_info(EVLIST_ENABLED_MSG);
2619		}
2620	}
2621
2622	err = event_enable_timer__start(rec->evlist->eet);
2623	if (err)
2624		goto out_child;
2625
2626	/* Debug message used by test scripts */
2627	pr_debug3("perf record has started\n");
2628	fflush(stderr);
2629
2630	trigger_ready(&auxtrace_snapshot_trigger);
2631	trigger_ready(&switch_output_trigger);
2632	perf_hooks__invoke_record_start();
2633
2634	/*
2635	 * Must write FINISHED_INIT so it will be seen after all other
2636	 * synthesized user events, but before any regular events.
2637	 */
2638	err = write_finished_init(rec, false);
2639	if (err < 0)
2640		goto out_child;
2641
2642	for (;;) {
2643		unsigned long long hits = thread->samples;
2644
2645		/*
2646		 * rec->evlist->bkw_mmap_state is possible to be
2647		 * BKW_MMAP_EMPTY here: when done == true and
2648		 * hits != rec->samples in previous round.
2649		 *
2650		 * evlist__toggle_bkw_mmap ensure we never
2651		 * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
2652		 */
2653		if (trigger_is_hit(&switch_output_trigger) || done || draining)
2654			evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
2655
2656		if (record__mmap_read_all(rec, false) < 0) {
2657			trigger_error(&auxtrace_snapshot_trigger);
2658			trigger_error(&switch_output_trigger);
2659			err = -1;
2660			goto out_child;
2661		}
2662
2663		if (auxtrace_record__snapshot_started) {
2664			auxtrace_record__snapshot_started = 0;
2665			if (!trigger_is_error(&auxtrace_snapshot_trigger))
2666				record__read_auxtrace_snapshot(rec, false);
2667			if (trigger_is_error(&auxtrace_snapshot_trigger)) {
2668				pr_err("AUX area tracing snapshot failed\n");
2669				err = -1;
2670				goto out_child;
2671			}
2672		}
2673
2674		if (trigger_is_hit(&switch_output_trigger)) {
2675			/*
2676			 * If switch_output_trigger is hit, the data in
2677			 * overwritable ring buffer should have been collected,
2678			 * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
2679			 *
2680			 * If SIGUSR2 raise after or during record__mmap_read_all(),
2681			 * record__mmap_read_all() didn't collect data from
2682			 * overwritable ring buffer. Read again.
2683			 */
2684			if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
2685				continue;
2686			trigger_ready(&switch_output_trigger);
2687
2688			/*
2689			 * Reenable events in overwrite ring buffer after
2690			 * record__mmap_read_all(): we should have collected
2691			 * data from it.
2692			 */
2693			evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
2694
2695			if (!quiet)
2696				fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
2697					record__waking(rec));
2698			thread->waking = 0;
2699			fd = record__switch_output(rec, false);
2700			if (fd < 0) {
2701				pr_err("Failed to switch to new file\n");
2702				trigger_error(&switch_output_trigger);
2703				err = fd;
2704				goto out_child;
2705			}
2706
2707			/* re-arm the alarm */
2708			if (rec->switch_output.time)
2709				alarm(rec->switch_output.time);
2710		}
2711
2712		if (hits == thread->samples) {
2713			if (done || draining)
2714				break;
2715			err = fdarray__poll(&thread->pollfd, -1);
2716			/*
2717			 * Propagate error, only if there's any. Ignore positive
2718			 * number of returned events and interrupt error.
2719			 */
2720			if (err > 0 || (err < 0 && errno == EINTR))
2721				err = 0;
2722			thread->waking++;
2723
2724			if (fdarray__filter(&thread->pollfd, POLLERR | POLLHUP,
2725					    record__thread_munmap_filtered, NULL) == 0)
2726				draining = true;
2727
2728			err = record__update_evlist_pollfd_from_thread(rec, rec->evlist, thread);
2729			if (err)
2730				goto out_child;
2731		}
2732
2733		if (evlist__ctlfd_process(rec->evlist, &cmd) > 0) {
2734			switch (cmd) {
2735			case EVLIST_CTL_CMD_SNAPSHOT:
2736				hit_auxtrace_snapshot_trigger(rec);
2737				evlist__ctlfd_ack(rec->evlist);
2738				break;
2739			case EVLIST_CTL_CMD_STOP:
2740				done = 1;
2741				break;
2742			case EVLIST_CTL_CMD_ACK:
2743			case EVLIST_CTL_CMD_UNSUPPORTED:
2744			case EVLIST_CTL_CMD_ENABLE:
2745			case EVLIST_CTL_CMD_DISABLE:
2746			case EVLIST_CTL_CMD_EVLIST:
2747			case EVLIST_CTL_CMD_PING:
2748			default:
2749				break;
2750			}
2751		}
2752
2753		err = event_enable_timer__process(rec->evlist->eet);
2754		if (err < 0)
2755			goto out_child;
2756		if (err) {
2757			err = 0;
2758			done = 1;
2759		}
2760
2761		/*
2762		 * When perf is starting the traced process, at the end events
2763		 * die with the process and we wait for that. Thus no need to
2764		 * disable events in this case.
2765		 */
2766		if (done && !disabled && !target__none(&opts->target)) {
2767			trigger_off(&auxtrace_snapshot_trigger);
2768			evlist__disable(rec->evlist);
2769			disabled = true;
2770		}
2771	}
2772
2773	trigger_off(&auxtrace_snapshot_trigger);
2774	trigger_off(&switch_output_trigger);
2775
2776	if (opts->auxtrace_snapshot_on_exit)
2777		record__auxtrace_snapshot_exit(rec);
2778
2779	if (forks && workload_exec_errno) {
2780		char msg[STRERR_BUFSIZE], strevsels[2048];
2781		const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
2782
2783		evlist__scnprintf_evsels(rec->evlist, sizeof(strevsels), strevsels);
2784
2785		pr_err("Failed to collect '%s' for the '%s' workload: %s\n",
2786			strevsels, argv[0], emsg);
2787		err = -1;
2788		goto out_child;
2789	}
2790
2791	if (!quiet)
2792		fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n",
2793			record__waking(rec));
2794
2795	write_finished_init(rec, true);
2796
2797	if (target__none(&rec->opts.target))
2798		record__synthesize_workload(rec, true);
2799
2800out_child:
2801	record__stop_threads(rec);
2802	record__mmap_read_all(rec, true);
2803out_free_threads:
2804	record__free_thread_data(rec);
2805	evlist__finalize_ctlfd(rec->evlist);
2806	record__aio_mmap_read_sync(rec);
2807
2808	if (rec->session->bytes_transferred && rec->session->bytes_compressed) {
2809		ratio = (float)rec->session->bytes_transferred/(float)rec->session->bytes_compressed;
2810		session->header.env.comp_ratio = ratio + 0.5;
2811	}
2812
2813	if (forks) {
2814		int exit_status;
2815
2816		if (!child_finished)
2817			kill(rec->evlist->workload.pid, SIGTERM);
2818
2819		wait(&exit_status);
2820
2821		if (err < 0)
2822			status = err;
2823		else if (WIFEXITED(exit_status))
2824			status = WEXITSTATUS(exit_status);
2825		else if (WIFSIGNALED(exit_status))
2826			signr = WTERMSIG(exit_status);
2827	} else
2828		status = err;
2829
2830	if (rec->off_cpu)
2831		rec->bytes_written += off_cpu_write(rec->session);
2832
2833	record__read_lost_samples(rec);
2834	record__synthesize(rec, true);
2835	/* this will be recalculated during process_buildids() */
2836	rec->samples = 0;
2837
2838	if (!err) {
2839		if (!rec->timestamp_filename) {
2840			record__finish_output(rec);
2841		} else {
2842			fd = record__switch_output(rec, true);
2843			if (fd < 0) {
2844				status = fd;
2845				goto out_delete_session;
2846			}
2847		}
2848	}
2849
2850	perf_hooks__invoke_record_end();
2851
2852	if (!err && !quiet) {
2853		char samples[128];
2854		const char *postfix = rec->timestamp_filename ?
2855					".<timestamp>" : "";
2856
2857		if (rec->samples && !rec->opts.full_auxtrace)
2858			scnprintf(samples, sizeof(samples),
2859				  " (%" PRIu64 " samples)", rec->samples);
2860		else
2861			samples[0] = '\0';
2862
2863		fprintf(stderr,	"[ perf record: Captured and wrote %.3f MB %s%s%s",
2864			perf_data__size(data) / 1024.0 / 1024.0,
2865			data->path, postfix, samples);
2866		if (ratio) {
2867			fprintf(stderr,	", compressed (original %.3f MB, ratio is %.3f)",
2868					rec->session->bytes_transferred / 1024.0 / 1024.0,
2869					ratio);
2870		}
2871		fprintf(stderr, " ]\n");
2872	}
2873
2874out_delete_session:
2875#ifdef HAVE_EVENTFD_SUPPORT
2876	if (done_fd >= 0) {
2877		fd = done_fd;
2878		done_fd = -1;
2879
2880		close(fd);
2881	}
2882#endif
2883	zstd_fini(&session->zstd_data);
2884	if (!opts->no_bpf_event)
2885		evlist__stop_sb_thread(rec->sb_evlist);
2886
2887	perf_session__delete(session);
2888	return status;
2889}
2890
2891static void callchain_debug(struct callchain_param *callchain)
2892{
2893	static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
2894
2895	pr_debug("callchain: type %s\n", str[callchain->record_mode]);
2896
2897	if (callchain->record_mode == CALLCHAIN_DWARF)
2898		pr_debug("callchain: stack dump size %d\n",
2899			 callchain->dump_size);
2900}
2901
2902int record_opts__parse_callchain(struct record_opts *record,
2903				 struct callchain_param *callchain,
2904				 const char *arg, bool unset)
2905{
2906	int ret;
2907	callchain->enabled = !unset;
2908
2909	/* --no-call-graph */
2910	if (unset) {
2911		callchain->record_mode = CALLCHAIN_NONE;
2912		pr_debug("callchain: disabled\n");
2913		return 0;
2914	}
2915
2916	ret = parse_callchain_record_opt(arg, callchain);
2917	if (!ret) {
2918		/* Enable data address sampling for DWARF unwind. */
2919		if (callchain->record_mode == CALLCHAIN_DWARF)
2920			record->sample_address = true;
2921		callchain_debug(callchain);
2922	}
2923
2924	return ret;
2925}
2926
2927int record_parse_callchain_opt(const struct option *opt,
2928			       const char *arg,
2929			       int unset)
2930{
2931	return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
2932}
2933
2934int record_callchain_opt(const struct option *opt,
2935			 const char *arg __maybe_unused,
2936			 int unset __maybe_unused)
2937{
2938	struct callchain_param *callchain = opt->value;
2939
2940	callchain->enabled = true;
2941
2942	if (callchain->record_mode == CALLCHAIN_NONE)
2943		callchain->record_mode = CALLCHAIN_FP;
2944
2945	callchain_debug(callchain);
2946	return 0;
2947}
2948
2949static int perf_record_config(const char *var, const char *value, void *cb)
2950{
2951	struct record *rec = cb;
2952
2953	if (!strcmp(var, "record.build-id")) {
2954		if (!strcmp(value, "cache"))
2955			rec->no_buildid_cache = false;
2956		else if (!strcmp(value, "no-cache"))
2957			rec->no_buildid_cache = true;
2958		else if (!strcmp(value, "skip"))
2959			rec->no_buildid = true;
2960		else if (!strcmp(value, "mmap"))
2961			rec->buildid_mmap = true;
2962		else
2963			return -1;
2964		return 0;
2965	}
2966	if (!strcmp(var, "record.call-graph")) {
2967		var = "call-graph.record-mode";
2968		return perf_default_config(var, value, cb);
2969	}
2970#ifdef HAVE_AIO_SUPPORT
2971	if (!strcmp(var, "record.aio")) {
2972		rec->opts.nr_cblocks = strtol(value, NULL, 0);
2973		if (!rec->opts.nr_cblocks)
2974			rec->opts.nr_cblocks = nr_cblocks_default;
2975	}
2976#endif
2977	if (!strcmp(var, "record.debuginfod")) {
2978		rec->debuginfod.urls = strdup(value);
2979		if (!rec->debuginfod.urls)
2980			return -ENOMEM;
2981		rec->debuginfod.set = true;
2982	}
2983
2984	return 0;
2985}
2986
2987static int record__parse_event_enable_time(const struct option *opt, const char *str, int unset)
2988{
2989	struct record *rec = (struct record *)opt->value;
2990
2991	return evlist__parse_event_enable_time(rec->evlist, &rec->opts, str, unset);
2992}
2993
2994static int record__parse_affinity(const struct option *opt, const char *str, int unset)
2995{
2996	struct record_opts *opts = (struct record_opts *)opt->value;
2997
2998	if (unset || !str)
2999		return 0;
3000
3001	if (!strcasecmp(str, "node"))
3002		opts->affinity = PERF_AFFINITY_NODE;
3003	else if (!strcasecmp(str, "cpu"))
3004		opts->affinity = PERF_AFFINITY_CPU;
3005
3006	return 0;
3007}
3008
3009static int record__mmap_cpu_mask_alloc(struct mmap_cpu_mask *mask, int nr_bits)
3010{
3011	mask->nbits = nr_bits;
3012	mask->bits = bitmap_zalloc(mask->nbits);
3013	if (!mask->bits)
3014		return -ENOMEM;
3015
3016	return 0;
3017}
3018
3019static void record__mmap_cpu_mask_free(struct mmap_cpu_mask *mask)
3020{
3021	bitmap_free(mask->bits);
3022	mask->nbits = 0;
3023}
3024
3025static int record__thread_mask_alloc(struct thread_mask *mask, int nr_bits)
3026{
3027	int ret;
3028
3029	ret = record__mmap_cpu_mask_alloc(&mask->maps, nr_bits);
3030	if (ret) {
3031		mask->affinity.bits = NULL;
3032		return ret;
3033	}
3034
3035	ret = record__mmap_cpu_mask_alloc(&mask->affinity, nr_bits);
3036	if (ret) {
3037		record__mmap_cpu_mask_free(&mask->maps);
3038		mask->maps.bits = NULL;
3039	}
3040
3041	return ret;
3042}
3043
3044static void record__thread_mask_free(struct thread_mask *mask)
3045{
3046	record__mmap_cpu_mask_free(&mask->maps);
3047	record__mmap_cpu_mask_free(&mask->affinity);
3048}
3049
3050static int record__parse_threads(const struct option *opt, const char *str, int unset)
3051{
3052	int s;
3053	struct record_opts *opts = opt->value;
3054
3055	if (unset || !str || !strlen(str)) {
3056		opts->threads_spec = THREAD_SPEC__CPU;
3057	} else {
3058		for (s = 1; s < THREAD_SPEC__MAX; s++) {
3059			if (s == THREAD_SPEC__USER) {
3060				opts->threads_user_spec = strdup(str);
3061				if (!opts->threads_user_spec)
3062					return -ENOMEM;
3063				opts->threads_spec = THREAD_SPEC__USER;
3064				break;
3065			}
3066			if (!strncasecmp(str, thread_spec_tags[s], strlen(thread_spec_tags[s]))) {
3067				opts->threads_spec = s;
3068				break;
3069			}
3070		}
3071	}
3072
3073	if (opts->threads_spec == THREAD_SPEC__USER)
3074		pr_debug("threads_spec: %s\n", opts->threads_user_spec);
3075	else
3076		pr_debug("threads_spec: %s\n", thread_spec_tags[opts->threads_spec]);
3077
3078	return 0;
3079}
3080
3081static int parse_output_max_size(const struct option *opt,
3082				 const char *str, int unset)
3083{
3084	unsigned long *s = (unsigned long *)opt->value;
3085	static struct parse_tag tags_size[] = {
3086		{ .tag  = 'B', .mult = 1       },
3087		{ .tag  = 'K', .mult = 1 << 10 },
3088		{ .tag  = 'M', .mult = 1 << 20 },
3089		{ .tag  = 'G', .mult = 1 << 30 },
3090		{ .tag  = 0 },
3091	};
3092	unsigned long val;
3093
3094	if (unset) {
3095		*s = 0;
3096		return 0;
3097	}
3098
3099	val = parse_tag_value(str, tags_size);
3100	if (val != (unsigned long) -1) {
3101		*s = val;
3102		return 0;
3103	}
3104
3105	return -1;
3106}
3107
3108static int record__parse_mmap_pages(const struct option *opt,
3109				    const char *str,
3110				    int unset __maybe_unused)
3111{
3112	struct record_opts *opts = opt->value;
3113	char *s, *p;
3114	unsigned int mmap_pages;
3115	int ret;
3116
3117	if (!str)
3118		return -EINVAL;
3119
3120	s = strdup(str);
3121	if (!s)
3122		return -ENOMEM;
3123
3124	p = strchr(s, ',');
3125	if (p)
3126		*p = '\0';
3127
3128	if (*s) {
3129		ret = __evlist__parse_mmap_pages(&mmap_pages, s);
3130		if (ret)
3131			goto out_free;
3132		opts->mmap_pages = mmap_pages;
3133	}
3134
3135	if (!p) {
3136		ret = 0;
3137		goto out_free;
3138	}
3139
3140	ret = __evlist__parse_mmap_pages(&mmap_pages, p + 1);
3141	if (ret)
3142		goto out_free;
3143
3144	opts->auxtrace_mmap_pages = mmap_pages;
3145
3146out_free:
3147	free(s);
3148	return ret;
3149}
3150
3151void __weak arch__add_leaf_frame_record_opts(struct record_opts *opts __maybe_unused)
3152{
3153}
3154
3155static int parse_control_option(const struct option *opt,
3156				const char *str,
3157				int unset __maybe_unused)
3158{
3159	struct record_opts *opts = opt->value;
3160
3161	return evlist__parse_control(str, &opts->ctl_fd, &opts->ctl_fd_ack, &opts->ctl_fd_close);
3162}
3163
3164static void switch_output_size_warn(struct record *rec)
3165{
3166	u64 wakeup_size = evlist__mmap_size(rec->opts.mmap_pages);
3167	struct switch_output *s = &rec->switch_output;
3168
3169	wakeup_size /= 2;
3170
3171	if (s->size < wakeup_size) {
3172		char buf[100];
3173
3174		unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
3175		pr_warning("WARNING: switch-output data size lower than "
3176			   "wakeup kernel buffer size (%s) "
3177			   "expect bigger perf.data sizes\n", buf);
3178	}
3179}
3180
3181static int switch_output_setup(struct record *rec)
3182{
3183	struct switch_output *s = &rec->switch_output;
3184	static struct parse_tag tags_size[] = {
3185		{ .tag  = 'B', .mult = 1       },
3186		{ .tag  = 'K', .mult = 1 << 10 },
3187		{ .tag  = 'M', .mult = 1 << 20 },
3188		{ .tag  = 'G', .mult = 1 << 30 },
3189		{ .tag  = 0 },
3190	};
3191	static struct parse_tag tags_time[] = {
3192		{ .tag  = 's', .mult = 1        },
3193		{ .tag  = 'm', .mult = 60       },
3194		{ .tag  = 'h', .mult = 60*60    },
3195		{ .tag  = 'd', .mult = 60*60*24 },
3196		{ .tag  = 0 },
3197	};
3198	unsigned long val;
3199
3200	/*
3201	 * If we're using --switch-output-events, then we imply its 
3202	 * --switch-output=signal, as we'll send a SIGUSR2 from the side band
3203	 *  thread to its parent.
3204	 */
3205	if (rec->switch_output_event_set) {
3206		if (record__threads_enabled(rec)) {
3207			pr_warning("WARNING: --switch-output-event option is not available in parallel streaming mode.\n");
3208			return 0;
3209		}
3210		goto do_signal;
3211	}
3212
3213	if (!s->set)
3214		return 0;
3215
3216	if (record__threads_enabled(rec)) {
3217		pr_warning("WARNING: --switch-output option is not available in parallel streaming mode.\n");
3218		return 0;
3219	}
3220
3221	if (!strcmp(s->str, "signal")) {
3222do_signal:
3223		s->signal = true;
3224		pr_debug("switch-output with SIGUSR2 signal\n");
3225		goto enabled;
3226	}
3227
3228	val = parse_tag_value(s->str, tags_size);
3229	if (val != (unsigned long) -1) {
3230		s->size = val;
3231		pr_debug("switch-output with %s size threshold\n", s->str);
3232		goto enabled;
3233	}
3234
3235	val = parse_tag_value(s->str, tags_time);
3236	if (val != (unsigned long) -1) {
3237		s->time = val;
3238		pr_debug("switch-output with %s time threshold (%lu seconds)\n",
3239			 s->str, s->time);
3240		goto enabled;
3241	}
3242
3243	return -1;
3244
3245enabled:
3246	rec->timestamp_filename = true;
3247	s->enabled              = true;
3248
3249	if (s->size && !rec->opts.no_buffering)
3250		switch_output_size_warn(rec);
3251
3252	return 0;
3253}
3254
3255static const char * const __record_usage[] = {
3256	"perf record [<options>] [<command>]",
3257	"perf record [<options>] -- <command> [<options>]",
3258	NULL
3259};
3260const char * const *record_usage = __record_usage;
3261
3262static int build_id__process_mmap(struct perf_tool *tool, union perf_event *event,
3263				  struct perf_sample *sample, struct machine *machine)
3264{
3265	/*
3266	 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
3267	 * no need to add them twice.
3268	 */
3269	if (!(event->header.misc & PERF_RECORD_MISC_USER))
3270		return 0;
3271	return perf_event__process_mmap(tool, event, sample, machine);
3272}
3273
3274static int build_id__process_mmap2(struct perf_tool *tool, union perf_event *event,
3275				   struct perf_sample *sample, struct machine *machine)
3276{
3277	/*
3278	 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
3279	 * no need to add them twice.
3280	 */
3281	if (!(event->header.misc & PERF_RECORD_MISC_USER))
3282		return 0;
3283
3284	return perf_event__process_mmap2(tool, event, sample, machine);
3285}
3286
3287static int process_timestamp_boundary(struct perf_tool *tool,
3288				      union perf_event *event __maybe_unused,
3289				      struct perf_sample *sample,
3290				      struct machine *machine __maybe_unused)
3291{
3292	struct record *rec = container_of(tool, struct record, tool);
3293
3294	set_timestamp_boundary(rec, sample->time);
3295	return 0;
3296}
3297
3298static int parse_record_synth_option(const struct option *opt,
3299				     const char *str,
3300				     int unset __maybe_unused)
3301{
3302	struct record_opts *opts = opt->value;
3303	char *p = strdup(str);
3304
3305	if (p == NULL)
3306		return -1;
3307
3308	opts->synth = parse_synth_opt(p);
3309	free(p);
3310
3311	if (opts->synth < 0) {
3312		pr_err("Invalid synth option: %s\n", str);
3313		return -1;
3314	}
3315	return 0;
3316}
3317
3318/*
3319 * XXX Ideally would be local to cmd_record() and passed to a record__new
3320 * because we need to have access to it in record__exit, that is called
3321 * after cmd_record() exits, but since record_options need to be accessible to
3322 * builtin-script, leave it here.
3323 *
3324 * At least we don't ouch it in all the other functions here directly.
3325 *
3326 * Just say no to tons of global variables, sigh.
3327 */
3328static struct record record = {
3329	.opts = {
3330		.sample_time	     = true,
3331		.mmap_pages	     = UINT_MAX,
3332		.user_freq	     = UINT_MAX,
3333		.user_interval	     = ULLONG_MAX,
3334		.freq		     = 4000,
3335		.target		     = {
3336			.uses_mmap   = true,
3337			.default_per_cpu = true,
3338		},
3339		.mmap_flush          = MMAP_FLUSH_DEFAULT,
3340		.nr_threads_synthesize = 1,
3341		.ctl_fd              = -1,
3342		.ctl_fd_ack          = -1,
3343		.synth               = PERF_SYNTH_ALL,
3344	},
3345	.tool = {
3346		.sample		= process_sample_event,
3347		.fork		= perf_event__process_fork,
3348		.exit		= perf_event__process_exit,
3349		.comm		= perf_event__process_comm,
3350		.namespaces	= perf_event__process_namespaces,
3351		.mmap		= build_id__process_mmap,
3352		.mmap2		= build_id__process_mmap2,
3353		.itrace_start	= process_timestamp_boundary,
3354		.aux		= process_timestamp_boundary,
3355		.ordered_events	= true,
3356	},
3357};
3358
3359const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
3360	"\n\t\t\t\tDefault: fp";
3361
3362static bool dry_run;
3363
3364static struct parse_events_option_args parse_events_option_args = {
3365	.evlistp = &record.evlist,
3366};
3367
3368static struct parse_events_option_args switch_output_parse_events_option_args = {
3369	.evlistp = &record.sb_evlist,
3370};
3371
3372/*
3373 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
3374 * with it and switch to use the library functions in perf_evlist that came
3375 * from builtin-record.c, i.e. use record_opts,
3376 * evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
3377 * using pipes, etc.
3378 */
3379static struct option __record_options[] = {
3380	OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
3381		     "event selector. use 'perf list' to list available events",
3382		     parse_events_option),
3383	OPT_CALLBACK(0, "filter", &record.evlist, "filter",
3384		     "event filter", parse_filter),
3385	OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
3386			   NULL, "don't record events from perf itself",
3387			   exclude_perf),
3388	OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
3389		    "record events on existing process id"),
3390	OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
3391		    "record events on existing thread id"),
3392	OPT_INTEGER('r', "realtime", &record.realtime_prio,
3393		    "collect data with this RT SCHED_FIFO priority"),
3394	OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
3395		    "collect data without buffering"),
3396	OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
3397		    "collect raw sample records from all opened counters"),
3398	OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
3399			    "system-wide collection from all CPUs"),
3400	OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
3401		    "list of cpus to monitor"),
3402	OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
3403	OPT_STRING('o', "output", &record.data.path, "file",
3404		    "output file name"),
3405	OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
3406			&record.opts.no_inherit_set,
3407			"child tasks do not inherit counters"),
3408	OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
3409		    "synthesize non-sample events at the end of output"),
3410	OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
3411	OPT_BOOLEAN(0, "no-bpf-event", &record.opts.no_bpf_event, "do not record bpf events"),
3412	OPT_BOOLEAN(0, "strict-freq", &record.opts.strict_freq,
3413		    "Fail if the specified frequency can't be used"),
3414	OPT_CALLBACK('F', "freq", &record.opts, "freq or 'max'",
3415		     "profile at this frequency",
3416		      record__parse_freq),
3417	OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
3418		     "number of mmap data pages and AUX area tracing mmap pages",
3419		     record__parse_mmap_pages),
3420	OPT_CALLBACK(0, "mmap-flush", &record.opts, "number",
3421		     "Minimal number of bytes that is extracted from mmap data pages (default: 1)",
3422		     record__mmap_flush_parse),
3423	OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
3424			   NULL, "enables call-graph recording" ,
3425			   &record_callchain_opt),
3426	OPT_CALLBACK(0, "call-graph", &record.opts,
3427		     "record_mode[,record_size]", record_callchain_help,
3428		     &record_parse_callchain_opt),
3429	OPT_INCR('v', "verbose", &verbose,
3430		    "be more verbose (show counter open errors, etc)"),
3431	OPT_BOOLEAN('q', "quiet", &quiet, "don't print any warnings or messages"),
3432	OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
3433		    "per thread counts"),
3434	OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
3435	OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
3436		    "Record the sample physical addresses"),
3437	OPT_BOOLEAN(0, "data-page-size", &record.opts.sample_data_page_size,
3438		    "Record the sampled data address data page size"),
3439	OPT_BOOLEAN(0, "code-page-size", &record.opts.sample_code_page_size,
3440		    "Record the sampled code address (ip) page size"),
3441	OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
3442	OPT_BOOLEAN(0, "sample-identifier", &record.opts.sample_identifier,
3443		    "Record the sample identifier"),
3444	OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
3445			&record.opts.sample_time_set,
3446			"Record the sample timestamps"),
3447	OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
3448			"Record the sample period"),
3449	OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
3450		    "don't sample"),
3451	OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
3452			&record.no_buildid_cache_set,
3453			"do not update the buildid cache"),
3454	OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
3455			&record.no_buildid_set,
3456			"do not collect buildids in perf.data"),
3457	OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
3458		     "monitor event in cgroup name only",
3459		     parse_cgroups),
3460	OPT_CALLBACK('D', "delay", &record, "ms",
3461		     "ms to wait before starting measurement after program start (-1: start with events disabled), "
3462		     "or ranges of time to enable events e.g. '-D 10-20,30-40'",
3463		     record__parse_event_enable_time),
3464	OPT_BOOLEAN(0, "kcore", &record.opts.kcore, "copy /proc/kcore"),
3465	OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
3466		   "user to profile"),
3467
3468	OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
3469		     "branch any", "sample any taken branches",
3470		     parse_branch_stack),
3471
3472	OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
3473		     "branch filter mask", "branch stack filter modes",
3474		     parse_branch_stack),
3475	OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
3476		    "sample by weight (on special events only)"),
3477	OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
3478		    "sample transaction flags (special events only)"),
3479	OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
3480		    "use per-thread mmaps"),
3481	OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
3482		    "sample selected machine registers on interrupt,"
3483		    " use '-I?' to list register names", parse_intr_regs),
3484	OPT_CALLBACK_OPTARG(0, "user-regs", &record.opts.sample_user_regs, NULL, "any register",
3485		    "sample selected machine registers on interrupt,"
3486		    " use '--user-regs=?' to list register names", parse_user_regs),
3487	OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
3488		    "Record running/enabled time of read (:S) events"),
3489	OPT_CALLBACK('k', "clockid", &record.opts,
3490	"clockid", "clockid to use for events, see clock_gettime()",
3491	parse_clockid),
3492	OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
3493			  "opts", "AUX area tracing Snapshot Mode", ""),
3494	OPT_STRING_OPTARG(0, "aux-sample", &record.opts.auxtrace_sample_opts,
3495			  "opts", "sample AUX area", ""),
3496	OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3497			"per thread proc mmap processing timeout in ms"),
3498	OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
3499		    "Record namespaces events"),
3500	OPT_BOOLEAN(0, "all-cgroups", &record.opts.record_cgroup,
3501		    "Record cgroup events"),
3502	OPT_BOOLEAN_SET(0, "switch-events", &record.opts.record_switch_events,
3503			&record.opts.record_switch_events_set,
3504			"Record context switch events"),
3505	OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
3506			 "Configure all used events to run in kernel space.",
3507			 PARSE_OPT_EXCLUSIVE),
3508	OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
3509			 "Configure all used events to run in user space.",
3510			 PARSE_OPT_EXCLUSIVE),
3511	OPT_BOOLEAN(0, "kernel-callchains", &record.opts.kernel_callchains,
3512		    "collect kernel callchains"),
3513	OPT_BOOLEAN(0, "user-callchains", &record.opts.user_callchains,
3514		    "collect user callchains"),
3515	OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
3516		   "file", "vmlinux pathname"),
3517	OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
3518		    "Record build-id of all DSOs regardless of hits"),
3519	OPT_BOOLEAN(0, "buildid-mmap", &record.buildid_mmap,
3520		    "Record build-id in map events"),
3521	OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
3522		    "append timestamp to output filename"),
3523	OPT_BOOLEAN(0, "timestamp-boundary", &record.timestamp_boundary,
3524		    "Record timestamp boundary (time of first/last samples)"),
3525	OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
3526			  &record.switch_output.set, "signal or size[BKMG] or time[smhd]",
3527			  "Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold",
3528			  "signal"),
3529	OPT_CALLBACK_SET(0, "switch-output-event", &switch_output_parse_events_option_args,
3530			 &record.switch_output_event_set, "switch output event",
3531			 "switch output event selector. use 'perf list' to list available events",
3532			 parse_events_option_new_evlist),
3533	OPT_INTEGER(0, "switch-max-files", &record.switch_output.num_files,
3534		   "Limit number of switch output generated files"),
3535	OPT_BOOLEAN(0, "dry-run", &dry_run,
3536		    "Parse options then exit"),
3537#ifdef HAVE_AIO_SUPPORT
3538	OPT_CALLBACK_OPTARG(0, "aio", &record.opts,
3539		     &nr_cblocks_default, "n", "Use <n> control blocks in asynchronous trace writing mode (default: 1, max: 4)",
3540		     record__aio_parse),
3541#endif
3542	OPT_CALLBACK(0, "affinity", &record.opts, "node|cpu",
3543		     "Set affinity mask of trace reading thread to NUMA node cpu mask or cpu of processed mmap buffer",
3544		     record__parse_affinity),
3545#ifdef HAVE_ZSTD_SUPPORT
3546	OPT_CALLBACK_OPTARG('z', "compression-level", &record.opts, &comp_level_default, "n",
3547			    "Compress records using specified level (default: 1 - fastest compression, 22 - greatest compression)",
3548			    record__parse_comp_level),
3549#endif
3550	OPT_CALLBACK(0, "max-size", &record.output_max_size,
3551		     "size", "Limit the maximum size of the output file", parse_output_max_size),
3552	OPT_UINTEGER(0, "num-thread-synthesize",
3553		     &record.opts.nr_threads_synthesize,
3554		     "number of threads to run for event synthesis"),
3555#ifdef HAVE_LIBPFM
3556	OPT_CALLBACK(0, "pfm-events", &record.evlist, "event",
3557		"libpfm4 event selector. use 'perf list' to list available events",
3558		parse_libpfm_events_option),
3559#endif
3560	OPT_CALLBACK(0, "control", &record.opts, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
3561		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events,\n"
3562		     "\t\t\t  'snapshot': AUX area tracing snapshot).\n"
3563		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
3564		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
3565		      parse_control_option),
3566	OPT_CALLBACK(0, "synth", &record.opts, "no|all|task|mmap|cgroup",
3567		     "Fine-tune event synthesis: default=all", parse_record_synth_option),
3568	OPT_STRING_OPTARG_SET(0, "debuginfod", &record.debuginfod.urls,
3569			  &record.debuginfod.set, "debuginfod urls",
3570			  "Enable debuginfod data retrieval from DEBUGINFOD_URLS or specified urls",
3571			  "system"),
3572	OPT_CALLBACK_OPTARG(0, "threads", &record.opts, NULL, "spec",
3573			    "write collected trace data into several data files using parallel threads",
3574			    record__parse_threads),
3575	OPT_BOOLEAN(0, "off-cpu", &record.off_cpu, "Enable off-cpu analysis"),
 
 
3576	OPT_END()
3577};
3578
3579struct option *record_options = __record_options;
3580
3581static int record__mmap_cpu_mask_init(struct mmap_cpu_mask *mask, struct perf_cpu_map *cpus)
3582{
3583	struct perf_cpu cpu;
3584	int idx;
3585
3586	if (cpu_map__is_dummy(cpus))
3587		return 0;
3588
3589	perf_cpu_map__for_each_cpu_skip_any(cpu, idx, cpus) {
3590		/* Return ENODEV is input cpu is greater than max cpu */
3591		if ((unsigned long)cpu.cpu > mask->nbits)
3592			return -ENODEV;
3593		__set_bit(cpu.cpu, mask->bits);
3594	}
3595
3596	return 0;
3597}
3598
3599static int record__mmap_cpu_mask_init_spec(struct mmap_cpu_mask *mask, const char *mask_spec)
3600{
3601	struct perf_cpu_map *cpus;
3602
3603	cpus = perf_cpu_map__new(mask_spec);
3604	if (!cpus)
3605		return -ENOMEM;
3606
3607	bitmap_zero(mask->bits, mask->nbits);
3608	if (record__mmap_cpu_mask_init(mask, cpus))
3609		return -ENODEV;
3610
3611	perf_cpu_map__put(cpus);
3612
3613	return 0;
3614}
3615
3616static void record__free_thread_masks(struct record *rec, int nr_threads)
3617{
3618	int t;
3619
3620	if (rec->thread_masks)
3621		for (t = 0; t < nr_threads; t++)
3622			record__thread_mask_free(&rec->thread_masks[t]);
3623
3624	zfree(&rec->thread_masks);
3625}
3626
3627static int record__alloc_thread_masks(struct record *rec, int nr_threads, int nr_bits)
3628{
3629	int t, ret;
3630
3631	rec->thread_masks = zalloc(nr_threads * sizeof(*(rec->thread_masks)));
3632	if (!rec->thread_masks) {
3633		pr_err("Failed to allocate thread masks\n");
3634		return -ENOMEM;
3635	}
3636
3637	for (t = 0; t < nr_threads; t++) {
3638		ret = record__thread_mask_alloc(&rec->thread_masks[t], nr_bits);
3639		if (ret) {
3640			pr_err("Failed to allocate thread masks[%d]\n", t);
3641			goto out_free;
3642		}
3643	}
3644
3645	return 0;
3646
3647out_free:
3648	record__free_thread_masks(rec, nr_threads);
3649
3650	return ret;
3651}
3652
3653static int record__init_thread_cpu_masks(struct record *rec, struct perf_cpu_map *cpus)
3654{
3655	int t, ret, nr_cpus = perf_cpu_map__nr(cpus);
3656
3657	ret = record__alloc_thread_masks(rec, nr_cpus, cpu__max_cpu().cpu);
3658	if (ret)
3659		return ret;
3660
3661	rec->nr_threads = nr_cpus;
3662	pr_debug("nr_threads: %d\n", rec->nr_threads);
3663
3664	for (t = 0; t < rec->nr_threads; t++) {
3665		__set_bit(perf_cpu_map__cpu(cpus, t).cpu, rec->thread_masks[t].maps.bits);
3666		__set_bit(perf_cpu_map__cpu(cpus, t).cpu, rec->thread_masks[t].affinity.bits);
3667		if (verbose > 0) {
3668			pr_debug("thread_masks[%d]: ", t);
3669			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].maps, "maps");
3670			pr_debug("thread_masks[%d]: ", t);
3671			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].affinity, "affinity");
3672		}
3673	}
3674
3675	return 0;
3676}
3677
3678static int record__init_thread_masks_spec(struct record *rec, struct perf_cpu_map *cpus,
3679					  const char **maps_spec, const char **affinity_spec,
3680					  u32 nr_spec)
3681{
3682	u32 s;
3683	int ret = 0, t = 0;
3684	struct mmap_cpu_mask cpus_mask;
3685	struct thread_mask thread_mask, full_mask, *thread_masks;
3686
3687	ret = record__mmap_cpu_mask_alloc(&cpus_mask, cpu__max_cpu().cpu);
3688	if (ret) {
3689		pr_err("Failed to allocate CPUs mask\n");
3690		return ret;
3691	}
3692
3693	ret = record__mmap_cpu_mask_init(&cpus_mask, cpus);
3694	if (ret) {
3695		pr_err("Failed to init cpu mask\n");
3696		goto out_free_cpu_mask;
3697	}
3698
3699	ret = record__thread_mask_alloc(&full_mask, cpu__max_cpu().cpu);
3700	if (ret) {
3701		pr_err("Failed to allocate full mask\n");
3702		goto out_free_cpu_mask;
3703	}
3704
3705	ret = record__thread_mask_alloc(&thread_mask, cpu__max_cpu().cpu);
3706	if (ret) {
3707		pr_err("Failed to allocate thread mask\n");
3708		goto out_free_full_and_cpu_masks;
3709	}
3710
3711	for (s = 0; s < nr_spec; s++) {
3712		ret = record__mmap_cpu_mask_init_spec(&thread_mask.maps, maps_spec[s]);
3713		if (ret) {
3714			pr_err("Failed to initialize maps thread mask\n");
3715			goto out_free;
3716		}
3717		ret = record__mmap_cpu_mask_init_spec(&thread_mask.affinity, affinity_spec[s]);
3718		if (ret) {
3719			pr_err("Failed to initialize affinity thread mask\n");
3720			goto out_free;
3721		}
3722
3723		/* ignore invalid CPUs but do not allow empty masks */
3724		if (!bitmap_and(thread_mask.maps.bits, thread_mask.maps.bits,
3725				cpus_mask.bits, thread_mask.maps.nbits)) {
3726			pr_err("Empty maps mask: %s\n", maps_spec[s]);
3727			ret = -EINVAL;
3728			goto out_free;
3729		}
3730		if (!bitmap_and(thread_mask.affinity.bits, thread_mask.affinity.bits,
3731				cpus_mask.bits, thread_mask.affinity.nbits)) {
3732			pr_err("Empty affinity mask: %s\n", affinity_spec[s]);
3733			ret = -EINVAL;
3734			goto out_free;
3735		}
3736
3737		/* do not allow intersection with other masks (full_mask) */
3738		if (bitmap_intersects(thread_mask.maps.bits, full_mask.maps.bits,
3739				      thread_mask.maps.nbits)) {
3740			pr_err("Intersecting maps mask: %s\n", maps_spec[s]);
3741			ret = -EINVAL;
3742			goto out_free;
3743		}
3744		if (bitmap_intersects(thread_mask.affinity.bits, full_mask.affinity.bits,
3745				      thread_mask.affinity.nbits)) {
3746			pr_err("Intersecting affinity mask: %s\n", affinity_spec[s]);
3747			ret = -EINVAL;
3748			goto out_free;
3749		}
3750
3751		bitmap_or(full_mask.maps.bits, full_mask.maps.bits,
3752			  thread_mask.maps.bits, full_mask.maps.nbits);
3753		bitmap_or(full_mask.affinity.bits, full_mask.affinity.bits,
3754			  thread_mask.affinity.bits, full_mask.maps.nbits);
3755
3756		thread_masks = realloc(rec->thread_masks, (t + 1) * sizeof(struct thread_mask));
3757		if (!thread_masks) {
3758			pr_err("Failed to reallocate thread masks\n");
3759			ret = -ENOMEM;
3760			goto out_free;
3761		}
3762		rec->thread_masks = thread_masks;
3763		rec->thread_masks[t] = thread_mask;
3764		if (verbose > 0) {
3765			pr_debug("thread_masks[%d]: ", t);
3766			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].maps, "maps");
3767			pr_debug("thread_masks[%d]: ", t);
3768			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].affinity, "affinity");
3769		}
3770		t++;
3771		ret = record__thread_mask_alloc(&thread_mask, cpu__max_cpu().cpu);
3772		if (ret) {
3773			pr_err("Failed to allocate thread mask\n");
3774			goto out_free_full_and_cpu_masks;
3775		}
3776	}
3777	rec->nr_threads = t;
3778	pr_debug("nr_threads: %d\n", rec->nr_threads);
3779	if (!rec->nr_threads)
3780		ret = -EINVAL;
3781
3782out_free:
3783	record__thread_mask_free(&thread_mask);
3784out_free_full_and_cpu_masks:
3785	record__thread_mask_free(&full_mask);
3786out_free_cpu_mask:
3787	record__mmap_cpu_mask_free(&cpus_mask);
3788
3789	return ret;
3790}
3791
3792static int record__init_thread_core_masks(struct record *rec, struct perf_cpu_map *cpus)
3793{
3794	int ret;
3795	struct cpu_topology *topo;
3796
3797	topo = cpu_topology__new();
3798	if (!topo) {
3799		pr_err("Failed to allocate CPU topology\n");
3800		return -ENOMEM;
3801	}
3802
3803	ret = record__init_thread_masks_spec(rec, cpus, topo->core_cpus_list,
3804					     topo->core_cpus_list, topo->core_cpus_lists);
3805	cpu_topology__delete(topo);
3806
3807	return ret;
3808}
3809
3810static int record__init_thread_package_masks(struct record *rec, struct perf_cpu_map *cpus)
3811{
3812	int ret;
3813	struct cpu_topology *topo;
3814
3815	topo = cpu_topology__new();
3816	if (!topo) {
3817		pr_err("Failed to allocate CPU topology\n");
3818		return -ENOMEM;
3819	}
3820
3821	ret = record__init_thread_masks_spec(rec, cpus, topo->package_cpus_list,
3822					     topo->package_cpus_list, topo->package_cpus_lists);
3823	cpu_topology__delete(topo);
3824
3825	return ret;
3826}
3827
3828static int record__init_thread_numa_masks(struct record *rec, struct perf_cpu_map *cpus)
3829{
3830	u32 s;
3831	int ret;
3832	const char **spec;
3833	struct numa_topology *topo;
3834
3835	topo = numa_topology__new();
3836	if (!topo) {
3837		pr_err("Failed to allocate NUMA topology\n");
3838		return -ENOMEM;
3839	}
3840
3841	spec = zalloc(topo->nr * sizeof(char *));
3842	if (!spec) {
3843		pr_err("Failed to allocate NUMA spec\n");
3844		ret = -ENOMEM;
3845		goto out_delete_topo;
3846	}
3847	for (s = 0; s < topo->nr; s++)
3848		spec[s] = topo->nodes[s].cpus;
3849
3850	ret = record__init_thread_masks_spec(rec, cpus, spec, spec, topo->nr);
3851
3852	zfree(&spec);
3853
3854out_delete_topo:
3855	numa_topology__delete(topo);
3856
3857	return ret;
3858}
3859
3860static int record__init_thread_user_masks(struct record *rec, struct perf_cpu_map *cpus)
3861{
3862	int t, ret;
3863	u32 s, nr_spec = 0;
3864	char **maps_spec = NULL, **affinity_spec = NULL, **tmp_spec;
3865	char *user_spec, *spec, *spec_ptr, *mask, *mask_ptr, *dup_mask = NULL;
3866
3867	for (t = 0, user_spec = (char *)rec->opts.threads_user_spec; ; t++, user_spec = NULL) {
3868		spec = strtok_r(user_spec, ":", &spec_ptr);
3869		if (spec == NULL)
3870			break;
3871		pr_debug2("threads_spec[%d]: %s\n", t, spec);
3872		mask = strtok_r(spec, "/", &mask_ptr);
3873		if (mask == NULL)
3874			break;
3875		pr_debug2("  maps mask: %s\n", mask);
3876		tmp_spec = realloc(maps_spec, (nr_spec + 1) * sizeof(char *));
3877		if (!tmp_spec) {
3878			pr_err("Failed to reallocate maps spec\n");
3879			ret = -ENOMEM;
3880			goto out_free;
3881		}
3882		maps_spec = tmp_spec;
3883		maps_spec[nr_spec] = dup_mask = strdup(mask);
3884		if (!maps_spec[nr_spec]) {
3885			pr_err("Failed to allocate maps spec[%d]\n", nr_spec);
3886			ret = -ENOMEM;
3887			goto out_free;
3888		}
3889		mask = strtok_r(NULL, "/", &mask_ptr);
3890		if (mask == NULL) {
3891			pr_err("Invalid thread maps or affinity specs\n");
3892			ret = -EINVAL;
3893			goto out_free;
3894		}
3895		pr_debug2("  affinity mask: %s\n", mask);
3896		tmp_spec = realloc(affinity_spec, (nr_spec + 1) * sizeof(char *));
3897		if (!tmp_spec) {
3898			pr_err("Failed to reallocate affinity spec\n");
3899			ret = -ENOMEM;
3900			goto out_free;
3901		}
3902		affinity_spec = tmp_spec;
3903		affinity_spec[nr_spec] = strdup(mask);
3904		if (!affinity_spec[nr_spec]) {
3905			pr_err("Failed to allocate affinity spec[%d]\n", nr_spec);
3906			ret = -ENOMEM;
3907			goto out_free;
3908		}
3909		dup_mask = NULL;
3910		nr_spec++;
3911	}
3912
3913	ret = record__init_thread_masks_spec(rec, cpus, (const char **)maps_spec,
3914					     (const char **)affinity_spec, nr_spec);
3915
3916out_free:
3917	free(dup_mask);
3918	for (s = 0; s < nr_spec; s++) {
3919		if (maps_spec)
3920			free(maps_spec[s]);
3921		if (affinity_spec)
3922			free(affinity_spec[s]);
3923	}
3924	free(affinity_spec);
3925	free(maps_spec);
3926
3927	return ret;
3928}
3929
3930static int record__init_thread_default_masks(struct record *rec, struct perf_cpu_map *cpus)
3931{
3932	int ret;
3933
3934	ret = record__alloc_thread_masks(rec, 1, cpu__max_cpu().cpu);
3935	if (ret)
3936		return ret;
3937
3938	if (record__mmap_cpu_mask_init(&rec->thread_masks->maps, cpus))
3939		return -ENODEV;
3940
3941	rec->nr_threads = 1;
3942
3943	return 0;
3944}
3945
3946static int record__init_thread_masks(struct record *rec)
3947{
3948	int ret = 0;
3949	struct perf_cpu_map *cpus = rec->evlist->core.all_cpus;
3950
3951	if (!record__threads_enabled(rec))
3952		return record__init_thread_default_masks(rec, cpus);
3953
3954	if (evlist__per_thread(rec->evlist)) {
3955		pr_err("--per-thread option is mutually exclusive to parallel streaming mode.\n");
3956		return -EINVAL;
3957	}
3958
3959	switch (rec->opts.threads_spec) {
3960	case THREAD_SPEC__CPU:
3961		ret = record__init_thread_cpu_masks(rec, cpus);
3962		break;
3963	case THREAD_SPEC__CORE:
3964		ret = record__init_thread_core_masks(rec, cpus);
3965		break;
3966	case THREAD_SPEC__PACKAGE:
3967		ret = record__init_thread_package_masks(rec, cpus);
3968		break;
3969	case THREAD_SPEC__NUMA:
3970		ret = record__init_thread_numa_masks(rec, cpus);
3971		break;
3972	case THREAD_SPEC__USER:
3973		ret = record__init_thread_user_masks(rec, cpus);
3974		break;
3975	default:
3976		break;
3977	}
3978
3979	return ret;
3980}
3981
3982int cmd_record(int argc, const char **argv)
3983{
3984	int err;
3985	struct record *rec = &record;
3986	char errbuf[BUFSIZ];
3987
3988	setlocale(LC_ALL, "");
3989
3990#ifndef HAVE_BPF_SKEL
3991# define set_nobuild(s, l, m, c) set_option_nobuild(record_options, s, l, m, c)
3992	set_nobuild('\0', "off-cpu", "no BUILD_BPF_SKEL=1", true);
3993# undef set_nobuild
3994#endif
3995
3996	/* Disable eager loading of kernel symbols that adds overhead to perf record. */
3997	symbol_conf.lazy_load_kernel_maps = true;
3998	rec->opts.affinity = PERF_AFFINITY_SYS;
3999
4000	rec->evlist = evlist__new();
4001	if (rec->evlist == NULL)
4002		return -ENOMEM;
4003
4004	err = perf_config(perf_record_config, rec);
4005	if (err)
4006		return err;
4007
4008	argc = parse_options(argc, argv, record_options, record_usage,
4009			    PARSE_OPT_STOP_AT_NON_OPTION);
4010	if (quiet)
4011		perf_quiet_option();
4012
4013	err = symbol__validate_sym_arguments();
4014	if (err)
4015		return err;
4016
4017	perf_debuginfod_setup(&record.debuginfod);
4018
4019	/* Make system wide (-a) the default target. */
4020	if (!argc && target__none(&rec->opts.target))
4021		rec->opts.target.system_wide = true;
4022
4023	if (nr_cgroups && !rec->opts.target.system_wide) {
4024		usage_with_options_msg(record_usage, record_options,
4025			"cgroup monitoring only available in system-wide mode");
4026
4027	}
4028
4029	if (rec->buildid_mmap) {
4030		if (!perf_can_record_build_id()) {
4031			pr_err("Failed: no support to record build id in mmap events, update your kernel.\n");
4032			err = -EINVAL;
4033			goto out_opts;
4034		}
4035		pr_debug("Enabling build id in mmap2 events.\n");
4036		/* Enable mmap build id synthesizing. */
4037		symbol_conf.buildid_mmap2 = true;
4038		/* Enable perf_event_attr::build_id bit. */
4039		rec->opts.build_id = true;
4040		/* Disable build id cache. */
4041		rec->no_buildid = true;
4042	}
4043
4044	if (rec->opts.record_cgroup && !perf_can_record_cgroup()) {
4045		pr_err("Kernel has no cgroup sampling support.\n");
4046		err = -EINVAL;
4047		goto out_opts;
4048	}
4049
4050	if (rec->opts.kcore)
4051		rec->opts.text_poke = true;
4052
4053	if (rec->opts.kcore || record__threads_enabled(rec))
4054		rec->data.is_dir = true;
4055
4056	if (record__threads_enabled(rec)) {
4057		if (rec->opts.affinity != PERF_AFFINITY_SYS) {
4058			pr_err("--affinity option is mutually exclusive to parallel streaming mode.\n");
4059			goto out_opts;
4060		}
4061		if (record__aio_enabled(rec)) {
4062			pr_err("Asynchronous streaming mode (--aio) is mutually exclusive to parallel streaming mode.\n");
4063			goto out_opts;
4064		}
4065	}
4066
4067	if (rec->opts.comp_level != 0) {
4068		pr_debug("Compression enabled, disabling build id collection at the end of the session.\n");
4069		rec->no_buildid = true;
4070	}
4071
4072	if (rec->opts.record_switch_events &&
4073	    !perf_can_record_switch_events()) {
4074		ui__error("kernel does not support recording context switch events\n");
4075		parse_options_usage(record_usage, record_options, "switch-events", 0);
4076		err = -EINVAL;
4077		goto out_opts;
4078	}
4079
4080	if (switch_output_setup(rec)) {
4081		parse_options_usage(record_usage, record_options, "switch-output", 0);
4082		err = -EINVAL;
4083		goto out_opts;
4084	}
4085
4086	if (rec->switch_output.time) {
4087		signal(SIGALRM, alarm_sig_handler);
4088		alarm(rec->switch_output.time);
4089	}
4090
4091	if (rec->switch_output.num_files) {
4092		rec->switch_output.filenames = calloc(rec->switch_output.num_files,
4093						      sizeof(char *));
4094		if (!rec->switch_output.filenames) {
4095			err = -EINVAL;
4096			goto out_opts;
4097		}
4098	}
4099
4100	if (rec->timestamp_filename && record__threads_enabled(rec)) {
4101		rec->timestamp_filename = false;
4102		pr_warning("WARNING: --timestamp-filename option is not available in parallel streaming mode.\n");
4103	}
4104
 
 
 
 
 
 
 
 
 
 
 
 
4105	/*
4106	 * Allow aliases to facilitate the lookup of symbols for address
4107	 * filters. Refer to auxtrace_parse_filters().
4108	 */
4109	symbol_conf.allow_aliases = true;
4110
4111	symbol__init(NULL);
4112
4113	err = record__auxtrace_init(rec);
4114	if (err)
4115		goto out;
4116
4117	if (dry_run)
4118		goto out;
4119
4120	err = -ENOMEM;
4121
4122	if (rec->no_buildid_cache || rec->no_buildid) {
4123		disable_buildid_cache();
4124	} else if (rec->switch_output.enabled) {
4125		/*
4126		 * In 'perf record --switch-output', disable buildid
4127		 * generation by default to reduce data file switching
4128		 * overhead. Still generate buildid if they are required
4129		 * explicitly using
4130		 *
4131		 *  perf record --switch-output --no-no-buildid \
4132		 *              --no-no-buildid-cache
4133		 *
4134		 * Following code equals to:
4135		 *
4136		 * if ((rec->no_buildid || !rec->no_buildid_set) &&
4137		 *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
4138		 *         disable_buildid_cache();
4139		 */
4140		bool disable = true;
4141
4142		if (rec->no_buildid_set && !rec->no_buildid)
4143			disable = false;
4144		if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
4145			disable = false;
4146		if (disable) {
4147			rec->no_buildid = true;
4148			rec->no_buildid_cache = true;
4149			disable_buildid_cache();
4150		}
4151	}
4152
4153	if (record.opts.overwrite)
4154		record.opts.tail_synthesize = true;
4155
4156	if (rec->evlist->core.nr_entries == 0) {
4157		bool can_profile_kernel = perf_event_paranoid_check(1);
4158
4159		err = parse_event(rec->evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu");
4160		if (err)
4161			goto out;
4162	}
4163
4164	if (rec->opts.target.tid && !rec->opts.no_inherit_set)
4165		rec->opts.no_inherit = true;
4166
4167	err = target__validate(&rec->opts.target);
4168	if (err) {
4169		target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
4170		ui__warning("%s\n", errbuf);
4171	}
4172
4173	err = target__parse_uid(&rec->opts.target);
4174	if (err) {
4175		int saved_errno = errno;
4176
4177		target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
4178		ui__error("%s", errbuf);
4179
4180		err = -saved_errno;
4181		goto out;
4182	}
4183
4184	/* Enable ignoring missing threads when -u/-p option is defined. */
4185	rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
4186
4187	evlist__warn_user_requested_cpus(rec->evlist, rec->opts.target.cpu_list);
4188
4189	if (callchain_param.enabled && callchain_param.record_mode == CALLCHAIN_FP)
4190		arch__add_leaf_frame_record_opts(&rec->opts);
4191
4192	err = -ENOMEM;
4193	if (evlist__create_maps(rec->evlist, &rec->opts.target) < 0) {
4194		if (rec->opts.target.pid != NULL) {
4195			pr_err("Couldn't create thread/CPU maps: %s\n",
4196				errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
4197			goto out;
4198		}
4199		else
4200			usage_with_options(record_usage, record_options);
4201	}
4202
4203	err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
4204	if (err)
4205		goto out;
4206
4207	/*
4208	 * We take all buildids when the file contains
4209	 * AUX area tracing data because we do not decode the
4210	 * trace because it would take too long.
4211	 */
4212	if (rec->opts.full_auxtrace)
4213		rec->buildid_all = true;
4214
4215	if (rec->opts.text_poke) {
4216		err = record__config_text_poke(rec->evlist);
4217		if (err) {
4218			pr_err("record__config_text_poke failed, error %d\n", err);
4219			goto out;
4220		}
4221	}
4222
4223	if (rec->off_cpu) {
4224		err = record__config_off_cpu(rec);
4225		if (err) {
4226			pr_err("record__config_off_cpu failed, error %d\n", err);
4227			goto out;
4228		}
4229	}
4230
4231	if (record_opts__config(&rec->opts)) {
4232		err = -EINVAL;
4233		goto out;
4234	}
4235
4236	err = record__config_tracking_events(rec);
4237	if (err) {
4238		pr_err("record__config_tracking_events failed, error %d\n", err);
4239		goto out;
4240	}
4241
4242	err = record__init_thread_masks(rec);
4243	if (err) {
4244		pr_err("Failed to initialize parallel data streaming masks\n");
4245		goto out;
4246	}
4247
4248	if (rec->opts.nr_cblocks > nr_cblocks_max)
4249		rec->opts.nr_cblocks = nr_cblocks_max;
4250	pr_debug("nr_cblocks: %d\n", rec->opts.nr_cblocks);
4251
4252	pr_debug("affinity: %s\n", affinity_tags[rec->opts.affinity]);
4253	pr_debug("mmap flush: %d\n", rec->opts.mmap_flush);
4254
4255	if (rec->opts.comp_level > comp_level_max)
4256		rec->opts.comp_level = comp_level_max;
4257	pr_debug("comp level: %d\n", rec->opts.comp_level);
4258
4259	err = __cmd_record(&record, argc, argv);
4260out:
4261	evlist__delete(rec->evlist);
 
4262	symbol__exit();
4263	auxtrace_record__free(rec->itr);
4264out_opts:
4265	record__free_thread_masks(rec, rec->nr_threads);
4266	rec->nr_threads = 0;
4267	evlist__close_control(rec->opts.ctl_fd, rec->opts.ctl_fd_ack, &rec->opts.ctl_fd_close);
 
4268	return err;
4269}
4270
4271static void snapshot_sig_handler(int sig __maybe_unused)
4272{
4273	struct record *rec = &record;
4274
4275	hit_auxtrace_snapshot_trigger(rec);
4276
4277	if (switch_output_signal(rec))
4278		trigger_hit(&switch_output_trigger);
4279}
4280
4281static void alarm_sig_handler(int sig __maybe_unused)
4282{
4283	struct record *rec = &record;
4284
4285	if (switch_output_time(rec))
4286		trigger_hit(&switch_output_trigger);
4287}
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * builtin-record.c
   4 *
   5 * Builtin record command: Record the profile of a workload
   6 * (or a CPU, or a PID) into the perf.data output file - for
   7 * later analysis via perf report.
   8 */
   9#include "builtin.h"
  10
  11#include "util/build-id.h"
  12#include <subcmd/parse-options.h>
  13#include <internal/xyarray.h>
  14#include "util/parse-events.h"
  15#include "util/config.h"
  16
  17#include "util/callchain.h"
  18#include "util/cgroup.h"
  19#include "util/header.h"
  20#include "util/event.h"
  21#include "util/evlist.h"
  22#include "util/evsel.h"
  23#include "util/debug.h"
  24#include "util/mmap.h"
  25#include "util/mutex.h"
  26#include "util/target.h"
  27#include "util/session.h"
  28#include "util/tool.h"
  29#include "util/symbol.h"
  30#include "util/record.h"
  31#include "util/cpumap.h"
  32#include "util/thread_map.h"
  33#include "util/data.h"
  34#include "util/perf_regs.h"
  35#include "util/auxtrace.h"
  36#include "util/tsc.h"
  37#include "util/parse-branch-options.h"
  38#include "util/parse-regs-options.h"
  39#include "util/perf_api_probe.h"
  40#include "util/trigger.h"
  41#include "util/perf-hooks.h"
  42#include "util/cpu-set-sched.h"
  43#include "util/synthetic-events.h"
  44#include "util/time-utils.h"
  45#include "util/units.h"
  46#include "util/bpf-event.h"
  47#include "util/util.h"
  48#include "util/pfm.h"
  49#include "util/pmu.h"
  50#include "util/pmus.h"
  51#include "util/clockid.h"
  52#include "util/off_cpu.h"
  53#include "util/bpf-filter.h"
  54#include "asm/bug.h"
  55#include "perf.h"
  56#include "cputopo.h"
  57
  58#include <errno.h>
  59#include <inttypes.h>
  60#include <locale.h>
  61#include <poll.h>
  62#include <pthread.h>
  63#include <unistd.h>
  64#ifndef HAVE_GETTID
  65#include <syscall.h>
  66#endif
  67#include <sched.h>
  68#include <signal.h>
  69#ifdef HAVE_EVENTFD_SUPPORT
  70#include <sys/eventfd.h>
  71#endif
  72#include <sys/mman.h>
  73#include <sys/wait.h>
  74#include <sys/types.h>
  75#include <sys/stat.h>
  76#include <fcntl.h>
  77#include <linux/err.h>
  78#include <linux/string.h>
  79#include <linux/time64.h>
  80#include <linux/zalloc.h>
  81#include <linux/bitmap.h>
  82#include <sys/time.h>
  83
  84struct switch_output {
  85	bool		 enabled;
  86	bool		 signal;
  87	unsigned long	 size;
  88	unsigned long	 time;
  89	const char	*str;
  90	bool		 set;
  91	char		 **filenames;
  92	int		 num_files;
  93	int		 cur_file;
  94};
  95
  96struct thread_mask {
  97	struct mmap_cpu_mask	maps;
  98	struct mmap_cpu_mask	affinity;
  99};
 100
 101struct record_thread {
 102	pid_t			tid;
 103	struct thread_mask	*mask;
 104	struct {
 105		int		msg[2];
 106		int		ack[2];
 107	} pipes;
 108	struct fdarray		pollfd;
 109	int			ctlfd_pos;
 110	int			nr_mmaps;
 111	struct mmap		**maps;
 112	struct mmap		**overwrite_maps;
 113	struct record		*rec;
 114	unsigned long long	samples;
 115	unsigned long		waking;
 116	u64			bytes_written;
 117	u64			bytes_transferred;
 118	u64			bytes_compressed;
 119};
 120
 121static __thread struct record_thread *thread;
 122
 123enum thread_msg {
 124	THREAD_MSG__UNDEFINED = 0,
 125	THREAD_MSG__READY,
 126	THREAD_MSG__MAX,
 127};
 128
 129static const char *thread_msg_tags[THREAD_MSG__MAX] = {
 130	"UNDEFINED", "READY"
 131};
 132
 133enum thread_spec {
 134	THREAD_SPEC__UNDEFINED = 0,
 135	THREAD_SPEC__CPU,
 136	THREAD_SPEC__CORE,
 137	THREAD_SPEC__PACKAGE,
 138	THREAD_SPEC__NUMA,
 139	THREAD_SPEC__USER,
 140	THREAD_SPEC__MAX,
 141};
 142
 143static const char *thread_spec_tags[THREAD_SPEC__MAX] = {
 144	"undefined", "cpu", "core", "package", "numa", "user"
 145};
 146
 147struct pollfd_index_map {
 148	int evlist_pollfd_index;
 149	int thread_pollfd_index;
 150};
 151
 152struct record {
 153	struct perf_tool	tool;
 154	struct record_opts	opts;
 155	u64			bytes_written;
 156	u64			thread_bytes_written;
 157	struct perf_data	data;
 158	struct auxtrace_record	*itr;
 159	struct evlist	*evlist;
 160	struct perf_session	*session;
 161	struct evlist		*sb_evlist;
 162	pthread_t		thread_id;
 163	int			realtime_prio;
 164	bool			switch_output_event_set;
 165	bool			no_buildid;
 166	bool			no_buildid_set;
 167	bool			no_buildid_cache;
 168	bool			no_buildid_cache_set;
 169	bool			buildid_all;
 170	bool			buildid_mmap;
 171	bool			timestamp_filename;
 172	bool			timestamp_boundary;
 173	bool			off_cpu;
 174	const char		*filter_action;
 175	struct switch_output	switch_output;
 176	unsigned long long	samples;
 177	unsigned long		output_max_size;	/* = 0: unlimited */
 178	struct perf_debuginfod	debuginfod;
 179	int			nr_threads;
 180	struct thread_mask	*thread_masks;
 181	struct record_thread	*thread_data;
 182	struct pollfd_index_map	*index_map;
 183	size_t			index_map_sz;
 184	size_t			index_map_cnt;
 185};
 186
 187static volatile int done;
 188
 189static volatile int auxtrace_record__snapshot_started;
 190static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
 191static DEFINE_TRIGGER(switch_output_trigger);
 192
 193static const char *affinity_tags[PERF_AFFINITY_MAX] = {
 194	"SYS", "NODE", "CPU"
 195};
 196
 197static int build_id__process_mmap(const struct perf_tool *tool, union perf_event *event,
 198				  struct perf_sample *sample, struct machine *machine);
 199static int build_id__process_mmap2(const struct perf_tool *tool, union perf_event *event,
 200				   struct perf_sample *sample, struct machine *machine);
 201static int process_timestamp_boundary(const struct perf_tool *tool,
 202				      union perf_event *event,
 203				      struct perf_sample *sample,
 204				      struct machine *machine);
 205
 206#ifndef HAVE_GETTID
 207static inline pid_t gettid(void)
 208{
 209	return (pid_t)syscall(__NR_gettid);
 210}
 211#endif
 212
 213static int record__threads_enabled(struct record *rec)
 214{
 215	return rec->opts.threads_spec;
 216}
 217
 218static bool switch_output_signal(struct record *rec)
 219{
 220	return rec->switch_output.signal &&
 221	       trigger_is_ready(&switch_output_trigger);
 222}
 223
 224static bool switch_output_size(struct record *rec)
 225{
 226	return rec->switch_output.size &&
 227	       trigger_is_ready(&switch_output_trigger) &&
 228	       (rec->bytes_written >= rec->switch_output.size);
 229}
 230
 231static bool switch_output_time(struct record *rec)
 232{
 233	return rec->switch_output.time &&
 234	       trigger_is_ready(&switch_output_trigger);
 235}
 236
 237static u64 record__bytes_written(struct record *rec)
 238{
 239	return rec->bytes_written + rec->thread_bytes_written;
 240}
 241
 242static bool record__output_max_size_exceeded(struct record *rec)
 243{
 244	return rec->output_max_size &&
 245	       (record__bytes_written(rec) >= rec->output_max_size);
 246}
 247
 248static int record__write(struct record *rec, struct mmap *map __maybe_unused,
 249			 void *bf, size_t size)
 250{
 251	struct perf_data_file *file = &rec->session->data->file;
 252
 253	if (map && map->file)
 254		file = map->file;
 255
 256	if (perf_data_file__write(file, bf, size) < 0) {
 257		pr_err("failed to write perf data, error: %m\n");
 258		return -1;
 259	}
 260
 261	if (map && map->file) {
 262		thread->bytes_written += size;
 263		rec->thread_bytes_written += size;
 264	} else {
 265		rec->bytes_written += size;
 266	}
 267
 268	if (record__output_max_size_exceeded(rec) && !done) {
 269		fprintf(stderr, "[ perf record: perf size limit reached (%" PRIu64 " KB),"
 270				" stopping session ]\n",
 271				record__bytes_written(rec) >> 10);
 272		done = 1;
 273	}
 274
 275	if (switch_output_size(rec))
 276		trigger_hit(&switch_output_trigger);
 277
 278	return 0;
 279}
 280
 281static int record__aio_enabled(struct record *rec);
 282static int record__comp_enabled(struct record *rec);
 283static ssize_t zstd_compress(struct perf_session *session, struct mmap *map,
 284			    void *dst, size_t dst_size, void *src, size_t src_size);
 285
 286#ifdef HAVE_AIO_SUPPORT
 287static int record__aio_write(struct aiocb *cblock, int trace_fd,
 288		void *buf, size_t size, off_t off)
 289{
 290	int rc;
 291
 292	cblock->aio_fildes = trace_fd;
 293	cblock->aio_buf    = buf;
 294	cblock->aio_nbytes = size;
 295	cblock->aio_offset = off;
 296	cblock->aio_sigevent.sigev_notify = SIGEV_NONE;
 297
 298	do {
 299		rc = aio_write(cblock);
 300		if (rc == 0) {
 301			break;
 302		} else if (errno != EAGAIN) {
 303			cblock->aio_fildes = -1;
 304			pr_err("failed to queue perf data, error: %m\n");
 305			break;
 306		}
 307	} while (1);
 308
 309	return rc;
 310}
 311
 312static int record__aio_complete(struct mmap *md, struct aiocb *cblock)
 313{
 314	void *rem_buf;
 315	off_t rem_off;
 316	size_t rem_size;
 317	int rc, aio_errno;
 318	ssize_t aio_ret, written;
 319
 320	aio_errno = aio_error(cblock);
 321	if (aio_errno == EINPROGRESS)
 322		return 0;
 323
 324	written = aio_ret = aio_return(cblock);
 325	if (aio_ret < 0) {
 326		if (aio_errno != EINTR)
 327			pr_err("failed to write perf data, error: %m\n");
 328		written = 0;
 329	}
 330
 331	rem_size = cblock->aio_nbytes - written;
 332
 333	if (rem_size == 0) {
 334		cblock->aio_fildes = -1;
 335		/*
 336		 * md->refcount is incremented in record__aio_pushfn() for
 337		 * every aio write request started in record__aio_push() so
 338		 * decrement it because the request is now complete.
 339		 */
 340		perf_mmap__put(&md->core);
 341		rc = 1;
 342	} else {
 343		/*
 344		 * aio write request may require restart with the
 345		 * remainder if the kernel didn't write whole
 346		 * chunk at once.
 347		 */
 348		rem_off = cblock->aio_offset + written;
 349		rem_buf = (void *)(cblock->aio_buf + written);
 350		record__aio_write(cblock, cblock->aio_fildes,
 351				rem_buf, rem_size, rem_off);
 352		rc = 0;
 353	}
 354
 355	return rc;
 356}
 357
 358static int record__aio_sync(struct mmap *md, bool sync_all)
 359{
 360	struct aiocb **aiocb = md->aio.aiocb;
 361	struct aiocb *cblocks = md->aio.cblocks;
 362	struct timespec timeout = { 0, 1000 * 1000  * 1 }; /* 1ms */
 363	int i, do_suspend;
 364
 365	do {
 366		do_suspend = 0;
 367		for (i = 0; i < md->aio.nr_cblocks; ++i) {
 368			if (cblocks[i].aio_fildes == -1 || record__aio_complete(md, &cblocks[i])) {
 369				if (sync_all)
 370					aiocb[i] = NULL;
 371				else
 372					return i;
 373			} else {
 374				/*
 375				 * Started aio write is not complete yet
 376				 * so it has to be waited before the
 377				 * next allocation.
 378				 */
 379				aiocb[i] = &cblocks[i];
 380				do_suspend = 1;
 381			}
 382		}
 383		if (!do_suspend)
 384			return -1;
 385
 386		while (aio_suspend((const struct aiocb **)aiocb, md->aio.nr_cblocks, &timeout)) {
 387			if (!(errno == EAGAIN || errno == EINTR))
 388				pr_err("failed to sync perf data, error: %m\n");
 389		}
 390	} while (1);
 391}
 392
 393struct record_aio {
 394	struct record	*rec;
 395	void		*data;
 396	size_t		size;
 397};
 398
 399static int record__aio_pushfn(struct mmap *map, void *to, void *buf, size_t size)
 400{
 401	struct record_aio *aio = to;
 402
 403	/*
 404	 * map->core.base data pointed by buf is copied into free map->aio.data[] buffer
 405	 * to release space in the kernel buffer as fast as possible, calling
 406	 * perf_mmap__consume() from perf_mmap__push() function.
 407	 *
 408	 * That lets the kernel to proceed with storing more profiling data into
 409	 * the kernel buffer earlier than other per-cpu kernel buffers are handled.
 410	 *
 411	 * Coping can be done in two steps in case the chunk of profiling data
 412	 * crosses the upper bound of the kernel buffer. In this case we first move
 413	 * part of data from map->start till the upper bound and then the remainder
 414	 * from the beginning of the kernel buffer till the end of the data chunk.
 415	 */
 416
 417	if (record__comp_enabled(aio->rec)) {
 418		ssize_t compressed = zstd_compress(aio->rec->session, NULL, aio->data + aio->size,
 419						   mmap__mmap_len(map) - aio->size,
 420						   buf, size);
 421		if (compressed < 0)
 422			return (int)compressed;
 423
 424		size = compressed;
 425	} else {
 426		memcpy(aio->data + aio->size, buf, size);
 427	}
 428
 429	if (!aio->size) {
 430		/*
 431		 * Increment map->refcount to guard map->aio.data[] buffer
 432		 * from premature deallocation because map object can be
 433		 * released earlier than aio write request started on
 434		 * map->aio.data[] buffer is complete.
 435		 *
 436		 * perf_mmap__put() is done at record__aio_complete()
 437		 * after started aio request completion or at record__aio_push()
 438		 * if the request failed to start.
 439		 */
 440		perf_mmap__get(&map->core);
 441	}
 442
 443	aio->size += size;
 444
 445	return size;
 446}
 447
 448static int record__aio_push(struct record *rec, struct mmap *map, off_t *off)
 449{
 450	int ret, idx;
 451	int trace_fd = rec->session->data->file.fd;
 452	struct record_aio aio = { .rec = rec, .size = 0 };
 453
 454	/*
 455	 * Call record__aio_sync() to wait till map->aio.data[] buffer
 456	 * becomes available after previous aio write operation.
 457	 */
 458
 459	idx = record__aio_sync(map, false);
 460	aio.data = map->aio.data[idx];
 461	ret = perf_mmap__push(map, &aio, record__aio_pushfn);
 462	if (ret != 0) /* ret > 0 - no data, ret < 0 - error */
 463		return ret;
 464
 465	rec->samples++;
 466	ret = record__aio_write(&(map->aio.cblocks[idx]), trace_fd, aio.data, aio.size, *off);
 467	if (!ret) {
 468		*off += aio.size;
 469		rec->bytes_written += aio.size;
 470		if (switch_output_size(rec))
 471			trigger_hit(&switch_output_trigger);
 472	} else {
 473		/*
 474		 * Decrement map->refcount incremented in record__aio_pushfn()
 475		 * back if record__aio_write() operation failed to start, otherwise
 476		 * map->refcount is decremented in record__aio_complete() after
 477		 * aio write operation finishes successfully.
 478		 */
 479		perf_mmap__put(&map->core);
 480	}
 481
 482	return ret;
 483}
 484
 485static off_t record__aio_get_pos(int trace_fd)
 486{
 487	return lseek(trace_fd, 0, SEEK_CUR);
 488}
 489
 490static void record__aio_set_pos(int trace_fd, off_t pos)
 491{
 492	lseek(trace_fd, pos, SEEK_SET);
 493}
 494
 495static void record__aio_mmap_read_sync(struct record *rec)
 496{
 497	int i;
 498	struct evlist *evlist = rec->evlist;
 499	struct mmap *maps = evlist->mmap;
 500
 501	if (!record__aio_enabled(rec))
 502		return;
 503
 504	for (i = 0; i < evlist->core.nr_mmaps; i++) {
 505		struct mmap *map = &maps[i];
 506
 507		if (map->core.base)
 508			record__aio_sync(map, true);
 509	}
 510}
 511
 512static int nr_cblocks_default = 1;
 513static int nr_cblocks_max = 4;
 514
 515static int record__aio_parse(const struct option *opt,
 516			     const char *str,
 517			     int unset)
 518{
 519	struct record_opts *opts = (struct record_opts *)opt->value;
 520
 521	if (unset) {
 522		opts->nr_cblocks = 0;
 523	} else {
 524		if (str)
 525			opts->nr_cblocks = strtol(str, NULL, 0);
 526		if (!opts->nr_cblocks)
 527			opts->nr_cblocks = nr_cblocks_default;
 528	}
 529
 530	return 0;
 531}
 532#else /* HAVE_AIO_SUPPORT */
 533static int nr_cblocks_max = 0;
 534
 535static int record__aio_push(struct record *rec __maybe_unused, struct mmap *map __maybe_unused,
 536			    off_t *off __maybe_unused)
 537{
 538	return -1;
 539}
 540
 541static off_t record__aio_get_pos(int trace_fd __maybe_unused)
 542{
 543	return -1;
 544}
 545
 546static void record__aio_set_pos(int trace_fd __maybe_unused, off_t pos __maybe_unused)
 547{
 548}
 549
 550static void record__aio_mmap_read_sync(struct record *rec __maybe_unused)
 551{
 552}
 553#endif
 554
 555static int record__aio_enabled(struct record *rec)
 556{
 557	return rec->opts.nr_cblocks > 0;
 558}
 559
 560#define MMAP_FLUSH_DEFAULT 1
 561static int record__mmap_flush_parse(const struct option *opt,
 562				    const char *str,
 563				    int unset)
 564{
 565	int flush_max;
 566	struct record_opts *opts = (struct record_opts *)opt->value;
 567	static struct parse_tag tags[] = {
 568			{ .tag  = 'B', .mult = 1       },
 569			{ .tag  = 'K', .mult = 1 << 10 },
 570			{ .tag  = 'M', .mult = 1 << 20 },
 571			{ .tag  = 'G', .mult = 1 << 30 },
 572			{ .tag  = 0 },
 573	};
 574
 575	if (unset)
 576		return 0;
 577
 578	if (str) {
 579		opts->mmap_flush = parse_tag_value(str, tags);
 580		if (opts->mmap_flush == (int)-1)
 581			opts->mmap_flush = strtol(str, NULL, 0);
 582	}
 583
 584	if (!opts->mmap_flush)
 585		opts->mmap_flush = MMAP_FLUSH_DEFAULT;
 586
 587	flush_max = evlist__mmap_size(opts->mmap_pages);
 588	flush_max /= 4;
 589	if (opts->mmap_flush > flush_max)
 590		opts->mmap_flush = flush_max;
 591
 592	return 0;
 593}
 594
 595#ifdef HAVE_ZSTD_SUPPORT
 596static unsigned int comp_level_default = 1;
 597
 598static int record__parse_comp_level(const struct option *opt, const char *str, int unset)
 599{
 600	struct record_opts *opts = opt->value;
 601
 602	if (unset) {
 603		opts->comp_level = 0;
 604	} else {
 605		if (str)
 606			opts->comp_level = strtol(str, NULL, 0);
 607		if (!opts->comp_level)
 608			opts->comp_level = comp_level_default;
 609	}
 610
 611	return 0;
 612}
 613#endif
 614static unsigned int comp_level_max = 22;
 615
 616static int record__comp_enabled(struct record *rec)
 617{
 618	return rec->opts.comp_level > 0;
 619}
 620
 621static int process_synthesized_event(const struct perf_tool *tool,
 622				     union perf_event *event,
 623				     struct perf_sample *sample __maybe_unused,
 624				     struct machine *machine __maybe_unused)
 625{
 626	struct record *rec = container_of(tool, struct record, tool);
 627	return record__write(rec, NULL, event, event->header.size);
 628}
 629
 630static struct mutex synth_lock;
 631
 632static int process_locked_synthesized_event(const struct perf_tool *tool,
 633				     union perf_event *event,
 634				     struct perf_sample *sample __maybe_unused,
 635				     struct machine *machine __maybe_unused)
 636{
 637	int ret;
 638
 639	mutex_lock(&synth_lock);
 640	ret = process_synthesized_event(tool, event, sample, machine);
 641	mutex_unlock(&synth_lock);
 642	return ret;
 643}
 644
 645static int record__pushfn(struct mmap *map, void *to, void *bf, size_t size)
 646{
 647	struct record *rec = to;
 648
 649	if (record__comp_enabled(rec)) {
 650		ssize_t compressed = zstd_compress(rec->session, map, map->data,
 651						   mmap__mmap_len(map), bf, size);
 652
 653		if (compressed < 0)
 654			return (int)compressed;
 655
 656		size = compressed;
 657		bf   = map->data;
 658	}
 659
 660	thread->samples++;
 661	return record__write(rec, map, bf, size);
 662}
 663
 664static volatile sig_atomic_t signr = -1;
 665static volatile sig_atomic_t child_finished;
 666#ifdef HAVE_EVENTFD_SUPPORT
 667static volatile sig_atomic_t done_fd = -1;
 668#endif
 669
 670static void sig_handler(int sig)
 671{
 672	if (sig == SIGCHLD)
 673		child_finished = 1;
 674	else
 675		signr = sig;
 676
 677	done = 1;
 678#ifdef HAVE_EVENTFD_SUPPORT
 679	if (done_fd >= 0) {
 680		u64 tmp = 1;
 681		int orig_errno = errno;
 682
 683		/*
 684		 * It is possible for this signal handler to run after done is
 685		 * checked in the main loop, but before the perf counter fds are
 686		 * polled. If this happens, the poll() will continue to wait
 687		 * even though done is set, and will only break out if either
 688		 * another signal is received, or the counters are ready for
 689		 * read. To ensure the poll() doesn't sleep when done is set,
 690		 * use an eventfd (done_fd) to wake up the poll().
 691		 */
 692		if (write(done_fd, &tmp, sizeof(tmp)) < 0)
 693			pr_err("failed to signal wakeup fd, error: %m\n");
 694
 695		errno = orig_errno;
 696	}
 697#endif // HAVE_EVENTFD_SUPPORT
 698}
 699
 700static void sigsegv_handler(int sig)
 701{
 702	perf_hooks__recover();
 703	sighandler_dump_stack(sig);
 704}
 705
 706static void record__sig_exit(void)
 707{
 708	if (signr == -1)
 709		return;
 710
 711	signal(signr, SIG_DFL);
 712	raise(signr);
 713}
 714
 715#ifdef HAVE_AUXTRACE_SUPPORT
 716
 717static int record__process_auxtrace(const struct perf_tool *tool,
 718				    struct mmap *map,
 719				    union perf_event *event, void *data1,
 720				    size_t len1, void *data2, size_t len2)
 721{
 722	struct record *rec = container_of(tool, struct record, tool);
 723	struct perf_data *data = &rec->data;
 724	size_t padding;
 725	u8 pad[8] = {0};
 726
 727	if (!perf_data__is_pipe(data) && perf_data__is_single_file(data)) {
 728		off_t file_offset;
 729		int fd = perf_data__fd(data);
 730		int err;
 731
 732		file_offset = lseek(fd, 0, SEEK_CUR);
 733		if (file_offset == -1)
 734			return -1;
 735		err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 736						     event, file_offset);
 737		if (err)
 738			return err;
 739	}
 740
 741	/* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 742	padding = (len1 + len2) & 7;
 743	if (padding)
 744		padding = 8 - padding;
 745
 746	record__write(rec, map, event, event->header.size);
 747	record__write(rec, map, data1, len1);
 748	if (len2)
 749		record__write(rec, map, data2, len2);
 750	record__write(rec, map, &pad, padding);
 751
 752	return 0;
 753}
 754
 755static int record__auxtrace_mmap_read(struct record *rec,
 756				      struct mmap *map)
 757{
 758	int ret;
 759
 760	ret = auxtrace_mmap__read(map, rec->itr, &rec->tool,
 761				  record__process_auxtrace);
 762	if (ret < 0)
 763		return ret;
 764
 765	if (ret)
 766		rec->samples++;
 767
 768	return 0;
 769}
 770
 771static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 772					       struct mmap *map)
 773{
 774	int ret;
 775
 776	ret = auxtrace_mmap__read_snapshot(map, rec->itr, &rec->tool,
 777					   record__process_auxtrace,
 778					   rec->opts.auxtrace_snapshot_size);
 779	if (ret < 0)
 780		return ret;
 781
 782	if (ret)
 783		rec->samples++;
 784
 785	return 0;
 786}
 787
 788static int record__auxtrace_read_snapshot_all(struct record *rec)
 789{
 790	int i;
 791	int rc = 0;
 792
 793	for (i = 0; i < rec->evlist->core.nr_mmaps; i++) {
 794		struct mmap *map = &rec->evlist->mmap[i];
 795
 796		if (!map->auxtrace_mmap.base)
 797			continue;
 798
 799		if (record__auxtrace_mmap_read_snapshot(rec, map) != 0) {
 800			rc = -1;
 801			goto out;
 802		}
 803	}
 804out:
 805	return rc;
 806}
 807
 808static void record__read_auxtrace_snapshot(struct record *rec, bool on_exit)
 809{
 810	pr_debug("Recording AUX area tracing snapshot\n");
 811	if (record__auxtrace_read_snapshot_all(rec) < 0) {
 812		trigger_error(&auxtrace_snapshot_trigger);
 813	} else {
 814		if (auxtrace_record__snapshot_finish(rec->itr, on_exit))
 815			trigger_error(&auxtrace_snapshot_trigger);
 816		else
 817			trigger_ready(&auxtrace_snapshot_trigger);
 818	}
 819}
 820
 821static int record__auxtrace_snapshot_exit(struct record *rec)
 822{
 823	if (trigger_is_error(&auxtrace_snapshot_trigger))
 824		return 0;
 825
 826	if (!auxtrace_record__snapshot_started &&
 827	    auxtrace_record__snapshot_start(rec->itr))
 828		return -1;
 829
 830	record__read_auxtrace_snapshot(rec, true);
 831	if (trigger_is_error(&auxtrace_snapshot_trigger))
 832		return -1;
 833
 834	return 0;
 835}
 836
 837static int record__auxtrace_init(struct record *rec)
 838{
 839	int err;
 840
 841	if ((rec->opts.auxtrace_snapshot_opts || rec->opts.auxtrace_sample_opts)
 842	    && record__threads_enabled(rec)) {
 843		pr_err("AUX area tracing options are not available in parallel streaming mode.\n");
 844		return -EINVAL;
 845	}
 846
 847	if (!rec->itr) {
 848		rec->itr = auxtrace_record__init(rec->evlist, &err);
 849		if (err)
 850			return err;
 851	}
 852
 853	err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
 854					      rec->opts.auxtrace_snapshot_opts);
 855	if (err)
 856		return err;
 857
 858	err = auxtrace_parse_sample_options(rec->itr, rec->evlist, &rec->opts,
 859					    rec->opts.auxtrace_sample_opts);
 860	if (err)
 861		return err;
 862
 863	auxtrace_regroup_aux_output(rec->evlist);
 864
 865	return auxtrace_parse_filters(rec->evlist);
 866}
 867
 868#else
 869
 870static inline
 871int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 872			       struct mmap *map __maybe_unused)
 873{
 874	return 0;
 875}
 876
 877static inline
 878void record__read_auxtrace_snapshot(struct record *rec __maybe_unused,
 879				    bool on_exit __maybe_unused)
 880{
 881}
 882
 883static inline
 884int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 885{
 886	return 0;
 887}
 888
 889static inline
 890int record__auxtrace_snapshot_exit(struct record *rec __maybe_unused)
 891{
 892	return 0;
 893}
 894
 895static int record__auxtrace_init(struct record *rec __maybe_unused)
 896{
 897	return 0;
 898}
 899
 900#endif
 901
 902static int record__config_text_poke(struct evlist *evlist)
 903{
 904	struct evsel *evsel;
 905
 906	/* Nothing to do if text poke is already configured */
 907	evlist__for_each_entry(evlist, evsel) {
 908		if (evsel->core.attr.text_poke)
 909			return 0;
 910	}
 911
 912	evsel = evlist__add_dummy_on_all_cpus(evlist);
 913	if (!evsel)
 914		return -ENOMEM;
 915
 916	evsel->core.attr.text_poke = 1;
 917	evsel->core.attr.ksymbol = 1;
 918	evsel->immediate = true;
 919	evsel__set_sample_bit(evsel, TIME);
 920
 921	return 0;
 922}
 923
 924static int record__config_off_cpu(struct record *rec)
 925{
 926	return off_cpu_prepare(rec->evlist, &rec->opts.target, &rec->opts);
 927}
 928
 929static bool record__tracking_system_wide(struct record *rec)
 930{
 931	struct evlist *evlist = rec->evlist;
 932	struct evsel *evsel;
 933
 934	/*
 935	 * If non-dummy evsel exists, system_wide sideband is need to
 936	 * help parse sample information.
 937	 * For example, PERF_EVENT_MMAP event to help parse symbol,
 938	 * and PERF_EVENT_COMM event to help parse task executable name.
 939	 */
 940	evlist__for_each_entry(evlist, evsel) {
 941		if (!evsel__is_dummy_event(evsel))
 942			return true;
 943	}
 944
 945	return false;
 946}
 947
 948static int record__config_tracking_events(struct record *rec)
 949{
 950	struct record_opts *opts = &rec->opts;
 951	struct evlist *evlist = rec->evlist;
 952	bool system_wide = false;
 953	struct evsel *evsel;
 954
 955	/*
 956	 * For initial_delay, system wide or a hybrid system, we need to add
 957	 * tracking event so that we can track PERF_RECORD_MMAP to cover the
 958	 * delay of waiting or event synthesis.
 959	 */
 960	if (opts->target.initial_delay || target__has_cpu(&opts->target) ||
 961	    perf_pmus__num_core_pmus() > 1) {
 962
 963		/*
 964		 * User space tasks can migrate between CPUs, so when tracing
 965		 * selected CPUs, sideband for all CPUs is still needed.
 966		 */
 967		if (!!opts->target.cpu_list && record__tracking_system_wide(rec))
 968			system_wide = true;
 969
 970		evsel = evlist__findnew_tracking_event(evlist, system_wide);
 971		if (!evsel)
 972			return -ENOMEM;
 973
 974		/*
 975		 * Enable the tracking event when the process is forked for
 976		 * initial_delay, immediately for system wide.
 977		 */
 978		if (opts->target.initial_delay && !evsel->immediate &&
 979		    !target__has_cpu(&opts->target))
 980			evsel->core.attr.enable_on_exec = 1;
 981		else
 982			evsel->immediate = 1;
 983	}
 984
 985	return 0;
 986}
 987
 988static bool record__kcore_readable(struct machine *machine)
 989{
 990	char kcore[PATH_MAX];
 991	int fd;
 992
 993	scnprintf(kcore, sizeof(kcore), "%s/proc/kcore", machine->root_dir);
 994
 995	fd = open(kcore, O_RDONLY);
 996	if (fd < 0)
 997		return false;
 998
 999	close(fd);
1000
1001	return true;
1002}
1003
1004static int record__kcore_copy(struct machine *machine, struct perf_data *data)
1005{
1006	char from_dir[PATH_MAX];
1007	char kcore_dir[PATH_MAX];
1008	int ret;
1009
1010	snprintf(from_dir, sizeof(from_dir), "%s/proc", machine->root_dir);
1011
1012	ret = perf_data__make_kcore_dir(data, kcore_dir, sizeof(kcore_dir));
1013	if (ret)
1014		return ret;
1015
1016	return kcore_copy(from_dir, kcore_dir);
1017}
1018
1019static void record__thread_data_init_pipes(struct record_thread *thread_data)
1020{
1021	thread_data->pipes.msg[0] = -1;
1022	thread_data->pipes.msg[1] = -1;
1023	thread_data->pipes.ack[0] = -1;
1024	thread_data->pipes.ack[1] = -1;
1025}
1026
1027static int record__thread_data_open_pipes(struct record_thread *thread_data)
1028{
1029	if (pipe(thread_data->pipes.msg))
1030		return -EINVAL;
1031
1032	if (pipe(thread_data->pipes.ack)) {
1033		close(thread_data->pipes.msg[0]);
1034		thread_data->pipes.msg[0] = -1;
1035		close(thread_data->pipes.msg[1]);
1036		thread_data->pipes.msg[1] = -1;
1037		return -EINVAL;
1038	}
1039
1040	pr_debug2("thread_data[%p]: msg=[%d,%d], ack=[%d,%d]\n", thread_data,
1041		 thread_data->pipes.msg[0], thread_data->pipes.msg[1],
1042		 thread_data->pipes.ack[0], thread_data->pipes.ack[1]);
1043
1044	return 0;
1045}
1046
1047static void record__thread_data_close_pipes(struct record_thread *thread_data)
1048{
1049	if (thread_data->pipes.msg[0] != -1) {
1050		close(thread_data->pipes.msg[0]);
1051		thread_data->pipes.msg[0] = -1;
1052	}
1053	if (thread_data->pipes.msg[1] != -1) {
1054		close(thread_data->pipes.msg[1]);
1055		thread_data->pipes.msg[1] = -1;
1056	}
1057	if (thread_data->pipes.ack[0] != -1) {
1058		close(thread_data->pipes.ack[0]);
1059		thread_data->pipes.ack[0] = -1;
1060	}
1061	if (thread_data->pipes.ack[1] != -1) {
1062		close(thread_data->pipes.ack[1]);
1063		thread_data->pipes.ack[1] = -1;
1064	}
1065}
1066
1067static bool evlist__per_thread(struct evlist *evlist)
1068{
1069	return cpu_map__is_dummy(evlist->core.user_requested_cpus);
1070}
1071
1072static int record__thread_data_init_maps(struct record_thread *thread_data, struct evlist *evlist)
1073{
1074	int m, tm, nr_mmaps = evlist->core.nr_mmaps;
1075	struct mmap *mmap = evlist->mmap;
1076	struct mmap *overwrite_mmap = evlist->overwrite_mmap;
1077	struct perf_cpu_map *cpus = evlist->core.all_cpus;
1078	bool per_thread = evlist__per_thread(evlist);
1079
1080	if (per_thread)
1081		thread_data->nr_mmaps = nr_mmaps;
1082	else
1083		thread_data->nr_mmaps = bitmap_weight(thread_data->mask->maps.bits,
1084						      thread_data->mask->maps.nbits);
1085	if (mmap) {
1086		thread_data->maps = zalloc(thread_data->nr_mmaps * sizeof(struct mmap *));
1087		if (!thread_data->maps)
1088			return -ENOMEM;
1089	}
1090	if (overwrite_mmap) {
1091		thread_data->overwrite_maps = zalloc(thread_data->nr_mmaps * sizeof(struct mmap *));
1092		if (!thread_data->overwrite_maps) {
1093			zfree(&thread_data->maps);
1094			return -ENOMEM;
1095		}
1096	}
1097	pr_debug2("thread_data[%p]: nr_mmaps=%d, maps=%p, ow_maps=%p\n", thread_data,
1098		 thread_data->nr_mmaps, thread_data->maps, thread_data->overwrite_maps);
1099
1100	for (m = 0, tm = 0; m < nr_mmaps && tm < thread_data->nr_mmaps; m++) {
1101		if (per_thread ||
1102		    test_bit(perf_cpu_map__cpu(cpus, m).cpu, thread_data->mask->maps.bits)) {
1103			if (thread_data->maps) {
1104				thread_data->maps[tm] = &mmap[m];
1105				pr_debug2("thread_data[%p]: cpu%d: maps[%d] -> mmap[%d]\n",
1106					  thread_data, perf_cpu_map__cpu(cpus, m).cpu, tm, m);
1107			}
1108			if (thread_data->overwrite_maps) {
1109				thread_data->overwrite_maps[tm] = &overwrite_mmap[m];
1110				pr_debug2("thread_data[%p]: cpu%d: ow_maps[%d] -> ow_mmap[%d]\n",
1111					  thread_data, perf_cpu_map__cpu(cpus, m).cpu, tm, m);
1112			}
1113			tm++;
1114		}
1115	}
1116
1117	return 0;
1118}
1119
1120static int record__thread_data_init_pollfd(struct record_thread *thread_data, struct evlist *evlist)
1121{
1122	int f, tm, pos;
1123	struct mmap *map, *overwrite_map;
1124
1125	fdarray__init(&thread_data->pollfd, 64);
1126
1127	for (tm = 0; tm < thread_data->nr_mmaps; tm++) {
1128		map = thread_data->maps ? thread_data->maps[tm] : NULL;
1129		overwrite_map = thread_data->overwrite_maps ?
1130				thread_data->overwrite_maps[tm] : NULL;
1131
1132		for (f = 0; f < evlist->core.pollfd.nr; f++) {
1133			void *ptr = evlist->core.pollfd.priv[f].ptr;
1134
1135			if ((map && ptr == map) || (overwrite_map && ptr == overwrite_map)) {
1136				pos = fdarray__dup_entry_from(&thread_data->pollfd, f,
1137							      &evlist->core.pollfd);
1138				if (pos < 0)
1139					return pos;
1140				pr_debug2("thread_data[%p]: pollfd[%d] <- event_fd=%d\n",
1141					 thread_data, pos, evlist->core.pollfd.entries[f].fd);
1142			}
1143		}
1144	}
1145
1146	return 0;
1147}
1148
1149static void record__free_thread_data(struct record *rec)
1150{
1151	int t;
1152	struct record_thread *thread_data = rec->thread_data;
1153
1154	if (thread_data == NULL)
1155		return;
1156
1157	for (t = 0; t < rec->nr_threads; t++) {
1158		record__thread_data_close_pipes(&thread_data[t]);
1159		zfree(&thread_data[t].maps);
1160		zfree(&thread_data[t].overwrite_maps);
1161		fdarray__exit(&thread_data[t].pollfd);
1162	}
1163
1164	zfree(&rec->thread_data);
1165}
1166
1167static int record__map_thread_evlist_pollfd_indexes(struct record *rec,
1168						    int evlist_pollfd_index,
1169						    int thread_pollfd_index)
1170{
1171	size_t x = rec->index_map_cnt;
1172
1173	if (realloc_array_as_needed(rec->index_map, rec->index_map_sz, x, NULL))
1174		return -ENOMEM;
1175	rec->index_map[x].evlist_pollfd_index = evlist_pollfd_index;
1176	rec->index_map[x].thread_pollfd_index = thread_pollfd_index;
1177	rec->index_map_cnt += 1;
1178	return 0;
1179}
1180
1181static int record__update_evlist_pollfd_from_thread(struct record *rec,
1182						    struct evlist *evlist,
1183						    struct record_thread *thread_data)
1184{
1185	struct pollfd *e_entries = evlist->core.pollfd.entries;
1186	struct pollfd *t_entries = thread_data->pollfd.entries;
1187	int err = 0;
1188	size_t i;
1189
1190	for (i = 0; i < rec->index_map_cnt; i++) {
1191		int e_pos = rec->index_map[i].evlist_pollfd_index;
1192		int t_pos = rec->index_map[i].thread_pollfd_index;
1193
1194		if (e_entries[e_pos].fd != t_entries[t_pos].fd ||
1195		    e_entries[e_pos].events != t_entries[t_pos].events) {
1196			pr_err("Thread and evlist pollfd index mismatch\n");
1197			err = -EINVAL;
1198			continue;
1199		}
1200		e_entries[e_pos].revents = t_entries[t_pos].revents;
1201	}
1202	return err;
1203}
1204
1205static int record__dup_non_perf_events(struct record *rec,
1206				       struct evlist *evlist,
1207				       struct record_thread *thread_data)
1208{
1209	struct fdarray *fda = &evlist->core.pollfd;
1210	int i, ret;
1211
1212	for (i = 0; i < fda->nr; i++) {
1213		if (!(fda->priv[i].flags & fdarray_flag__non_perf_event))
1214			continue;
1215		ret = fdarray__dup_entry_from(&thread_data->pollfd, i, fda);
1216		if (ret < 0) {
1217			pr_err("Failed to duplicate descriptor in main thread pollfd\n");
1218			return ret;
1219		}
1220		pr_debug2("thread_data[%p]: pollfd[%d] <- non_perf_event fd=%d\n",
1221			  thread_data, ret, fda->entries[i].fd);
1222		ret = record__map_thread_evlist_pollfd_indexes(rec, i, ret);
1223		if (ret < 0) {
1224			pr_err("Failed to map thread and evlist pollfd indexes\n");
1225			return ret;
1226		}
1227	}
1228	return 0;
1229}
1230
1231static int record__alloc_thread_data(struct record *rec, struct evlist *evlist)
1232{
1233	int t, ret;
1234	struct record_thread *thread_data;
1235
1236	rec->thread_data = zalloc(rec->nr_threads * sizeof(*(rec->thread_data)));
1237	if (!rec->thread_data) {
1238		pr_err("Failed to allocate thread data\n");
1239		return -ENOMEM;
1240	}
1241	thread_data = rec->thread_data;
1242
1243	for (t = 0; t < rec->nr_threads; t++)
1244		record__thread_data_init_pipes(&thread_data[t]);
1245
1246	for (t = 0; t < rec->nr_threads; t++) {
1247		thread_data[t].rec = rec;
1248		thread_data[t].mask = &rec->thread_masks[t];
1249		ret = record__thread_data_init_maps(&thread_data[t], evlist);
1250		if (ret) {
1251			pr_err("Failed to initialize thread[%d] maps\n", t);
1252			goto out_free;
1253		}
1254		ret = record__thread_data_init_pollfd(&thread_data[t], evlist);
1255		if (ret) {
1256			pr_err("Failed to initialize thread[%d] pollfd\n", t);
1257			goto out_free;
1258		}
1259		if (t) {
1260			thread_data[t].tid = -1;
1261			ret = record__thread_data_open_pipes(&thread_data[t]);
1262			if (ret) {
1263				pr_err("Failed to open thread[%d] communication pipes\n", t);
1264				goto out_free;
1265			}
1266			ret = fdarray__add(&thread_data[t].pollfd, thread_data[t].pipes.msg[0],
1267					   POLLIN | POLLERR | POLLHUP, fdarray_flag__nonfilterable);
1268			if (ret < 0) {
1269				pr_err("Failed to add descriptor to thread[%d] pollfd\n", t);
1270				goto out_free;
1271			}
1272			thread_data[t].ctlfd_pos = ret;
1273			pr_debug2("thread_data[%p]: pollfd[%d] <- ctl_fd=%d\n",
1274				 thread_data, thread_data[t].ctlfd_pos,
1275				 thread_data[t].pipes.msg[0]);
1276		} else {
1277			thread_data[t].tid = gettid();
1278
1279			ret = record__dup_non_perf_events(rec, evlist, &thread_data[t]);
1280			if (ret < 0)
1281				goto out_free;
1282
1283			thread_data[t].ctlfd_pos = -1; /* Not used */
1284		}
1285	}
1286
1287	return 0;
1288
1289out_free:
1290	record__free_thread_data(rec);
1291
1292	return ret;
1293}
1294
1295static int record__mmap_evlist(struct record *rec,
1296			       struct evlist *evlist)
1297{
1298	int i, ret;
1299	struct record_opts *opts = &rec->opts;
1300	bool auxtrace_overwrite = opts->auxtrace_snapshot_mode ||
1301				  opts->auxtrace_sample_mode;
1302	char msg[512];
1303
1304	if (opts->affinity != PERF_AFFINITY_SYS)
1305		cpu__setup_cpunode_map();
1306
1307	if (evlist__mmap_ex(evlist, opts->mmap_pages,
1308				 opts->auxtrace_mmap_pages,
1309				 auxtrace_overwrite,
1310				 opts->nr_cblocks, opts->affinity,
1311				 opts->mmap_flush, opts->comp_level) < 0) {
1312		if (errno == EPERM) {
1313			pr_err("Permission error mapping pages.\n"
1314			       "Consider increasing "
1315			       "/proc/sys/kernel/perf_event_mlock_kb,\n"
1316			       "or try again with a smaller value of -m/--mmap_pages.\n"
1317			       "(current value: %u,%u)\n",
1318			       opts->mmap_pages, opts->auxtrace_mmap_pages);
1319			return -errno;
1320		} else {
1321			pr_err("failed to mmap with %d (%s)\n", errno,
1322				str_error_r(errno, msg, sizeof(msg)));
1323			if (errno)
1324				return -errno;
1325			else
1326				return -EINVAL;
1327		}
1328	}
1329
1330	if (evlist__initialize_ctlfd(evlist, opts->ctl_fd, opts->ctl_fd_ack))
1331		return -1;
1332
1333	ret = record__alloc_thread_data(rec, evlist);
1334	if (ret)
1335		return ret;
1336
1337	if (record__threads_enabled(rec)) {
1338		ret = perf_data__create_dir(&rec->data, evlist->core.nr_mmaps);
1339		if (ret) {
1340			pr_err("Failed to create data directory: %s\n", strerror(-ret));
1341			return ret;
1342		}
1343		for (i = 0; i < evlist->core.nr_mmaps; i++) {
1344			if (evlist->mmap)
1345				evlist->mmap[i].file = &rec->data.dir.files[i];
1346			if (evlist->overwrite_mmap)
1347				evlist->overwrite_mmap[i].file = &rec->data.dir.files[i];
1348		}
1349	}
1350
1351	return 0;
1352}
1353
1354static int record__mmap(struct record *rec)
1355{
1356	return record__mmap_evlist(rec, rec->evlist);
1357}
1358
1359static int record__open(struct record *rec)
1360{
1361	char msg[BUFSIZ];
1362	struct evsel *pos;
1363	struct evlist *evlist = rec->evlist;
1364	struct perf_session *session = rec->session;
1365	struct record_opts *opts = &rec->opts;
1366	int rc = 0;
1367
1368	evlist__for_each_entry(evlist, pos) {
1369try_again:
1370		if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
1371			if (evsel__fallback(pos, &opts->target, errno, msg, sizeof(msg))) {
1372				if (verbose > 0)
1373					ui__warning("%s\n", msg);
1374				goto try_again;
1375			}
1376			if ((errno == EINVAL || errno == EBADF) &&
1377			    pos->core.leader != &pos->core &&
1378			    pos->weak_group) {
1379			        pos = evlist__reset_weak_group(evlist, pos, true);
1380				goto try_again;
1381			}
1382			rc = -errno;
1383			evsel__open_strerror(pos, &opts->target, errno, msg, sizeof(msg));
1384			ui__error("%s\n", msg);
1385			goto out;
1386		}
1387
1388		pos->supported = true;
1389	}
1390
1391	if (symbol_conf.kptr_restrict && !evlist__exclude_kernel(evlist)) {
1392		pr_warning(
1393"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1394"check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1395"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1396"file is not found in the buildid cache or in the vmlinux path.\n\n"
1397"Samples in kernel modules won't be resolved at all.\n\n"
1398"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1399"even with a suitable vmlinux or kallsyms file.\n\n");
1400	}
1401
1402	if (evlist__apply_filters(evlist, &pos, &opts->target)) {
1403		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
1404			pos->filter ?: "BPF", evsel__name(pos), errno,
1405			str_error_r(errno, msg, sizeof(msg)));
1406		rc = -1;
1407		goto out;
1408	}
1409
1410	rc = record__mmap(rec);
1411	if (rc)
1412		goto out;
1413
1414	session->evlist = evlist;
1415	perf_session__set_id_hdr_size(session);
1416out:
1417	return rc;
1418}
1419
1420static void set_timestamp_boundary(struct record *rec, u64 sample_time)
1421{
1422	if (rec->evlist->first_sample_time == 0)
1423		rec->evlist->first_sample_time = sample_time;
1424
1425	if (sample_time)
1426		rec->evlist->last_sample_time = sample_time;
1427}
1428
1429static int process_sample_event(const struct perf_tool *tool,
1430				union perf_event *event,
1431				struct perf_sample *sample,
1432				struct evsel *evsel,
1433				struct machine *machine)
1434{
1435	struct record *rec = container_of(tool, struct record, tool);
1436
1437	set_timestamp_boundary(rec, sample->time);
1438
1439	if (rec->buildid_all)
1440		return 0;
1441
1442	rec->samples++;
1443	return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
1444}
1445
1446static int process_buildids(struct record *rec)
1447{
1448	struct perf_session *session = rec->session;
1449
1450	if (perf_data__size(&rec->data) == 0)
1451		return 0;
1452
1453	/*
1454	 * During this process, it'll load kernel map and replace the
1455	 * dso->long_name to a real pathname it found.  In this case
1456	 * we prefer the vmlinux path like
1457	 *   /lib/modules/3.16.4/build/vmlinux
1458	 *
1459	 * rather than build-id path (in debug directory).
1460	 *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
1461	 */
1462	symbol_conf.ignore_vmlinux_buildid = true;
1463
1464	/*
1465	 * If --buildid-all is given, it marks all DSO regardless of hits,
1466	 * so no need to process samples. But if timestamp_boundary is enabled,
1467	 * it still needs to walk on all samples to get the timestamps of
1468	 * first/last samples.
1469	 */
1470	if (rec->buildid_all && !rec->timestamp_boundary)
1471		rec->tool.sample = process_event_sample_stub;
1472
1473	return perf_session__process_events(session);
1474}
1475
1476static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
1477{
1478	int err;
1479	struct perf_tool *tool = data;
1480	/*
1481	 *As for guest kernel when processing subcommand record&report,
1482	 *we arrange module mmap prior to guest kernel mmap and trigger
1483	 *a preload dso because default guest module symbols are loaded
1484	 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
1485	 *method is used to avoid symbol missing when the first addr is
1486	 *in module instead of in guest kernel.
1487	 */
1488	err = perf_event__synthesize_modules(tool, process_synthesized_event,
1489					     machine);
1490	if (err < 0)
1491		pr_err("Couldn't record guest kernel [%d]'s reference"
1492		       " relocation symbol.\n", machine->pid);
1493
1494	/*
1495	 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
1496	 * have no _text sometimes.
1497	 */
1498	err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1499						 machine);
1500	if (err < 0)
1501		pr_err("Couldn't record guest kernel [%d]'s reference"
1502		       " relocation symbol.\n", machine->pid);
1503}
1504
1505static struct perf_event_header finished_round_event = {
1506	.size = sizeof(struct perf_event_header),
1507	.type = PERF_RECORD_FINISHED_ROUND,
1508};
1509
1510static struct perf_event_header finished_init_event = {
1511	.size = sizeof(struct perf_event_header),
1512	.type = PERF_RECORD_FINISHED_INIT,
1513};
1514
1515static void record__adjust_affinity(struct record *rec, struct mmap *map)
1516{
1517	if (rec->opts.affinity != PERF_AFFINITY_SYS &&
1518	    !bitmap_equal(thread->mask->affinity.bits, map->affinity_mask.bits,
1519			  thread->mask->affinity.nbits)) {
1520		bitmap_zero(thread->mask->affinity.bits, thread->mask->affinity.nbits);
1521		bitmap_or(thread->mask->affinity.bits, thread->mask->affinity.bits,
1522			  map->affinity_mask.bits, thread->mask->affinity.nbits);
1523		sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&thread->mask->affinity),
1524					(cpu_set_t *)thread->mask->affinity.bits);
1525		if (verbose == 2) {
1526			pr_debug("threads[%d]: running on cpu%d: ", thread->tid, sched_getcpu());
1527			mmap_cpu_mask__scnprintf(&thread->mask->affinity, "affinity");
1528		}
1529	}
1530}
1531
1532static size_t process_comp_header(void *record, size_t increment)
1533{
1534	struct perf_record_compressed *event = record;
1535	size_t size = sizeof(*event);
1536
1537	if (increment) {
1538		event->header.size += increment;
1539		return increment;
1540	}
1541
1542	event->header.type = PERF_RECORD_COMPRESSED;
1543	event->header.size = size;
1544
1545	return size;
1546}
1547
1548static ssize_t zstd_compress(struct perf_session *session, struct mmap *map,
1549			    void *dst, size_t dst_size, void *src, size_t src_size)
1550{
1551	ssize_t compressed;
1552	size_t max_record_size = PERF_SAMPLE_MAX_SIZE - sizeof(struct perf_record_compressed) - 1;
1553	struct zstd_data *zstd_data = &session->zstd_data;
1554
1555	if (map && map->file)
1556		zstd_data = &map->zstd_data;
1557
1558	compressed = zstd_compress_stream_to_records(zstd_data, dst, dst_size, src, src_size,
1559						     max_record_size, process_comp_header);
1560	if (compressed < 0)
1561		return compressed;
1562
1563	if (map && map->file) {
1564		thread->bytes_transferred += src_size;
1565		thread->bytes_compressed  += compressed;
1566	} else {
1567		session->bytes_transferred += src_size;
1568		session->bytes_compressed  += compressed;
1569	}
1570
1571	return compressed;
1572}
1573
1574static int record__mmap_read_evlist(struct record *rec, struct evlist *evlist,
1575				    bool overwrite, bool synch)
1576{
1577	u64 bytes_written = rec->bytes_written;
1578	int i;
1579	int rc = 0;
1580	int nr_mmaps;
1581	struct mmap **maps;
1582	int trace_fd = rec->data.file.fd;
1583	off_t off = 0;
1584
1585	if (!evlist)
1586		return 0;
1587
1588	nr_mmaps = thread->nr_mmaps;
1589	maps = overwrite ? thread->overwrite_maps : thread->maps;
1590
1591	if (!maps)
1592		return 0;
1593
1594	if (overwrite && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
1595		return 0;
1596
1597	if (record__aio_enabled(rec))
1598		off = record__aio_get_pos(trace_fd);
1599
1600	for (i = 0; i < nr_mmaps; i++) {
1601		u64 flush = 0;
1602		struct mmap *map = maps[i];
1603
1604		if (map->core.base) {
1605			record__adjust_affinity(rec, map);
1606			if (synch) {
1607				flush = map->core.flush;
1608				map->core.flush = 1;
1609			}
1610			if (!record__aio_enabled(rec)) {
1611				if (perf_mmap__push(map, rec, record__pushfn) < 0) {
1612					if (synch)
1613						map->core.flush = flush;
1614					rc = -1;
1615					goto out;
1616				}
1617			} else {
1618				if (record__aio_push(rec, map, &off) < 0) {
1619					record__aio_set_pos(trace_fd, off);
1620					if (synch)
1621						map->core.flush = flush;
1622					rc = -1;
1623					goto out;
1624				}
1625			}
1626			if (synch)
1627				map->core.flush = flush;
1628		}
1629
1630		if (map->auxtrace_mmap.base && !rec->opts.auxtrace_snapshot_mode &&
1631		    !rec->opts.auxtrace_sample_mode &&
1632		    record__auxtrace_mmap_read(rec, map) != 0) {
1633			rc = -1;
1634			goto out;
1635		}
1636	}
1637
1638	if (record__aio_enabled(rec))
1639		record__aio_set_pos(trace_fd, off);
1640
1641	/*
1642	 * Mark the round finished in case we wrote
1643	 * at least one event.
1644	 *
1645	 * No need for round events in directory mode,
1646	 * because per-cpu maps and files have data
1647	 * sorted by kernel.
1648	 */
1649	if (!record__threads_enabled(rec) && bytes_written != rec->bytes_written)
1650		rc = record__write(rec, NULL, &finished_round_event, sizeof(finished_round_event));
1651
1652	if (overwrite)
1653		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
1654out:
1655	return rc;
1656}
1657
1658static int record__mmap_read_all(struct record *rec, bool synch)
1659{
1660	int err;
1661
1662	err = record__mmap_read_evlist(rec, rec->evlist, false, synch);
1663	if (err)
1664		return err;
1665
1666	return record__mmap_read_evlist(rec, rec->evlist, true, synch);
1667}
1668
1669static void record__thread_munmap_filtered(struct fdarray *fda, int fd,
1670					   void *arg __maybe_unused)
1671{
1672	struct perf_mmap *map = fda->priv[fd].ptr;
1673
1674	if (map)
1675		perf_mmap__put(map);
1676}
1677
1678static void *record__thread(void *arg)
1679{
1680	enum thread_msg msg = THREAD_MSG__READY;
1681	bool terminate = false;
1682	struct fdarray *pollfd;
1683	int err, ctlfd_pos;
1684
1685	thread = arg;
1686	thread->tid = gettid();
1687
1688	err = write(thread->pipes.ack[1], &msg, sizeof(msg));
1689	if (err == -1)
1690		pr_warning("threads[%d]: failed to notify on start: %s\n",
1691			   thread->tid, strerror(errno));
1692
1693	pr_debug("threads[%d]: started on cpu%d\n", thread->tid, sched_getcpu());
1694
1695	pollfd = &thread->pollfd;
1696	ctlfd_pos = thread->ctlfd_pos;
1697
1698	for (;;) {
1699		unsigned long long hits = thread->samples;
1700
1701		if (record__mmap_read_all(thread->rec, false) < 0 || terminate)
1702			break;
1703
1704		if (hits == thread->samples) {
1705
1706			err = fdarray__poll(pollfd, -1);
1707			/*
1708			 * Propagate error, only if there's any. Ignore positive
1709			 * number of returned events and interrupt error.
1710			 */
1711			if (err > 0 || (err < 0 && errno == EINTR))
1712				err = 0;
1713			thread->waking++;
1714
1715			if (fdarray__filter(pollfd, POLLERR | POLLHUP,
1716					    record__thread_munmap_filtered, NULL) == 0)
1717				break;
1718		}
1719
1720		if (pollfd->entries[ctlfd_pos].revents & POLLHUP) {
1721			terminate = true;
1722			close(thread->pipes.msg[0]);
1723			thread->pipes.msg[0] = -1;
1724			pollfd->entries[ctlfd_pos].fd = -1;
1725			pollfd->entries[ctlfd_pos].events = 0;
1726		}
1727
1728		pollfd->entries[ctlfd_pos].revents = 0;
1729	}
1730	record__mmap_read_all(thread->rec, true);
1731
1732	err = write(thread->pipes.ack[1], &msg, sizeof(msg));
1733	if (err == -1)
1734		pr_warning("threads[%d]: failed to notify on termination: %s\n",
1735			   thread->tid, strerror(errno));
1736
1737	return NULL;
1738}
1739
1740static void record__init_features(struct record *rec)
1741{
1742	struct perf_session *session = rec->session;
1743	int feat;
1744
1745	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1746		perf_header__set_feat(&session->header, feat);
1747
1748	if (rec->no_buildid)
1749		perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1750
1751#ifdef HAVE_LIBTRACEEVENT
1752	if (!have_tracepoints(&rec->evlist->core.entries))
1753		perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1754#endif
1755
1756	if (!rec->opts.branch_stack)
1757		perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1758
1759	if (!rec->opts.full_auxtrace)
1760		perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1761
1762	if (!(rec->opts.use_clockid && rec->opts.clockid_res_ns))
1763		perf_header__clear_feat(&session->header, HEADER_CLOCKID);
1764
1765	if (!rec->opts.use_clockid)
1766		perf_header__clear_feat(&session->header, HEADER_CLOCK_DATA);
1767
1768	if (!record__threads_enabled(rec))
1769		perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1770
1771	if (!record__comp_enabled(rec))
1772		perf_header__clear_feat(&session->header, HEADER_COMPRESSED);
1773
1774	perf_header__clear_feat(&session->header, HEADER_STAT);
1775}
1776
1777static void
1778record__finish_output(struct record *rec)
1779{
1780	int i;
1781	struct perf_data *data = &rec->data;
1782	int fd = perf_data__fd(data);
1783
1784	if (data->is_pipe) {
1785		/* Just to display approx. size */
1786		data->file.size = rec->bytes_written;
1787		return;
1788	}
1789
1790	rec->session->header.data_size += rec->bytes_written;
1791	data->file.size = lseek(perf_data__fd(data), 0, SEEK_CUR);
1792	if (record__threads_enabled(rec)) {
1793		for (i = 0; i < data->dir.nr; i++)
1794			data->dir.files[i].size = lseek(data->dir.files[i].fd, 0, SEEK_CUR);
1795	}
1796
1797	if (!rec->no_buildid) {
1798		process_buildids(rec);
1799
1800		if (rec->buildid_all)
1801			perf_session__dsos_hit_all(rec->session);
1802	}
1803	perf_session__write_header(rec->session, rec->evlist, fd, true);
1804
1805	return;
1806}
1807
1808static int record__synthesize_workload(struct record *rec, bool tail)
1809{
1810	int err;
1811	struct perf_thread_map *thread_map;
1812	bool needs_mmap = rec->opts.synth & PERF_SYNTH_MMAP;
1813
1814	if (rec->opts.tail_synthesize != tail)
1815		return 0;
1816
1817	thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
1818	if (thread_map == NULL)
1819		return -1;
1820
1821	err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
1822						 process_synthesized_event,
1823						 &rec->session->machines.host,
1824						 needs_mmap,
1825						 rec->opts.sample_address);
1826	perf_thread_map__put(thread_map);
1827	return err;
1828}
1829
1830static int write_finished_init(struct record *rec, bool tail)
1831{
1832	if (rec->opts.tail_synthesize != tail)
1833		return 0;
1834
1835	return record__write(rec, NULL, &finished_init_event, sizeof(finished_init_event));
1836}
1837
1838static int record__synthesize(struct record *rec, bool tail);
1839
1840static int
1841record__switch_output(struct record *rec, bool at_exit)
1842{
1843	struct perf_data *data = &rec->data;
1844	char *new_filename = NULL;
1845	int fd, err;
1846
1847	/* Same Size:      "2015122520103046"*/
1848	char timestamp[] = "InvalidTimestamp";
1849
1850	record__aio_mmap_read_sync(rec);
1851
1852	write_finished_init(rec, true);
1853
1854	record__synthesize(rec, true);
1855	if (target__none(&rec->opts.target))
1856		record__synthesize_workload(rec, true);
1857
1858	rec->samples = 0;
1859	record__finish_output(rec);
1860	err = fetch_current_timestamp(timestamp, sizeof(timestamp));
1861	if (err) {
1862		pr_err("Failed to get current timestamp\n");
1863		return -EINVAL;
1864	}
1865
1866	fd = perf_data__switch(data, timestamp,
1867			       rec->session->header.data_offset,
1868			       at_exit, &new_filename);
1869	if (fd >= 0 && !at_exit) {
1870		rec->bytes_written = 0;
1871		rec->session->header.data_size = 0;
1872	}
1873
1874	if (!quiet) {
1875		fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
1876			data->path, timestamp);
1877	}
1878
1879	if (rec->switch_output.num_files) {
1880		int n = rec->switch_output.cur_file + 1;
1881
1882		if (n >= rec->switch_output.num_files)
1883			n = 0;
1884		rec->switch_output.cur_file = n;
1885		if (rec->switch_output.filenames[n]) {
1886			remove(rec->switch_output.filenames[n]);
1887			zfree(&rec->switch_output.filenames[n]);
1888		}
1889		rec->switch_output.filenames[n] = new_filename;
1890	} else {
1891		free(new_filename);
1892	}
1893
1894	/* Output tracking events */
1895	if (!at_exit) {
1896		record__synthesize(rec, false);
1897
1898		/*
1899		 * In 'perf record --switch-output' without -a,
1900		 * record__synthesize() in record__switch_output() won't
1901		 * generate tracking events because there's no thread_map
1902		 * in evlist. Which causes newly created perf.data doesn't
1903		 * contain map and comm information.
1904		 * Create a fake thread_map and directly call
1905		 * perf_event__synthesize_thread_map() for those events.
1906		 */
1907		if (target__none(&rec->opts.target))
1908			record__synthesize_workload(rec, false);
1909		write_finished_init(rec, false);
1910	}
1911	return fd;
1912}
1913
1914static void __record__save_lost_samples(struct record *rec, struct evsel *evsel,
1915					struct perf_record_lost_samples *lost,
1916					int cpu_idx, int thread_idx, u64 lost_count,
1917					u16 misc_flag)
1918{
1919	struct perf_sample_id *sid;
1920	struct perf_sample sample = {};
1921	int id_hdr_size;
1922
1923	lost->lost = lost_count;
1924	if (evsel->core.ids) {
1925		sid = xyarray__entry(evsel->core.sample_id, cpu_idx, thread_idx);
1926		sample.id = sid->id;
1927	}
1928
1929	id_hdr_size = perf_event__synthesize_id_sample((void *)(lost + 1),
1930						       evsel->core.attr.sample_type, &sample);
1931	lost->header.size = sizeof(*lost) + id_hdr_size;
1932	lost->header.misc = misc_flag;
1933	record__write(rec, NULL, lost, lost->header.size);
1934}
1935
1936static void record__read_lost_samples(struct record *rec)
1937{
1938	struct perf_session *session = rec->session;
1939	struct perf_record_lost_samples_and_ids lost;
1940	struct evsel *evsel;
1941
1942	/* there was an error during record__open */
1943	if (session->evlist == NULL)
1944		return;
1945
1946	evlist__for_each_entry(session->evlist, evsel) {
1947		struct xyarray *xy = evsel->core.sample_id;
1948		u64 lost_count;
1949
1950		if (xy == NULL || evsel->core.fd == NULL)
1951			continue;
1952		if (xyarray__max_x(evsel->core.fd) != xyarray__max_x(xy) ||
1953		    xyarray__max_y(evsel->core.fd) != xyarray__max_y(xy)) {
1954			pr_debug("Unmatched FD vs. sample ID: skip reading LOST count\n");
1955			continue;
1956		}
1957
1958		for (int x = 0; x < xyarray__max_x(xy); x++) {
1959			for (int y = 0; y < xyarray__max_y(xy); y++) {
1960				struct perf_counts_values count;
1961
1962				if (perf_evsel__read(&evsel->core, x, y, &count) < 0) {
1963					pr_debug("read LOST count failed\n");
1964					return;
1965				}
1966
1967				if (count.lost) {
1968					memset(&lost, 0, sizeof(lost));
1969					lost.lost.header.type = PERF_RECORD_LOST_SAMPLES;
1970					__record__save_lost_samples(rec, evsel, &lost.lost,
 
 
 
 
 
 
 
1971								    x, y, count.lost, 0);
1972				}
1973			}
1974		}
1975
1976		lost_count = perf_bpf_filter__lost_count(evsel);
1977		if (lost_count) {
1978			memset(&lost, 0, sizeof(lost));
1979			lost.lost.header.type = PERF_RECORD_LOST_SAMPLES;
1980			__record__save_lost_samples(rec, evsel, &lost.lost, 0, 0, lost_count,
 
 
 
 
 
 
 
1981						    PERF_RECORD_MISC_LOST_SAMPLES_BPF);
1982		}
1983	}
 
 
1984}
1985
1986static volatile sig_atomic_t workload_exec_errno;
1987
1988/*
1989 * evlist__prepare_workload will send a SIGUSR1
1990 * if the fork fails, since we asked by setting its
1991 * want_signal to true.
1992 */
1993static void workload_exec_failed_signal(int signo __maybe_unused,
1994					siginfo_t *info,
1995					void *ucontext __maybe_unused)
1996{
1997	workload_exec_errno = info->si_value.sival_int;
1998	done = 1;
1999	child_finished = 1;
2000}
2001
2002static void snapshot_sig_handler(int sig);
2003static void alarm_sig_handler(int sig);
2004
2005static const struct perf_event_mmap_page *evlist__pick_pc(struct evlist *evlist)
2006{
2007	if (evlist) {
2008		if (evlist->mmap && evlist->mmap[0].core.base)
2009			return evlist->mmap[0].core.base;
2010		if (evlist->overwrite_mmap && evlist->overwrite_mmap[0].core.base)
2011			return evlist->overwrite_mmap[0].core.base;
2012	}
2013	return NULL;
2014}
2015
2016static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
2017{
2018	const struct perf_event_mmap_page *pc = evlist__pick_pc(rec->evlist);
2019	if (pc)
2020		return pc;
2021	return NULL;
2022}
2023
2024static int record__synthesize(struct record *rec, bool tail)
2025{
2026	struct perf_session *session = rec->session;
2027	struct machine *machine = &session->machines.host;
2028	struct perf_data *data = &rec->data;
2029	struct record_opts *opts = &rec->opts;
2030	struct perf_tool *tool = &rec->tool;
2031	int err = 0;
2032	event_op f = process_synthesized_event;
2033
2034	if (rec->opts.tail_synthesize != tail)
2035		return 0;
2036
2037	if (data->is_pipe) {
2038		err = perf_event__synthesize_for_pipe(tool, session, data,
2039						      process_synthesized_event);
2040		if (err < 0)
2041			goto out;
2042
2043		rec->bytes_written += err;
2044	}
2045
2046	err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
2047					  process_synthesized_event, machine);
2048	if (err)
2049		goto out;
2050
2051	/* Synthesize id_index before auxtrace_info */
2052	err = perf_event__synthesize_id_index(tool,
2053					      process_synthesized_event,
2054					      session->evlist, machine);
2055	if (err)
2056		goto out;
2057
2058	if (rec->opts.full_auxtrace) {
2059		err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
2060					session, process_synthesized_event);
2061		if (err)
2062			goto out;
2063	}
2064
2065	if (!evlist__exclude_kernel(rec->evlist)) {
2066		err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
2067							 machine);
2068		WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
2069				   "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
2070				   "Check /proc/kallsyms permission or run as root.\n");
2071
2072		err = perf_event__synthesize_modules(tool, process_synthesized_event,
2073						     machine);
2074		WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
2075				   "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
2076				   "Check /proc/modules permission or run as root.\n");
2077	}
2078
2079	if (perf_guest) {
2080		machines__process_guests(&session->machines,
2081					 perf_event__synthesize_guest_os, tool);
2082	}
2083
2084	err = perf_event__synthesize_extra_attr(&rec->tool,
2085						rec->evlist,
2086						process_synthesized_event,
2087						data->is_pipe);
2088	if (err)
2089		goto out;
2090
2091	err = perf_event__synthesize_thread_map2(&rec->tool, rec->evlist->core.threads,
2092						 process_synthesized_event,
2093						NULL);
2094	if (err < 0) {
2095		pr_err("Couldn't synthesize thread map.\n");
2096		return err;
2097	}
2098
2099	err = perf_event__synthesize_cpu_map(&rec->tool, rec->evlist->core.all_cpus,
2100					     process_synthesized_event, NULL);
2101	if (err < 0) {
2102		pr_err("Couldn't synthesize cpu map.\n");
2103		return err;
2104	}
2105
2106	err = perf_event__synthesize_bpf_events(session, process_synthesized_event,
2107						machine, opts);
2108	if (err < 0) {
2109		pr_warning("Couldn't synthesize bpf events.\n");
2110		err = 0;
2111	}
2112
2113	if (rec->opts.synth & PERF_SYNTH_CGROUP) {
2114		err = perf_event__synthesize_cgroups(tool, process_synthesized_event,
2115						     machine);
2116		if (err < 0) {
2117			pr_warning("Couldn't synthesize cgroup events.\n");
2118			err = 0;
2119		}
2120	}
2121
2122	if (rec->opts.nr_threads_synthesize > 1) {
2123		mutex_init(&synth_lock);
2124		perf_set_multithreaded();
2125		f = process_locked_synthesized_event;
2126	}
2127
2128	if (rec->opts.synth & PERF_SYNTH_TASK) {
2129		bool needs_mmap = rec->opts.synth & PERF_SYNTH_MMAP;
2130
2131		err = __machine__synthesize_threads(machine, tool, &opts->target,
2132						    rec->evlist->core.threads,
2133						    f, needs_mmap, opts->sample_address,
2134						    rec->opts.nr_threads_synthesize);
2135	}
2136
2137	if (rec->opts.nr_threads_synthesize > 1) {
2138		perf_set_singlethreaded();
2139		mutex_destroy(&synth_lock);
2140	}
2141
2142out:
2143	return err;
2144}
2145
2146static int record__process_signal_event(union perf_event *event __maybe_unused, void *data)
2147{
2148	struct record *rec = data;
2149	pthread_kill(rec->thread_id, SIGUSR2);
2150	return 0;
2151}
2152
2153static int record__setup_sb_evlist(struct record *rec)
2154{
2155	struct record_opts *opts = &rec->opts;
2156
2157	if (rec->sb_evlist != NULL) {
2158		/*
2159		 * We get here if --switch-output-event populated the
2160		 * sb_evlist, so associate a callback that will send a SIGUSR2
2161		 * to the main thread.
2162		 */
2163		evlist__set_cb(rec->sb_evlist, record__process_signal_event, rec);
2164		rec->thread_id = pthread_self();
2165	}
2166#ifdef HAVE_LIBBPF_SUPPORT
2167	if (!opts->no_bpf_event) {
2168		if (rec->sb_evlist == NULL) {
2169			rec->sb_evlist = evlist__new();
2170
2171			if (rec->sb_evlist == NULL) {
2172				pr_err("Couldn't create side band evlist.\n.");
2173				return -1;
2174			}
2175		}
2176
2177		if (evlist__add_bpf_sb_event(rec->sb_evlist, &rec->session->header.env)) {
2178			pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
2179			return -1;
2180		}
2181	}
2182#endif
2183	if (evlist__start_sb_thread(rec->sb_evlist, &rec->opts.target)) {
2184		pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
2185		opts->no_bpf_event = true;
2186	}
2187
2188	return 0;
2189}
2190
2191static int record__init_clock(struct record *rec)
2192{
2193	struct perf_session *session = rec->session;
2194	struct timespec ref_clockid;
2195	struct timeval ref_tod;
2196	u64 ref;
2197
2198	if (!rec->opts.use_clockid)
2199		return 0;
2200
2201	if (rec->opts.use_clockid && rec->opts.clockid_res_ns)
2202		session->header.env.clock.clockid_res_ns = rec->opts.clockid_res_ns;
2203
2204	session->header.env.clock.clockid = rec->opts.clockid;
2205
2206	if (gettimeofday(&ref_tod, NULL) != 0) {
2207		pr_err("gettimeofday failed, cannot set reference time.\n");
2208		return -1;
2209	}
2210
2211	if (clock_gettime(rec->opts.clockid, &ref_clockid)) {
2212		pr_err("clock_gettime failed, cannot set reference time.\n");
2213		return -1;
2214	}
2215
2216	ref = (u64) ref_tod.tv_sec * NSEC_PER_SEC +
2217	      (u64) ref_tod.tv_usec * NSEC_PER_USEC;
2218
2219	session->header.env.clock.tod_ns = ref;
2220
2221	ref = (u64) ref_clockid.tv_sec * NSEC_PER_SEC +
2222	      (u64) ref_clockid.tv_nsec;
2223
2224	session->header.env.clock.clockid_ns = ref;
2225	return 0;
2226}
2227
2228static void hit_auxtrace_snapshot_trigger(struct record *rec)
2229{
2230	if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
2231		trigger_hit(&auxtrace_snapshot_trigger);
2232		auxtrace_record__snapshot_started = 1;
2233		if (auxtrace_record__snapshot_start(rec->itr))
2234			trigger_error(&auxtrace_snapshot_trigger);
2235	}
2236}
2237
2238static int record__terminate_thread(struct record_thread *thread_data)
2239{
2240	int err;
2241	enum thread_msg ack = THREAD_MSG__UNDEFINED;
2242	pid_t tid = thread_data->tid;
2243
2244	close(thread_data->pipes.msg[1]);
2245	thread_data->pipes.msg[1] = -1;
2246	err = read(thread_data->pipes.ack[0], &ack, sizeof(ack));
2247	if (err > 0)
2248		pr_debug2("threads[%d]: sent %s\n", tid, thread_msg_tags[ack]);
2249	else
2250		pr_warning("threads[%d]: failed to receive termination notification from %d\n",
2251			   thread->tid, tid);
2252
2253	return 0;
2254}
2255
2256static int record__start_threads(struct record *rec)
2257{
2258	int t, tt, err, ret = 0, nr_threads = rec->nr_threads;
2259	struct record_thread *thread_data = rec->thread_data;
2260	sigset_t full, mask;
2261	pthread_t handle;
2262	pthread_attr_t attrs;
2263
2264	thread = &thread_data[0];
2265
2266	if (!record__threads_enabled(rec))
2267		return 0;
2268
2269	sigfillset(&full);
2270	if (sigprocmask(SIG_SETMASK, &full, &mask)) {
2271		pr_err("Failed to block signals on threads start: %s\n", strerror(errno));
2272		return -1;
2273	}
2274
2275	pthread_attr_init(&attrs);
2276	pthread_attr_setdetachstate(&attrs, PTHREAD_CREATE_DETACHED);
2277
2278	for (t = 1; t < nr_threads; t++) {
2279		enum thread_msg msg = THREAD_MSG__UNDEFINED;
2280
2281#ifdef HAVE_PTHREAD_ATTR_SETAFFINITY_NP
2282		pthread_attr_setaffinity_np(&attrs,
2283					    MMAP_CPU_MASK_BYTES(&(thread_data[t].mask->affinity)),
2284					    (cpu_set_t *)(thread_data[t].mask->affinity.bits));
2285#endif
2286		if (pthread_create(&handle, &attrs, record__thread, &thread_data[t])) {
2287			for (tt = 1; tt < t; tt++)
2288				record__terminate_thread(&thread_data[t]);
2289			pr_err("Failed to start threads: %s\n", strerror(errno));
2290			ret = -1;
2291			goto out_err;
2292		}
2293
2294		err = read(thread_data[t].pipes.ack[0], &msg, sizeof(msg));
2295		if (err > 0)
2296			pr_debug2("threads[%d]: sent %s\n", rec->thread_data[t].tid,
2297				  thread_msg_tags[msg]);
2298		else
2299			pr_warning("threads[%d]: failed to receive start notification from %d\n",
2300				   thread->tid, rec->thread_data[t].tid);
2301	}
2302
2303	sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&thread->mask->affinity),
2304			(cpu_set_t *)thread->mask->affinity.bits);
2305
2306	pr_debug("threads[%d]: started on cpu%d\n", thread->tid, sched_getcpu());
2307
2308out_err:
2309	pthread_attr_destroy(&attrs);
2310
2311	if (sigprocmask(SIG_SETMASK, &mask, NULL)) {
2312		pr_err("Failed to unblock signals on threads start: %s\n", strerror(errno));
2313		ret = -1;
2314	}
2315
2316	return ret;
2317}
2318
2319static int record__stop_threads(struct record *rec)
2320{
2321	int t;
2322	struct record_thread *thread_data = rec->thread_data;
2323
2324	for (t = 1; t < rec->nr_threads; t++)
2325		record__terminate_thread(&thread_data[t]);
2326
2327	for (t = 0; t < rec->nr_threads; t++) {
2328		rec->samples += thread_data[t].samples;
2329		if (!record__threads_enabled(rec))
2330			continue;
2331		rec->session->bytes_transferred += thread_data[t].bytes_transferred;
2332		rec->session->bytes_compressed += thread_data[t].bytes_compressed;
2333		pr_debug("threads[%d]: samples=%lld, wakes=%ld, ", thread_data[t].tid,
2334			 thread_data[t].samples, thread_data[t].waking);
2335		if (thread_data[t].bytes_transferred && thread_data[t].bytes_compressed)
2336			pr_debug("transferred=%" PRIu64 ", compressed=%" PRIu64 "\n",
2337				 thread_data[t].bytes_transferred, thread_data[t].bytes_compressed);
2338		else
2339			pr_debug("written=%" PRIu64 "\n", thread_data[t].bytes_written);
2340	}
2341
2342	return 0;
2343}
2344
2345static unsigned long record__waking(struct record *rec)
2346{
2347	int t;
2348	unsigned long waking = 0;
2349	struct record_thread *thread_data = rec->thread_data;
2350
2351	for (t = 0; t < rec->nr_threads; t++)
2352		waking += thread_data[t].waking;
2353
2354	return waking;
2355}
2356
2357static int __cmd_record(struct record *rec, int argc, const char **argv)
2358{
2359	int err;
2360	int status = 0;
2361	const bool forks = argc > 0;
2362	struct perf_tool *tool = &rec->tool;
2363	struct record_opts *opts = &rec->opts;
2364	struct perf_data *data = &rec->data;
2365	struct perf_session *session;
2366	bool disabled = false, draining = false;
2367	int fd;
2368	float ratio = 0;
2369	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
2370
2371	atexit(record__sig_exit);
2372	signal(SIGCHLD, sig_handler);
2373	signal(SIGINT, sig_handler);
2374	signal(SIGTERM, sig_handler);
2375	signal(SIGSEGV, sigsegv_handler);
2376
 
 
 
2377	if (rec->opts.record_cgroup) {
2378#ifndef HAVE_FILE_HANDLE
 
 
2379		pr_err("cgroup tracking is not supported\n");
2380		return -1;
2381#endif
2382	}
2383
2384	if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
2385		signal(SIGUSR2, snapshot_sig_handler);
2386		if (rec->opts.auxtrace_snapshot_mode)
2387			trigger_on(&auxtrace_snapshot_trigger);
2388		if (rec->switch_output.enabled)
2389			trigger_on(&switch_output_trigger);
2390	} else {
2391		signal(SIGUSR2, SIG_IGN);
2392	}
2393
2394	perf_tool__init(tool, /*ordered_events=*/true);
2395	tool->sample		= process_sample_event;
2396	tool->fork		= perf_event__process_fork;
2397	tool->exit		= perf_event__process_exit;
2398	tool->comm		= perf_event__process_comm;
2399	tool->namespaces	= perf_event__process_namespaces;
2400	tool->mmap		= build_id__process_mmap;
2401	tool->mmap2		= build_id__process_mmap2;
2402	tool->itrace_start	= process_timestamp_boundary;
2403	tool->aux		= process_timestamp_boundary;
2404	tool->namespace_events	= rec->opts.record_namespaces;
2405	tool->cgroup_events	= rec->opts.record_cgroup;
2406	session = perf_session__new(data, tool);
2407	if (IS_ERR(session)) {
2408		pr_err("Perf session creation failed.\n");
2409		return PTR_ERR(session);
2410	}
2411
2412	if (record__threads_enabled(rec)) {
2413		if (perf_data__is_pipe(&rec->data)) {
2414			pr_err("Parallel trace streaming is not available in pipe mode.\n");
2415			return -1;
2416		}
2417		if (rec->opts.full_auxtrace) {
2418			pr_err("Parallel trace streaming is not available in AUX area tracing mode.\n");
2419			return -1;
2420		}
2421	}
2422
2423	fd = perf_data__fd(data);
2424	rec->session = session;
2425
2426	if (zstd_init(&session->zstd_data, rec->opts.comp_level) < 0) {
2427		pr_err("Compression initialization failed.\n");
2428		return -1;
2429	}
2430#ifdef HAVE_EVENTFD_SUPPORT
2431	done_fd = eventfd(0, EFD_NONBLOCK);
2432	if (done_fd < 0) {
2433		pr_err("Failed to create wakeup eventfd, error: %m\n");
2434		status = -1;
2435		goto out_delete_session;
2436	}
2437	err = evlist__add_wakeup_eventfd(rec->evlist, done_fd);
2438	if (err < 0) {
2439		pr_err("Failed to add wakeup eventfd to poll list\n");
2440		status = err;
2441		goto out_delete_session;
2442	}
2443#endif // HAVE_EVENTFD_SUPPORT
2444
2445	session->header.env.comp_type  = PERF_COMP_ZSTD;
2446	session->header.env.comp_level = rec->opts.comp_level;
2447
2448	if (rec->opts.kcore &&
2449	    !record__kcore_readable(&session->machines.host)) {
2450		pr_err("ERROR: kcore is not readable.\n");
2451		return -1;
2452	}
2453
2454	if (record__init_clock(rec))
2455		return -1;
2456
2457	record__init_features(rec);
2458
2459	if (forks) {
2460		err = evlist__prepare_workload(rec->evlist, &opts->target, argv, data->is_pipe,
2461					       workload_exec_failed_signal);
2462		if (err < 0) {
2463			pr_err("Couldn't run the workload!\n");
2464			status = err;
2465			goto out_delete_session;
2466		}
2467	}
2468
2469	/*
2470	 * If we have just single event and are sending data
2471	 * through pipe, we need to force the ids allocation,
2472	 * because we synthesize event name through the pipe
2473	 * and need the id for that.
2474	 */
2475	if (data->is_pipe && rec->evlist->core.nr_entries == 1)
2476		rec->opts.sample_id = true;
2477
2478	if (rec->timestamp_filename && perf_data__is_pipe(data)) {
2479		rec->timestamp_filename = false;
2480		pr_warning("WARNING: --timestamp-filename option is not available in pipe mode.\n");
2481	}
2482
2483	evlist__uniquify_name(rec->evlist);
2484
2485	evlist__config(rec->evlist, opts, &callchain_param);
2486
2487	/* Debug message used by test scripts */
2488	pr_debug3("perf record opening and mmapping events\n");
2489	if (record__open(rec) != 0) {
2490		err = -1;
2491		goto out_free_threads;
2492	}
2493	/* Debug message used by test scripts */
2494	pr_debug3("perf record done opening and mmapping events\n");
2495	session->header.env.comp_mmap_len = session->evlist->core.mmap_len;
2496
2497	if (rec->opts.kcore) {
2498		err = record__kcore_copy(&session->machines.host, data);
2499		if (err) {
2500			pr_err("ERROR: Failed to copy kcore\n");
2501			goto out_free_threads;
2502		}
2503	}
2504
2505	/*
2506	 * Normally perf_session__new would do this, but it doesn't have the
2507	 * evlist.
2508	 */
2509	if (rec->tool.ordered_events && !evlist__sample_id_all(rec->evlist)) {
2510		pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
2511		rec->tool.ordered_events = false;
2512	}
2513
2514	if (evlist__nr_groups(rec->evlist) == 0)
2515		perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
2516
2517	if (data->is_pipe) {
2518		err = perf_header__write_pipe(fd);
2519		if (err < 0)
2520			goto out_free_threads;
2521	} else {
2522		err = perf_session__write_header(session, rec->evlist, fd, false);
2523		if (err < 0)
2524			goto out_free_threads;
2525	}
2526
2527	err = -1;
2528	if (!rec->no_buildid
2529	    && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
2530		pr_err("Couldn't generate buildids. "
2531		       "Use --no-buildid to profile anyway.\n");
2532		goto out_free_threads;
2533	}
2534
2535	err = record__setup_sb_evlist(rec);
2536	if (err)
2537		goto out_free_threads;
2538
2539	err = record__synthesize(rec, false);
2540	if (err < 0)
2541		goto out_free_threads;
2542
2543	if (rec->realtime_prio) {
2544		struct sched_param param;
2545
2546		param.sched_priority = rec->realtime_prio;
2547		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
2548			pr_err("Could not set realtime priority.\n");
2549			err = -1;
2550			goto out_free_threads;
2551		}
2552	}
2553
2554	if (record__start_threads(rec))
2555		goto out_free_threads;
2556
2557	/*
2558	 * When perf is starting the traced process, all the events
2559	 * (apart from group members) have enable_on_exec=1 set,
2560	 * so don't spoil it by prematurely enabling them.
2561	 */
2562	if (!target__none(&opts->target) && !opts->target.initial_delay)
2563		evlist__enable(rec->evlist);
2564
2565	/*
2566	 * Let the child rip
2567	 */
2568	if (forks) {
2569		struct machine *machine = &session->machines.host;
2570		union perf_event *event;
2571		pid_t tgid;
2572
2573		event = malloc(sizeof(event->comm) + machine->id_hdr_size);
2574		if (event == NULL) {
2575			err = -ENOMEM;
2576			goto out_child;
2577		}
2578
2579		/*
2580		 * Some H/W events are generated before COMM event
2581		 * which is emitted during exec(), so perf script
2582		 * cannot see a correct process name for those events.
2583		 * Synthesize COMM event to prevent it.
2584		 */
2585		tgid = perf_event__synthesize_comm(tool, event,
2586						   rec->evlist->workload.pid,
2587						   process_synthesized_event,
2588						   machine);
2589		free(event);
2590
2591		if (tgid == -1)
2592			goto out_child;
2593
2594		event = malloc(sizeof(event->namespaces) +
2595			       (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
2596			       machine->id_hdr_size);
2597		if (event == NULL) {
2598			err = -ENOMEM;
2599			goto out_child;
2600		}
2601
2602		/*
2603		 * Synthesize NAMESPACES event for the command specified.
2604		 */
2605		perf_event__synthesize_namespaces(tool, event,
2606						  rec->evlist->workload.pid,
2607						  tgid, process_synthesized_event,
2608						  machine);
2609		free(event);
2610
2611		evlist__start_workload(rec->evlist);
2612	}
2613
2614	if (opts->target.initial_delay) {
2615		pr_info(EVLIST_DISABLED_MSG);
2616		if (opts->target.initial_delay > 0) {
2617			usleep(opts->target.initial_delay * USEC_PER_MSEC);
2618			evlist__enable(rec->evlist);
2619			pr_info(EVLIST_ENABLED_MSG);
2620		}
2621	}
2622
2623	err = event_enable_timer__start(rec->evlist->eet);
2624	if (err)
2625		goto out_child;
2626
2627	/* Debug message used by test scripts */
2628	pr_debug3("perf record has started\n");
2629	fflush(stderr);
2630
2631	trigger_ready(&auxtrace_snapshot_trigger);
2632	trigger_ready(&switch_output_trigger);
2633	perf_hooks__invoke_record_start();
2634
2635	/*
2636	 * Must write FINISHED_INIT so it will be seen after all other
2637	 * synthesized user events, but before any regular events.
2638	 */
2639	err = write_finished_init(rec, false);
2640	if (err < 0)
2641		goto out_child;
2642
2643	for (;;) {
2644		unsigned long long hits = thread->samples;
2645
2646		/*
2647		 * rec->evlist->bkw_mmap_state is possible to be
2648		 * BKW_MMAP_EMPTY here: when done == true and
2649		 * hits != rec->samples in previous round.
2650		 *
2651		 * evlist__toggle_bkw_mmap ensure we never
2652		 * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
2653		 */
2654		if (trigger_is_hit(&switch_output_trigger) || done || draining)
2655			evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
2656
2657		if (record__mmap_read_all(rec, false) < 0) {
2658			trigger_error(&auxtrace_snapshot_trigger);
2659			trigger_error(&switch_output_trigger);
2660			err = -1;
2661			goto out_child;
2662		}
2663
2664		if (auxtrace_record__snapshot_started) {
2665			auxtrace_record__snapshot_started = 0;
2666			if (!trigger_is_error(&auxtrace_snapshot_trigger))
2667				record__read_auxtrace_snapshot(rec, false);
2668			if (trigger_is_error(&auxtrace_snapshot_trigger)) {
2669				pr_err("AUX area tracing snapshot failed\n");
2670				err = -1;
2671				goto out_child;
2672			}
2673		}
2674
2675		if (trigger_is_hit(&switch_output_trigger)) {
2676			/*
2677			 * If switch_output_trigger is hit, the data in
2678			 * overwritable ring buffer should have been collected,
2679			 * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
2680			 *
2681			 * If SIGUSR2 raise after or during record__mmap_read_all(),
2682			 * record__mmap_read_all() didn't collect data from
2683			 * overwritable ring buffer. Read again.
2684			 */
2685			if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
2686				continue;
2687			trigger_ready(&switch_output_trigger);
2688
2689			/*
2690			 * Reenable events in overwrite ring buffer after
2691			 * record__mmap_read_all(): we should have collected
2692			 * data from it.
2693			 */
2694			evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
2695
2696			if (!quiet)
2697				fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
2698					record__waking(rec));
2699			thread->waking = 0;
2700			fd = record__switch_output(rec, false);
2701			if (fd < 0) {
2702				pr_err("Failed to switch to new file\n");
2703				trigger_error(&switch_output_trigger);
2704				err = fd;
2705				goto out_child;
2706			}
2707
2708			/* re-arm the alarm */
2709			if (rec->switch_output.time)
2710				alarm(rec->switch_output.time);
2711		}
2712
2713		if (hits == thread->samples) {
2714			if (done || draining)
2715				break;
2716			err = fdarray__poll(&thread->pollfd, -1);
2717			/*
2718			 * Propagate error, only if there's any. Ignore positive
2719			 * number of returned events and interrupt error.
2720			 */
2721			if (err > 0 || (err < 0 && errno == EINTR))
2722				err = 0;
2723			thread->waking++;
2724
2725			if (fdarray__filter(&thread->pollfd, POLLERR | POLLHUP,
2726					    record__thread_munmap_filtered, NULL) == 0)
2727				draining = true;
2728
2729			err = record__update_evlist_pollfd_from_thread(rec, rec->evlist, thread);
2730			if (err)
2731				goto out_child;
2732		}
2733
2734		if (evlist__ctlfd_process(rec->evlist, &cmd) > 0) {
2735			switch (cmd) {
2736			case EVLIST_CTL_CMD_SNAPSHOT:
2737				hit_auxtrace_snapshot_trigger(rec);
2738				evlist__ctlfd_ack(rec->evlist);
2739				break;
2740			case EVLIST_CTL_CMD_STOP:
2741				done = 1;
2742				break;
2743			case EVLIST_CTL_CMD_ACK:
2744			case EVLIST_CTL_CMD_UNSUPPORTED:
2745			case EVLIST_CTL_CMD_ENABLE:
2746			case EVLIST_CTL_CMD_DISABLE:
2747			case EVLIST_CTL_CMD_EVLIST:
2748			case EVLIST_CTL_CMD_PING:
2749			default:
2750				break;
2751			}
2752		}
2753
2754		err = event_enable_timer__process(rec->evlist->eet);
2755		if (err < 0)
2756			goto out_child;
2757		if (err) {
2758			err = 0;
2759			done = 1;
2760		}
2761
2762		/*
2763		 * When perf is starting the traced process, at the end events
2764		 * die with the process and we wait for that. Thus no need to
2765		 * disable events in this case.
2766		 */
2767		if (done && !disabled && !target__none(&opts->target)) {
2768			trigger_off(&auxtrace_snapshot_trigger);
2769			evlist__disable(rec->evlist);
2770			disabled = true;
2771		}
2772	}
2773
2774	trigger_off(&auxtrace_snapshot_trigger);
2775	trigger_off(&switch_output_trigger);
2776
2777	if (opts->auxtrace_snapshot_on_exit)
2778		record__auxtrace_snapshot_exit(rec);
2779
2780	if (forks && workload_exec_errno) {
2781		char msg[STRERR_BUFSIZE], strevsels[2048];
2782		const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
2783
2784		evlist__scnprintf_evsels(rec->evlist, sizeof(strevsels), strevsels);
2785
2786		pr_err("Failed to collect '%s' for the '%s' workload: %s\n",
2787			strevsels, argv[0], emsg);
2788		err = -1;
2789		goto out_child;
2790	}
2791
2792	if (!quiet)
2793		fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n",
2794			record__waking(rec));
2795
2796	write_finished_init(rec, true);
2797
2798	if (target__none(&rec->opts.target))
2799		record__synthesize_workload(rec, true);
2800
2801out_child:
2802	record__stop_threads(rec);
2803	record__mmap_read_all(rec, true);
2804out_free_threads:
2805	record__free_thread_data(rec);
2806	evlist__finalize_ctlfd(rec->evlist);
2807	record__aio_mmap_read_sync(rec);
2808
2809	if (rec->session->bytes_transferred && rec->session->bytes_compressed) {
2810		ratio = (float)rec->session->bytes_transferred/(float)rec->session->bytes_compressed;
2811		session->header.env.comp_ratio = ratio + 0.5;
2812	}
2813
2814	if (forks) {
2815		int exit_status;
2816
2817		if (!child_finished)
2818			kill(rec->evlist->workload.pid, SIGTERM);
2819
2820		wait(&exit_status);
2821
2822		if (err < 0)
2823			status = err;
2824		else if (WIFEXITED(exit_status))
2825			status = WEXITSTATUS(exit_status);
2826		else if (WIFSIGNALED(exit_status))
2827			signr = WTERMSIG(exit_status);
2828	} else
2829		status = err;
2830
2831	if (rec->off_cpu)
2832		rec->bytes_written += off_cpu_write(rec->session);
2833
2834	record__read_lost_samples(rec);
2835	record__synthesize(rec, true);
2836	/* this will be recalculated during process_buildids() */
2837	rec->samples = 0;
2838
2839	if (!err) {
2840		if (!rec->timestamp_filename) {
2841			record__finish_output(rec);
2842		} else {
2843			fd = record__switch_output(rec, true);
2844			if (fd < 0) {
2845				status = fd;
2846				goto out_delete_session;
2847			}
2848		}
2849	}
2850
2851	perf_hooks__invoke_record_end();
2852
2853	if (!err && !quiet) {
2854		char samples[128];
2855		const char *postfix = rec->timestamp_filename ?
2856					".<timestamp>" : "";
2857
2858		if (rec->samples && !rec->opts.full_auxtrace)
2859			scnprintf(samples, sizeof(samples),
2860				  " (%" PRIu64 " samples)", rec->samples);
2861		else
2862			samples[0] = '\0';
2863
2864		fprintf(stderr,	"[ perf record: Captured and wrote %.3f MB %s%s%s",
2865			perf_data__size(data) / 1024.0 / 1024.0,
2866			data->path, postfix, samples);
2867		if (ratio) {
2868			fprintf(stderr,	", compressed (original %.3f MB, ratio is %.3f)",
2869					rec->session->bytes_transferred / 1024.0 / 1024.0,
2870					ratio);
2871		}
2872		fprintf(stderr, " ]\n");
2873	}
2874
2875out_delete_session:
2876#ifdef HAVE_EVENTFD_SUPPORT
2877	if (done_fd >= 0) {
2878		fd = done_fd;
2879		done_fd = -1;
2880
2881		close(fd);
2882	}
2883#endif
2884	zstd_fini(&session->zstd_data);
2885	if (!opts->no_bpf_event)
2886		evlist__stop_sb_thread(rec->sb_evlist);
2887
2888	perf_session__delete(session);
2889	return status;
2890}
2891
2892static void callchain_debug(struct callchain_param *callchain)
2893{
2894	static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
2895
2896	pr_debug("callchain: type %s\n", str[callchain->record_mode]);
2897
2898	if (callchain->record_mode == CALLCHAIN_DWARF)
2899		pr_debug("callchain: stack dump size %d\n",
2900			 callchain->dump_size);
2901}
2902
2903int record_opts__parse_callchain(struct record_opts *record,
2904				 struct callchain_param *callchain,
2905				 const char *arg, bool unset)
2906{
2907	int ret;
2908	callchain->enabled = !unset;
2909
2910	/* --no-call-graph */
2911	if (unset) {
2912		callchain->record_mode = CALLCHAIN_NONE;
2913		pr_debug("callchain: disabled\n");
2914		return 0;
2915	}
2916
2917	ret = parse_callchain_record_opt(arg, callchain);
2918	if (!ret) {
2919		/* Enable data address sampling for DWARF unwind. */
2920		if (callchain->record_mode == CALLCHAIN_DWARF)
2921			record->sample_address = true;
2922		callchain_debug(callchain);
2923	}
2924
2925	return ret;
2926}
2927
2928int record_parse_callchain_opt(const struct option *opt,
2929			       const char *arg,
2930			       int unset)
2931{
2932	return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
2933}
2934
2935int record_callchain_opt(const struct option *opt,
2936			 const char *arg __maybe_unused,
2937			 int unset __maybe_unused)
2938{
2939	struct callchain_param *callchain = opt->value;
2940
2941	callchain->enabled = true;
2942
2943	if (callchain->record_mode == CALLCHAIN_NONE)
2944		callchain->record_mode = CALLCHAIN_FP;
2945
2946	callchain_debug(callchain);
2947	return 0;
2948}
2949
2950static int perf_record_config(const char *var, const char *value, void *cb)
2951{
2952	struct record *rec = cb;
2953
2954	if (!strcmp(var, "record.build-id")) {
2955		if (!strcmp(value, "cache"))
2956			rec->no_buildid_cache = false;
2957		else if (!strcmp(value, "no-cache"))
2958			rec->no_buildid_cache = true;
2959		else if (!strcmp(value, "skip"))
2960			rec->no_buildid = true;
2961		else if (!strcmp(value, "mmap"))
2962			rec->buildid_mmap = true;
2963		else
2964			return -1;
2965		return 0;
2966	}
2967	if (!strcmp(var, "record.call-graph")) {
2968		var = "call-graph.record-mode";
2969		return perf_default_config(var, value, cb);
2970	}
2971#ifdef HAVE_AIO_SUPPORT
2972	if (!strcmp(var, "record.aio")) {
2973		rec->opts.nr_cblocks = strtol(value, NULL, 0);
2974		if (!rec->opts.nr_cblocks)
2975			rec->opts.nr_cblocks = nr_cblocks_default;
2976	}
2977#endif
2978	if (!strcmp(var, "record.debuginfod")) {
2979		rec->debuginfod.urls = strdup(value);
2980		if (!rec->debuginfod.urls)
2981			return -ENOMEM;
2982		rec->debuginfod.set = true;
2983	}
2984
2985	return 0;
2986}
2987
2988static int record__parse_event_enable_time(const struct option *opt, const char *str, int unset)
2989{
2990	struct record *rec = (struct record *)opt->value;
2991
2992	return evlist__parse_event_enable_time(rec->evlist, &rec->opts, str, unset);
2993}
2994
2995static int record__parse_affinity(const struct option *opt, const char *str, int unset)
2996{
2997	struct record_opts *opts = (struct record_opts *)opt->value;
2998
2999	if (unset || !str)
3000		return 0;
3001
3002	if (!strcasecmp(str, "node"))
3003		opts->affinity = PERF_AFFINITY_NODE;
3004	else if (!strcasecmp(str, "cpu"))
3005		opts->affinity = PERF_AFFINITY_CPU;
3006
3007	return 0;
3008}
3009
3010static int record__mmap_cpu_mask_alloc(struct mmap_cpu_mask *mask, int nr_bits)
3011{
3012	mask->nbits = nr_bits;
3013	mask->bits = bitmap_zalloc(mask->nbits);
3014	if (!mask->bits)
3015		return -ENOMEM;
3016
3017	return 0;
3018}
3019
3020static void record__mmap_cpu_mask_free(struct mmap_cpu_mask *mask)
3021{
3022	bitmap_free(mask->bits);
3023	mask->nbits = 0;
3024}
3025
3026static int record__thread_mask_alloc(struct thread_mask *mask, int nr_bits)
3027{
3028	int ret;
3029
3030	ret = record__mmap_cpu_mask_alloc(&mask->maps, nr_bits);
3031	if (ret) {
3032		mask->affinity.bits = NULL;
3033		return ret;
3034	}
3035
3036	ret = record__mmap_cpu_mask_alloc(&mask->affinity, nr_bits);
3037	if (ret) {
3038		record__mmap_cpu_mask_free(&mask->maps);
3039		mask->maps.bits = NULL;
3040	}
3041
3042	return ret;
3043}
3044
3045static void record__thread_mask_free(struct thread_mask *mask)
3046{
3047	record__mmap_cpu_mask_free(&mask->maps);
3048	record__mmap_cpu_mask_free(&mask->affinity);
3049}
3050
3051static int record__parse_threads(const struct option *opt, const char *str, int unset)
3052{
3053	int s;
3054	struct record_opts *opts = opt->value;
3055
3056	if (unset || !str || !strlen(str)) {
3057		opts->threads_spec = THREAD_SPEC__CPU;
3058	} else {
3059		for (s = 1; s < THREAD_SPEC__MAX; s++) {
3060			if (s == THREAD_SPEC__USER) {
3061				opts->threads_user_spec = strdup(str);
3062				if (!opts->threads_user_spec)
3063					return -ENOMEM;
3064				opts->threads_spec = THREAD_SPEC__USER;
3065				break;
3066			}
3067			if (!strncasecmp(str, thread_spec_tags[s], strlen(thread_spec_tags[s]))) {
3068				opts->threads_spec = s;
3069				break;
3070			}
3071		}
3072	}
3073
3074	if (opts->threads_spec == THREAD_SPEC__USER)
3075		pr_debug("threads_spec: %s\n", opts->threads_user_spec);
3076	else
3077		pr_debug("threads_spec: %s\n", thread_spec_tags[opts->threads_spec]);
3078
3079	return 0;
3080}
3081
3082static int parse_output_max_size(const struct option *opt,
3083				 const char *str, int unset)
3084{
3085	unsigned long *s = (unsigned long *)opt->value;
3086	static struct parse_tag tags_size[] = {
3087		{ .tag  = 'B', .mult = 1       },
3088		{ .tag  = 'K', .mult = 1 << 10 },
3089		{ .tag  = 'M', .mult = 1 << 20 },
3090		{ .tag  = 'G', .mult = 1 << 30 },
3091		{ .tag  = 0 },
3092	};
3093	unsigned long val;
3094
3095	if (unset) {
3096		*s = 0;
3097		return 0;
3098	}
3099
3100	val = parse_tag_value(str, tags_size);
3101	if (val != (unsigned long) -1) {
3102		*s = val;
3103		return 0;
3104	}
3105
3106	return -1;
3107}
3108
3109static int record__parse_mmap_pages(const struct option *opt,
3110				    const char *str,
3111				    int unset __maybe_unused)
3112{
3113	struct record_opts *opts = opt->value;
3114	char *s, *p;
3115	unsigned int mmap_pages;
3116	int ret;
3117
3118	if (!str)
3119		return -EINVAL;
3120
3121	s = strdup(str);
3122	if (!s)
3123		return -ENOMEM;
3124
3125	p = strchr(s, ',');
3126	if (p)
3127		*p = '\0';
3128
3129	if (*s) {
3130		ret = __evlist__parse_mmap_pages(&mmap_pages, s);
3131		if (ret)
3132			goto out_free;
3133		opts->mmap_pages = mmap_pages;
3134	}
3135
3136	if (!p) {
3137		ret = 0;
3138		goto out_free;
3139	}
3140
3141	ret = __evlist__parse_mmap_pages(&mmap_pages, p + 1);
3142	if (ret)
3143		goto out_free;
3144
3145	opts->auxtrace_mmap_pages = mmap_pages;
3146
3147out_free:
3148	free(s);
3149	return ret;
3150}
3151
3152void __weak arch__add_leaf_frame_record_opts(struct record_opts *opts __maybe_unused)
3153{
3154}
3155
3156static int parse_control_option(const struct option *opt,
3157				const char *str,
3158				int unset __maybe_unused)
3159{
3160	struct record_opts *opts = opt->value;
3161
3162	return evlist__parse_control(str, &opts->ctl_fd, &opts->ctl_fd_ack, &opts->ctl_fd_close);
3163}
3164
3165static void switch_output_size_warn(struct record *rec)
3166{
3167	u64 wakeup_size = evlist__mmap_size(rec->opts.mmap_pages);
3168	struct switch_output *s = &rec->switch_output;
3169
3170	wakeup_size /= 2;
3171
3172	if (s->size < wakeup_size) {
3173		char buf[100];
3174
3175		unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
3176		pr_warning("WARNING: switch-output data size lower than "
3177			   "wakeup kernel buffer size (%s) "
3178			   "expect bigger perf.data sizes\n", buf);
3179	}
3180}
3181
3182static int switch_output_setup(struct record *rec)
3183{
3184	struct switch_output *s = &rec->switch_output;
3185	static struct parse_tag tags_size[] = {
3186		{ .tag  = 'B', .mult = 1       },
3187		{ .tag  = 'K', .mult = 1 << 10 },
3188		{ .tag  = 'M', .mult = 1 << 20 },
3189		{ .tag  = 'G', .mult = 1 << 30 },
3190		{ .tag  = 0 },
3191	};
3192	static struct parse_tag tags_time[] = {
3193		{ .tag  = 's', .mult = 1        },
3194		{ .tag  = 'm', .mult = 60       },
3195		{ .tag  = 'h', .mult = 60*60    },
3196		{ .tag  = 'd', .mult = 60*60*24 },
3197		{ .tag  = 0 },
3198	};
3199	unsigned long val;
3200
3201	/*
3202	 * If we're using --switch-output-events, then we imply its
3203	 * --switch-output=signal, as we'll send a SIGUSR2 from the side band
3204	 *  thread to its parent.
3205	 */
3206	if (rec->switch_output_event_set) {
3207		if (record__threads_enabled(rec)) {
3208			pr_warning("WARNING: --switch-output-event option is not available in parallel streaming mode.\n");
3209			return 0;
3210		}
3211		goto do_signal;
3212	}
3213
3214	if (!s->set)
3215		return 0;
3216
3217	if (record__threads_enabled(rec)) {
3218		pr_warning("WARNING: --switch-output option is not available in parallel streaming mode.\n");
3219		return 0;
3220	}
3221
3222	if (!strcmp(s->str, "signal")) {
3223do_signal:
3224		s->signal = true;
3225		pr_debug("switch-output with SIGUSR2 signal\n");
3226		goto enabled;
3227	}
3228
3229	val = parse_tag_value(s->str, tags_size);
3230	if (val != (unsigned long) -1) {
3231		s->size = val;
3232		pr_debug("switch-output with %s size threshold\n", s->str);
3233		goto enabled;
3234	}
3235
3236	val = parse_tag_value(s->str, tags_time);
3237	if (val != (unsigned long) -1) {
3238		s->time = val;
3239		pr_debug("switch-output with %s time threshold (%lu seconds)\n",
3240			 s->str, s->time);
3241		goto enabled;
3242	}
3243
3244	return -1;
3245
3246enabled:
3247	rec->timestamp_filename = true;
3248	s->enabled              = true;
3249
3250	if (s->size && !rec->opts.no_buffering)
3251		switch_output_size_warn(rec);
3252
3253	return 0;
3254}
3255
3256static const char * const __record_usage[] = {
3257	"perf record [<options>] [<command>]",
3258	"perf record [<options>] -- <command> [<options>]",
3259	NULL
3260};
3261const char * const *record_usage = __record_usage;
3262
3263static int build_id__process_mmap(const struct perf_tool *tool, union perf_event *event,
3264				  struct perf_sample *sample, struct machine *machine)
3265{
3266	/*
3267	 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
3268	 * no need to add them twice.
3269	 */
3270	if (!(event->header.misc & PERF_RECORD_MISC_USER))
3271		return 0;
3272	return perf_event__process_mmap(tool, event, sample, machine);
3273}
3274
3275static int build_id__process_mmap2(const struct perf_tool *tool, union perf_event *event,
3276				   struct perf_sample *sample, struct machine *machine)
3277{
3278	/*
3279	 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
3280	 * no need to add them twice.
3281	 */
3282	if (!(event->header.misc & PERF_RECORD_MISC_USER))
3283		return 0;
3284
3285	return perf_event__process_mmap2(tool, event, sample, machine);
3286}
3287
3288static int process_timestamp_boundary(const struct perf_tool *tool,
3289				      union perf_event *event __maybe_unused,
3290				      struct perf_sample *sample,
3291				      struct machine *machine __maybe_unused)
3292{
3293	struct record *rec = container_of(tool, struct record, tool);
3294
3295	set_timestamp_boundary(rec, sample->time);
3296	return 0;
3297}
3298
3299static int parse_record_synth_option(const struct option *opt,
3300				     const char *str,
3301				     int unset __maybe_unused)
3302{
3303	struct record_opts *opts = opt->value;
3304	char *p = strdup(str);
3305
3306	if (p == NULL)
3307		return -1;
3308
3309	opts->synth = parse_synth_opt(p);
3310	free(p);
3311
3312	if (opts->synth < 0) {
3313		pr_err("Invalid synth option: %s\n", str);
3314		return -1;
3315	}
3316	return 0;
3317}
3318
3319/*
3320 * XXX Ideally would be local to cmd_record() and passed to a record__new
3321 * because we need to have access to it in record__exit, that is called
3322 * after cmd_record() exits, but since record_options need to be accessible to
3323 * builtin-script, leave it here.
3324 *
3325 * At least we don't ouch it in all the other functions here directly.
3326 *
3327 * Just say no to tons of global variables, sigh.
3328 */
3329static struct record record = {
3330	.opts = {
3331		.sample_time	     = true,
3332		.mmap_pages	     = UINT_MAX,
3333		.user_freq	     = UINT_MAX,
3334		.user_interval	     = ULLONG_MAX,
3335		.freq		     = 4000,
3336		.target		     = {
3337			.uses_mmap   = true,
3338			.default_per_cpu = true,
3339		},
3340		.mmap_flush          = MMAP_FLUSH_DEFAULT,
3341		.nr_threads_synthesize = 1,
3342		.ctl_fd              = -1,
3343		.ctl_fd_ack          = -1,
3344		.synth               = PERF_SYNTH_ALL,
3345	},
 
 
 
 
 
 
 
 
 
 
 
 
3346};
3347
3348const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
3349	"\n\t\t\t\tDefault: fp";
3350
3351static bool dry_run;
3352
3353static struct parse_events_option_args parse_events_option_args = {
3354	.evlistp = &record.evlist,
3355};
3356
3357static struct parse_events_option_args switch_output_parse_events_option_args = {
3358	.evlistp = &record.sb_evlist,
3359};
3360
3361/*
3362 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
3363 * with it and switch to use the library functions in perf_evlist that came
3364 * from builtin-record.c, i.e. use record_opts,
3365 * evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
3366 * using pipes, etc.
3367 */
3368static struct option __record_options[] = {
3369	OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
3370		     "event selector. use 'perf list' to list available events",
3371		     parse_events_option),
3372	OPT_CALLBACK(0, "filter", &record.evlist, "filter",
3373		     "event filter", parse_filter),
3374	OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
3375			   NULL, "don't record events from perf itself",
3376			   exclude_perf),
3377	OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
3378		    "record events on existing process id"),
3379	OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
3380		    "record events on existing thread id"),
3381	OPT_INTEGER('r', "realtime", &record.realtime_prio,
3382		    "collect data with this RT SCHED_FIFO priority"),
3383	OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
3384		    "collect data without buffering"),
3385	OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
3386		    "collect raw sample records from all opened counters"),
3387	OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
3388			    "system-wide collection from all CPUs"),
3389	OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
3390		    "list of cpus to monitor"),
3391	OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
3392	OPT_STRING('o', "output", &record.data.path, "file",
3393		    "output file name"),
3394	OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
3395			&record.opts.no_inherit_set,
3396			"child tasks do not inherit counters"),
3397	OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
3398		    "synthesize non-sample events at the end of output"),
3399	OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
3400	OPT_BOOLEAN(0, "no-bpf-event", &record.opts.no_bpf_event, "do not record bpf events"),
3401	OPT_BOOLEAN(0, "strict-freq", &record.opts.strict_freq,
3402		    "Fail if the specified frequency can't be used"),
3403	OPT_CALLBACK('F', "freq", &record.opts, "freq or 'max'",
3404		     "profile at this frequency",
3405		      record__parse_freq),
3406	OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
3407		     "number of mmap data pages and AUX area tracing mmap pages",
3408		     record__parse_mmap_pages),
3409	OPT_CALLBACK(0, "mmap-flush", &record.opts, "number",
3410		     "Minimal number of bytes that is extracted from mmap data pages (default: 1)",
3411		     record__mmap_flush_parse),
3412	OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
3413			   NULL, "enables call-graph recording" ,
3414			   &record_callchain_opt),
3415	OPT_CALLBACK(0, "call-graph", &record.opts,
3416		     "record_mode[,record_size]", record_callchain_help,
3417		     &record_parse_callchain_opt),
3418	OPT_INCR('v', "verbose", &verbose,
3419		    "be more verbose (show counter open errors, etc)"),
3420	OPT_BOOLEAN('q', "quiet", &quiet, "don't print any warnings or messages"),
3421	OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
3422		    "per thread counts"),
3423	OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
3424	OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
3425		    "Record the sample physical addresses"),
3426	OPT_BOOLEAN(0, "data-page-size", &record.opts.sample_data_page_size,
3427		    "Record the sampled data address data page size"),
3428	OPT_BOOLEAN(0, "code-page-size", &record.opts.sample_code_page_size,
3429		    "Record the sampled code address (ip) page size"),
3430	OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
3431	OPT_BOOLEAN(0, "sample-identifier", &record.opts.sample_identifier,
3432		    "Record the sample identifier"),
3433	OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
3434			&record.opts.sample_time_set,
3435			"Record the sample timestamps"),
3436	OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
3437			"Record the sample period"),
3438	OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
3439		    "don't sample"),
3440	OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
3441			&record.no_buildid_cache_set,
3442			"do not update the buildid cache"),
3443	OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
3444			&record.no_buildid_set,
3445			"do not collect buildids in perf.data"),
3446	OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
3447		     "monitor event in cgroup name only",
3448		     parse_cgroups),
3449	OPT_CALLBACK('D', "delay", &record, "ms",
3450		     "ms to wait before starting measurement after program start (-1: start with events disabled), "
3451		     "or ranges of time to enable events e.g. '-D 10-20,30-40'",
3452		     record__parse_event_enable_time),
3453	OPT_BOOLEAN(0, "kcore", &record.opts.kcore, "copy /proc/kcore"),
3454	OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
3455		   "user to profile"),
3456
3457	OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
3458		     "branch any", "sample any taken branches",
3459		     parse_branch_stack),
3460
3461	OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
3462		     "branch filter mask", "branch stack filter modes",
3463		     parse_branch_stack),
3464	OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
3465		    "sample by weight (on special events only)"),
3466	OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
3467		    "sample transaction flags (special events only)"),
3468	OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
3469		    "use per-thread mmaps"),
3470	OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
3471		    "sample selected machine registers on interrupt,"
3472		    " use '-I?' to list register names", parse_intr_regs),
3473	OPT_CALLBACK_OPTARG(0, "user-regs", &record.opts.sample_user_regs, NULL, "any register",
3474		    "sample selected machine registers on interrupt,"
3475		    " use '--user-regs=?' to list register names", parse_user_regs),
3476	OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
3477		    "Record running/enabled time of read (:S) events"),
3478	OPT_CALLBACK('k', "clockid", &record.opts,
3479	"clockid", "clockid to use for events, see clock_gettime()",
3480	parse_clockid),
3481	OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
3482			  "opts", "AUX area tracing Snapshot Mode", ""),
3483	OPT_STRING_OPTARG(0, "aux-sample", &record.opts.auxtrace_sample_opts,
3484			  "opts", "sample AUX area", ""),
3485	OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3486			"per thread proc mmap processing timeout in ms"),
3487	OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
3488		    "Record namespaces events"),
3489	OPT_BOOLEAN(0, "all-cgroups", &record.opts.record_cgroup,
3490		    "Record cgroup events"),
3491	OPT_BOOLEAN_SET(0, "switch-events", &record.opts.record_switch_events,
3492			&record.opts.record_switch_events_set,
3493			"Record context switch events"),
3494	OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
3495			 "Configure all used events to run in kernel space.",
3496			 PARSE_OPT_EXCLUSIVE),
3497	OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
3498			 "Configure all used events to run in user space.",
3499			 PARSE_OPT_EXCLUSIVE),
3500	OPT_BOOLEAN(0, "kernel-callchains", &record.opts.kernel_callchains,
3501		    "collect kernel callchains"),
3502	OPT_BOOLEAN(0, "user-callchains", &record.opts.user_callchains,
3503		    "collect user callchains"),
3504	OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
3505		   "file", "vmlinux pathname"),
3506	OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
3507		    "Record build-id of all DSOs regardless of hits"),
3508	OPT_BOOLEAN(0, "buildid-mmap", &record.buildid_mmap,
3509		    "Record build-id in map events"),
3510	OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
3511		    "append timestamp to output filename"),
3512	OPT_BOOLEAN(0, "timestamp-boundary", &record.timestamp_boundary,
3513		    "Record timestamp boundary (time of first/last samples)"),
3514	OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
3515			  &record.switch_output.set, "signal or size[BKMG] or time[smhd]",
3516			  "Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold",
3517			  "signal"),
3518	OPT_CALLBACK_SET(0, "switch-output-event", &switch_output_parse_events_option_args,
3519			 &record.switch_output_event_set, "switch output event",
3520			 "switch output event selector. use 'perf list' to list available events",
3521			 parse_events_option_new_evlist),
3522	OPT_INTEGER(0, "switch-max-files", &record.switch_output.num_files,
3523		   "Limit number of switch output generated files"),
3524	OPT_BOOLEAN(0, "dry-run", &dry_run,
3525		    "Parse options then exit"),
3526#ifdef HAVE_AIO_SUPPORT
3527	OPT_CALLBACK_OPTARG(0, "aio", &record.opts,
3528		     &nr_cblocks_default, "n", "Use <n> control blocks in asynchronous trace writing mode (default: 1, max: 4)",
3529		     record__aio_parse),
3530#endif
3531	OPT_CALLBACK(0, "affinity", &record.opts, "node|cpu",
3532		     "Set affinity mask of trace reading thread to NUMA node cpu mask or cpu of processed mmap buffer",
3533		     record__parse_affinity),
3534#ifdef HAVE_ZSTD_SUPPORT
3535	OPT_CALLBACK_OPTARG('z', "compression-level", &record.opts, &comp_level_default, "n",
3536			    "Compress records using specified level (default: 1 - fastest compression, 22 - greatest compression)",
3537			    record__parse_comp_level),
3538#endif
3539	OPT_CALLBACK(0, "max-size", &record.output_max_size,
3540		     "size", "Limit the maximum size of the output file", parse_output_max_size),
3541	OPT_UINTEGER(0, "num-thread-synthesize",
3542		     &record.opts.nr_threads_synthesize,
3543		     "number of threads to run for event synthesis"),
3544#ifdef HAVE_LIBPFM
3545	OPT_CALLBACK(0, "pfm-events", &record.evlist, "event",
3546		"libpfm4 event selector. use 'perf list' to list available events",
3547		parse_libpfm_events_option),
3548#endif
3549	OPT_CALLBACK(0, "control", &record.opts, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
3550		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events,\n"
3551		     "\t\t\t  'snapshot': AUX area tracing snapshot).\n"
3552		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
3553		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
3554		      parse_control_option),
3555	OPT_CALLBACK(0, "synth", &record.opts, "no|all|task|mmap|cgroup",
3556		     "Fine-tune event synthesis: default=all", parse_record_synth_option),
3557	OPT_STRING_OPTARG_SET(0, "debuginfod", &record.debuginfod.urls,
3558			  &record.debuginfod.set, "debuginfod urls",
3559			  "Enable debuginfod data retrieval from DEBUGINFOD_URLS or specified urls",
3560			  "system"),
3561	OPT_CALLBACK_OPTARG(0, "threads", &record.opts, NULL, "spec",
3562			    "write collected trace data into several data files using parallel threads",
3563			    record__parse_threads),
3564	OPT_BOOLEAN(0, "off-cpu", &record.off_cpu, "Enable off-cpu analysis"),
3565	OPT_STRING(0, "setup-filter", &record.filter_action, "pin|unpin",
3566		   "BPF filter action"),
3567	OPT_END()
3568};
3569
3570struct option *record_options = __record_options;
3571
3572static int record__mmap_cpu_mask_init(struct mmap_cpu_mask *mask, struct perf_cpu_map *cpus)
3573{
3574	struct perf_cpu cpu;
3575	int idx;
3576
3577	if (cpu_map__is_dummy(cpus))
3578		return 0;
3579
3580	perf_cpu_map__for_each_cpu_skip_any(cpu, idx, cpus) {
3581		/* Return ENODEV is input cpu is greater than max cpu */
3582		if ((unsigned long)cpu.cpu > mask->nbits)
3583			return -ENODEV;
3584		__set_bit(cpu.cpu, mask->bits);
3585	}
3586
3587	return 0;
3588}
3589
3590static int record__mmap_cpu_mask_init_spec(struct mmap_cpu_mask *mask, const char *mask_spec)
3591{
3592	struct perf_cpu_map *cpus;
3593
3594	cpus = perf_cpu_map__new(mask_spec);
3595	if (!cpus)
3596		return -ENOMEM;
3597
3598	bitmap_zero(mask->bits, mask->nbits);
3599	if (record__mmap_cpu_mask_init(mask, cpus))
3600		return -ENODEV;
3601
3602	perf_cpu_map__put(cpus);
3603
3604	return 0;
3605}
3606
3607static void record__free_thread_masks(struct record *rec, int nr_threads)
3608{
3609	int t;
3610
3611	if (rec->thread_masks)
3612		for (t = 0; t < nr_threads; t++)
3613			record__thread_mask_free(&rec->thread_masks[t]);
3614
3615	zfree(&rec->thread_masks);
3616}
3617
3618static int record__alloc_thread_masks(struct record *rec, int nr_threads, int nr_bits)
3619{
3620	int t, ret;
3621
3622	rec->thread_masks = zalloc(nr_threads * sizeof(*(rec->thread_masks)));
3623	if (!rec->thread_masks) {
3624		pr_err("Failed to allocate thread masks\n");
3625		return -ENOMEM;
3626	}
3627
3628	for (t = 0; t < nr_threads; t++) {
3629		ret = record__thread_mask_alloc(&rec->thread_masks[t], nr_bits);
3630		if (ret) {
3631			pr_err("Failed to allocate thread masks[%d]\n", t);
3632			goto out_free;
3633		}
3634	}
3635
3636	return 0;
3637
3638out_free:
3639	record__free_thread_masks(rec, nr_threads);
3640
3641	return ret;
3642}
3643
3644static int record__init_thread_cpu_masks(struct record *rec, struct perf_cpu_map *cpus)
3645{
3646	int t, ret, nr_cpus = perf_cpu_map__nr(cpus);
3647
3648	ret = record__alloc_thread_masks(rec, nr_cpus, cpu__max_cpu().cpu);
3649	if (ret)
3650		return ret;
3651
3652	rec->nr_threads = nr_cpus;
3653	pr_debug("nr_threads: %d\n", rec->nr_threads);
3654
3655	for (t = 0; t < rec->nr_threads; t++) {
3656		__set_bit(perf_cpu_map__cpu(cpus, t).cpu, rec->thread_masks[t].maps.bits);
3657		__set_bit(perf_cpu_map__cpu(cpus, t).cpu, rec->thread_masks[t].affinity.bits);
3658		if (verbose > 0) {
3659			pr_debug("thread_masks[%d]: ", t);
3660			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].maps, "maps");
3661			pr_debug("thread_masks[%d]: ", t);
3662			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].affinity, "affinity");
3663		}
3664	}
3665
3666	return 0;
3667}
3668
3669static int record__init_thread_masks_spec(struct record *rec, struct perf_cpu_map *cpus,
3670					  const char **maps_spec, const char **affinity_spec,
3671					  u32 nr_spec)
3672{
3673	u32 s;
3674	int ret = 0, t = 0;
3675	struct mmap_cpu_mask cpus_mask;
3676	struct thread_mask thread_mask, full_mask, *thread_masks;
3677
3678	ret = record__mmap_cpu_mask_alloc(&cpus_mask, cpu__max_cpu().cpu);
3679	if (ret) {
3680		pr_err("Failed to allocate CPUs mask\n");
3681		return ret;
3682	}
3683
3684	ret = record__mmap_cpu_mask_init(&cpus_mask, cpus);
3685	if (ret) {
3686		pr_err("Failed to init cpu mask\n");
3687		goto out_free_cpu_mask;
3688	}
3689
3690	ret = record__thread_mask_alloc(&full_mask, cpu__max_cpu().cpu);
3691	if (ret) {
3692		pr_err("Failed to allocate full mask\n");
3693		goto out_free_cpu_mask;
3694	}
3695
3696	ret = record__thread_mask_alloc(&thread_mask, cpu__max_cpu().cpu);
3697	if (ret) {
3698		pr_err("Failed to allocate thread mask\n");
3699		goto out_free_full_and_cpu_masks;
3700	}
3701
3702	for (s = 0; s < nr_spec; s++) {
3703		ret = record__mmap_cpu_mask_init_spec(&thread_mask.maps, maps_spec[s]);
3704		if (ret) {
3705			pr_err("Failed to initialize maps thread mask\n");
3706			goto out_free;
3707		}
3708		ret = record__mmap_cpu_mask_init_spec(&thread_mask.affinity, affinity_spec[s]);
3709		if (ret) {
3710			pr_err("Failed to initialize affinity thread mask\n");
3711			goto out_free;
3712		}
3713
3714		/* ignore invalid CPUs but do not allow empty masks */
3715		if (!bitmap_and(thread_mask.maps.bits, thread_mask.maps.bits,
3716				cpus_mask.bits, thread_mask.maps.nbits)) {
3717			pr_err("Empty maps mask: %s\n", maps_spec[s]);
3718			ret = -EINVAL;
3719			goto out_free;
3720		}
3721		if (!bitmap_and(thread_mask.affinity.bits, thread_mask.affinity.bits,
3722				cpus_mask.bits, thread_mask.affinity.nbits)) {
3723			pr_err("Empty affinity mask: %s\n", affinity_spec[s]);
3724			ret = -EINVAL;
3725			goto out_free;
3726		}
3727
3728		/* do not allow intersection with other masks (full_mask) */
3729		if (bitmap_intersects(thread_mask.maps.bits, full_mask.maps.bits,
3730				      thread_mask.maps.nbits)) {
3731			pr_err("Intersecting maps mask: %s\n", maps_spec[s]);
3732			ret = -EINVAL;
3733			goto out_free;
3734		}
3735		if (bitmap_intersects(thread_mask.affinity.bits, full_mask.affinity.bits,
3736				      thread_mask.affinity.nbits)) {
3737			pr_err("Intersecting affinity mask: %s\n", affinity_spec[s]);
3738			ret = -EINVAL;
3739			goto out_free;
3740		}
3741
3742		bitmap_or(full_mask.maps.bits, full_mask.maps.bits,
3743			  thread_mask.maps.bits, full_mask.maps.nbits);
3744		bitmap_or(full_mask.affinity.bits, full_mask.affinity.bits,
3745			  thread_mask.affinity.bits, full_mask.maps.nbits);
3746
3747		thread_masks = realloc(rec->thread_masks, (t + 1) * sizeof(struct thread_mask));
3748		if (!thread_masks) {
3749			pr_err("Failed to reallocate thread masks\n");
3750			ret = -ENOMEM;
3751			goto out_free;
3752		}
3753		rec->thread_masks = thread_masks;
3754		rec->thread_masks[t] = thread_mask;
3755		if (verbose > 0) {
3756			pr_debug("thread_masks[%d]: ", t);
3757			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].maps, "maps");
3758			pr_debug("thread_masks[%d]: ", t);
3759			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].affinity, "affinity");
3760		}
3761		t++;
3762		ret = record__thread_mask_alloc(&thread_mask, cpu__max_cpu().cpu);
3763		if (ret) {
3764			pr_err("Failed to allocate thread mask\n");
3765			goto out_free_full_and_cpu_masks;
3766		}
3767	}
3768	rec->nr_threads = t;
3769	pr_debug("nr_threads: %d\n", rec->nr_threads);
3770	if (!rec->nr_threads)
3771		ret = -EINVAL;
3772
3773out_free:
3774	record__thread_mask_free(&thread_mask);
3775out_free_full_and_cpu_masks:
3776	record__thread_mask_free(&full_mask);
3777out_free_cpu_mask:
3778	record__mmap_cpu_mask_free(&cpus_mask);
3779
3780	return ret;
3781}
3782
3783static int record__init_thread_core_masks(struct record *rec, struct perf_cpu_map *cpus)
3784{
3785	int ret;
3786	struct cpu_topology *topo;
3787
3788	topo = cpu_topology__new();
3789	if (!topo) {
3790		pr_err("Failed to allocate CPU topology\n");
3791		return -ENOMEM;
3792	}
3793
3794	ret = record__init_thread_masks_spec(rec, cpus, topo->core_cpus_list,
3795					     topo->core_cpus_list, topo->core_cpus_lists);
3796	cpu_topology__delete(topo);
3797
3798	return ret;
3799}
3800
3801static int record__init_thread_package_masks(struct record *rec, struct perf_cpu_map *cpus)
3802{
3803	int ret;
3804	struct cpu_topology *topo;
3805
3806	topo = cpu_topology__new();
3807	if (!topo) {
3808		pr_err("Failed to allocate CPU topology\n");
3809		return -ENOMEM;
3810	}
3811
3812	ret = record__init_thread_masks_spec(rec, cpus, topo->package_cpus_list,
3813					     topo->package_cpus_list, topo->package_cpus_lists);
3814	cpu_topology__delete(topo);
3815
3816	return ret;
3817}
3818
3819static int record__init_thread_numa_masks(struct record *rec, struct perf_cpu_map *cpus)
3820{
3821	u32 s;
3822	int ret;
3823	const char **spec;
3824	struct numa_topology *topo;
3825
3826	topo = numa_topology__new();
3827	if (!topo) {
3828		pr_err("Failed to allocate NUMA topology\n");
3829		return -ENOMEM;
3830	}
3831
3832	spec = zalloc(topo->nr * sizeof(char *));
3833	if (!spec) {
3834		pr_err("Failed to allocate NUMA spec\n");
3835		ret = -ENOMEM;
3836		goto out_delete_topo;
3837	}
3838	for (s = 0; s < topo->nr; s++)
3839		spec[s] = topo->nodes[s].cpus;
3840
3841	ret = record__init_thread_masks_spec(rec, cpus, spec, spec, topo->nr);
3842
3843	zfree(&spec);
3844
3845out_delete_topo:
3846	numa_topology__delete(topo);
3847
3848	return ret;
3849}
3850
3851static int record__init_thread_user_masks(struct record *rec, struct perf_cpu_map *cpus)
3852{
3853	int t, ret;
3854	u32 s, nr_spec = 0;
3855	char **maps_spec = NULL, **affinity_spec = NULL, **tmp_spec;
3856	char *user_spec, *spec, *spec_ptr, *mask, *mask_ptr, *dup_mask = NULL;
3857
3858	for (t = 0, user_spec = (char *)rec->opts.threads_user_spec; ; t++, user_spec = NULL) {
3859		spec = strtok_r(user_spec, ":", &spec_ptr);
3860		if (spec == NULL)
3861			break;
3862		pr_debug2("threads_spec[%d]: %s\n", t, spec);
3863		mask = strtok_r(spec, "/", &mask_ptr);
3864		if (mask == NULL)
3865			break;
3866		pr_debug2("  maps mask: %s\n", mask);
3867		tmp_spec = realloc(maps_spec, (nr_spec + 1) * sizeof(char *));
3868		if (!tmp_spec) {
3869			pr_err("Failed to reallocate maps spec\n");
3870			ret = -ENOMEM;
3871			goto out_free;
3872		}
3873		maps_spec = tmp_spec;
3874		maps_spec[nr_spec] = dup_mask = strdup(mask);
3875		if (!maps_spec[nr_spec]) {
3876			pr_err("Failed to allocate maps spec[%d]\n", nr_spec);
3877			ret = -ENOMEM;
3878			goto out_free;
3879		}
3880		mask = strtok_r(NULL, "/", &mask_ptr);
3881		if (mask == NULL) {
3882			pr_err("Invalid thread maps or affinity specs\n");
3883			ret = -EINVAL;
3884			goto out_free;
3885		}
3886		pr_debug2("  affinity mask: %s\n", mask);
3887		tmp_spec = realloc(affinity_spec, (nr_spec + 1) * sizeof(char *));
3888		if (!tmp_spec) {
3889			pr_err("Failed to reallocate affinity spec\n");
3890			ret = -ENOMEM;
3891			goto out_free;
3892		}
3893		affinity_spec = tmp_spec;
3894		affinity_spec[nr_spec] = strdup(mask);
3895		if (!affinity_spec[nr_spec]) {
3896			pr_err("Failed to allocate affinity spec[%d]\n", nr_spec);
3897			ret = -ENOMEM;
3898			goto out_free;
3899		}
3900		dup_mask = NULL;
3901		nr_spec++;
3902	}
3903
3904	ret = record__init_thread_masks_spec(rec, cpus, (const char **)maps_spec,
3905					     (const char **)affinity_spec, nr_spec);
3906
3907out_free:
3908	free(dup_mask);
3909	for (s = 0; s < nr_spec; s++) {
3910		if (maps_spec)
3911			free(maps_spec[s]);
3912		if (affinity_spec)
3913			free(affinity_spec[s]);
3914	}
3915	free(affinity_spec);
3916	free(maps_spec);
3917
3918	return ret;
3919}
3920
3921static int record__init_thread_default_masks(struct record *rec, struct perf_cpu_map *cpus)
3922{
3923	int ret;
3924
3925	ret = record__alloc_thread_masks(rec, 1, cpu__max_cpu().cpu);
3926	if (ret)
3927		return ret;
3928
3929	if (record__mmap_cpu_mask_init(&rec->thread_masks->maps, cpus))
3930		return -ENODEV;
3931
3932	rec->nr_threads = 1;
3933
3934	return 0;
3935}
3936
3937static int record__init_thread_masks(struct record *rec)
3938{
3939	int ret = 0;
3940	struct perf_cpu_map *cpus = rec->evlist->core.all_cpus;
3941
3942	if (!record__threads_enabled(rec))
3943		return record__init_thread_default_masks(rec, cpus);
3944
3945	if (evlist__per_thread(rec->evlist)) {
3946		pr_err("--per-thread option is mutually exclusive to parallel streaming mode.\n");
3947		return -EINVAL;
3948	}
3949
3950	switch (rec->opts.threads_spec) {
3951	case THREAD_SPEC__CPU:
3952		ret = record__init_thread_cpu_masks(rec, cpus);
3953		break;
3954	case THREAD_SPEC__CORE:
3955		ret = record__init_thread_core_masks(rec, cpus);
3956		break;
3957	case THREAD_SPEC__PACKAGE:
3958		ret = record__init_thread_package_masks(rec, cpus);
3959		break;
3960	case THREAD_SPEC__NUMA:
3961		ret = record__init_thread_numa_masks(rec, cpus);
3962		break;
3963	case THREAD_SPEC__USER:
3964		ret = record__init_thread_user_masks(rec, cpus);
3965		break;
3966	default:
3967		break;
3968	}
3969
3970	return ret;
3971}
3972
3973int cmd_record(int argc, const char **argv)
3974{
3975	int err;
3976	struct record *rec = &record;
3977	char errbuf[BUFSIZ];
3978
3979	setlocale(LC_ALL, "");
3980
3981#ifndef HAVE_BPF_SKEL
3982# define set_nobuild(s, l, m, c) set_option_nobuild(record_options, s, l, m, c)
3983	set_nobuild('\0', "off-cpu", "no BUILD_BPF_SKEL=1", true);
3984# undef set_nobuild
3985#endif
3986
3987	/* Disable eager loading of kernel symbols that adds overhead to perf record. */
3988	symbol_conf.lazy_load_kernel_maps = true;
3989	rec->opts.affinity = PERF_AFFINITY_SYS;
3990
3991	rec->evlist = evlist__new();
3992	if (rec->evlist == NULL)
3993		return -ENOMEM;
3994
3995	err = perf_config(perf_record_config, rec);
3996	if (err)
3997		return err;
3998
3999	argc = parse_options(argc, argv, record_options, record_usage,
4000			    PARSE_OPT_STOP_AT_NON_OPTION);
4001	if (quiet)
4002		perf_quiet_option();
4003
4004	err = symbol__validate_sym_arguments();
4005	if (err)
4006		return err;
4007
4008	perf_debuginfod_setup(&record.debuginfod);
4009
4010	/* Make system wide (-a) the default target. */
4011	if (!argc && target__none(&rec->opts.target))
4012		rec->opts.target.system_wide = true;
4013
4014	if (nr_cgroups && !rec->opts.target.system_wide) {
4015		usage_with_options_msg(record_usage, record_options,
4016			"cgroup monitoring only available in system-wide mode");
4017
4018	}
4019
4020	if (rec->buildid_mmap) {
4021		if (!perf_can_record_build_id()) {
4022			pr_err("Failed: no support to record build id in mmap events, update your kernel.\n");
4023			err = -EINVAL;
4024			goto out_opts;
4025		}
4026		pr_debug("Enabling build id in mmap2 events.\n");
4027		/* Enable mmap build id synthesizing. */
4028		symbol_conf.buildid_mmap2 = true;
4029		/* Enable perf_event_attr::build_id bit. */
4030		rec->opts.build_id = true;
4031		/* Disable build id cache. */
4032		rec->no_buildid = true;
4033	}
4034
4035	if (rec->opts.record_cgroup && !perf_can_record_cgroup()) {
4036		pr_err("Kernel has no cgroup sampling support.\n");
4037		err = -EINVAL;
4038		goto out_opts;
4039	}
4040
4041	if (rec->opts.kcore)
4042		rec->opts.text_poke = true;
4043
4044	if (rec->opts.kcore || record__threads_enabled(rec))
4045		rec->data.is_dir = true;
4046
4047	if (record__threads_enabled(rec)) {
4048		if (rec->opts.affinity != PERF_AFFINITY_SYS) {
4049			pr_err("--affinity option is mutually exclusive to parallel streaming mode.\n");
4050			goto out_opts;
4051		}
4052		if (record__aio_enabled(rec)) {
4053			pr_err("Asynchronous streaming mode (--aio) is mutually exclusive to parallel streaming mode.\n");
4054			goto out_opts;
4055		}
4056	}
4057
4058	if (rec->opts.comp_level != 0) {
4059		pr_debug("Compression enabled, disabling build id collection at the end of the session.\n");
4060		rec->no_buildid = true;
4061	}
4062
4063	if (rec->opts.record_switch_events &&
4064	    !perf_can_record_switch_events()) {
4065		ui__error("kernel does not support recording context switch events\n");
4066		parse_options_usage(record_usage, record_options, "switch-events", 0);
4067		err = -EINVAL;
4068		goto out_opts;
4069	}
4070
4071	if (switch_output_setup(rec)) {
4072		parse_options_usage(record_usage, record_options, "switch-output", 0);
4073		err = -EINVAL;
4074		goto out_opts;
4075	}
4076
4077	if (rec->switch_output.time) {
4078		signal(SIGALRM, alarm_sig_handler);
4079		alarm(rec->switch_output.time);
4080	}
4081
4082	if (rec->switch_output.num_files) {
4083		rec->switch_output.filenames = calloc(rec->switch_output.num_files,
4084						      sizeof(char *));
4085		if (!rec->switch_output.filenames) {
4086			err = -EINVAL;
4087			goto out_opts;
4088		}
4089	}
4090
4091	if (rec->timestamp_filename && record__threads_enabled(rec)) {
4092		rec->timestamp_filename = false;
4093		pr_warning("WARNING: --timestamp-filename option is not available in parallel streaming mode.\n");
4094	}
4095
4096	if (rec->filter_action) {
4097		if (!strcmp(rec->filter_action, "pin"))
4098			err = perf_bpf_filter__pin();
4099		else if (!strcmp(rec->filter_action, "unpin"))
4100			err = perf_bpf_filter__unpin();
4101		else {
4102			pr_warning("Unknown BPF filter action: %s\n", rec->filter_action);
4103			err = -EINVAL;
4104		}
4105		goto out_opts;
4106	}
4107
4108	/*
4109	 * Allow aliases to facilitate the lookup of symbols for address
4110	 * filters. Refer to auxtrace_parse_filters().
4111	 */
4112	symbol_conf.allow_aliases = true;
4113
4114	symbol__init(NULL);
4115
4116	err = record__auxtrace_init(rec);
4117	if (err)
4118		goto out;
4119
4120	if (dry_run)
4121		goto out;
4122
4123	err = -ENOMEM;
4124
4125	if (rec->no_buildid_cache || rec->no_buildid) {
4126		disable_buildid_cache();
4127	} else if (rec->switch_output.enabled) {
4128		/*
4129		 * In 'perf record --switch-output', disable buildid
4130		 * generation by default to reduce data file switching
4131		 * overhead. Still generate buildid if they are required
4132		 * explicitly using
4133		 *
4134		 *  perf record --switch-output --no-no-buildid \
4135		 *              --no-no-buildid-cache
4136		 *
4137		 * Following code equals to:
4138		 *
4139		 * if ((rec->no_buildid || !rec->no_buildid_set) &&
4140		 *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
4141		 *         disable_buildid_cache();
4142		 */
4143		bool disable = true;
4144
4145		if (rec->no_buildid_set && !rec->no_buildid)
4146			disable = false;
4147		if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
4148			disable = false;
4149		if (disable) {
4150			rec->no_buildid = true;
4151			rec->no_buildid_cache = true;
4152			disable_buildid_cache();
4153		}
4154	}
4155
4156	if (record.opts.overwrite)
4157		record.opts.tail_synthesize = true;
4158
4159	if (rec->evlist->core.nr_entries == 0) {
4160		err = parse_event(rec->evlist, "cycles:P");
 
 
4161		if (err)
4162			goto out;
4163	}
4164
4165	if (rec->opts.target.tid && !rec->opts.no_inherit_set)
4166		rec->opts.no_inherit = true;
4167
4168	err = target__validate(&rec->opts.target);
4169	if (err) {
4170		target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
4171		ui__warning("%s\n", errbuf);
4172	}
4173
4174	err = target__parse_uid(&rec->opts.target);
4175	if (err) {
4176		int saved_errno = errno;
4177
4178		target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
4179		ui__error("%s", errbuf);
4180
4181		err = -saved_errno;
4182		goto out;
4183	}
4184
4185	/* Enable ignoring missing threads when -u/-p option is defined. */
4186	rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
4187
4188	evlist__warn_user_requested_cpus(rec->evlist, rec->opts.target.cpu_list);
4189
4190	if (callchain_param.enabled && callchain_param.record_mode == CALLCHAIN_FP)
4191		arch__add_leaf_frame_record_opts(&rec->opts);
4192
4193	err = -ENOMEM;
4194	if (evlist__create_maps(rec->evlist, &rec->opts.target) < 0) {
4195		if (rec->opts.target.pid != NULL) {
4196			pr_err("Couldn't create thread/CPU maps: %s\n",
4197				errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
4198			goto out;
4199		}
4200		else
4201			usage_with_options(record_usage, record_options);
4202	}
4203
4204	err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
4205	if (err)
4206		goto out;
4207
4208	/*
4209	 * We take all buildids when the file contains
4210	 * AUX area tracing data because we do not decode the
4211	 * trace because it would take too long.
4212	 */
4213	if (rec->opts.full_auxtrace)
4214		rec->buildid_all = true;
4215
4216	if (rec->opts.text_poke) {
4217		err = record__config_text_poke(rec->evlist);
4218		if (err) {
4219			pr_err("record__config_text_poke failed, error %d\n", err);
4220			goto out;
4221		}
4222	}
4223
4224	if (rec->off_cpu) {
4225		err = record__config_off_cpu(rec);
4226		if (err) {
4227			pr_err("record__config_off_cpu failed, error %d\n", err);
4228			goto out;
4229		}
4230	}
4231
4232	if (record_opts__config(&rec->opts)) {
4233		err = -EINVAL;
4234		goto out;
4235	}
4236
4237	err = record__config_tracking_events(rec);
4238	if (err) {
4239		pr_err("record__config_tracking_events failed, error %d\n", err);
4240		goto out;
4241	}
4242
4243	err = record__init_thread_masks(rec);
4244	if (err) {
4245		pr_err("Failed to initialize parallel data streaming masks\n");
4246		goto out;
4247	}
4248
4249	if (rec->opts.nr_cblocks > nr_cblocks_max)
4250		rec->opts.nr_cblocks = nr_cblocks_max;
4251	pr_debug("nr_cblocks: %d\n", rec->opts.nr_cblocks);
4252
4253	pr_debug("affinity: %s\n", affinity_tags[rec->opts.affinity]);
4254	pr_debug("mmap flush: %d\n", rec->opts.mmap_flush);
4255
4256	if (rec->opts.comp_level > comp_level_max)
4257		rec->opts.comp_level = comp_level_max;
4258	pr_debug("comp level: %d\n", rec->opts.comp_level);
4259
4260	err = __cmd_record(&record, argc, argv);
4261out:
4262	record__free_thread_masks(rec, rec->nr_threads);
4263	rec->nr_threads = 0;
4264	symbol__exit();
4265	auxtrace_record__free(rec->itr);
4266out_opts:
 
 
4267	evlist__close_control(rec->opts.ctl_fd, rec->opts.ctl_fd_ack, &rec->opts.ctl_fd_close);
4268	evlist__delete(rec->evlist);
4269	return err;
4270}
4271
4272static void snapshot_sig_handler(int sig __maybe_unused)
4273{
4274	struct record *rec = &record;
4275
4276	hit_auxtrace_snapshot_trigger(rec);
4277
4278	if (switch_output_signal(rec))
4279		trigger_hit(&switch_output_trigger);
4280}
4281
4282static void alarm_sig_handler(int sig __maybe_unused)
4283{
4284	struct record *rec = &record;
4285
4286	if (switch_output_time(rec))
4287		trigger_hit(&switch_output_trigger);
4288}