Linux Audio

Check our new training course

Loading...
v4.6
 
   1/*
   2 * builtin-record.c
   3 *
   4 * Builtin record command: Record the profile of a workload
   5 * (or a CPU, or a PID) into the perf.data output file - for
   6 * later analysis via perf report.
   7 */
   8#include "builtin.h"
   9
  10#include "perf.h"
  11
  12#include "util/build-id.h"
  13#include "util/util.h"
  14#include <subcmd/parse-options.h>
 
  15#include "util/parse-events.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/session.h"
  25#include "util/tool.h"
  26#include "util/symbol.h"
 
  27#include "util/cpumap.h"
  28#include "util/thread_map.h"
  29#include "util/data.h"
  30#include "util/perf_regs.h"
  31#include "util/auxtrace.h"
 
  32#include "util/parse-branch-options.h"
  33#include "util/parse-regs-options.h"
 
  34#include "util/llvm-utils.h"
  35#include "util/bpf-loader.h"
 
 
 
 
 
 
 
 
 
 
 
 
 
  36#include "asm/bug.h"
 
 
  37
 
 
 
 
 
  38#include <unistd.h>
 
 
 
  39#include <sched.h>
 
 
 
 
  40#include <sys/mman.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  41
 
 
 
 
 
 
 
 
  42
  43struct record {
  44	struct perf_tool	tool;
  45	struct record_opts	opts;
  46	u64			bytes_written;
  47	struct perf_data_file	file;
  48	struct auxtrace_record	*itr;
  49	struct perf_evlist	*evlist;
  50	struct perf_session	*session;
  51	const char		*progname;
 
  52	int			realtime_prio;
 
  53	bool			no_buildid;
  54	bool			no_buildid_set;
  55	bool			no_buildid_cache;
  56	bool			no_buildid_cache_set;
  57	bool			buildid_all;
 
 
 
 
 
  58	unsigned long long	samples;
 
 
 
 
 
 
 
 
  59};
  60
  61static int record__write(struct record *rec, void *bf, size_t size)
 
 
 
 
 
 
 
 
 
 
 
  62{
  63	if (perf_data_file__write(rec->session->file, bf, size) < 0) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  64		pr_err("failed to write perf data, error: %m\n");
  65		return -1;
  66	}
  67
  68	rec->bytes_written += size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  69	return 0;
  70}
  71
  72static int process_synthesized_event(struct perf_tool *tool,
  73				     union perf_event *event,
  74				     struct perf_sample *sample __maybe_unused,
  75				     struct machine *machine __maybe_unused)
 
 
 
 
  76{
  77	struct record *rec = container_of(tool, struct record, tool);
  78	return record__write(rec, event, event->header.size);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  79}
  80
  81static int record__mmap_read(struct record *rec, int idx)
  82{
  83	struct perf_mmap *md = &rec->evlist->mmap[idx];
  84	u64 head = perf_mmap__read_head(md);
  85	u64 old = md->prev;
  86	unsigned char *data = md->base + page_size;
  87	unsigned long size;
  88	void *buf;
  89	int rc = 0;
  90
  91	if (old == head)
 
  92		return 0;
  93
  94	rec->samples++;
 
 
 
 
 
  95
  96	size = head - old;
  97
  98	if ((old & md->mask) + size != (head & md->mask)) {
  99		buf = &data[old & md->mask];
 100		size = md->mask + 1 - (old & md->mask);
 101		old += size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 102
 103		if (record__write(rec, buf, size) < 0) {
 104			rc = -1;
 105			goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 106		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 107	}
 108
 109	buf = &data[old & md->mask];
 110	size = head - old;
 111	old += size;
 112
 113	if (record__write(rec, buf, size) < 0) {
 114		rc = -1;
 115		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 116	}
 117
 118	md->prev = old;
 119	perf_evlist__mmap_consume(rec->evlist, idx);
 120out:
 121	return rc;
 122}
 123
 124static volatile int done;
 125static volatile int signr = -1;
 126static volatile int child_finished;
 127static volatile int auxtrace_snapshot_enabled;
 128static volatile int auxtrace_snapshot_err;
 129static volatile int auxtrace_record__snapshot_started;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 130
 131static void sig_handler(int sig)
 132{
 133	if (sig == SIGCHLD)
 134		child_finished = 1;
 135	else
 136		signr = sig;
 137
 138	done = 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 139}
 140
 141static void record__sig_exit(void)
 142{
 143	if (signr == -1)
 144		return;
 145
 146	signal(signr, SIG_DFL);
 147	raise(signr);
 148}
 149
 150#ifdef HAVE_AUXTRACE_SUPPORT
 151
 152static int record__process_auxtrace(struct perf_tool *tool,
 
 153				    union perf_event *event, void *data1,
 154				    size_t len1, void *data2, size_t len2)
 155{
 156	struct record *rec = container_of(tool, struct record, tool);
 157	struct perf_data_file *file = &rec->file;
 158	size_t padding;
 159	u8 pad[8] = {0};
 160
 161	if (!perf_data_file__is_pipe(file)) {
 162		off_t file_offset;
 163		int fd = perf_data_file__fd(file);
 164		int err;
 165
 166		file_offset = lseek(fd, 0, SEEK_CUR);
 167		if (file_offset == -1)
 168			return -1;
 169		err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 170						     event, file_offset);
 171		if (err)
 172			return err;
 173	}
 174
 175	/* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 176	padding = (len1 + len2) & 7;
 177	if (padding)
 178		padding = 8 - padding;
 179
 180	record__write(rec, event, event->header.size);
 181	record__write(rec, data1, len1);
 182	if (len2)
 183		record__write(rec, data2, len2);
 184	record__write(rec, &pad, padding);
 185
 186	return 0;
 187}
 188
 189static int record__auxtrace_mmap_read(struct record *rec,
 190				      struct auxtrace_mmap *mm)
 191{
 192	int ret;
 193
 194	ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
 195				  record__process_auxtrace);
 196	if (ret < 0)
 197		return ret;
 198
 199	if (ret)
 200		rec->samples++;
 201
 202	return 0;
 203}
 204
 205static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 206					       struct auxtrace_mmap *mm)
 207{
 208	int ret;
 209
 210	ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
 211					   record__process_auxtrace,
 212					   rec->opts.auxtrace_snapshot_size);
 213	if (ret < 0)
 214		return ret;
 215
 216	if (ret)
 217		rec->samples++;
 218
 219	return 0;
 220}
 221
 222static int record__auxtrace_read_snapshot_all(struct record *rec)
 223{
 224	int i;
 225	int rc = 0;
 226
 227	for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 228		struct auxtrace_mmap *mm =
 229				&rec->evlist->mmap[i].auxtrace_mmap;
 230
 231		if (!mm->base)
 232			continue;
 233
 234		if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
 235			rc = -1;
 236			goto out;
 237		}
 238	}
 239out:
 240	return rc;
 241}
 242
 243static void record__read_auxtrace_snapshot(struct record *rec)
 244{
 245	pr_debug("Recording AUX area tracing snapshot\n");
 246	if (record__auxtrace_read_snapshot_all(rec) < 0) {
 247		auxtrace_snapshot_err = -1;
 248	} else {
 249		auxtrace_snapshot_err = auxtrace_record__snapshot_finish(rec->itr);
 250		if (!auxtrace_snapshot_err)
 251			auxtrace_snapshot_enabled = 1;
 
 252	}
 253}
 254
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 255#else
 256
 257static inline
 258int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 259			       struct auxtrace_mmap *mm __maybe_unused)
 260{
 261	return 0;
 262}
 263
 264static inline
 265void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
 
 266{
 267}
 268
 269static inline
 270int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 271{
 272	return 0;
 273}
 274
 
 
 
 
 
 
 
 
 
 
 
 275#endif
 276
 277static int record__open(struct record *rec)
 278{
 279	char msg[512];
 280	struct perf_evsel *pos;
 281	struct perf_evlist *evlist = rec->evlist;
 282	struct perf_session *session = rec->session;
 283	struct record_opts *opts = &rec->opts;
 284	int rc = 0;
 
 285
 286	perf_evlist__config(evlist, opts);
 
 
 287
 288	evlist__for_each(evlist, pos) {
 289try_again:
 290		if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
 291			if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 292				if (verbose)
 293					ui__warning("%s\n", msg);
 294				goto try_again;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 295			}
 
 
 
 296
 297			rc = -errno;
 298			perf_evsel__open_strerror(pos, &opts->target,
 299						  errno, msg, sizeof(msg));
 300			ui__error("%s\n", msg);
 301			goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 302		}
 303	}
 304
 305	if (perf_evlist__apply_filters(evlist, &pos)) {
 306		error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 307			pos->filter, perf_evsel__name(pos), errno,
 308			strerror_r(errno, msg, sizeof(msg)));
 309		rc = -1;
 310		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 311	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 312
 313	if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
 314				 opts->auxtrace_mmap_pages,
 315				 opts->auxtrace_snapshot_mode) < 0) {
 
 
 316		if (errno == EPERM) {
 317			pr_err("Permission error mapping pages.\n"
 318			       "Consider increasing "
 319			       "/proc/sys/kernel/perf_event_mlock_kb,\n"
 320			       "or try again with a smaller value of -m/--mmap_pages.\n"
 321			       "(current value: %u,%u)\n",
 322			       opts->mmap_pages, opts->auxtrace_mmap_pages);
 323			rc = -errno;
 324		} else {
 325			pr_err("failed to mmap with %d (%s)\n", errno,
 326				strerror_r(errno, msg, sizeof(msg)));
 327			if (errno)
 328				rc = -errno;
 329			else
 330				rc = -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 331		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 332		goto out;
 333	}
 334
 
 
 
 
 335	session->evlist = evlist;
 336	perf_session__set_id_hdr_size(session);
 337out:
 338	return rc;
 339}
 340
 
 
 
 
 
 
 
 
 
 341static int process_sample_event(struct perf_tool *tool,
 342				union perf_event *event,
 343				struct perf_sample *sample,
 344				struct perf_evsel *evsel,
 345				struct machine *machine)
 346{
 347	struct record *rec = container_of(tool, struct record, tool);
 348
 349	rec->samples++;
 350
 
 
 
 
 351	return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 352}
 353
 354static int process_buildids(struct record *rec)
 355{
 356	struct perf_data_file *file  = &rec->file;
 357	struct perf_session *session = rec->session;
 358
 359	if (file->size == 0)
 360		return 0;
 361
 362	/*
 363	 * During this process, it'll load kernel map and replace the
 364	 * dso->long_name to a real pathname it found.  In this case
 365	 * we prefer the vmlinux path like
 366	 *   /lib/modules/3.16.4/build/vmlinux
 367	 *
 368	 * rather than build-id path (in debug directory).
 369	 *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 370	 */
 371	symbol_conf.ignore_vmlinux_buildid = true;
 372
 373	/*
 374	 * If --buildid-all is given, it marks all DSO regardless of hits,
 375	 * so no need to process samples.
 
 
 376	 */
 377	if (rec->buildid_all)
 378		rec->tool.sample = NULL;
 379
 380	return perf_session__process_events(session);
 381}
 382
 383static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 384{
 385	int err;
 386	struct perf_tool *tool = data;
 387	/*
 388	 *As for guest kernel when processing subcommand record&report,
 389	 *we arrange module mmap prior to guest kernel mmap and trigger
 390	 *a preload dso because default guest module symbols are loaded
 391	 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 392	 *method is used to avoid symbol missing when the first addr is
 393	 *in module instead of in guest kernel.
 394	 */
 395	err = perf_event__synthesize_modules(tool, process_synthesized_event,
 396					     machine);
 397	if (err < 0)
 398		pr_err("Couldn't record guest kernel [%d]'s reference"
 399		       " relocation symbol.\n", machine->pid);
 400
 401	/*
 402	 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 403	 * have no _text sometimes.
 404	 */
 405	err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 406						 machine);
 407	if (err < 0)
 408		pr_err("Couldn't record guest kernel [%d]'s reference"
 409		       " relocation symbol.\n", machine->pid);
 410}
 411
 412static struct perf_event_header finished_round_event = {
 413	.size = sizeof(struct perf_event_header),
 414	.type = PERF_RECORD_FINISHED_ROUND,
 415};
 416
 417static int record__mmap_read_all(struct record *rec)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 418{
 419	u64 bytes_written = rec->bytes_written;
 420	int i;
 421	int rc = 0;
 
 
 
 
 
 
 
 422
 423	for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 424		struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
 
 
 
 425
 426		if (rec->evlist->mmap[i].base) {
 427			if (record__mmap_read(rec, i) != 0) {
 428				rc = -1;
 429				goto out;
 
 
 
 
 
 
 
 
 
 
 
 430			}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 431		}
 432
 433		if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 434		    record__auxtrace_mmap_read(rec, mm) != 0) {
 
 435			rc = -1;
 436			goto out;
 437		}
 438	}
 439
 
 
 
 440	/*
 441	 * Mark the round finished in case we wrote
 442	 * at least one event.
 
 
 
 
 443	 */
 444	if (bytes_written != rec->bytes_written)
 445		rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 446
 
 
 447out:
 448	return rc;
 449}
 450
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 451static void record__init_features(struct record *rec)
 452{
 453	struct perf_session *session = rec->session;
 454	int feat;
 455
 456	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 457		perf_header__set_feat(&session->header, feat);
 458
 459	if (rec->no_buildid)
 460		perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 461
 462	if (!have_tracepoints(&rec->evlist->entries))
 
 463		perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 
 464
 465	if (!rec->opts.branch_stack)
 466		perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 467
 468	if (!rec->opts.full_auxtrace)
 469		perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 470
 
 
 
 
 
 
 
 
 
 
 
 
 471	perf_header__clear_feat(&session->header, HEADER_STAT);
 472}
 473
 474static void
 475record__finish_output(struct record *rec)
 476{
 477	struct perf_data_file *file = &rec->file;
 478	int fd = perf_data_file__fd(file);
 
 479
 480	if (file->is_pipe)
 481		return;
 482
 483	rec->session->header.data_size += rec->bytes_written;
 484	file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 
 
 
 
 485
 486	if (!rec->no_buildid) {
 487		process_buildids(rec);
 488
 489		if (rec->buildid_all)
 490			dsos__hit_all(rec->session);
 491	}
 492	perf_session__write_header(rec->session, rec->evlist, fd, true);
 493
 494	return;
 495}
 496
 497static volatile int workload_exec_errno;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 498
 499/*
 500 * perf_evlist__prepare_workload will send a SIGUSR1
 501 * if the fork fails, since we asked by setting its
 502 * want_signal to true.
 503 */
 504static void workload_exec_failed_signal(int signo __maybe_unused,
 505					siginfo_t *info,
 506					void *ucontext __maybe_unused)
 507{
 508	workload_exec_errno = info->si_value.sival_int;
 509	done = 1;
 510	child_finished = 1;
 511}
 512
 513static void snapshot_sig_handler(int sig);
 
 514
 515static int record__synthesize(struct record *rec)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 516{
 517	struct perf_session *session = rec->session;
 518	struct machine *machine = &session->machines.host;
 519	struct perf_data_file *file = &rec->file;
 520	struct record_opts *opts = &rec->opts;
 521	struct perf_tool *tool = &rec->tool;
 522	int fd = perf_data_file__fd(file);
 523	int err = 0;
 
 524
 525	if (file->is_pipe) {
 526		err = perf_event__synthesize_attrs(tool, session,
 527						   process_synthesized_event);
 528		if (err < 0) {
 529			pr_err("Couldn't synthesize attrs.\n");
 
 
 530			goto out;
 531		}
 532
 533		if (have_tracepoints(&rec->evlist->entries)) {
 534			/*
 535			 * FIXME err <= 0 here actually means that
 536			 * there were no tracepoints so its not really
 537			 * an error, just that we don't need to
 538			 * synthesize anything.  We really have to
 539			 * return this more properly and also
 540			 * propagate errors that now are calling die()
 541			 */
 542			err = perf_event__synthesize_tracing_data(tool,	fd, rec->evlist,
 543								  process_synthesized_event);
 544			if (err <= 0) {
 545				pr_err("Couldn't record tracing data.\n");
 546				goto out;
 547			}
 548			rec->bytes_written += err;
 549		}
 550	}
 551
 
 
 
 
 
 
 
 
 
 
 
 
 552	if (rec->opts.full_auxtrace) {
 553		err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 554					session, process_synthesized_event);
 555		if (err)
 556			goto out;
 557	}
 558
 559	err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 560						 machine);
 561	WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
 562			   "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 563			   "Check /proc/kallsyms permission or run as root.\n");
 564
 565	err = perf_event__synthesize_modules(tool, process_synthesized_event,
 566					     machine);
 567	WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
 568			   "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 569			   "Check /proc/modules permission or run as root.\n");
 
 
 570
 571	if (perf_guest) {
 572		machines__process_guests(&session->machines,
 573					 perf_event__synthesize_guest_os, tool);
 574	}
 575
 576	err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 577					    process_synthesized_event, opts->sample_address,
 578					    opts->proc_map_timeout);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 579out:
 580	return err;
 581}
 582
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 583static int __cmd_record(struct record *rec, int argc, const char **argv)
 584{
 585	int err;
 586	int status = 0;
 587	unsigned long waking = 0;
 588	const bool forks = argc > 0;
 589	struct machine *machine;
 590	struct perf_tool *tool = &rec->tool;
 591	struct record_opts *opts = &rec->opts;
 592	struct perf_data_file *file = &rec->file;
 593	struct perf_session *session;
 594	bool disabled = false, draining = false;
 595	int fd;
 596
 597	rec->progname = argv[0];
 598
 599	atexit(record__sig_exit);
 600	signal(SIGCHLD, sig_handler);
 601	signal(SIGINT, sig_handler);
 602	signal(SIGTERM, sig_handler);
 603	if (rec->opts.auxtrace_snapshot_mode)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 604		signal(SIGUSR2, snapshot_sig_handler);
 605	else
 
 
 
 
 606		signal(SIGUSR2, SIG_IGN);
 
 607
 608	session = perf_session__new(file, false, tool);
 609	if (session == NULL) {
 610		pr_err("Perf session creation failed.\n");
 611		return -1;
 
 
 
 
 
 
 
 
 
 
 
 612	}
 613
 614	fd = perf_data_file__fd(file);
 615	rec->session = session;
 616
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 617	record__init_features(rec);
 618
 619	if (forks) {
 620		err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 621						    argv, file->is_pipe,
 622						    workload_exec_failed_signal);
 623		if (err < 0) {
 624			pr_err("Couldn't run the workload!\n");
 625			status = err;
 626			goto out_delete_session;
 627		}
 628	}
 629
 
 
 
 
 
 
 
 
 
 
 
 
 
 630	if (record__open(rec) != 0) {
 631		err = -1;
 632		goto out_child;
 
 
 
 
 
 
 
 
 
 
 
 633	}
 634
 635	err = bpf__apply_obj_config();
 636	if (err) {
 637		char errbuf[BUFSIZ];
 638
 639		bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
 640		pr_err("ERROR: Apply config to BPF failed: %s\n",
 641			 errbuf);
 642		goto out_child;
 643	}
 644
 645	/*
 646	 * Normally perf_session__new would do this, but it doesn't have the
 647	 * evlist.
 648	 */
 649	if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 650		pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 651		rec->tool.ordered_events = false;
 652	}
 653
 654	if (!rec->evlist->nr_groups)
 655		perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 656
 657	if (file->is_pipe) {
 658		err = perf_header__write_pipe(fd);
 659		if (err < 0)
 660			goto out_child;
 661	} else {
 662		err = perf_session__write_header(session, rec->evlist, fd, false);
 663		if (err < 0)
 664			goto out_child;
 665	}
 666
 
 667	if (!rec->no_buildid
 668	    && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 669		pr_err("Couldn't generate buildids. "
 670		       "Use --no-buildid to profile anyway.\n");
 671		err = -1;
 672		goto out_child;
 673	}
 674
 675	machine = &session->machines.host;
 
 
 676
 677	err = record__synthesize(rec);
 678	if (err < 0)
 679		goto out_child;
 680
 681	if (rec->realtime_prio) {
 682		struct sched_param param;
 683
 684		param.sched_priority = rec->realtime_prio;
 685		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 686			pr_err("Could not set realtime priority.\n");
 687			err = -1;
 688			goto out_child;
 689		}
 690	}
 691
 
 
 
 692	/*
 693	 * When perf is starting the traced process, all the events
 694	 * (apart from group members) have enable_on_exec=1 set,
 695	 * so don't spoil it by prematurely enabling them.
 696	 */
 697	if (!target__none(&opts->target) && !opts->initial_delay)
 698		perf_evlist__enable(rec->evlist);
 699
 700	/*
 701	 * Let the child rip
 702	 */
 703	if (forks) {
 
 704		union perf_event *event;
 
 705
 706		event = malloc(sizeof(event->comm) + machine->id_hdr_size);
 707		if (event == NULL) {
 708			err = -ENOMEM;
 709			goto out_child;
 710		}
 711
 712		/*
 713		 * Some H/W events are generated before COMM event
 714		 * which is emitted during exec(), so perf script
 715		 * cannot see a correct process name for those events.
 716		 * Synthesize COMM event to prevent it.
 717		 */
 718		perf_event__synthesize_comm(tool, event,
 719					    rec->evlist->workload.pid,
 720					    process_synthesized_event,
 721					    machine);
 722		free(event);
 723
 724		perf_evlist__start_workload(rec->evlist);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 725	}
 726
 727	if (opts->initial_delay) {
 728		usleep(opts->initial_delay * 1000);
 729		perf_evlist__enable(rec->evlist);
 
 
 
 
 730	}
 731
 732	auxtrace_snapshot_enabled = 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 733	for (;;) {
 734		unsigned long long hits = rec->samples;
 735
 736		if (record__mmap_read_all(rec) < 0) {
 737			auxtrace_snapshot_enabled = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 738			err = -1;
 739			goto out_child;
 740		}
 741
 742		if (auxtrace_record__snapshot_started) {
 743			auxtrace_record__snapshot_started = 0;
 744			if (!auxtrace_snapshot_err)
 745				record__read_auxtrace_snapshot(rec);
 746			if (auxtrace_snapshot_err) {
 747				pr_err("AUX area tracing snapshot failed\n");
 748				err = -1;
 749				goto out_child;
 750			}
 751		}
 752
 753		if (hits == rec->samples) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 754			if (done || draining)
 755				break;
 756			err = perf_evlist__poll(rec->evlist, -1);
 757			/*
 758			 * Propagate error, only if there's any. Ignore positive
 759			 * number of returned events and interrupt error.
 760			 */
 761			if (err > 0 || (err < 0 && errno == EINTR))
 762				err = 0;
 763			waking++;
 764
 765			if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 
 766				draining = true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 767		}
 768
 769		/*
 770		 * When perf is starting the traced process, at the end events
 771		 * die with the process and we wait for that. Thus no need to
 772		 * disable events in this case.
 773		 */
 774		if (done && !disabled && !target__none(&opts->target)) {
 775			auxtrace_snapshot_enabled = 0;
 776			perf_evlist__disable(rec->evlist);
 777			disabled = true;
 778		}
 779	}
 780	auxtrace_snapshot_enabled = 0;
 
 
 
 
 
 781
 782	if (forks && workload_exec_errno) {
 783		char msg[STRERR_BUFSIZE];
 784		const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 785		pr_err("Workload failed: %s\n", emsg);
 
 
 
 
 786		err = -1;
 787		goto out_child;
 788	}
 789
 790	if (!quiet)
 791		fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 
 
 
 
 
 
 792
 793out_child:
 
 
 
 
 
 
 
 
 
 
 
 
 794	if (forks) {
 795		int exit_status;
 796
 797		if (!child_finished)
 798			kill(rec->evlist->workload.pid, SIGTERM);
 799
 800		wait(&exit_status);
 801
 802		if (err < 0)
 803			status = err;
 804		else if (WIFEXITED(exit_status))
 805			status = WEXITSTATUS(exit_status);
 806		else if (WIFSIGNALED(exit_status))
 807			signr = WTERMSIG(exit_status);
 808	} else
 809		status = err;
 810
 
 
 
 
 
 811	/* this will be recalculated during process_buildids() */
 812	rec->samples = 0;
 813
 814	if (!err)
 815		record__finish_output(rec);
 
 
 
 
 
 
 
 
 
 
 
 816
 817	if (!err && !quiet) {
 818		char samples[128];
 
 
 819
 820		if (rec->samples && !rec->opts.full_auxtrace)
 821			scnprintf(samples, sizeof(samples),
 822				  " (%" PRIu64 " samples)", rec->samples);
 823		else
 824			samples[0] = '\0';
 825
 826		fprintf(stderr,	"[ perf record: Captured and wrote %.3f MB %s%s ]\n",
 827			perf_data_file__size(file) / 1024.0 / 1024.0,
 828			file->path, samples);
 
 
 
 
 
 
 829	}
 830
 831out_delete_session:
 
 
 
 
 
 
 
 
 
 832	perf_session__delete(session);
 
 
 
 833	return status;
 834}
 835
 836static void callchain_debug(void)
 837{
 838	static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
 839
 840	pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
 841
 842	if (callchain_param.record_mode == CALLCHAIN_DWARF)
 843		pr_debug("callchain: stack dump size %d\n",
 844			 callchain_param.dump_size);
 845}
 846
 847int record_parse_callchain_opt(const struct option *opt,
 848			       const char *arg,
 849			       int unset)
 850{
 851	int ret;
 852	struct record_opts *record = (struct record_opts *)opt->value;
 853
 854	record->callgraph_set = true;
 855	callchain_param.enabled = !unset;
 856
 857	/* --no-call-graph */
 858	if (unset) {
 859		callchain_param.record_mode = CALLCHAIN_NONE;
 860		pr_debug("callchain: disabled\n");
 861		return 0;
 862	}
 863
 864	ret = parse_callchain_record_opt(arg, &callchain_param);
 865	if (!ret) {
 866		/* Enable data address sampling for DWARF unwind. */
 867		if (callchain_param.record_mode == CALLCHAIN_DWARF)
 868			record->sample_address = true;
 869		callchain_debug();
 870	}
 871
 872	return ret;
 873}
 874
 
 
 
 
 
 
 
 875int record_callchain_opt(const struct option *opt,
 876			 const char *arg __maybe_unused,
 877			 int unset __maybe_unused)
 878{
 879	struct record_opts *record = (struct record_opts *)opt->value;
 880
 881	record->callgraph_set = true;
 882	callchain_param.enabled = true;
 883
 884	if (callchain_param.record_mode == CALLCHAIN_NONE)
 885		callchain_param.record_mode = CALLCHAIN_FP;
 886
 887	callchain_debug();
 888	return 0;
 889}
 890
 891static int perf_record_config(const char *var, const char *value, void *cb)
 892{
 893	struct record *rec = cb;
 894
 895	if (!strcmp(var, "record.build-id")) {
 896		if (!strcmp(value, "cache"))
 897			rec->no_buildid_cache = false;
 898		else if (!strcmp(value, "no-cache"))
 899			rec->no_buildid_cache = true;
 900		else if (!strcmp(value, "skip"))
 901			rec->no_buildid = true;
 
 
 902		else
 903			return -1;
 904		return 0;
 905	}
 906	if (!strcmp(var, "record.call-graph"))
 907		var = "call-graph.record-mode"; /* fall-through */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 908
 909	return perf_default_config(var, value, cb);
 910}
 911
 912struct clockid_map {
 913	const char *name;
 914	int clockid;
 915};
 916
 917#define CLOCKID_MAP(n, c)	\
 918	{ .name = n, .clockid = (c), }
 919
 920#define CLOCKID_END	{ .name = NULL, }
 
 
 921
 
 
 922
 923/*
 924 * Add the missing ones, we need to build on many distros...
 925 */
 926#ifndef CLOCK_MONOTONIC_RAW
 927#define CLOCK_MONOTONIC_RAW 4
 928#endif
 929#ifndef CLOCK_BOOTTIME
 930#define CLOCK_BOOTTIME 7
 931#endif
 932#ifndef CLOCK_TAI
 933#define CLOCK_TAI 11
 934#endif
 935
 936static const struct clockid_map clockids[] = {
 937	/* available for all events, NMI safe */
 938	CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
 939	CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
 940
 941	/* available for some events */
 942	CLOCKID_MAP("realtime", CLOCK_REALTIME),
 943	CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
 944	CLOCKID_MAP("tai", CLOCK_TAI),
 945
 946	/* available for the lazy */
 947	CLOCKID_MAP("mono", CLOCK_MONOTONIC),
 948	CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
 949	CLOCKID_MAP("real", CLOCK_REALTIME),
 950	CLOCKID_MAP("boot", CLOCK_BOOTTIME),
 951
 952	CLOCKID_END,
 953};
 
 
 
 
 
 
 
 954
 955static int parse_clockid(const struct option *opt, const char *str, int unset)
 956{
 957	struct record_opts *opts = (struct record_opts *)opt->value;
 958	const struct clockid_map *cm;
 959	const char *ostr = str;
 960
 961	if (unset) {
 962		opts->use_clockid = 0;
 963		return 0;
 964	}
 965
 966	/* no arg passed */
 967	if (!str)
 968		return 0;
 
 
 969
 970	/* no setting it twice */
 971	if (opts->use_clockid)
 972		return -1;
 
 
 973
 974	opts->use_clockid = true;
 
 975
 976	/* if its a number, we're done */
 977	if (sscanf(str, "%d", &opts->clockid) == 1)
 978		return 0;
 
 
 979
 980	/* allow a "CLOCK_" prefix to the name */
 981	if (!strncasecmp(str, "CLOCK_", 6))
 982		str += 6;
 
 983
 984	for (cm = clockids; cm->name; cm++) {
 985		if (!strcasecmp(str, cm->name)) {
 986			opts->clockid = cm->clockid;
 987			return 0;
 
 
 
 
 
 
 
 
 
 
 
 988		}
 989	}
 990
 991	opts->use_clockid = false;
 992	ui__warning("unknown clockid %s, check man page\n", ostr);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 993	return -1;
 994}
 995
 996static int record__parse_mmap_pages(const struct option *opt,
 997				    const char *str,
 998				    int unset __maybe_unused)
 999{
1000	struct record_opts *opts = opt->value;
1001	char *s, *p;
1002	unsigned int mmap_pages;
1003	int ret;
1004
1005	if (!str)
1006		return -EINVAL;
1007
1008	s = strdup(str);
1009	if (!s)
1010		return -ENOMEM;
1011
1012	p = strchr(s, ',');
1013	if (p)
1014		*p = '\0';
1015
1016	if (*s) {
1017		ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1018		if (ret)
1019			goto out_free;
1020		opts->mmap_pages = mmap_pages;
1021	}
1022
1023	if (!p) {
1024		ret = 0;
1025		goto out_free;
1026	}
1027
1028	ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1029	if (ret)
1030		goto out_free;
1031
1032	opts->auxtrace_mmap_pages = mmap_pages;
1033
1034out_free:
1035	free(s);
1036	return ret;
1037}
1038
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1039static const char * const __record_usage[] = {
1040	"perf record [<options>] [<command>]",
1041	"perf record [<options>] -- <command> [<options>]",
1042	NULL
1043};
1044const char * const *record_usage = __record_usage;
1045
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1046/*
1047 * XXX Ideally would be local to cmd_record() and passed to a record__new
1048 * because we need to have access to it in record__exit, that is called
1049 * after cmd_record() exits, but since record_options need to be accessible to
1050 * builtin-script, leave it here.
1051 *
1052 * At least we don't ouch it in all the other functions here directly.
1053 *
1054 * Just say no to tons of global variables, sigh.
1055 */
1056static struct record record = {
1057	.opts = {
1058		.sample_time	     = true,
1059		.mmap_pages	     = UINT_MAX,
1060		.user_freq	     = UINT_MAX,
1061		.user_interval	     = ULLONG_MAX,
1062		.freq		     = 4000,
1063		.target		     = {
1064			.uses_mmap   = true,
1065			.default_per_cpu = true,
1066		},
1067		.proc_map_timeout     = 500,
 
 
 
 
1068	},
1069	.tool = {
1070		.sample		= process_sample_event,
1071		.fork		= perf_event__process_fork,
1072		.exit		= perf_event__process_exit,
1073		.comm		= perf_event__process_comm,
1074		.mmap		= perf_event__process_mmap,
1075		.mmap2		= perf_event__process_mmap2,
 
 
 
1076		.ordered_events	= true,
1077	},
1078};
1079
1080const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1081	"\n\t\t\t\tDefault: fp";
1082
 
 
1083/*
1084 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1085 * with it and switch to use the library functions in perf_evlist that came
1086 * from builtin-record.c, i.e. use record_opts,
1087 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1088 * using pipes, etc.
1089 */
1090struct option __record_options[] = {
1091	OPT_CALLBACK('e', "event", &record.evlist, "event",
1092		     "event selector. use 'perf list' to list available events",
1093		     parse_events_option),
1094	OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1095		     "event filter", parse_filter),
1096	OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1097			   NULL, "don't record events from perf itself",
1098			   exclude_perf),
1099	OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1100		    "record events on existing process id"),
1101	OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1102		    "record events on existing thread id"),
1103	OPT_INTEGER('r', "realtime", &record.realtime_prio,
1104		    "collect data with this RT SCHED_FIFO priority"),
1105	OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1106		    "collect data without buffering"),
1107	OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1108		    "collect raw sample records from all opened counters"),
1109	OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1110			    "system-wide collection from all CPUs"),
1111	OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1112		    "list of cpus to monitor"),
1113	OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1114	OPT_STRING('o', "output", &record.file.path, "file",
1115		    "output file name"),
1116	OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1117			&record.opts.no_inherit_set,
1118			"child tasks do not inherit counters"),
1119	OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 
 
 
 
 
 
 
 
1120	OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1121		     "number of mmap data pages and AUX area tracing mmap pages",
1122		     record__parse_mmap_pages),
1123	OPT_BOOLEAN(0, "group", &record.opts.group,
1124		    "put the counters into a counter group"),
1125	OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
 
1126			   NULL, "enables call-graph recording" ,
1127			   &record_callchain_opt),
1128	OPT_CALLBACK(0, "call-graph", &record.opts,
1129		     "record_mode[,record_size]", record_callchain_help,
1130		     &record_parse_callchain_opt),
1131	OPT_INCR('v', "verbose", &verbose,
1132		    "be more verbose (show counter open errors, etc)"),
1133	OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1134	OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1135		    "per thread counts"),
1136	OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
 
 
 
 
 
 
 
 
 
1137	OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1138			&record.opts.sample_time_set,
1139			"Record the sample timestamps"),
1140	OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
 
1141	OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1142		    "don't sample"),
1143	OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1144			&record.no_buildid_cache_set,
1145			"do not update the buildid cache"),
1146	OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1147			&record.no_buildid_set,
1148			"do not collect buildids in perf.data"),
1149	OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1150		     "monitor event in cgroup name only",
1151		     parse_cgroups),
1152	OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1153		  "ms to wait before starting measurement after program start"),
 
 
 
1154	OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1155		   "user to profile"),
1156
1157	OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1158		     "branch any", "sample any taken branches",
1159		     parse_branch_stack),
1160
1161	OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1162		     "branch filter mask", "branch stack filter modes",
1163		     parse_branch_stack),
1164	OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1165		    "sample by weight (on special events only)"),
1166	OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1167		    "sample transaction flags (special events only)"),
1168	OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1169		    "use per-thread mmaps"),
1170	OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1171		    "sample selected machine registers on interrupt,"
1172		    " use -I ? to list register names", parse_regs),
 
 
 
1173	OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1174		    "Record running/enabled time of read (:S) events"),
1175	OPT_CALLBACK('k', "clockid", &record.opts,
1176	"clockid", "clockid to use for events, see clock_gettime()",
1177	parse_clockid),
1178	OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1179			  "opts", "AUX area tracing Snapshot Mode", ""),
1180	OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
 
 
1181			"per thread proc mmap processing timeout in ms"),
1182	OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1183		    "Record context switch events"),
 
 
 
 
 
1184	OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1185			 "Configure all used events to run in kernel space.",
1186			 PARSE_OPT_EXCLUSIVE),
1187	OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1188			 "Configure all used events to run in user space.",
1189			 PARSE_OPT_EXCLUSIVE),
 
 
 
 
1190	OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1191		   "clang binary to use for compiling BPF scriptlets"),
1192	OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1193		   "options passed to clang when compiling BPF scriptlets"),
1194	OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1195		   "file", "vmlinux pathname"),
1196	OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1197		    "Record build-id of all DSOs regardless of hits"),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1198	OPT_END()
1199};
1200
1201struct option *record_options = __record_options;
1202
1203int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1204{
1205	int err;
1206	struct record *rec = &record;
1207	char errbuf[BUFSIZ];
1208
 
 
1209#ifndef HAVE_LIBBPF_SUPPORT
1210# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1211	set_nobuild('\0', "clang-path", true);
1212	set_nobuild('\0', "clang-opt", true);
1213# undef set_nobuild
1214#endif
1215
1216#ifndef HAVE_BPF_PROLOGUE
1217# if !defined (HAVE_DWARF_SUPPORT)
1218#  define REASON  "NO_DWARF=1"
1219# elif !defined (HAVE_LIBBPF_SUPPORT)
1220#  define REASON  "NO_LIBBPF=1"
1221# else
1222#  define REASON  "this architecture doesn't support BPF prologue"
1223# endif
1224# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1225	set_nobuild('\0', "vmlinux", true);
1226# undef set_nobuild
1227# undef REASON
1228#endif
1229
1230	rec->evlist = perf_evlist__new();
 
 
 
 
 
 
 
 
1231	if (rec->evlist == NULL)
1232		return -ENOMEM;
1233
1234	perf_config(perf_record_config, rec);
 
 
1235
1236	argc = parse_options(argc, argv, record_options, record_usage,
1237			    PARSE_OPT_STOP_AT_NON_OPTION);
 
 
 
 
 
 
 
 
 
 
1238	if (!argc && target__none(&rec->opts.target))
1239		usage_with_options(record_usage, record_options);
1240
1241	if (nr_cgroups && !rec->opts.target.system_wide) {
1242		usage_with_options_msg(record_usage, record_options,
1243			"cgroup monitoring only available in system-wide mode");
1244
1245	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1246	if (rec->opts.record_switch_events &&
1247	    !perf_can_record_switch_events()) {
1248		ui__error("kernel does not support recording context switch events\n");
1249		parse_options_usage(record_usage, record_options, "switch-events", 0);
1250		return -EINVAL;
 
1251	}
1252
1253	if (!rec->itr) {
1254		rec->itr = auxtrace_record__init(rec->evlist, &err);
1255		if (err)
1256			return err;
1257	}
1258
1259	err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1260					      rec->opts.auxtrace_snapshot_opts);
1261	if (err)
1262		return err;
1263
1264	err = -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1265
1266	symbol__init(NULL);
1267
1268	if (symbol_conf.kptr_restrict)
1269		pr_warning(
1270"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1271"check /proc/sys/kernel/kptr_restrict.\n\n"
1272"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1273"file is not found in the buildid cache or in the vmlinux path.\n\n"
1274"Samples in kernel modules won't be resolved at all.\n\n"
1275"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1276"even with a suitable vmlinux or kallsyms file.\n\n");
 
 
 
 
 
 
 
1277
1278	if (rec->no_buildid_cache || rec->no_buildid)
1279		disable_buildid_cache();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1280
1281	if (rec->evlist->nr_entries == 0 &&
1282	    perf_evlist__add_default(rec->evlist) < 0) {
1283		pr_err("Not enough memory for event selector list\n");
1284		goto out_symbol_exit;
1285	}
1286
1287	if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1288		rec->opts.no_inherit = true;
1289
1290	err = target__validate(&rec->opts.target);
1291	if (err) {
1292		target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1293		ui__warning("%s", errbuf);
1294	}
1295
1296	err = target__parse_uid(&rec->opts.target);
1297	if (err) {
1298		int saved_errno = errno;
1299
1300		target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1301		ui__error("%s", errbuf);
1302
1303		err = -saved_errno;
1304		goto out_symbol_exit;
1305	}
1306
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1307	err = -ENOMEM;
1308	if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1309		usage_with_options(record_usage, record_options);
 
 
 
 
 
 
 
1310
1311	err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1312	if (err)
1313		goto out_symbol_exit;
1314
1315	/*
1316	 * We take all buildids when the file contains
1317	 * AUX area tracing data because we do not decode the
1318	 * trace because it would take too long.
1319	 */
1320	if (rec->opts.full_auxtrace)
1321		rec->buildid_all = true;
1322
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1323	if (record_opts__config(&rec->opts)) {
1324		err = -EINVAL;
1325		goto out_symbol_exit;
1326	}
1327
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1328	err = __cmd_record(&record, argc, argv);
1329out_symbol_exit:
1330	perf_evlist__delete(rec->evlist);
1331	symbol__exit();
1332	auxtrace_record__free(rec->itr);
 
 
 
 
1333	return err;
1334}
1335
1336static void snapshot_sig_handler(int sig __maybe_unused)
1337{
1338	if (!auxtrace_snapshot_enabled)
1339		return;
1340	auxtrace_snapshot_enabled = 0;
1341	auxtrace_snapshot_err = auxtrace_record__snapshot_start(record.itr);
1342	auxtrace_record__snapshot_started = 1;
 
 
 
 
 
 
 
 
 
1343}
v6.2
   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/llvm-utils.h"
  41#include "util/bpf-loader.h"
  42#include "util/trigger.h"
  43#include "util/perf-hooks.h"
  44#include "util/cpu-set-sched.h"
  45#include "util/synthetic-events.h"
  46#include "util/time-utils.h"
  47#include "util/units.h"
  48#include "util/bpf-event.h"
  49#include "util/util.h"
  50#include "util/pfm.h"
  51#include "util/clockid.h"
  52#include "util/pmu-hybrid.h"
  53#include "util/evlist-hybrid.h"
  54#include "util/off_cpu.h"
  55#include "asm/bug.h"
  56#include "perf.h"
  57#include "cputopo.h"
  58
  59#include <errno.h>
  60#include <inttypes.h>
  61#include <locale.h>
  62#include <poll.h>
  63#include <pthread.h>
  64#include <unistd.h>
  65#ifndef HAVE_GETTID
  66#include <syscall.h>
  67#endif
  68#include <sched.h>
  69#include <signal.h>
  70#ifdef HAVE_EVENTFD_SUPPORT
  71#include <sys/eventfd.h>
  72#endif
  73#include <sys/mman.h>
  74#include <sys/wait.h>
  75#include <sys/types.h>
  76#include <sys/stat.h>
  77#include <fcntl.h>
  78#include <linux/err.h>
  79#include <linux/string.h>
  80#include <linux/time64.h>
  81#include <linux/zalloc.h>
  82#include <linux/bitmap.h>
  83#include <sys/time.h>
  84
  85struct switch_output {
  86	bool		 enabled;
  87	bool		 signal;
  88	unsigned long	 size;
  89	unsigned long	 time;
  90	const char	*str;
  91	bool		 set;
  92	char		 **filenames;
  93	int		 num_files;
  94	int		 cur_file;
  95};
  96
  97struct thread_mask {
  98	struct mmap_cpu_mask	maps;
  99	struct mmap_cpu_mask	affinity;
 100};
 101
 102struct record_thread {
 103	pid_t			tid;
 104	struct thread_mask	*mask;
 105	struct {
 106		int		msg[2];
 107		int		ack[2];
 108	} pipes;
 109	struct fdarray		pollfd;
 110	int			ctlfd_pos;
 111	int			nr_mmaps;
 112	struct mmap		**maps;
 113	struct mmap		**overwrite_maps;
 114	struct record		*rec;
 115	unsigned long long	samples;
 116	unsigned long		waking;
 117	u64			bytes_written;
 118	u64			bytes_transferred;
 119	u64			bytes_compressed;
 120};
 121
 122static __thread struct record_thread *thread;
 123
 124enum thread_msg {
 125	THREAD_MSG__UNDEFINED = 0,
 126	THREAD_MSG__READY,
 127	THREAD_MSG__MAX,
 128};
 129
 130static const char *thread_msg_tags[THREAD_MSG__MAX] = {
 131	"UNDEFINED", "READY"
 132};
 133
 134enum thread_spec {
 135	THREAD_SPEC__UNDEFINED = 0,
 136	THREAD_SPEC__CPU,
 137	THREAD_SPEC__CORE,
 138	THREAD_SPEC__PACKAGE,
 139	THREAD_SPEC__NUMA,
 140	THREAD_SPEC__USER,
 141	THREAD_SPEC__MAX,
 142};
 143
 144static const char *thread_spec_tags[THREAD_SPEC__MAX] = {
 145	"undefined", "cpu", "core", "package", "numa", "user"
 146};
 147
 148struct pollfd_index_map {
 149	int evlist_pollfd_index;
 150	int thread_pollfd_index;
 151};
 152
 153struct record {
 154	struct perf_tool	tool;
 155	struct record_opts	opts;
 156	u64			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	int t;
 230	u64 bytes_written = rec->bytes_written;
 231	struct record_thread *thread_data = rec->thread_data;
 232
 233	for (t = 0; t < rec->nr_threads; t++)
 234		bytes_written += thread_data[t].bytes_written;
 235
 236	return bytes_written;
 237}
 238
 239static bool record__output_max_size_exceeded(struct record *rec)
 240{
 241	return rec->output_max_size &&
 242	       (record__bytes_written(rec) >= rec->output_max_size);
 243}
 244
 245static int record__write(struct record *rec, struct mmap *map __maybe_unused,
 246			 void *bf, size_t size)
 247{
 248	struct perf_data_file *file = &rec->session->data->file;
 249
 250	if (map && map->file)
 251		file = map->file;
 252
 253	if (perf_data_file__write(file, bf, size) < 0) {
 254		pr_err("failed to write perf data, error: %m\n");
 255		return -1;
 256	}
 257
 258	if (map && map->file)
 259		thread->bytes_written += size;
 260	else
 261		rec->bytes_written += size;
 262
 263	if (record__output_max_size_exceeded(rec) && !done) {
 264		fprintf(stderr, "[ perf record: perf size limit reached (%" PRIu64 " KB),"
 265				" stopping session ]\n",
 266				record__bytes_written(rec) >> 10);
 267		done = 1;
 268	}
 269
 270	if (switch_output_size(rec))
 271		trigger_hit(&switch_output_trigger);
 272
 273	return 0;
 274}
 275
 276static int record__aio_enabled(struct record *rec);
 277static int record__comp_enabled(struct record *rec);
 278static size_t zstd_compress(struct perf_session *session, struct mmap *map,
 279			    void *dst, size_t dst_size, void *src, size_t src_size);
 280
 281#ifdef HAVE_AIO_SUPPORT
 282static int record__aio_write(struct aiocb *cblock, int trace_fd,
 283		void *buf, size_t size, off_t off)
 284{
 285	int rc;
 286
 287	cblock->aio_fildes = trace_fd;
 288	cblock->aio_buf    = buf;
 289	cblock->aio_nbytes = size;
 290	cblock->aio_offset = off;
 291	cblock->aio_sigevent.sigev_notify = SIGEV_NONE;
 292
 293	do {
 294		rc = aio_write(cblock);
 295		if (rc == 0) {
 296			break;
 297		} else if (errno != EAGAIN) {
 298			cblock->aio_fildes = -1;
 299			pr_err("failed to queue perf data, error: %m\n");
 300			break;
 301		}
 302	} while (1);
 303
 304	return rc;
 305}
 306
 307static int record__aio_complete(struct mmap *md, struct aiocb *cblock)
 308{
 309	void *rem_buf;
 310	off_t rem_off;
 311	size_t rem_size;
 312	int rc, aio_errno;
 313	ssize_t aio_ret, written;
 
 
 314
 315	aio_errno = aio_error(cblock);
 316	if (aio_errno == EINPROGRESS)
 317		return 0;
 318
 319	written = aio_ret = aio_return(cblock);
 320	if (aio_ret < 0) {
 321		if (aio_errno != EINTR)
 322			pr_err("failed to write perf data, error: %m\n");
 323		written = 0;
 324	}
 325
 326	rem_size = cblock->aio_nbytes - written;
 327
 328	if (rem_size == 0) {
 329		cblock->aio_fildes = -1;
 330		/*
 331		 * md->refcount is incremented in record__aio_pushfn() for
 332		 * every aio write request started in record__aio_push() so
 333		 * decrement it because the request is now complete.
 334		 */
 335		perf_mmap__put(&md->core);
 336		rc = 1;
 337	} else {
 338		/*
 339		 * aio write request may require restart with the
 340		 * reminder if the kernel didn't write whole
 341		 * chunk at once.
 342		 */
 343		rem_off = cblock->aio_offset + written;
 344		rem_buf = (void *)(cblock->aio_buf + written);
 345		record__aio_write(cblock, cblock->aio_fildes,
 346				rem_buf, rem_size, rem_off);
 347		rc = 0;
 348	}
 349
 350	return rc;
 351}
 352
 353static int record__aio_sync(struct mmap *md, bool sync_all)
 354{
 355	struct aiocb **aiocb = md->aio.aiocb;
 356	struct aiocb *cblocks = md->aio.cblocks;
 357	struct timespec timeout = { 0, 1000 * 1000  * 1 }; /* 1ms */
 358	int i, do_suspend;
 359
 360	do {
 361		do_suspend = 0;
 362		for (i = 0; i < md->aio.nr_cblocks; ++i) {
 363			if (cblocks[i].aio_fildes == -1 || record__aio_complete(md, &cblocks[i])) {
 364				if (sync_all)
 365					aiocb[i] = NULL;
 366				else
 367					return i;
 368			} else {
 369				/*
 370				 * Started aio write is not complete yet
 371				 * so it has to be waited before the
 372				 * next allocation.
 373				 */
 374				aiocb[i] = &cblocks[i];
 375				do_suspend = 1;
 376			}
 377		}
 378		if (!do_suspend)
 379			return -1;
 380
 381		while (aio_suspend((const struct aiocb **)aiocb, md->aio.nr_cblocks, &timeout)) {
 382			if (!(errno == EAGAIN || errno == EINTR))
 383				pr_err("failed to sync perf data, error: %m\n");
 384		}
 385	} while (1);
 386}
 387
 388struct record_aio {
 389	struct record	*rec;
 390	void		*data;
 391	size_t		size;
 392};
 393
 394static int record__aio_pushfn(struct mmap *map, void *to, void *buf, size_t size)
 395{
 396	struct record_aio *aio = to;
 397
 398	/*
 399	 * map->core.base data pointed by buf is copied into free map->aio.data[] buffer
 400	 * to release space in the kernel buffer as fast as possible, calling
 401	 * perf_mmap__consume() from perf_mmap__push() function.
 402	 *
 403	 * That lets the kernel to proceed with storing more profiling data into
 404	 * the kernel buffer earlier than other per-cpu kernel buffers are handled.
 405	 *
 406	 * Coping can be done in two steps in case the chunk of profiling data
 407	 * crosses the upper bound of the kernel buffer. In this case we first move
 408	 * part of data from map->start till the upper bound and then the reminder
 409	 * from the beginning of the kernel buffer till the end of the data chunk.
 410	 */
 411
 412	if (record__comp_enabled(aio->rec)) {
 413		size = zstd_compress(aio->rec->session, NULL, aio->data + aio->size,
 414				     mmap__mmap_len(map) - aio->size,
 415				     buf, size);
 416	} else {
 417		memcpy(aio->data + aio->size, buf, size);
 418	}
 419
 420	if (!aio->size) {
 421		/*
 422		 * Increment map->refcount to guard map->aio.data[] buffer
 423		 * from premature deallocation because map object can be
 424		 * released earlier than aio write request started on
 425		 * map->aio.data[] buffer is complete.
 426		 *
 427		 * perf_mmap__put() is done at record__aio_complete()
 428		 * after started aio request completion or at record__aio_push()
 429		 * if the request failed to start.
 430		 */
 431		perf_mmap__get(&map->core);
 432	}
 433
 434	aio->size += size;
 
 
 435
 436	return size;
 437}
 438
 439static int record__aio_push(struct record *rec, struct mmap *map, off_t *off)
 440{
 441	int ret, idx;
 442	int trace_fd = rec->session->data->file.fd;
 443	struct record_aio aio = { .rec = rec, .size = 0 };
 444
 445	/*
 446	 * Call record__aio_sync() to wait till map->aio.data[] buffer
 447	 * becomes available after previous aio write operation.
 448	 */
 449
 450	idx = record__aio_sync(map, false);
 451	aio.data = map->aio.data[idx];
 452	ret = perf_mmap__push(map, &aio, record__aio_pushfn);
 453	if (ret != 0) /* ret > 0 - no data, ret < 0 - error */
 454		return ret;
 455
 456	rec->samples++;
 457	ret = record__aio_write(&(map->aio.cblocks[idx]), trace_fd, aio.data, aio.size, *off);
 458	if (!ret) {
 459		*off += aio.size;
 460		rec->bytes_written += aio.size;
 461		if (switch_output_size(rec))
 462			trigger_hit(&switch_output_trigger);
 463	} else {
 464		/*
 465		 * Decrement map->refcount incremented in record__aio_pushfn()
 466		 * back if record__aio_write() operation failed to start, otherwise
 467		 * map->refcount is decremented in record__aio_complete() after
 468		 * aio write operation finishes successfully.
 469		 */
 470		perf_mmap__put(&map->core);
 471	}
 472
 473	return ret;
 
 
 
 474}
 475
 476static off_t record__aio_get_pos(int trace_fd)
 477{
 478	return lseek(trace_fd, 0, SEEK_CUR);
 479}
 480
 481static void record__aio_set_pos(int trace_fd, off_t pos)
 482{
 483	lseek(trace_fd, pos, SEEK_SET);
 484}
 485
 486static void record__aio_mmap_read_sync(struct record *rec)
 487{
 488	int i;
 489	struct evlist *evlist = rec->evlist;
 490	struct mmap *maps = evlist->mmap;
 491
 492	if (!record__aio_enabled(rec))
 493		return;
 494
 495	for (i = 0; i < evlist->core.nr_mmaps; i++) {
 496		struct mmap *map = &maps[i];
 497
 498		if (map->core.base)
 499			record__aio_sync(map, true);
 500	}
 501}
 502
 503static int nr_cblocks_default = 1;
 504static int nr_cblocks_max = 4;
 505
 506static int record__aio_parse(const struct option *opt,
 507			     const char *str,
 508			     int unset)
 509{
 510	struct record_opts *opts = (struct record_opts *)opt->value;
 511
 512	if (unset) {
 513		opts->nr_cblocks = 0;
 514	} else {
 515		if (str)
 516			opts->nr_cblocks = strtol(str, NULL, 0);
 517		if (!opts->nr_cblocks)
 518			opts->nr_cblocks = nr_cblocks_default;
 519	}
 520
 521	return 0;
 522}
 523#else /* HAVE_AIO_SUPPORT */
 524static int nr_cblocks_max = 0;
 525
 526static int record__aio_push(struct record *rec __maybe_unused, struct mmap *map __maybe_unused,
 527			    off_t *off __maybe_unused)
 528{
 529	return -1;
 530}
 531
 532static off_t record__aio_get_pos(int trace_fd __maybe_unused)
 533{
 534	return -1;
 535}
 536
 537static void record__aio_set_pos(int trace_fd __maybe_unused, off_t pos __maybe_unused)
 538{
 539}
 540
 541static void record__aio_mmap_read_sync(struct record *rec __maybe_unused)
 542{
 543}
 544#endif
 545
 546static int record__aio_enabled(struct record *rec)
 547{
 548	return rec->opts.nr_cblocks > 0;
 549}
 550
 551#define MMAP_FLUSH_DEFAULT 1
 552static int record__mmap_flush_parse(const struct option *opt,
 553				    const char *str,
 554				    int unset)
 555{
 556	int flush_max;
 557	struct record_opts *opts = (struct record_opts *)opt->value;
 558	static struct parse_tag tags[] = {
 559			{ .tag  = 'B', .mult = 1       },
 560			{ .tag  = 'K', .mult = 1 << 10 },
 561			{ .tag  = 'M', .mult = 1 << 20 },
 562			{ .tag  = 'G', .mult = 1 << 30 },
 563			{ .tag  = 0 },
 564	};
 565
 566	if (unset)
 567		return 0;
 568
 569	if (str) {
 570		opts->mmap_flush = parse_tag_value(str, tags);
 571		if (opts->mmap_flush == (int)-1)
 572			opts->mmap_flush = strtol(str, NULL, 0);
 573	}
 574
 575	if (!opts->mmap_flush)
 576		opts->mmap_flush = MMAP_FLUSH_DEFAULT;
 577
 578	flush_max = evlist__mmap_size(opts->mmap_pages);
 579	flush_max /= 4;
 580	if (opts->mmap_flush > flush_max)
 581		opts->mmap_flush = flush_max;
 582
 583	return 0;
 584}
 585
 586#ifdef HAVE_ZSTD_SUPPORT
 587static unsigned int comp_level_default = 1;
 588
 589static int record__parse_comp_level(const struct option *opt, const char *str, int unset)
 590{
 591	struct record_opts *opts = opt->value;
 592
 593	if (unset) {
 594		opts->comp_level = 0;
 595	} else {
 596		if (str)
 597			opts->comp_level = strtol(str, NULL, 0);
 598		if (!opts->comp_level)
 599			opts->comp_level = comp_level_default;
 600	}
 601
 602	return 0;
 603}
 604#endif
 605static unsigned int comp_level_max = 22;
 606
 607static int record__comp_enabled(struct record *rec)
 608{
 609	return rec->opts.comp_level > 0;
 610}
 611
 612static int process_synthesized_event(struct perf_tool *tool,
 613				     union perf_event *event,
 614				     struct perf_sample *sample __maybe_unused,
 615				     struct machine *machine __maybe_unused)
 616{
 617	struct record *rec = container_of(tool, struct record, tool);
 618	return record__write(rec, NULL, event, event->header.size);
 619}
 620
 621static struct mutex synth_lock;
 622
 623static int process_locked_synthesized_event(struct perf_tool *tool,
 624				     union perf_event *event,
 625				     struct perf_sample *sample __maybe_unused,
 626				     struct machine *machine __maybe_unused)
 627{
 628	int ret;
 629
 630	mutex_lock(&synth_lock);
 631	ret = process_synthesized_event(tool, event, sample, machine);
 632	mutex_unlock(&synth_lock);
 633	return ret;
 634}
 635
 636static int record__pushfn(struct mmap *map, void *to, void *bf, size_t size)
 637{
 638	struct record *rec = to;
 639
 640	if (record__comp_enabled(rec)) {
 641		size = zstd_compress(rec->session, map, map->data, mmap__mmap_len(map), bf, size);
 642		bf   = map->data;
 643	}
 644
 645	thread->samples++;
 646	return record__write(rec, map, bf, size);
 647}
 648
 649static volatile sig_atomic_t signr = -1;
 650static volatile sig_atomic_t child_finished;
 651#ifdef HAVE_EVENTFD_SUPPORT
 652static volatile sig_atomic_t done_fd = -1;
 653#endif
 654
 655static void sig_handler(int sig)
 656{
 657	if (sig == SIGCHLD)
 658		child_finished = 1;
 659	else
 660		signr = sig;
 661
 662	done = 1;
 663#ifdef HAVE_EVENTFD_SUPPORT
 664	if (done_fd >= 0) {
 665		u64 tmp = 1;
 666		int orig_errno = errno;
 667
 668		/*
 669		 * It is possible for this signal handler to run after done is
 670		 * checked in the main loop, but before the perf counter fds are
 671		 * polled. If this happens, the poll() will continue to wait
 672		 * even though done is set, and will only break out if either
 673		 * another signal is received, or the counters are ready for
 674		 * read. To ensure the poll() doesn't sleep when done is set,
 675		 * use an eventfd (done_fd) to wake up the poll().
 676		 */
 677		if (write(done_fd, &tmp, sizeof(tmp)) < 0)
 678			pr_err("failed to signal wakeup fd, error: %m\n");
 679
 680		errno = orig_errno;
 681	}
 682#endif // HAVE_EVENTFD_SUPPORT
 683}
 684
 685static void sigsegv_handler(int sig)
 686{
 687	perf_hooks__recover();
 688	sighandler_dump_stack(sig);
 689}
 690
 691static void record__sig_exit(void)
 692{
 693	if (signr == -1)
 694		return;
 695
 696	signal(signr, SIG_DFL);
 697	raise(signr);
 698}
 699
 700#ifdef HAVE_AUXTRACE_SUPPORT
 701
 702static int record__process_auxtrace(struct perf_tool *tool,
 703				    struct mmap *map,
 704				    union perf_event *event, void *data1,
 705				    size_t len1, void *data2, size_t len2)
 706{
 707	struct record *rec = container_of(tool, struct record, tool);
 708	struct perf_data *data = &rec->data;
 709	size_t padding;
 710	u8 pad[8] = {0};
 711
 712	if (!perf_data__is_pipe(data) && perf_data__is_single_file(data)) {
 713		off_t file_offset;
 714		int fd = perf_data__fd(data);
 715		int err;
 716
 717		file_offset = lseek(fd, 0, SEEK_CUR);
 718		if (file_offset == -1)
 719			return -1;
 720		err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 721						     event, file_offset);
 722		if (err)
 723			return err;
 724	}
 725
 726	/* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 727	padding = (len1 + len2) & 7;
 728	if (padding)
 729		padding = 8 - padding;
 730
 731	record__write(rec, map, event, event->header.size);
 732	record__write(rec, map, data1, len1);
 733	if (len2)
 734		record__write(rec, map, data2, len2);
 735	record__write(rec, map, &pad, padding);
 736
 737	return 0;
 738}
 739
 740static int record__auxtrace_mmap_read(struct record *rec,
 741				      struct mmap *map)
 742{
 743	int ret;
 744
 745	ret = auxtrace_mmap__read(map, rec->itr, &rec->tool,
 746				  record__process_auxtrace);
 747	if (ret < 0)
 748		return ret;
 749
 750	if (ret)
 751		rec->samples++;
 752
 753	return 0;
 754}
 755
 756static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 757					       struct mmap *map)
 758{
 759	int ret;
 760
 761	ret = auxtrace_mmap__read_snapshot(map, rec->itr, &rec->tool,
 762					   record__process_auxtrace,
 763					   rec->opts.auxtrace_snapshot_size);
 764	if (ret < 0)
 765		return ret;
 766
 767	if (ret)
 768		rec->samples++;
 769
 770	return 0;
 771}
 772
 773static int record__auxtrace_read_snapshot_all(struct record *rec)
 774{
 775	int i;
 776	int rc = 0;
 777
 778	for (i = 0; i < rec->evlist->core.nr_mmaps; i++) {
 779		struct mmap *map = &rec->evlist->mmap[i];
 
 780
 781		if (!map->auxtrace_mmap.base)
 782			continue;
 783
 784		if (record__auxtrace_mmap_read_snapshot(rec, map) != 0) {
 785			rc = -1;
 786			goto out;
 787		}
 788	}
 789out:
 790	return rc;
 791}
 792
 793static void record__read_auxtrace_snapshot(struct record *rec, bool on_exit)
 794{
 795	pr_debug("Recording AUX area tracing snapshot\n");
 796	if (record__auxtrace_read_snapshot_all(rec) < 0) {
 797		trigger_error(&auxtrace_snapshot_trigger);
 798	} else {
 799		if (auxtrace_record__snapshot_finish(rec->itr, on_exit))
 800			trigger_error(&auxtrace_snapshot_trigger);
 801		else
 802			trigger_ready(&auxtrace_snapshot_trigger);
 803	}
 804}
 805
 806static int record__auxtrace_snapshot_exit(struct record *rec)
 807{
 808	if (trigger_is_error(&auxtrace_snapshot_trigger))
 809		return 0;
 810
 811	if (!auxtrace_record__snapshot_started &&
 812	    auxtrace_record__snapshot_start(rec->itr))
 813		return -1;
 814
 815	record__read_auxtrace_snapshot(rec, true);
 816	if (trigger_is_error(&auxtrace_snapshot_trigger))
 817		return -1;
 818
 819	return 0;
 820}
 821
 822static int record__auxtrace_init(struct record *rec)
 823{
 824	int err;
 825
 826	if ((rec->opts.auxtrace_snapshot_opts || rec->opts.auxtrace_sample_opts)
 827	    && record__threads_enabled(rec)) {
 828		pr_err("AUX area tracing options are not available in parallel streaming mode.\n");
 829		return -EINVAL;
 830	}
 831
 832	if (!rec->itr) {
 833		rec->itr = auxtrace_record__init(rec->evlist, &err);
 834		if (err)
 835			return err;
 836	}
 837
 838	err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
 839					      rec->opts.auxtrace_snapshot_opts);
 840	if (err)
 841		return err;
 842
 843	err = auxtrace_parse_sample_options(rec->itr, rec->evlist, &rec->opts,
 844					    rec->opts.auxtrace_sample_opts);
 845	if (err)
 846		return err;
 847
 848	auxtrace_regroup_aux_output(rec->evlist);
 849
 850	return auxtrace_parse_filters(rec->evlist);
 851}
 852
 853#else
 854
 855static inline
 856int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 857			       struct mmap *map __maybe_unused)
 858{
 859	return 0;
 860}
 861
 862static inline
 863void record__read_auxtrace_snapshot(struct record *rec __maybe_unused,
 864				    bool on_exit __maybe_unused)
 865{
 866}
 867
 868static inline
 869int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 870{
 871	return 0;
 872}
 873
 874static inline
 875int record__auxtrace_snapshot_exit(struct record *rec __maybe_unused)
 876{
 877	return 0;
 878}
 879
 880static int record__auxtrace_init(struct record *rec __maybe_unused)
 881{
 882	return 0;
 883}
 884
 885#endif
 886
 887static int record__config_text_poke(struct evlist *evlist)
 888{
 889	struct evsel *evsel;
 890
 891	/* Nothing to do if text poke is already configured */
 892	evlist__for_each_entry(evlist, evsel) {
 893		if (evsel->core.attr.text_poke)
 894			return 0;
 895	}
 896
 897	evsel = evlist__add_dummy_on_all_cpus(evlist);
 898	if (!evsel)
 899		return -ENOMEM;
 900
 901	evsel->core.attr.text_poke = 1;
 902	evsel->core.attr.ksymbol = 1;
 903	evsel->immediate = true;
 904	evsel__set_sample_bit(evsel, TIME);
 905
 906	return 0;
 907}
 908
 909static int record__config_off_cpu(struct record *rec)
 910{
 911	return off_cpu_prepare(rec->evlist, &rec->opts.target, &rec->opts);
 912}
 913
 914static bool record__kcore_readable(struct machine *machine)
 915{
 916	char kcore[PATH_MAX];
 917	int fd;
 918
 919	scnprintf(kcore, sizeof(kcore), "%s/proc/kcore", machine->root_dir);
 920
 921	fd = open(kcore, O_RDONLY);
 922	if (fd < 0)
 923		return false;
 924
 925	close(fd);
 926
 927	return true;
 928}
 929
 930static int record__kcore_copy(struct machine *machine, struct perf_data *data)
 931{
 932	char from_dir[PATH_MAX];
 933	char kcore_dir[PATH_MAX];
 934	int ret;
 935
 936	snprintf(from_dir, sizeof(from_dir), "%s/proc", machine->root_dir);
 937
 938	ret = perf_data__make_kcore_dir(data, kcore_dir, sizeof(kcore_dir));
 939	if (ret)
 940		return ret;
 941
 942	return kcore_copy(from_dir, kcore_dir);
 943}
 944
 945static void record__thread_data_init_pipes(struct record_thread *thread_data)
 946{
 947	thread_data->pipes.msg[0] = -1;
 948	thread_data->pipes.msg[1] = -1;
 949	thread_data->pipes.ack[0] = -1;
 950	thread_data->pipes.ack[1] = -1;
 951}
 952
 953static int record__thread_data_open_pipes(struct record_thread *thread_data)
 954{
 955	if (pipe(thread_data->pipes.msg))
 956		return -EINVAL;
 957
 958	if (pipe(thread_data->pipes.ack)) {
 959		close(thread_data->pipes.msg[0]);
 960		thread_data->pipes.msg[0] = -1;
 961		close(thread_data->pipes.msg[1]);
 962		thread_data->pipes.msg[1] = -1;
 963		return -EINVAL;
 964	}
 965
 966	pr_debug2("thread_data[%p]: msg=[%d,%d], ack=[%d,%d]\n", thread_data,
 967		 thread_data->pipes.msg[0], thread_data->pipes.msg[1],
 968		 thread_data->pipes.ack[0], thread_data->pipes.ack[1]);
 969
 970	return 0;
 971}
 972
 973static void record__thread_data_close_pipes(struct record_thread *thread_data)
 974{
 975	if (thread_data->pipes.msg[0] != -1) {
 976		close(thread_data->pipes.msg[0]);
 977		thread_data->pipes.msg[0] = -1;
 978	}
 979	if (thread_data->pipes.msg[1] != -1) {
 980		close(thread_data->pipes.msg[1]);
 981		thread_data->pipes.msg[1] = -1;
 982	}
 983	if (thread_data->pipes.ack[0] != -1) {
 984		close(thread_data->pipes.ack[0]);
 985		thread_data->pipes.ack[0] = -1;
 986	}
 987	if (thread_data->pipes.ack[1] != -1) {
 988		close(thread_data->pipes.ack[1]);
 989		thread_data->pipes.ack[1] = -1;
 990	}
 991}
 992
 993static bool evlist__per_thread(struct evlist *evlist)
 994{
 995	return cpu_map__is_dummy(evlist->core.user_requested_cpus);
 996}
 997
 998static int record__thread_data_init_maps(struct record_thread *thread_data, struct evlist *evlist)
 999{
1000	int m, tm, nr_mmaps = evlist->core.nr_mmaps;
1001	struct mmap *mmap = evlist->mmap;
1002	struct mmap *overwrite_mmap = evlist->overwrite_mmap;
1003	struct perf_cpu_map *cpus = evlist->core.all_cpus;
1004	bool per_thread = evlist__per_thread(evlist);
1005
1006	if (per_thread)
1007		thread_data->nr_mmaps = nr_mmaps;
1008	else
1009		thread_data->nr_mmaps = bitmap_weight(thread_data->mask->maps.bits,
1010						      thread_data->mask->maps.nbits);
1011	if (mmap) {
1012		thread_data->maps = zalloc(thread_data->nr_mmaps * sizeof(struct mmap *));
1013		if (!thread_data->maps)
1014			return -ENOMEM;
1015	}
1016	if (overwrite_mmap) {
1017		thread_data->overwrite_maps = zalloc(thread_data->nr_mmaps * sizeof(struct mmap *));
1018		if (!thread_data->overwrite_maps) {
1019			zfree(&thread_data->maps);
1020			return -ENOMEM;
1021		}
1022	}
1023	pr_debug2("thread_data[%p]: nr_mmaps=%d, maps=%p, ow_maps=%p\n", thread_data,
1024		 thread_data->nr_mmaps, thread_data->maps, thread_data->overwrite_maps);
1025
1026	for (m = 0, tm = 0; m < nr_mmaps && tm < thread_data->nr_mmaps; m++) {
1027		if (per_thread ||
1028		    test_bit(perf_cpu_map__cpu(cpus, m).cpu, thread_data->mask->maps.bits)) {
1029			if (thread_data->maps) {
1030				thread_data->maps[tm] = &mmap[m];
1031				pr_debug2("thread_data[%p]: cpu%d: maps[%d] -> mmap[%d]\n",
1032					  thread_data, perf_cpu_map__cpu(cpus, m).cpu, tm, m);
1033			}
1034			if (thread_data->overwrite_maps) {
1035				thread_data->overwrite_maps[tm] = &overwrite_mmap[m];
1036				pr_debug2("thread_data[%p]: cpu%d: ow_maps[%d] -> ow_mmap[%d]\n",
1037					  thread_data, perf_cpu_map__cpu(cpus, m).cpu, tm, m);
1038			}
1039			tm++;
1040		}
1041	}
1042
1043	return 0;
1044}
1045
1046static int record__thread_data_init_pollfd(struct record_thread *thread_data, struct evlist *evlist)
1047{
1048	int f, tm, pos;
1049	struct mmap *map, *overwrite_map;
1050
1051	fdarray__init(&thread_data->pollfd, 64);
1052
1053	for (tm = 0; tm < thread_data->nr_mmaps; tm++) {
1054		map = thread_data->maps ? thread_data->maps[tm] : NULL;
1055		overwrite_map = thread_data->overwrite_maps ?
1056				thread_data->overwrite_maps[tm] : NULL;
1057
1058		for (f = 0; f < evlist->core.pollfd.nr; f++) {
1059			void *ptr = evlist->core.pollfd.priv[f].ptr;
1060
1061			if ((map && ptr == map) || (overwrite_map && ptr == overwrite_map)) {
1062				pos = fdarray__dup_entry_from(&thread_data->pollfd, f,
1063							      &evlist->core.pollfd);
1064				if (pos < 0)
1065					return pos;
1066				pr_debug2("thread_data[%p]: pollfd[%d] <- event_fd=%d\n",
1067					 thread_data, pos, evlist->core.pollfd.entries[f].fd);
1068			}
1069		}
1070	}
1071
1072	return 0;
1073}
1074
1075static void record__free_thread_data(struct record *rec)
1076{
1077	int t;
1078	struct record_thread *thread_data = rec->thread_data;
1079
1080	if (thread_data == NULL)
1081		return;
1082
1083	for (t = 0; t < rec->nr_threads; t++) {
1084		record__thread_data_close_pipes(&thread_data[t]);
1085		zfree(&thread_data[t].maps);
1086		zfree(&thread_data[t].overwrite_maps);
1087		fdarray__exit(&thread_data[t].pollfd);
1088	}
1089
1090	zfree(&rec->thread_data);
1091}
1092
1093static int record__map_thread_evlist_pollfd_indexes(struct record *rec,
1094						    int evlist_pollfd_index,
1095						    int thread_pollfd_index)
1096{
1097	size_t x = rec->index_map_cnt;
1098
1099	if (realloc_array_as_needed(rec->index_map, rec->index_map_sz, x, NULL))
1100		return -ENOMEM;
1101	rec->index_map[x].evlist_pollfd_index = evlist_pollfd_index;
1102	rec->index_map[x].thread_pollfd_index = thread_pollfd_index;
1103	rec->index_map_cnt += 1;
1104	return 0;
1105}
1106
1107static int record__update_evlist_pollfd_from_thread(struct record *rec,
1108						    struct evlist *evlist,
1109						    struct record_thread *thread_data)
1110{
1111	struct pollfd *e_entries = evlist->core.pollfd.entries;
1112	struct pollfd *t_entries = thread_data->pollfd.entries;
1113	int err = 0;
1114	size_t i;
1115
1116	for (i = 0; i < rec->index_map_cnt; i++) {
1117		int e_pos = rec->index_map[i].evlist_pollfd_index;
1118		int t_pos = rec->index_map[i].thread_pollfd_index;
1119
1120		if (e_entries[e_pos].fd != t_entries[t_pos].fd ||
1121		    e_entries[e_pos].events != t_entries[t_pos].events) {
1122			pr_err("Thread and evlist pollfd index mismatch\n");
1123			err = -EINVAL;
1124			continue;
1125		}
1126		e_entries[e_pos].revents = t_entries[t_pos].revents;
1127	}
1128	return err;
1129}
1130
1131static int record__dup_non_perf_events(struct record *rec,
1132				       struct evlist *evlist,
1133				       struct record_thread *thread_data)
1134{
1135	struct fdarray *fda = &evlist->core.pollfd;
1136	int i, ret;
1137
1138	for (i = 0; i < fda->nr; i++) {
1139		if (!(fda->priv[i].flags & fdarray_flag__non_perf_event))
1140			continue;
1141		ret = fdarray__dup_entry_from(&thread_data->pollfd, i, fda);
1142		if (ret < 0) {
1143			pr_err("Failed to duplicate descriptor in main thread pollfd\n");
1144			return ret;
1145		}
1146		pr_debug2("thread_data[%p]: pollfd[%d] <- non_perf_event fd=%d\n",
1147			  thread_data, ret, fda->entries[i].fd);
1148		ret = record__map_thread_evlist_pollfd_indexes(rec, i, ret);
1149		if (ret < 0) {
1150			pr_err("Failed to map thread and evlist pollfd indexes\n");
1151			return ret;
1152		}
1153	}
1154	return 0;
1155}
1156
1157static int record__alloc_thread_data(struct record *rec, struct evlist *evlist)
1158{
1159	int t, ret;
1160	struct record_thread *thread_data;
1161
1162	rec->thread_data = zalloc(rec->nr_threads * sizeof(*(rec->thread_data)));
1163	if (!rec->thread_data) {
1164		pr_err("Failed to allocate thread data\n");
1165		return -ENOMEM;
1166	}
1167	thread_data = rec->thread_data;
1168
1169	for (t = 0; t < rec->nr_threads; t++)
1170		record__thread_data_init_pipes(&thread_data[t]);
1171
1172	for (t = 0; t < rec->nr_threads; t++) {
1173		thread_data[t].rec = rec;
1174		thread_data[t].mask = &rec->thread_masks[t];
1175		ret = record__thread_data_init_maps(&thread_data[t], evlist);
1176		if (ret) {
1177			pr_err("Failed to initialize thread[%d] maps\n", t);
1178			goto out_free;
1179		}
1180		ret = record__thread_data_init_pollfd(&thread_data[t], evlist);
1181		if (ret) {
1182			pr_err("Failed to initialize thread[%d] pollfd\n", t);
1183			goto out_free;
1184		}
1185		if (t) {
1186			thread_data[t].tid = -1;
1187			ret = record__thread_data_open_pipes(&thread_data[t]);
1188			if (ret) {
1189				pr_err("Failed to open thread[%d] communication pipes\n", t);
1190				goto out_free;
1191			}
1192			ret = fdarray__add(&thread_data[t].pollfd, thread_data[t].pipes.msg[0],
1193					   POLLIN | POLLERR | POLLHUP, fdarray_flag__nonfilterable);
1194			if (ret < 0) {
1195				pr_err("Failed to add descriptor to thread[%d] pollfd\n", t);
1196				goto out_free;
1197			}
1198			thread_data[t].ctlfd_pos = ret;
1199			pr_debug2("thread_data[%p]: pollfd[%d] <- ctl_fd=%d\n",
1200				 thread_data, thread_data[t].ctlfd_pos,
1201				 thread_data[t].pipes.msg[0]);
1202		} else {
1203			thread_data[t].tid = gettid();
1204
1205			ret = record__dup_non_perf_events(rec, evlist, &thread_data[t]);
1206			if (ret < 0)
1207				goto out_free;
1208
1209			thread_data[t].ctlfd_pos = -1; /* Not used */
1210		}
1211	}
1212
1213	return 0;
1214
1215out_free:
1216	record__free_thread_data(rec);
1217
1218	return ret;
1219}
1220
1221static int record__mmap_evlist(struct record *rec,
1222			       struct evlist *evlist)
1223{
1224	int i, ret;
1225	struct record_opts *opts = &rec->opts;
1226	bool auxtrace_overwrite = opts->auxtrace_snapshot_mode ||
1227				  opts->auxtrace_sample_mode;
1228	char msg[512];
1229
1230	if (opts->affinity != PERF_AFFINITY_SYS)
1231		cpu__setup_cpunode_map();
1232
1233	if (evlist__mmap_ex(evlist, opts->mmap_pages,
1234				 opts->auxtrace_mmap_pages,
1235				 auxtrace_overwrite,
1236				 opts->nr_cblocks, opts->affinity,
1237				 opts->mmap_flush, opts->comp_level) < 0) {
1238		if (errno == EPERM) {
1239			pr_err("Permission error mapping pages.\n"
1240			       "Consider increasing "
1241			       "/proc/sys/kernel/perf_event_mlock_kb,\n"
1242			       "or try again with a smaller value of -m/--mmap_pages.\n"
1243			       "(current value: %u,%u)\n",
1244			       opts->mmap_pages, opts->auxtrace_mmap_pages);
1245			return -errno;
1246		} else {
1247			pr_err("failed to mmap with %d (%s)\n", errno,
1248				str_error_r(errno, msg, sizeof(msg)));
1249			if (errno)
1250				return -errno;
1251			else
1252				return -EINVAL;
1253		}
1254	}
1255
1256	if (evlist__initialize_ctlfd(evlist, opts->ctl_fd, opts->ctl_fd_ack))
1257		return -1;
1258
1259	ret = record__alloc_thread_data(rec, evlist);
1260	if (ret)
1261		return ret;
1262
1263	if (record__threads_enabled(rec)) {
1264		ret = perf_data__create_dir(&rec->data, evlist->core.nr_mmaps);
1265		if (ret) {
1266			pr_err("Failed to create data directory: %s\n", strerror(-ret));
1267			return ret;
1268		}
1269		for (i = 0; i < evlist->core.nr_mmaps; i++) {
1270			if (evlist->mmap)
1271				evlist->mmap[i].file = &rec->data.dir.files[i];
1272			if (evlist->overwrite_mmap)
1273				evlist->overwrite_mmap[i].file = &rec->data.dir.files[i];
1274		}
1275	}
1276
1277	return 0;
1278}
1279
1280static int record__mmap(struct record *rec)
1281{
1282	return record__mmap_evlist(rec, rec->evlist);
1283}
1284
1285static int record__open(struct record *rec)
1286{
1287	char msg[BUFSIZ];
1288	struct evsel *pos;
1289	struct evlist *evlist = rec->evlist;
1290	struct perf_session *session = rec->session;
1291	struct record_opts *opts = &rec->opts;
1292	int rc = 0;
1293
1294	/*
1295	 * For initial_delay, system wide or a hybrid system, we need to add a
1296	 * dummy event so that we can track PERF_RECORD_MMAP to cover the delay
1297	 * of waiting or event synthesis.
1298	 */
1299	if (opts->initial_delay || target__has_cpu(&opts->target) ||
1300	    perf_pmu__has_hybrid()) {
1301		pos = evlist__get_tracking_event(evlist);
1302		if (!evsel__is_dummy_event(pos)) {
1303			/* Set up dummy event. */
1304			if (evlist__add_dummy(evlist))
1305				return -ENOMEM;
1306			pos = evlist__last(evlist);
1307			evlist__set_tracking_event(evlist, pos);
1308		}
1309
1310		/*
1311		 * Enable the dummy event when the process is forked for
1312		 * initial_delay, immediately for system wide.
1313		 */
1314		if (opts->initial_delay && !pos->immediate &&
1315		    !target__has_cpu(&opts->target))
1316			pos->core.attr.enable_on_exec = 1;
1317		else
1318			pos->immediate = 1;
1319	}
1320
1321	evlist__config(evlist, opts, &callchain_param);
1322
1323	evlist__for_each_entry(evlist, pos) {
1324try_again:
1325		if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
1326			if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
1327				if (verbose > 0)
1328					ui__warning("%s\n", msg);
1329				goto try_again;
1330			}
1331			if ((errno == EINVAL || errno == EBADF) &&
1332			    pos->core.leader != &pos->core &&
1333			    pos->weak_group) {
1334			        pos = evlist__reset_weak_group(evlist, pos, true);
1335				goto try_again;
1336			}
1337			rc = -errno;
1338			evsel__open_strerror(pos, &opts->target, errno, msg, sizeof(msg));
1339			ui__error("%s\n", msg);
1340			goto out;
1341		}
1342
1343		pos->supported = true;
1344	}
1345
1346	if (symbol_conf.kptr_restrict && !evlist__exclude_kernel(evlist)) {
1347		pr_warning(
1348"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1349"check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1350"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1351"file is not found in the buildid cache or in the vmlinux path.\n\n"
1352"Samples in kernel modules won't be resolved at all.\n\n"
1353"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1354"even with a suitable vmlinux or kallsyms file.\n\n");
1355	}
1356
1357	if (evlist__apply_filters(evlist, &pos)) {
1358		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
1359			pos->filter, evsel__name(pos), errno,
1360			str_error_r(errno, msg, sizeof(msg)));
1361		rc = -1;
1362		goto out;
1363	}
1364
1365	rc = record__mmap(rec);
1366	if (rc)
1367		goto out;
1368
1369	session->evlist = evlist;
1370	perf_session__set_id_hdr_size(session);
1371out:
1372	return rc;
1373}
1374
1375static void set_timestamp_boundary(struct record *rec, u64 sample_time)
1376{
1377	if (rec->evlist->first_sample_time == 0)
1378		rec->evlist->first_sample_time = sample_time;
1379
1380	if (sample_time)
1381		rec->evlist->last_sample_time = sample_time;
1382}
1383
1384static int process_sample_event(struct perf_tool *tool,
1385				union perf_event *event,
1386				struct perf_sample *sample,
1387				struct evsel *evsel,
1388				struct machine *machine)
1389{
1390	struct record *rec = container_of(tool, struct record, tool);
1391
1392	set_timestamp_boundary(rec, sample->time);
1393
1394	if (rec->buildid_all)
1395		return 0;
1396
1397	rec->samples++;
1398	return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
1399}
1400
1401static int process_buildids(struct record *rec)
1402{
 
1403	struct perf_session *session = rec->session;
1404
1405	if (perf_data__size(&rec->data) == 0)
1406		return 0;
1407
1408	/*
1409	 * During this process, it'll load kernel map and replace the
1410	 * dso->long_name to a real pathname it found.  In this case
1411	 * we prefer the vmlinux path like
1412	 *   /lib/modules/3.16.4/build/vmlinux
1413	 *
1414	 * rather than build-id path (in debug directory).
1415	 *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
1416	 */
1417	symbol_conf.ignore_vmlinux_buildid = true;
1418
1419	/*
1420	 * If --buildid-all is given, it marks all DSO regardless of hits,
1421	 * so no need to process samples. But if timestamp_boundary is enabled,
1422	 * it still needs to walk on all samples to get the timestamps of
1423	 * first/last samples.
1424	 */
1425	if (rec->buildid_all && !rec->timestamp_boundary)
1426		rec->tool.sample = NULL;
1427
1428	return perf_session__process_events(session);
1429}
1430
1431static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
1432{
1433	int err;
1434	struct perf_tool *tool = data;
1435	/*
1436	 *As for guest kernel when processing subcommand record&report,
1437	 *we arrange module mmap prior to guest kernel mmap and trigger
1438	 *a preload dso because default guest module symbols are loaded
1439	 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
1440	 *method is used to avoid symbol missing when the first addr is
1441	 *in module instead of in guest kernel.
1442	 */
1443	err = perf_event__synthesize_modules(tool, process_synthesized_event,
1444					     machine);
1445	if (err < 0)
1446		pr_err("Couldn't record guest kernel [%d]'s reference"
1447		       " relocation symbol.\n", machine->pid);
1448
1449	/*
1450	 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
1451	 * have no _text sometimes.
1452	 */
1453	err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1454						 machine);
1455	if (err < 0)
1456		pr_err("Couldn't record guest kernel [%d]'s reference"
1457		       " relocation symbol.\n", machine->pid);
1458}
1459
1460static struct perf_event_header finished_round_event = {
1461	.size = sizeof(struct perf_event_header),
1462	.type = PERF_RECORD_FINISHED_ROUND,
1463};
1464
1465static struct perf_event_header finished_init_event = {
1466	.size = sizeof(struct perf_event_header),
1467	.type = PERF_RECORD_FINISHED_INIT,
1468};
1469
1470static void record__adjust_affinity(struct record *rec, struct mmap *map)
1471{
1472	if (rec->opts.affinity != PERF_AFFINITY_SYS &&
1473	    !bitmap_equal(thread->mask->affinity.bits, map->affinity_mask.bits,
1474			  thread->mask->affinity.nbits)) {
1475		bitmap_zero(thread->mask->affinity.bits, thread->mask->affinity.nbits);
1476		bitmap_or(thread->mask->affinity.bits, thread->mask->affinity.bits,
1477			  map->affinity_mask.bits, thread->mask->affinity.nbits);
1478		sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&thread->mask->affinity),
1479					(cpu_set_t *)thread->mask->affinity.bits);
1480		if (verbose == 2) {
1481			pr_debug("threads[%d]: running on cpu%d: ", thread->tid, sched_getcpu());
1482			mmap_cpu_mask__scnprintf(&thread->mask->affinity, "affinity");
1483		}
1484	}
1485}
1486
1487static size_t process_comp_header(void *record, size_t increment)
1488{
1489	struct perf_record_compressed *event = record;
1490	size_t size = sizeof(*event);
1491
1492	if (increment) {
1493		event->header.size += increment;
1494		return increment;
1495	}
1496
1497	event->header.type = PERF_RECORD_COMPRESSED;
1498	event->header.size = size;
1499
1500	return size;
1501}
1502
1503static size_t zstd_compress(struct perf_session *session, struct mmap *map,
1504			    void *dst, size_t dst_size, void *src, size_t src_size)
1505{
1506	size_t compressed;
1507	size_t max_record_size = PERF_SAMPLE_MAX_SIZE - sizeof(struct perf_record_compressed) - 1;
1508	struct zstd_data *zstd_data = &session->zstd_data;
1509
1510	if (map && map->file)
1511		zstd_data = &map->zstd_data;
1512
1513	compressed = zstd_compress_stream_to_records(zstd_data, dst, dst_size, src, src_size,
1514						     max_record_size, process_comp_header);
1515
1516	if (map && map->file) {
1517		thread->bytes_transferred += src_size;
1518		thread->bytes_compressed  += compressed;
1519	} else {
1520		session->bytes_transferred += src_size;
1521		session->bytes_compressed  += compressed;
1522	}
1523
1524	return compressed;
1525}
1526
1527static int record__mmap_read_evlist(struct record *rec, struct evlist *evlist,
1528				    bool overwrite, bool synch)
1529{
1530	u64 bytes_written = rec->bytes_written;
1531	int i;
1532	int rc = 0;
1533	int nr_mmaps;
1534	struct mmap **maps;
1535	int trace_fd = rec->data.file.fd;
1536	off_t off = 0;
1537
1538	if (!evlist)
1539		return 0;
1540
1541	nr_mmaps = thread->nr_mmaps;
1542	maps = overwrite ? thread->overwrite_maps : thread->maps;
1543
1544	if (!maps)
1545		return 0;
1546
1547	if (overwrite && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
1548		return 0;
1549
1550	if (record__aio_enabled(rec))
1551		off = record__aio_get_pos(trace_fd);
1552
1553	for (i = 0; i < nr_mmaps; i++) {
1554		u64 flush = 0;
1555		struct mmap *map = maps[i];
1556
1557		if (map->core.base) {
1558			record__adjust_affinity(rec, map);
1559			if (synch) {
1560				flush = map->core.flush;
1561				map->core.flush = 1;
1562			}
1563			if (!record__aio_enabled(rec)) {
1564				if (perf_mmap__push(map, rec, record__pushfn) < 0) {
1565					if (synch)
1566						map->core.flush = flush;
1567					rc = -1;
1568					goto out;
1569				}
1570			} else {
1571				if (record__aio_push(rec, map, &off) < 0) {
1572					record__aio_set_pos(trace_fd, off);
1573					if (synch)
1574						map->core.flush = flush;
1575					rc = -1;
1576					goto out;
1577				}
1578			}
1579			if (synch)
1580				map->core.flush = flush;
1581		}
1582
1583		if (map->auxtrace_mmap.base && !rec->opts.auxtrace_snapshot_mode &&
1584		    !rec->opts.auxtrace_sample_mode &&
1585		    record__auxtrace_mmap_read(rec, map) != 0) {
1586			rc = -1;
1587			goto out;
1588		}
1589	}
1590
1591	if (record__aio_enabled(rec))
1592		record__aio_set_pos(trace_fd, off);
1593
1594	/*
1595	 * Mark the round finished in case we wrote
1596	 * at least one event.
1597	 *
1598	 * No need for round events in directory mode,
1599	 * because per-cpu maps and files have data
1600	 * sorted by kernel.
1601	 */
1602	if (!record__threads_enabled(rec) && bytes_written != rec->bytes_written)
1603		rc = record__write(rec, NULL, &finished_round_event, sizeof(finished_round_event));
1604
1605	if (overwrite)
1606		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
1607out:
1608	return rc;
1609}
1610
1611static int record__mmap_read_all(struct record *rec, bool synch)
1612{
1613	int err;
1614
1615	err = record__mmap_read_evlist(rec, rec->evlist, false, synch);
1616	if (err)
1617		return err;
1618
1619	return record__mmap_read_evlist(rec, rec->evlist, true, synch);
1620}
1621
1622static void record__thread_munmap_filtered(struct fdarray *fda, int fd,
1623					   void *arg __maybe_unused)
1624{
1625	struct perf_mmap *map = fda->priv[fd].ptr;
1626
1627	if (map)
1628		perf_mmap__put(map);
1629}
1630
1631static void *record__thread(void *arg)
1632{
1633	enum thread_msg msg = THREAD_MSG__READY;
1634	bool terminate = false;
1635	struct fdarray *pollfd;
1636	int err, ctlfd_pos;
1637
1638	thread = arg;
1639	thread->tid = gettid();
1640
1641	err = write(thread->pipes.ack[1], &msg, sizeof(msg));
1642	if (err == -1)
1643		pr_warning("threads[%d]: failed to notify on start: %s\n",
1644			   thread->tid, strerror(errno));
1645
1646	pr_debug("threads[%d]: started on cpu%d\n", thread->tid, sched_getcpu());
1647
1648	pollfd = &thread->pollfd;
1649	ctlfd_pos = thread->ctlfd_pos;
1650
1651	for (;;) {
1652		unsigned long long hits = thread->samples;
1653
1654		if (record__mmap_read_all(thread->rec, false) < 0 || terminate)
1655			break;
1656
1657		if (hits == thread->samples) {
1658
1659			err = fdarray__poll(pollfd, -1);
1660			/*
1661			 * Propagate error, only if there's any. Ignore positive
1662			 * number of returned events and interrupt error.
1663			 */
1664			if (err > 0 || (err < 0 && errno == EINTR))
1665				err = 0;
1666			thread->waking++;
1667
1668			if (fdarray__filter(pollfd, POLLERR | POLLHUP,
1669					    record__thread_munmap_filtered, NULL) == 0)
1670				break;
1671		}
1672
1673		if (pollfd->entries[ctlfd_pos].revents & POLLHUP) {
1674			terminate = true;
1675			close(thread->pipes.msg[0]);
1676			thread->pipes.msg[0] = -1;
1677			pollfd->entries[ctlfd_pos].fd = -1;
1678			pollfd->entries[ctlfd_pos].events = 0;
1679		}
1680
1681		pollfd->entries[ctlfd_pos].revents = 0;
1682	}
1683	record__mmap_read_all(thread->rec, true);
1684
1685	err = write(thread->pipes.ack[1], &msg, sizeof(msg));
1686	if (err == -1)
1687		pr_warning("threads[%d]: failed to notify on termination: %s\n",
1688			   thread->tid, strerror(errno));
1689
1690	return NULL;
1691}
1692
1693static void record__init_features(struct record *rec)
1694{
1695	struct perf_session *session = rec->session;
1696	int feat;
1697
1698	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1699		perf_header__set_feat(&session->header, feat);
1700
1701	if (rec->no_buildid)
1702		perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1703
1704#ifdef HAVE_LIBTRACEEVENT
1705	if (!have_tracepoints(&rec->evlist->core.entries))
1706		perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1707#endif
1708
1709	if (!rec->opts.branch_stack)
1710		perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1711
1712	if (!rec->opts.full_auxtrace)
1713		perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1714
1715	if (!(rec->opts.use_clockid && rec->opts.clockid_res_ns))
1716		perf_header__clear_feat(&session->header, HEADER_CLOCKID);
1717
1718	if (!rec->opts.use_clockid)
1719		perf_header__clear_feat(&session->header, HEADER_CLOCK_DATA);
1720
1721	if (!record__threads_enabled(rec))
1722		perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1723
1724	if (!record__comp_enabled(rec))
1725		perf_header__clear_feat(&session->header, HEADER_COMPRESSED);
1726
1727	perf_header__clear_feat(&session->header, HEADER_STAT);
1728}
1729
1730static void
1731record__finish_output(struct record *rec)
1732{
1733	int i;
1734	struct perf_data *data = &rec->data;
1735	int fd = perf_data__fd(data);
1736
1737	if (data->is_pipe)
1738		return;
1739
1740	rec->session->header.data_size += rec->bytes_written;
1741	data->file.size = lseek(perf_data__fd(data), 0, SEEK_CUR);
1742	if (record__threads_enabled(rec)) {
1743		for (i = 0; i < data->dir.nr; i++)
1744			data->dir.files[i].size = lseek(data->dir.files[i].fd, 0, SEEK_CUR);
1745	}
1746
1747	if (!rec->no_buildid) {
1748		process_buildids(rec);
1749
1750		if (rec->buildid_all)
1751			dsos__hit_all(rec->session);
1752	}
1753	perf_session__write_header(rec->session, rec->evlist, fd, true);
1754
1755	return;
1756}
1757
1758static int record__synthesize_workload(struct record *rec, bool tail)
1759{
1760	int err;
1761	struct perf_thread_map *thread_map;
1762	bool needs_mmap = rec->opts.synth & PERF_SYNTH_MMAP;
1763
1764	if (rec->opts.tail_synthesize != tail)
1765		return 0;
1766
1767	thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
1768	if (thread_map == NULL)
1769		return -1;
1770
1771	err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
1772						 process_synthesized_event,
1773						 &rec->session->machines.host,
1774						 needs_mmap,
1775						 rec->opts.sample_address);
1776	perf_thread_map__put(thread_map);
1777	return err;
1778}
1779
1780static int write_finished_init(struct record *rec, bool tail)
1781{
1782	if (rec->opts.tail_synthesize != tail)
1783		return 0;
1784
1785	return record__write(rec, NULL, &finished_init_event, sizeof(finished_init_event));
1786}
1787
1788static int record__synthesize(struct record *rec, bool tail);
1789
1790static int
1791record__switch_output(struct record *rec, bool at_exit)
1792{
1793	struct perf_data *data = &rec->data;
1794	int fd, err;
1795	char *new_filename;
1796
1797	/* Same Size:      "2015122520103046"*/
1798	char timestamp[] = "InvalidTimestamp";
1799
1800	record__aio_mmap_read_sync(rec);
1801
1802	write_finished_init(rec, true);
1803
1804	record__synthesize(rec, true);
1805	if (target__none(&rec->opts.target))
1806		record__synthesize_workload(rec, true);
1807
1808	rec->samples = 0;
1809	record__finish_output(rec);
1810	err = fetch_current_timestamp(timestamp, sizeof(timestamp));
1811	if (err) {
1812		pr_err("Failed to get current timestamp\n");
1813		return -EINVAL;
1814	}
1815
1816	fd = perf_data__switch(data, timestamp,
1817				    rec->session->header.data_offset,
1818				    at_exit, &new_filename);
1819	if (fd >= 0 && !at_exit) {
1820		rec->bytes_written = 0;
1821		rec->session->header.data_size = 0;
1822	}
1823
1824	if (!quiet)
1825		fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
1826			data->path, timestamp);
1827
1828	if (rec->switch_output.num_files) {
1829		int n = rec->switch_output.cur_file + 1;
1830
1831		if (n >= rec->switch_output.num_files)
1832			n = 0;
1833		rec->switch_output.cur_file = n;
1834		if (rec->switch_output.filenames[n]) {
1835			remove(rec->switch_output.filenames[n]);
1836			zfree(&rec->switch_output.filenames[n]);
1837		}
1838		rec->switch_output.filenames[n] = new_filename;
1839	} else {
1840		free(new_filename);
1841	}
1842
1843	/* Output tracking events */
1844	if (!at_exit) {
1845		record__synthesize(rec, false);
1846
1847		/*
1848		 * In 'perf record --switch-output' without -a,
1849		 * record__synthesize() in record__switch_output() won't
1850		 * generate tracking events because there's no thread_map
1851		 * in evlist. Which causes newly created perf.data doesn't
1852		 * contain map and comm information.
1853		 * Create a fake thread_map and directly call
1854		 * perf_event__synthesize_thread_map() for those events.
1855		 */
1856		if (target__none(&rec->opts.target))
1857			record__synthesize_workload(rec, false);
1858		write_finished_init(rec, false);
1859	}
1860	return fd;
1861}
1862
1863static void __record__read_lost_samples(struct record *rec, struct evsel *evsel,
1864					struct perf_record_lost_samples *lost,
1865					int cpu_idx, int thread_idx)
1866{
1867	struct perf_counts_values count;
1868	struct perf_sample_id *sid;
1869	struct perf_sample sample = {};
1870	int id_hdr_size;
1871
1872	if (perf_evsel__read(&evsel->core, cpu_idx, thread_idx, &count) < 0) {
1873		pr_err("read LOST count failed\n");
1874		return;
1875	}
1876
1877	if (count.lost == 0)
1878		return;
1879
1880	lost->lost = count.lost;
1881	if (evsel->core.ids) {
1882		sid = xyarray__entry(evsel->core.sample_id, cpu_idx, thread_idx);
1883		sample.id = sid->id;
1884	}
1885
1886	id_hdr_size = perf_event__synthesize_id_sample((void *)(lost + 1),
1887						       evsel->core.attr.sample_type, &sample);
1888	lost->header.size = sizeof(*lost) + id_hdr_size;
1889	record__write(rec, NULL, lost, lost->header.size);
1890}
1891
1892static void record__read_lost_samples(struct record *rec)
1893{
1894	struct perf_session *session = rec->session;
1895	struct perf_record_lost_samples *lost;
1896	struct evsel *evsel;
1897
1898	/* there was an error during record__open */
1899	if (session->evlist == NULL)
1900		return;
1901
1902	lost = zalloc(PERF_SAMPLE_MAX_SIZE);
1903	if (lost == NULL) {
1904		pr_debug("Memory allocation failed\n");
1905		return;
1906	}
1907
1908	lost->header.type = PERF_RECORD_LOST_SAMPLES;
1909
1910	evlist__for_each_entry(session->evlist, evsel) {
1911		struct xyarray *xy = evsel->core.sample_id;
1912
1913		if (xy == NULL || evsel->core.fd == NULL)
1914			continue;
1915		if (xyarray__max_x(evsel->core.fd) != xyarray__max_x(xy) ||
1916		    xyarray__max_y(evsel->core.fd) != xyarray__max_y(xy)) {
1917			pr_debug("Unmatched FD vs. sample ID: skip reading LOST count\n");
1918			continue;
1919		}
1920
1921		for (int x = 0; x < xyarray__max_x(xy); x++) {
1922			for (int y = 0; y < xyarray__max_y(xy); y++) {
1923				__record__read_lost_samples(rec, evsel, lost, x, y);
1924			}
1925		}
1926	}
1927	free(lost);
1928
1929}
1930
1931static volatile sig_atomic_t workload_exec_errno;
1932
1933/*
1934 * evlist__prepare_workload will send a SIGUSR1
1935 * if the fork fails, since we asked by setting its
1936 * want_signal to true.
1937 */
1938static void workload_exec_failed_signal(int signo __maybe_unused,
1939					siginfo_t *info,
1940					void *ucontext __maybe_unused)
1941{
1942	workload_exec_errno = info->si_value.sival_int;
1943	done = 1;
1944	child_finished = 1;
1945}
1946
1947static void snapshot_sig_handler(int sig);
1948static void alarm_sig_handler(int sig);
1949
1950static const struct perf_event_mmap_page *evlist__pick_pc(struct evlist *evlist)
1951{
1952	if (evlist) {
1953		if (evlist->mmap && evlist->mmap[0].core.base)
1954			return evlist->mmap[0].core.base;
1955		if (evlist->overwrite_mmap && evlist->overwrite_mmap[0].core.base)
1956			return evlist->overwrite_mmap[0].core.base;
1957	}
1958	return NULL;
1959}
1960
1961static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
1962{
1963	const struct perf_event_mmap_page *pc = evlist__pick_pc(rec->evlist);
1964	if (pc)
1965		return pc;
1966	return NULL;
1967}
1968
1969static int record__synthesize(struct record *rec, bool tail)
1970{
1971	struct perf_session *session = rec->session;
1972	struct machine *machine = &session->machines.host;
1973	struct perf_data *data = &rec->data;
1974	struct record_opts *opts = &rec->opts;
1975	struct perf_tool *tool = &rec->tool;
 
1976	int err = 0;
1977	event_op f = process_synthesized_event;
1978
1979	if (rec->opts.tail_synthesize != tail)
1980		return 0;
1981
1982	if (data->is_pipe) {
1983		err = perf_event__synthesize_for_pipe(tool, session, data,
1984						      process_synthesized_event);
1985		if (err < 0)
1986			goto out;
 
1987
1988		rec->bytes_written += err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1989	}
1990
1991	err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
1992					  process_synthesized_event, machine);
1993	if (err)
1994		goto out;
1995
1996	/* Synthesize id_index before auxtrace_info */
1997	err = perf_event__synthesize_id_index(tool,
1998					      process_synthesized_event,
1999					      session->evlist, machine);
2000	if (err)
2001		goto out;
2002
2003	if (rec->opts.full_auxtrace) {
2004		err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
2005					session, process_synthesized_event);
2006		if (err)
2007			goto out;
2008	}
2009
2010	if (!evlist__exclude_kernel(rec->evlist)) {
2011		err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
2012							 machine);
2013		WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
2014				   "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
2015				   "Check /proc/kallsyms permission or run as root.\n");
2016
2017		err = perf_event__synthesize_modules(tool, process_synthesized_event,
2018						     machine);
2019		WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
2020				   "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
2021				   "Check /proc/modules permission or run as root.\n");
2022	}
2023
2024	if (perf_guest) {
2025		machines__process_guests(&session->machines,
2026					 perf_event__synthesize_guest_os, tool);
2027	}
2028
2029	err = perf_event__synthesize_extra_attr(&rec->tool,
2030						rec->evlist,
2031						process_synthesized_event,
2032						data->is_pipe);
2033	if (err)
2034		goto out;
2035
2036	err = perf_event__synthesize_thread_map2(&rec->tool, rec->evlist->core.threads,
2037						 process_synthesized_event,
2038						NULL);
2039	if (err < 0) {
2040		pr_err("Couldn't synthesize thread map.\n");
2041		return err;
2042	}
2043
2044	err = perf_event__synthesize_cpu_map(&rec->tool, rec->evlist->core.all_cpus,
2045					     process_synthesized_event, NULL);
2046	if (err < 0) {
2047		pr_err("Couldn't synthesize cpu map.\n");
2048		return err;
2049	}
2050
2051	err = perf_event__synthesize_bpf_events(session, process_synthesized_event,
2052						machine, opts);
2053	if (err < 0) {
2054		pr_warning("Couldn't synthesize bpf events.\n");
2055		err = 0;
2056	}
2057
2058	if (rec->opts.synth & PERF_SYNTH_CGROUP) {
2059		err = perf_event__synthesize_cgroups(tool, process_synthesized_event,
2060						     machine);
2061		if (err < 0) {
2062			pr_warning("Couldn't synthesize cgroup events.\n");
2063			err = 0;
2064		}
2065	}
2066
2067	if (rec->opts.nr_threads_synthesize > 1) {
2068		mutex_init(&synth_lock);
2069		perf_set_multithreaded();
2070		f = process_locked_synthesized_event;
2071	}
2072
2073	if (rec->opts.synth & PERF_SYNTH_TASK) {
2074		bool needs_mmap = rec->opts.synth & PERF_SYNTH_MMAP;
2075
2076		err = __machine__synthesize_threads(machine, tool, &opts->target,
2077						    rec->evlist->core.threads,
2078						    f, needs_mmap, opts->sample_address,
2079						    rec->opts.nr_threads_synthesize);
2080	}
2081
2082	if (rec->opts.nr_threads_synthesize > 1) {
2083		perf_set_singlethreaded();
2084		mutex_destroy(&synth_lock);
2085	}
2086
2087out:
2088	return err;
2089}
2090
2091static int record__process_signal_event(union perf_event *event __maybe_unused, void *data)
2092{
2093	struct record *rec = data;
2094	pthread_kill(rec->thread_id, SIGUSR2);
2095	return 0;
2096}
2097
2098static int record__setup_sb_evlist(struct record *rec)
2099{
2100	struct record_opts *opts = &rec->opts;
2101
2102	if (rec->sb_evlist != NULL) {
2103		/*
2104		 * We get here if --switch-output-event populated the
2105		 * sb_evlist, so associate a callback that will send a SIGUSR2
2106		 * to the main thread.
2107		 */
2108		evlist__set_cb(rec->sb_evlist, record__process_signal_event, rec);
2109		rec->thread_id = pthread_self();
2110	}
2111#ifdef HAVE_LIBBPF_SUPPORT
2112	if (!opts->no_bpf_event) {
2113		if (rec->sb_evlist == NULL) {
2114			rec->sb_evlist = evlist__new();
2115
2116			if (rec->sb_evlist == NULL) {
2117				pr_err("Couldn't create side band evlist.\n.");
2118				return -1;
2119			}
2120		}
2121
2122		if (evlist__add_bpf_sb_event(rec->sb_evlist, &rec->session->header.env)) {
2123			pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
2124			return -1;
2125		}
2126	}
2127#endif
2128	if (evlist__start_sb_thread(rec->sb_evlist, &rec->opts.target)) {
2129		pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
2130		opts->no_bpf_event = true;
2131	}
2132
2133	return 0;
2134}
2135
2136static int record__init_clock(struct record *rec)
2137{
2138	struct perf_session *session = rec->session;
2139	struct timespec ref_clockid;
2140	struct timeval ref_tod;
2141	u64 ref;
2142
2143	if (!rec->opts.use_clockid)
2144		return 0;
2145
2146	if (rec->opts.use_clockid && rec->opts.clockid_res_ns)
2147		session->header.env.clock.clockid_res_ns = rec->opts.clockid_res_ns;
2148
2149	session->header.env.clock.clockid = rec->opts.clockid;
2150
2151	if (gettimeofday(&ref_tod, NULL) != 0) {
2152		pr_err("gettimeofday failed, cannot set reference time.\n");
2153		return -1;
2154	}
2155
2156	if (clock_gettime(rec->opts.clockid, &ref_clockid)) {
2157		pr_err("clock_gettime failed, cannot set reference time.\n");
2158		return -1;
2159	}
2160
2161	ref = (u64) ref_tod.tv_sec * NSEC_PER_SEC +
2162	      (u64) ref_tod.tv_usec * NSEC_PER_USEC;
2163
2164	session->header.env.clock.tod_ns = ref;
2165
2166	ref = (u64) ref_clockid.tv_sec * NSEC_PER_SEC +
2167	      (u64) ref_clockid.tv_nsec;
2168
2169	session->header.env.clock.clockid_ns = ref;
2170	return 0;
2171}
2172
2173static void hit_auxtrace_snapshot_trigger(struct record *rec)
2174{
2175	if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
2176		trigger_hit(&auxtrace_snapshot_trigger);
2177		auxtrace_record__snapshot_started = 1;
2178		if (auxtrace_record__snapshot_start(rec->itr))
2179			trigger_error(&auxtrace_snapshot_trigger);
2180	}
2181}
2182
2183static void record__uniquify_name(struct record *rec)
2184{
2185	struct evsel *pos;
2186	struct evlist *evlist = rec->evlist;
2187	char *new_name;
2188	int ret;
2189
2190	if (!perf_pmu__has_hybrid())
2191		return;
2192
2193	evlist__for_each_entry(evlist, pos) {
2194		if (!evsel__is_hybrid(pos))
2195			continue;
2196
2197		if (strchr(pos->name, '/'))
2198			continue;
2199
2200		ret = asprintf(&new_name, "%s/%s/",
2201			       pos->pmu_name, pos->name);
2202		if (ret) {
2203			free(pos->name);
2204			pos->name = new_name;
2205		}
2206	}
2207}
2208
2209static int record__terminate_thread(struct record_thread *thread_data)
2210{
2211	int err;
2212	enum thread_msg ack = THREAD_MSG__UNDEFINED;
2213	pid_t tid = thread_data->tid;
2214
2215	close(thread_data->pipes.msg[1]);
2216	thread_data->pipes.msg[1] = -1;
2217	err = read(thread_data->pipes.ack[0], &ack, sizeof(ack));
2218	if (err > 0)
2219		pr_debug2("threads[%d]: sent %s\n", tid, thread_msg_tags[ack]);
2220	else
2221		pr_warning("threads[%d]: failed to receive termination notification from %d\n",
2222			   thread->tid, tid);
2223
2224	return 0;
2225}
2226
2227static int record__start_threads(struct record *rec)
2228{
2229	int t, tt, err, ret = 0, nr_threads = rec->nr_threads;
2230	struct record_thread *thread_data = rec->thread_data;
2231	sigset_t full, mask;
2232	pthread_t handle;
2233	pthread_attr_t attrs;
2234
2235	thread = &thread_data[0];
2236
2237	if (!record__threads_enabled(rec))
2238		return 0;
2239
2240	sigfillset(&full);
2241	if (sigprocmask(SIG_SETMASK, &full, &mask)) {
2242		pr_err("Failed to block signals on threads start: %s\n", strerror(errno));
2243		return -1;
2244	}
2245
2246	pthread_attr_init(&attrs);
2247	pthread_attr_setdetachstate(&attrs, PTHREAD_CREATE_DETACHED);
2248
2249	for (t = 1; t < nr_threads; t++) {
2250		enum thread_msg msg = THREAD_MSG__UNDEFINED;
2251
2252#ifdef HAVE_PTHREAD_ATTR_SETAFFINITY_NP
2253		pthread_attr_setaffinity_np(&attrs,
2254					    MMAP_CPU_MASK_BYTES(&(thread_data[t].mask->affinity)),
2255					    (cpu_set_t *)(thread_data[t].mask->affinity.bits));
2256#endif
2257		if (pthread_create(&handle, &attrs, record__thread, &thread_data[t])) {
2258			for (tt = 1; tt < t; tt++)
2259				record__terminate_thread(&thread_data[t]);
2260			pr_err("Failed to start threads: %s\n", strerror(errno));
2261			ret = -1;
2262			goto out_err;
2263		}
2264
2265		err = read(thread_data[t].pipes.ack[0], &msg, sizeof(msg));
2266		if (err > 0)
2267			pr_debug2("threads[%d]: sent %s\n", rec->thread_data[t].tid,
2268				  thread_msg_tags[msg]);
2269		else
2270			pr_warning("threads[%d]: failed to receive start notification from %d\n",
2271				   thread->tid, rec->thread_data[t].tid);
2272	}
2273
2274	sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&thread->mask->affinity),
2275			(cpu_set_t *)thread->mask->affinity.bits);
2276
2277	pr_debug("threads[%d]: started on cpu%d\n", thread->tid, sched_getcpu());
2278
2279out_err:
2280	pthread_attr_destroy(&attrs);
2281
2282	if (sigprocmask(SIG_SETMASK, &mask, NULL)) {
2283		pr_err("Failed to unblock signals on threads start: %s\n", strerror(errno));
2284		ret = -1;
2285	}
2286
2287	return ret;
2288}
2289
2290static int record__stop_threads(struct record *rec)
2291{
2292	int t;
2293	struct record_thread *thread_data = rec->thread_data;
2294
2295	for (t = 1; t < rec->nr_threads; t++)
2296		record__terminate_thread(&thread_data[t]);
2297
2298	for (t = 0; t < rec->nr_threads; t++) {
2299		rec->samples += thread_data[t].samples;
2300		if (!record__threads_enabled(rec))
2301			continue;
2302		rec->session->bytes_transferred += thread_data[t].bytes_transferred;
2303		rec->session->bytes_compressed += thread_data[t].bytes_compressed;
2304		pr_debug("threads[%d]: samples=%lld, wakes=%ld, ", thread_data[t].tid,
2305			 thread_data[t].samples, thread_data[t].waking);
2306		if (thread_data[t].bytes_transferred && thread_data[t].bytes_compressed)
2307			pr_debug("transferred=%" PRIu64 ", compressed=%" PRIu64 "\n",
2308				 thread_data[t].bytes_transferred, thread_data[t].bytes_compressed);
2309		else
2310			pr_debug("written=%" PRIu64 "\n", thread_data[t].bytes_written);
2311	}
2312
2313	return 0;
2314}
2315
2316static unsigned long record__waking(struct record *rec)
2317{
2318	int t;
2319	unsigned long waking = 0;
2320	struct record_thread *thread_data = rec->thread_data;
2321
2322	for (t = 0; t < rec->nr_threads; t++)
2323		waking += thread_data[t].waking;
2324
2325	return waking;
2326}
2327
2328static int __cmd_record(struct record *rec, int argc, const char **argv)
2329{
2330	int err;
2331	int status = 0;
 
2332	const bool forks = argc > 0;
 
2333	struct perf_tool *tool = &rec->tool;
2334	struct record_opts *opts = &rec->opts;
2335	struct perf_data *data = &rec->data;
2336	struct perf_session *session;
2337	bool disabled = false, draining = false;
2338	int fd;
2339	float ratio = 0;
2340	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
2341
2342	atexit(record__sig_exit);
2343	signal(SIGCHLD, sig_handler);
2344	signal(SIGINT, sig_handler);
2345	signal(SIGTERM, sig_handler);
2346	signal(SIGSEGV, sigsegv_handler);
2347
2348	if (rec->opts.record_namespaces)
2349		tool->namespace_events = true;
2350
2351	if (rec->opts.record_cgroup) {
2352#ifdef HAVE_FILE_HANDLE
2353		tool->cgroup_events = true;
2354#else
2355		pr_err("cgroup tracking is not supported\n");
2356		return -1;
2357#endif
2358	}
2359
2360	if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
2361		signal(SIGUSR2, snapshot_sig_handler);
2362		if (rec->opts.auxtrace_snapshot_mode)
2363			trigger_on(&auxtrace_snapshot_trigger);
2364		if (rec->switch_output.enabled)
2365			trigger_on(&switch_output_trigger);
2366	} else {
2367		signal(SIGUSR2, SIG_IGN);
2368	}
2369
2370	session = perf_session__new(data, tool);
2371	if (IS_ERR(session)) {
2372		pr_err("Perf session creation failed.\n");
2373		return PTR_ERR(session);
2374	}
2375
2376	if (record__threads_enabled(rec)) {
2377		if (perf_data__is_pipe(&rec->data)) {
2378			pr_err("Parallel trace streaming is not available in pipe mode.\n");
2379			return -1;
2380		}
2381		if (rec->opts.full_auxtrace) {
2382			pr_err("Parallel trace streaming is not available in AUX area tracing mode.\n");
2383			return -1;
2384		}
2385	}
2386
2387	fd = perf_data__fd(data);
2388	rec->session = session;
2389
2390	if (zstd_init(&session->zstd_data, rec->opts.comp_level) < 0) {
2391		pr_err("Compression initialization failed.\n");
2392		return -1;
2393	}
2394#ifdef HAVE_EVENTFD_SUPPORT
2395	done_fd = eventfd(0, EFD_NONBLOCK);
2396	if (done_fd < 0) {
2397		pr_err("Failed to create wakeup eventfd, error: %m\n");
2398		status = -1;
2399		goto out_delete_session;
2400	}
2401	err = evlist__add_wakeup_eventfd(rec->evlist, done_fd);
2402	if (err < 0) {
2403		pr_err("Failed to add wakeup eventfd to poll list\n");
2404		status = err;
2405		goto out_delete_session;
2406	}
2407#endif // HAVE_EVENTFD_SUPPORT
2408
2409	session->header.env.comp_type  = PERF_COMP_ZSTD;
2410	session->header.env.comp_level = rec->opts.comp_level;
2411
2412	if (rec->opts.kcore &&
2413	    !record__kcore_readable(&session->machines.host)) {
2414		pr_err("ERROR: kcore is not readable.\n");
2415		return -1;
2416	}
2417
2418	if (record__init_clock(rec))
2419		return -1;
2420
2421	record__init_features(rec);
2422
2423	if (forks) {
2424		err = evlist__prepare_workload(rec->evlist, &opts->target, argv, data->is_pipe,
2425					       workload_exec_failed_signal);
 
2426		if (err < 0) {
2427			pr_err("Couldn't run the workload!\n");
2428			status = err;
2429			goto out_delete_session;
2430		}
2431	}
2432
2433	/*
2434	 * If we have just single event and are sending data
2435	 * through pipe, we need to force the ids allocation,
2436	 * because we synthesize event name through the pipe
2437	 * and need the id for that.
2438	 */
2439	if (data->is_pipe && rec->evlist->core.nr_entries == 1)
2440		rec->opts.sample_id = true;
2441
2442	record__uniquify_name(rec);
2443
2444	/* Debug message used by test scripts */
2445	pr_debug3("perf record opening and mmapping events\n");
2446	if (record__open(rec) != 0) {
2447		err = -1;
2448		goto out_free_threads;
2449	}
2450	/* Debug message used by test scripts */
2451	pr_debug3("perf record done opening and mmapping events\n");
2452	session->header.env.comp_mmap_len = session->evlist->core.mmap_len;
2453
2454	if (rec->opts.kcore) {
2455		err = record__kcore_copy(&session->machines.host, data);
2456		if (err) {
2457			pr_err("ERROR: Failed to copy kcore\n");
2458			goto out_free_threads;
2459		}
2460	}
2461
2462	err = bpf__apply_obj_config();
2463	if (err) {
2464		char errbuf[BUFSIZ];
2465
2466		bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2467		pr_err("ERROR: Apply config to BPF failed: %s\n",
2468			 errbuf);
2469		goto out_free_threads;
2470	}
2471
2472	/*
2473	 * Normally perf_session__new would do this, but it doesn't have the
2474	 * evlist.
2475	 */
2476	if (rec->tool.ordered_events && !evlist__sample_id_all(rec->evlist)) {
2477		pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
2478		rec->tool.ordered_events = false;
2479	}
2480
2481	if (!rec->evlist->core.nr_groups)
2482		perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
2483
2484	if (data->is_pipe) {
2485		err = perf_header__write_pipe(fd);
2486		if (err < 0)
2487			goto out_free_threads;
2488	} else {
2489		err = perf_session__write_header(session, rec->evlist, fd, false);
2490		if (err < 0)
2491			goto out_free_threads;
2492	}
2493
2494	err = -1;
2495	if (!rec->no_buildid
2496	    && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
2497		pr_err("Couldn't generate buildids. "
2498		       "Use --no-buildid to profile anyway.\n");
2499		goto out_free_threads;
 
2500	}
2501
2502	err = record__setup_sb_evlist(rec);
2503	if (err)
2504		goto out_free_threads;
2505
2506	err = record__synthesize(rec, false);
2507	if (err < 0)
2508		goto out_free_threads;
2509
2510	if (rec->realtime_prio) {
2511		struct sched_param param;
2512
2513		param.sched_priority = rec->realtime_prio;
2514		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
2515			pr_err("Could not set realtime priority.\n");
2516			err = -1;
2517			goto out_free_threads;
2518		}
2519	}
2520
2521	if (record__start_threads(rec))
2522		goto out_free_threads;
2523
2524	/*
2525	 * When perf is starting the traced process, all the events
2526	 * (apart from group members) have enable_on_exec=1 set,
2527	 * so don't spoil it by prematurely enabling them.
2528	 */
2529	if (!target__none(&opts->target) && !opts->initial_delay)
2530		evlist__enable(rec->evlist);
2531
2532	/*
2533	 * Let the child rip
2534	 */
2535	if (forks) {
2536		struct machine *machine = &session->machines.host;
2537		union perf_event *event;
2538		pid_t tgid;
2539
2540		event = malloc(sizeof(event->comm) + machine->id_hdr_size);
2541		if (event == NULL) {
2542			err = -ENOMEM;
2543			goto out_child;
2544		}
2545
2546		/*
2547		 * Some H/W events are generated before COMM event
2548		 * which is emitted during exec(), so perf script
2549		 * cannot see a correct process name for those events.
2550		 * Synthesize COMM event to prevent it.
2551		 */
2552		tgid = perf_event__synthesize_comm(tool, event,
2553						   rec->evlist->workload.pid,
2554						   process_synthesized_event,
2555						   machine);
2556		free(event);
2557
2558		if (tgid == -1)
2559			goto out_child;
2560
2561		event = malloc(sizeof(event->namespaces) +
2562			       (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
2563			       machine->id_hdr_size);
2564		if (event == NULL) {
2565			err = -ENOMEM;
2566			goto out_child;
2567		}
2568
2569		/*
2570		 * Synthesize NAMESPACES event for the command specified.
2571		 */
2572		perf_event__synthesize_namespaces(tool, event,
2573						  rec->evlist->workload.pid,
2574						  tgid, process_synthesized_event,
2575						  machine);
2576		free(event);
2577
2578		evlist__start_workload(rec->evlist);
2579	}
2580
2581	if (opts->initial_delay) {
2582		pr_info(EVLIST_DISABLED_MSG);
2583		if (opts->initial_delay > 0) {
2584			usleep(opts->initial_delay * USEC_PER_MSEC);
2585			evlist__enable(rec->evlist);
2586			pr_info(EVLIST_ENABLED_MSG);
2587		}
2588	}
2589
2590	err = event_enable_timer__start(rec->evlist->eet);
2591	if (err)
2592		goto out_child;
2593
2594	/* Debug message used by test scripts */
2595	pr_debug3("perf record has started\n");
2596	fflush(stderr);
2597
2598	trigger_ready(&auxtrace_snapshot_trigger);
2599	trigger_ready(&switch_output_trigger);
2600	perf_hooks__invoke_record_start();
2601
2602	/*
2603	 * Must write FINISHED_INIT so it will be seen after all other
2604	 * synthesized user events, but before any regular events.
2605	 */
2606	err = write_finished_init(rec, false);
2607	if (err < 0)
2608		goto out_child;
2609
2610	for (;;) {
2611		unsigned long long hits = thread->samples;
2612
2613		/*
2614		 * rec->evlist->bkw_mmap_state is possible to be
2615		 * BKW_MMAP_EMPTY here: when done == true and
2616		 * hits != rec->samples in previous round.
2617		 *
2618		 * evlist__toggle_bkw_mmap ensure we never
2619		 * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
2620		 */
2621		if (trigger_is_hit(&switch_output_trigger) || done || draining)
2622			evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
2623
2624		if (record__mmap_read_all(rec, false) < 0) {
2625			trigger_error(&auxtrace_snapshot_trigger);
2626			trigger_error(&switch_output_trigger);
2627			err = -1;
2628			goto out_child;
2629		}
2630
2631		if (auxtrace_record__snapshot_started) {
2632			auxtrace_record__snapshot_started = 0;
2633			if (!trigger_is_error(&auxtrace_snapshot_trigger))
2634				record__read_auxtrace_snapshot(rec, false);
2635			if (trigger_is_error(&auxtrace_snapshot_trigger)) {
2636				pr_err("AUX area tracing snapshot failed\n");
2637				err = -1;
2638				goto out_child;
2639			}
2640		}
2641
2642		if (trigger_is_hit(&switch_output_trigger)) {
2643			/*
2644			 * If switch_output_trigger is hit, the data in
2645			 * overwritable ring buffer should have been collected,
2646			 * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
2647			 *
2648			 * If SIGUSR2 raise after or during record__mmap_read_all(),
2649			 * record__mmap_read_all() didn't collect data from
2650			 * overwritable ring buffer. Read again.
2651			 */
2652			if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
2653				continue;
2654			trigger_ready(&switch_output_trigger);
2655
2656			/*
2657			 * Reenable events in overwrite ring buffer after
2658			 * record__mmap_read_all(): we should have collected
2659			 * data from it.
2660			 */
2661			evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
2662
2663			if (!quiet)
2664				fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
2665					record__waking(rec));
2666			thread->waking = 0;
2667			fd = record__switch_output(rec, false);
2668			if (fd < 0) {
2669				pr_err("Failed to switch to new file\n");
2670				trigger_error(&switch_output_trigger);
2671				err = fd;
2672				goto out_child;
2673			}
2674
2675			/* re-arm the alarm */
2676			if (rec->switch_output.time)
2677				alarm(rec->switch_output.time);
2678		}
2679
2680		if (hits == thread->samples) {
2681			if (done || draining)
2682				break;
2683			err = fdarray__poll(&thread->pollfd, -1);
2684			/*
2685			 * Propagate error, only if there's any. Ignore positive
2686			 * number of returned events and interrupt error.
2687			 */
2688			if (err > 0 || (err < 0 && errno == EINTR))
2689				err = 0;
2690			thread->waking++;
2691
2692			if (fdarray__filter(&thread->pollfd, POLLERR | POLLHUP,
2693					    record__thread_munmap_filtered, NULL) == 0)
2694				draining = true;
2695
2696			err = record__update_evlist_pollfd_from_thread(rec, rec->evlist, thread);
2697			if (err)
2698				goto out_child;
2699		}
2700
2701		if (evlist__ctlfd_process(rec->evlist, &cmd) > 0) {
2702			switch (cmd) {
2703			case EVLIST_CTL_CMD_SNAPSHOT:
2704				hit_auxtrace_snapshot_trigger(rec);
2705				evlist__ctlfd_ack(rec->evlist);
2706				break;
2707			case EVLIST_CTL_CMD_STOP:
2708				done = 1;
2709				break;
2710			case EVLIST_CTL_CMD_ACK:
2711			case EVLIST_CTL_CMD_UNSUPPORTED:
2712			case EVLIST_CTL_CMD_ENABLE:
2713			case EVLIST_CTL_CMD_DISABLE:
2714			case EVLIST_CTL_CMD_EVLIST:
2715			case EVLIST_CTL_CMD_PING:
2716			default:
2717				break;
2718			}
2719		}
2720
2721		err = event_enable_timer__process(rec->evlist->eet);
2722		if (err < 0)
2723			goto out_child;
2724		if (err) {
2725			err = 0;
2726			done = 1;
2727		}
2728
2729		/*
2730		 * When perf is starting the traced process, at the end events
2731		 * die with the process and we wait for that. Thus no need to
2732		 * disable events in this case.
2733		 */
2734		if (done && !disabled && !target__none(&opts->target)) {
2735			trigger_off(&auxtrace_snapshot_trigger);
2736			evlist__disable(rec->evlist);
2737			disabled = true;
2738		}
2739	}
2740
2741	trigger_off(&auxtrace_snapshot_trigger);
2742	trigger_off(&switch_output_trigger);
2743
2744	if (opts->auxtrace_snapshot_on_exit)
2745		record__auxtrace_snapshot_exit(rec);
2746
2747	if (forks && workload_exec_errno) {
2748		char msg[STRERR_BUFSIZE], strevsels[2048];
2749		const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
2750
2751		evlist__scnprintf_evsels(rec->evlist, sizeof(strevsels), strevsels);
2752
2753		pr_err("Failed to collect '%s' for the '%s' workload: %s\n",
2754			strevsels, argv[0], emsg);
2755		err = -1;
2756		goto out_child;
2757	}
2758
2759	if (!quiet)
2760		fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n",
2761			record__waking(rec));
2762
2763	write_finished_init(rec, true);
2764
2765	if (target__none(&rec->opts.target))
2766		record__synthesize_workload(rec, true);
2767
2768out_child:
2769	record__stop_threads(rec);
2770	record__mmap_read_all(rec, true);
2771out_free_threads:
2772	record__free_thread_data(rec);
2773	evlist__finalize_ctlfd(rec->evlist);
2774	record__aio_mmap_read_sync(rec);
2775
2776	if (rec->session->bytes_transferred && rec->session->bytes_compressed) {
2777		ratio = (float)rec->session->bytes_transferred/(float)rec->session->bytes_compressed;
2778		session->header.env.comp_ratio = ratio + 0.5;
2779	}
2780
2781	if (forks) {
2782		int exit_status;
2783
2784		if (!child_finished)
2785			kill(rec->evlist->workload.pid, SIGTERM);
2786
2787		wait(&exit_status);
2788
2789		if (err < 0)
2790			status = err;
2791		else if (WIFEXITED(exit_status))
2792			status = WEXITSTATUS(exit_status);
2793		else if (WIFSIGNALED(exit_status))
2794			signr = WTERMSIG(exit_status);
2795	} else
2796		status = err;
2797
2798	if (rec->off_cpu)
2799		rec->bytes_written += off_cpu_write(rec->session);
2800
2801	record__read_lost_samples(rec);
2802	record__synthesize(rec, true);
2803	/* this will be recalculated during process_buildids() */
2804	rec->samples = 0;
2805
2806	if (!err) {
2807		if (!rec->timestamp_filename) {
2808			record__finish_output(rec);
2809		} else {
2810			fd = record__switch_output(rec, true);
2811			if (fd < 0) {
2812				status = fd;
2813				goto out_delete_session;
2814			}
2815		}
2816	}
2817
2818	perf_hooks__invoke_record_end();
2819
2820	if (!err && !quiet) {
2821		char samples[128];
2822		const char *postfix = rec->timestamp_filename ?
2823					".<timestamp>" : "";
2824
2825		if (rec->samples && !rec->opts.full_auxtrace)
2826			scnprintf(samples, sizeof(samples),
2827				  " (%" PRIu64 " samples)", rec->samples);
2828		else
2829			samples[0] = '\0';
2830
2831		fprintf(stderr,	"[ perf record: Captured and wrote %.3f MB %s%s%s",
2832			perf_data__size(data) / 1024.0 / 1024.0,
2833			data->path, postfix, samples);
2834		if (ratio) {
2835			fprintf(stderr,	", compressed (original %.3f MB, ratio is %.3f)",
2836					rec->session->bytes_transferred / 1024.0 / 1024.0,
2837					ratio);
2838		}
2839		fprintf(stderr, " ]\n");
2840	}
2841
2842out_delete_session:
2843#ifdef HAVE_EVENTFD_SUPPORT
2844	if (done_fd >= 0) {
2845		fd = done_fd;
2846		done_fd = -1;
2847
2848		close(fd);
2849	}
2850#endif
2851	zstd_fini(&session->zstd_data);
2852	perf_session__delete(session);
2853
2854	if (!opts->no_bpf_event)
2855		evlist__stop_sb_thread(rec->sb_evlist);
2856	return status;
2857}
2858
2859static void callchain_debug(struct callchain_param *callchain)
2860{
2861	static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
2862
2863	pr_debug("callchain: type %s\n", str[callchain->record_mode]);
2864
2865	if (callchain->record_mode == CALLCHAIN_DWARF)
2866		pr_debug("callchain: stack dump size %d\n",
2867			 callchain->dump_size);
2868}
2869
2870int record_opts__parse_callchain(struct record_opts *record,
2871				 struct callchain_param *callchain,
2872				 const char *arg, bool unset)
2873{
2874	int ret;
2875	callchain->enabled = !unset;
 
 
 
2876
2877	/* --no-call-graph */
2878	if (unset) {
2879		callchain->record_mode = CALLCHAIN_NONE;
2880		pr_debug("callchain: disabled\n");
2881		return 0;
2882	}
2883
2884	ret = parse_callchain_record_opt(arg, callchain);
2885	if (!ret) {
2886		/* Enable data address sampling for DWARF unwind. */
2887		if (callchain->record_mode == CALLCHAIN_DWARF)
2888			record->sample_address = true;
2889		callchain_debug(callchain);
2890	}
2891
2892	return ret;
2893}
2894
2895int record_parse_callchain_opt(const struct option *opt,
2896			       const char *arg,
2897			       int unset)
2898{
2899	return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
2900}
2901
2902int record_callchain_opt(const struct option *opt,
2903			 const char *arg __maybe_unused,
2904			 int unset __maybe_unused)
2905{
2906	struct callchain_param *callchain = opt->value;
2907
2908	callchain->enabled = true;
 
2909
2910	if (callchain->record_mode == CALLCHAIN_NONE)
2911		callchain->record_mode = CALLCHAIN_FP;
2912
2913	callchain_debug(callchain);
2914	return 0;
2915}
2916
2917static int perf_record_config(const char *var, const char *value, void *cb)
2918{
2919	struct record *rec = cb;
2920
2921	if (!strcmp(var, "record.build-id")) {
2922		if (!strcmp(value, "cache"))
2923			rec->no_buildid_cache = false;
2924		else if (!strcmp(value, "no-cache"))
2925			rec->no_buildid_cache = true;
2926		else if (!strcmp(value, "skip"))
2927			rec->no_buildid = true;
2928		else if (!strcmp(value, "mmap"))
2929			rec->buildid_mmap = true;
2930		else
2931			return -1;
2932		return 0;
2933	}
2934	if (!strcmp(var, "record.call-graph")) {
2935		var = "call-graph.record-mode";
2936		return perf_default_config(var, value, cb);
2937	}
2938#ifdef HAVE_AIO_SUPPORT
2939	if (!strcmp(var, "record.aio")) {
2940		rec->opts.nr_cblocks = strtol(value, NULL, 0);
2941		if (!rec->opts.nr_cblocks)
2942			rec->opts.nr_cblocks = nr_cblocks_default;
2943	}
2944#endif
2945	if (!strcmp(var, "record.debuginfod")) {
2946		rec->debuginfod.urls = strdup(value);
2947		if (!rec->debuginfod.urls)
2948			return -ENOMEM;
2949		rec->debuginfod.set = true;
2950	}
2951
2952	return 0;
2953}
2954
2955static int record__parse_event_enable_time(const struct option *opt, const char *str, int unset)
2956{
2957	struct record *rec = (struct record *)opt->value;
 
2958
2959	return evlist__parse_event_enable_time(rec->evlist, &rec->opts, str, unset);
2960}
2961
2962static int record__parse_affinity(const struct option *opt, const char *str, int unset)
2963{
2964	struct record_opts *opts = (struct record_opts *)opt->value;
2965
2966	if (unset || !str)
2967		return 0;
2968
2969	if (!strcasecmp(str, "node"))
2970		opts->affinity = PERF_AFFINITY_NODE;
2971	else if (!strcasecmp(str, "cpu"))
2972		opts->affinity = PERF_AFFINITY_CPU;
 
 
 
 
 
 
 
 
2973
2974	return 0;
2975}
 
 
 
 
 
 
 
 
 
 
 
 
 
2976
2977static int record__mmap_cpu_mask_alloc(struct mmap_cpu_mask *mask, int nr_bits)
2978{
2979	mask->nbits = nr_bits;
2980	mask->bits = bitmap_zalloc(mask->nbits);
2981	if (!mask->bits)
2982		return -ENOMEM;
2983
2984	return 0;
2985}
2986
2987static void record__mmap_cpu_mask_free(struct mmap_cpu_mask *mask)
2988{
2989	bitmap_free(mask->bits);
2990	mask->nbits = 0;
2991}
2992
2993static int record__thread_mask_alloc(struct thread_mask *mask, int nr_bits)
2994{
2995	int ret;
 
2996
2997	ret = record__mmap_cpu_mask_alloc(&mask->maps, nr_bits);
2998	if (ret) {
2999		mask->affinity.bits = NULL;
3000		return ret;
3001	}
3002
3003	ret = record__mmap_cpu_mask_alloc(&mask->affinity, nr_bits);
3004	if (ret) {
3005		record__mmap_cpu_mask_free(&mask->maps);
3006		mask->maps.bits = NULL;
3007	}
3008
3009	return ret;
3010}
3011
3012static void record__thread_mask_free(struct thread_mask *mask)
3013{
3014	record__mmap_cpu_mask_free(&mask->maps);
3015	record__mmap_cpu_mask_free(&mask->affinity);
3016}
3017
3018static int record__parse_threads(const struct option *opt, const char *str, int unset)
3019{
3020	int s;
3021	struct record_opts *opts = opt->value;
3022
3023	if (unset || !str || !strlen(str)) {
3024		opts->threads_spec = THREAD_SPEC__CPU;
3025	} else {
3026		for (s = 1; s < THREAD_SPEC__MAX; s++) {
3027			if (s == THREAD_SPEC__USER) {
3028				opts->threads_user_spec = strdup(str);
3029				if (!opts->threads_user_spec)
3030					return -ENOMEM;
3031				opts->threads_spec = THREAD_SPEC__USER;
3032				break;
3033			}
3034			if (!strncasecmp(str, thread_spec_tags[s], strlen(thread_spec_tags[s]))) {
3035				opts->threads_spec = s;
3036				break;
3037			}
3038		}
3039	}
3040
3041	if (opts->threads_spec == THREAD_SPEC__USER)
3042		pr_debug("threads_spec: %s\n", opts->threads_user_spec);
3043	else
3044		pr_debug("threads_spec: %s\n", thread_spec_tags[opts->threads_spec]);
3045
3046	return 0;
3047}
3048
3049static int parse_output_max_size(const struct option *opt,
3050				 const char *str, int unset)
3051{
3052	unsigned long *s = (unsigned long *)opt->value;
3053	static struct parse_tag tags_size[] = {
3054		{ .tag  = 'B', .mult = 1       },
3055		{ .tag  = 'K', .mult = 1 << 10 },
3056		{ .tag  = 'M', .mult = 1 << 20 },
3057		{ .tag  = 'G', .mult = 1 << 30 },
3058		{ .tag  = 0 },
3059	};
3060	unsigned long val;
3061
3062	if (unset) {
3063		*s = 0;
3064		return 0;
3065	}
3066
3067	val = parse_tag_value(str, tags_size);
3068	if (val != (unsigned long) -1) {
3069		*s = val;
3070		return 0;
3071	}
3072
3073	return -1;
3074}
3075
3076static int record__parse_mmap_pages(const struct option *opt,
3077				    const char *str,
3078				    int unset __maybe_unused)
3079{
3080	struct record_opts *opts = opt->value;
3081	char *s, *p;
3082	unsigned int mmap_pages;
3083	int ret;
3084
3085	if (!str)
3086		return -EINVAL;
3087
3088	s = strdup(str);
3089	if (!s)
3090		return -ENOMEM;
3091
3092	p = strchr(s, ',');
3093	if (p)
3094		*p = '\0';
3095
3096	if (*s) {
3097		ret = __evlist__parse_mmap_pages(&mmap_pages, s);
3098		if (ret)
3099			goto out_free;
3100		opts->mmap_pages = mmap_pages;
3101	}
3102
3103	if (!p) {
3104		ret = 0;
3105		goto out_free;
3106	}
3107
3108	ret = __evlist__parse_mmap_pages(&mmap_pages, p + 1);
3109	if (ret)
3110		goto out_free;
3111
3112	opts->auxtrace_mmap_pages = mmap_pages;
3113
3114out_free:
3115	free(s);
3116	return ret;
3117}
3118
3119void __weak arch__add_leaf_frame_record_opts(struct record_opts *opts __maybe_unused)
3120{
3121}
3122
3123static int parse_control_option(const struct option *opt,
3124				const char *str,
3125				int unset __maybe_unused)
3126{
3127	struct record_opts *opts = opt->value;
3128
3129	return evlist__parse_control(str, &opts->ctl_fd, &opts->ctl_fd_ack, &opts->ctl_fd_close);
3130}
3131
3132static void switch_output_size_warn(struct record *rec)
3133{
3134	u64 wakeup_size = evlist__mmap_size(rec->opts.mmap_pages);
3135	struct switch_output *s = &rec->switch_output;
3136
3137	wakeup_size /= 2;
3138
3139	if (s->size < wakeup_size) {
3140		char buf[100];
3141
3142		unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
3143		pr_warning("WARNING: switch-output data size lower than "
3144			   "wakeup kernel buffer size (%s) "
3145			   "expect bigger perf.data sizes\n", buf);
3146	}
3147}
3148
3149static int switch_output_setup(struct record *rec)
3150{
3151	struct switch_output *s = &rec->switch_output;
3152	static struct parse_tag tags_size[] = {
3153		{ .tag  = 'B', .mult = 1       },
3154		{ .tag  = 'K', .mult = 1 << 10 },
3155		{ .tag  = 'M', .mult = 1 << 20 },
3156		{ .tag  = 'G', .mult = 1 << 30 },
3157		{ .tag  = 0 },
3158	};
3159	static struct parse_tag tags_time[] = {
3160		{ .tag  = 's', .mult = 1        },
3161		{ .tag  = 'm', .mult = 60       },
3162		{ .tag  = 'h', .mult = 60*60    },
3163		{ .tag  = 'd', .mult = 60*60*24 },
3164		{ .tag  = 0 },
3165	};
3166	unsigned long val;
3167
3168	/*
3169	 * If we're using --switch-output-events, then we imply its 
3170	 * --switch-output=signal, as we'll send a SIGUSR2 from the side band
3171	 *  thread to its parent.
3172	 */
3173	if (rec->switch_output_event_set) {
3174		if (record__threads_enabled(rec)) {
3175			pr_warning("WARNING: --switch-output-event option is not available in parallel streaming mode.\n");
3176			return 0;
3177		}
3178		goto do_signal;
3179	}
3180
3181	if (!s->set)
3182		return 0;
3183
3184	if (record__threads_enabled(rec)) {
3185		pr_warning("WARNING: --switch-output option is not available in parallel streaming mode.\n");
3186		return 0;
3187	}
3188
3189	if (!strcmp(s->str, "signal")) {
3190do_signal:
3191		s->signal = true;
3192		pr_debug("switch-output with SIGUSR2 signal\n");
3193		goto enabled;
3194	}
3195
3196	val = parse_tag_value(s->str, tags_size);
3197	if (val != (unsigned long) -1) {
3198		s->size = val;
3199		pr_debug("switch-output with %s size threshold\n", s->str);
3200		goto enabled;
3201	}
3202
3203	val = parse_tag_value(s->str, tags_time);
3204	if (val != (unsigned long) -1) {
3205		s->time = val;
3206		pr_debug("switch-output with %s time threshold (%lu seconds)\n",
3207			 s->str, s->time);
3208		goto enabled;
3209	}
3210
3211	return -1;
3212
3213enabled:
3214	rec->timestamp_filename = true;
3215	s->enabled              = true;
3216
3217	if (s->size && !rec->opts.no_buffering)
3218		switch_output_size_warn(rec);
3219
3220	return 0;
3221}
3222
3223static const char * const __record_usage[] = {
3224	"perf record [<options>] [<command>]",
3225	"perf record [<options>] -- <command> [<options>]",
3226	NULL
3227};
3228const char * const *record_usage = __record_usage;
3229
3230static int build_id__process_mmap(struct perf_tool *tool, union perf_event *event,
3231				  struct perf_sample *sample, struct machine *machine)
3232{
3233	/*
3234	 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
3235	 * no need to add them twice.
3236	 */
3237	if (!(event->header.misc & PERF_RECORD_MISC_USER))
3238		return 0;
3239	return perf_event__process_mmap(tool, event, sample, machine);
3240}
3241
3242static int build_id__process_mmap2(struct perf_tool *tool, union perf_event *event,
3243				   struct perf_sample *sample, struct machine *machine)
3244{
3245	/*
3246	 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
3247	 * no need to add them twice.
3248	 */
3249	if (!(event->header.misc & PERF_RECORD_MISC_USER))
3250		return 0;
3251
3252	return perf_event__process_mmap2(tool, event, sample, machine);
3253}
3254
3255static int process_timestamp_boundary(struct perf_tool *tool,
3256				      union perf_event *event __maybe_unused,
3257				      struct perf_sample *sample,
3258				      struct machine *machine __maybe_unused)
3259{
3260	struct record *rec = container_of(tool, struct record, tool);
3261
3262	set_timestamp_boundary(rec, sample->time);
3263	return 0;
3264}
3265
3266static int parse_record_synth_option(const struct option *opt,
3267				     const char *str,
3268				     int unset __maybe_unused)
3269{
3270	struct record_opts *opts = opt->value;
3271	char *p = strdup(str);
3272
3273	if (p == NULL)
3274		return -1;
3275
3276	opts->synth = parse_synth_opt(p);
3277	free(p);
3278
3279	if (opts->synth < 0) {
3280		pr_err("Invalid synth option: %s\n", str);
3281		return -1;
3282	}
3283	return 0;
3284}
3285
3286/*
3287 * XXX Ideally would be local to cmd_record() and passed to a record__new
3288 * because we need to have access to it in record__exit, that is called
3289 * after cmd_record() exits, but since record_options need to be accessible to
3290 * builtin-script, leave it here.
3291 *
3292 * At least we don't ouch it in all the other functions here directly.
3293 *
3294 * Just say no to tons of global variables, sigh.
3295 */
3296static struct record record = {
3297	.opts = {
3298		.sample_time	     = true,
3299		.mmap_pages	     = UINT_MAX,
3300		.user_freq	     = UINT_MAX,
3301		.user_interval	     = ULLONG_MAX,
3302		.freq		     = 4000,
3303		.target		     = {
3304			.uses_mmap   = true,
3305			.default_per_cpu = true,
3306		},
3307		.mmap_flush          = MMAP_FLUSH_DEFAULT,
3308		.nr_threads_synthesize = 1,
3309		.ctl_fd              = -1,
3310		.ctl_fd_ack          = -1,
3311		.synth               = PERF_SYNTH_ALL,
3312	},
3313	.tool = {
3314		.sample		= process_sample_event,
3315		.fork		= perf_event__process_fork,
3316		.exit		= perf_event__process_exit,
3317		.comm		= perf_event__process_comm,
3318		.namespaces	= perf_event__process_namespaces,
3319		.mmap		= build_id__process_mmap,
3320		.mmap2		= build_id__process_mmap2,
3321		.itrace_start	= process_timestamp_boundary,
3322		.aux		= process_timestamp_boundary,
3323		.ordered_events	= true,
3324	},
3325};
3326
3327const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
3328	"\n\t\t\t\tDefault: fp";
3329
3330static bool dry_run;
3331
3332/*
3333 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
3334 * with it and switch to use the library functions in perf_evlist that came
3335 * from builtin-record.c, i.e. use record_opts,
3336 * evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
3337 * using pipes, etc.
3338 */
3339static struct option __record_options[] = {
3340	OPT_CALLBACK('e', "event", &record.evlist, "event",
3341		     "event selector. use 'perf list' to list available events",
3342		     parse_events_option),
3343	OPT_CALLBACK(0, "filter", &record.evlist, "filter",
3344		     "event filter", parse_filter),
3345	OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
3346			   NULL, "don't record events from perf itself",
3347			   exclude_perf),
3348	OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
3349		    "record events on existing process id"),
3350	OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
3351		    "record events on existing thread id"),
3352	OPT_INTEGER('r', "realtime", &record.realtime_prio,
3353		    "collect data with this RT SCHED_FIFO priority"),
3354	OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
3355		    "collect data without buffering"),
3356	OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
3357		    "collect raw sample records from all opened counters"),
3358	OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
3359			    "system-wide collection from all CPUs"),
3360	OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
3361		    "list of cpus to monitor"),
3362	OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
3363	OPT_STRING('o', "output", &record.data.path, "file",
3364		    "output file name"),
3365	OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
3366			&record.opts.no_inherit_set,
3367			"child tasks do not inherit counters"),
3368	OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
3369		    "synthesize non-sample events at the end of output"),
3370	OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
3371	OPT_BOOLEAN(0, "no-bpf-event", &record.opts.no_bpf_event, "do not record bpf events"),
3372	OPT_BOOLEAN(0, "strict-freq", &record.opts.strict_freq,
3373		    "Fail if the specified frequency can't be used"),
3374	OPT_CALLBACK('F', "freq", &record.opts, "freq or 'max'",
3375		     "profile at this frequency",
3376		      record__parse_freq),
3377	OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
3378		     "number of mmap data pages and AUX area tracing mmap pages",
3379		     record__parse_mmap_pages),
3380	OPT_CALLBACK(0, "mmap-flush", &record.opts, "number",
3381		     "Minimal number of bytes that is extracted from mmap data pages (default: 1)",
3382		     record__mmap_flush_parse),
3383	OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
3384			   NULL, "enables call-graph recording" ,
3385			   &record_callchain_opt),
3386	OPT_CALLBACK(0, "call-graph", &record.opts,
3387		     "record_mode[,record_size]", record_callchain_help,
3388		     &record_parse_callchain_opt),
3389	OPT_INCR('v', "verbose", &verbose,
3390		    "be more verbose (show counter open errors, etc)"),
3391	OPT_BOOLEAN('q', "quiet", &quiet, "don't print any warnings or messages"),
3392	OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
3393		    "per thread counts"),
3394	OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
3395	OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
3396		    "Record the sample physical addresses"),
3397	OPT_BOOLEAN(0, "data-page-size", &record.opts.sample_data_page_size,
3398		    "Record the sampled data address data page size"),
3399	OPT_BOOLEAN(0, "code-page-size", &record.opts.sample_code_page_size,
3400		    "Record the sampled code address (ip) page size"),
3401	OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
3402	OPT_BOOLEAN(0, "sample-identifier", &record.opts.sample_identifier,
3403		    "Record the sample identifier"),
3404	OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
3405			&record.opts.sample_time_set,
3406			"Record the sample timestamps"),
3407	OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
3408			"Record the sample period"),
3409	OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
3410		    "don't sample"),
3411	OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
3412			&record.no_buildid_cache_set,
3413			"do not update the buildid cache"),
3414	OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
3415			&record.no_buildid_set,
3416			"do not collect buildids in perf.data"),
3417	OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
3418		     "monitor event in cgroup name only",
3419		     parse_cgroups),
3420	OPT_CALLBACK('D', "delay", &record, "ms",
3421		     "ms to wait before starting measurement after program start (-1: start with events disabled), "
3422		     "or ranges of time to enable events e.g. '-D 10-20,30-40'",
3423		     record__parse_event_enable_time),
3424	OPT_BOOLEAN(0, "kcore", &record.opts.kcore, "copy /proc/kcore"),
3425	OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
3426		   "user to profile"),
3427
3428	OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
3429		     "branch any", "sample any taken branches",
3430		     parse_branch_stack),
3431
3432	OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
3433		     "branch filter mask", "branch stack filter modes",
3434		     parse_branch_stack),
3435	OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
3436		    "sample by weight (on special events only)"),
3437	OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
3438		    "sample transaction flags (special events only)"),
3439	OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
3440		    "use per-thread mmaps"),
3441	OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
3442		    "sample selected machine registers on interrupt,"
3443		    " use '-I?' to list register names", parse_intr_regs),
3444	OPT_CALLBACK_OPTARG(0, "user-regs", &record.opts.sample_user_regs, NULL, "any register",
3445		    "sample selected machine registers on interrupt,"
3446		    " use '--user-regs=?' to list register names", parse_user_regs),
3447	OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
3448		    "Record running/enabled time of read (:S) events"),
3449	OPT_CALLBACK('k', "clockid", &record.opts,
3450	"clockid", "clockid to use for events, see clock_gettime()",
3451	parse_clockid),
3452	OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
3453			  "opts", "AUX area tracing Snapshot Mode", ""),
3454	OPT_STRING_OPTARG(0, "aux-sample", &record.opts.auxtrace_sample_opts,
3455			  "opts", "sample AUX area", ""),
3456	OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3457			"per thread proc mmap processing timeout in ms"),
3458	OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
3459		    "Record namespaces events"),
3460	OPT_BOOLEAN(0, "all-cgroups", &record.opts.record_cgroup,
3461		    "Record cgroup events"),
3462	OPT_BOOLEAN_SET(0, "switch-events", &record.opts.record_switch_events,
3463			&record.opts.record_switch_events_set,
3464			"Record context switch events"),
3465	OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
3466			 "Configure all used events to run in kernel space.",
3467			 PARSE_OPT_EXCLUSIVE),
3468	OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
3469			 "Configure all used events to run in user space.",
3470			 PARSE_OPT_EXCLUSIVE),
3471	OPT_BOOLEAN(0, "kernel-callchains", &record.opts.kernel_callchains,
3472		    "collect kernel callchains"),
3473	OPT_BOOLEAN(0, "user-callchains", &record.opts.user_callchains,
3474		    "collect user callchains"),
3475	OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
3476		   "clang binary to use for compiling BPF scriptlets"),
3477	OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
3478		   "options passed to clang when compiling BPF scriptlets"),
3479	OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
3480		   "file", "vmlinux pathname"),
3481	OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
3482		    "Record build-id of all DSOs regardless of hits"),
3483	OPT_BOOLEAN(0, "buildid-mmap", &record.buildid_mmap,
3484		    "Record build-id in map events"),
3485	OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
3486		    "append timestamp to output filename"),
3487	OPT_BOOLEAN(0, "timestamp-boundary", &record.timestamp_boundary,
3488		    "Record timestamp boundary (time of first/last samples)"),
3489	OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
3490			  &record.switch_output.set, "signal or size[BKMG] or time[smhd]",
3491			  "Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold",
3492			  "signal"),
3493	OPT_CALLBACK_SET(0, "switch-output-event", &record.sb_evlist, &record.switch_output_event_set, "switch output event",
3494			 "switch output event selector. use 'perf list' to list available events",
3495			 parse_events_option_new_evlist),
3496	OPT_INTEGER(0, "switch-max-files", &record.switch_output.num_files,
3497		   "Limit number of switch output generated files"),
3498	OPT_BOOLEAN(0, "dry-run", &dry_run,
3499		    "Parse options then exit"),
3500#ifdef HAVE_AIO_SUPPORT
3501	OPT_CALLBACK_OPTARG(0, "aio", &record.opts,
3502		     &nr_cblocks_default, "n", "Use <n> control blocks in asynchronous trace writing mode (default: 1, max: 4)",
3503		     record__aio_parse),
3504#endif
3505	OPT_CALLBACK(0, "affinity", &record.opts, "node|cpu",
3506		     "Set affinity mask of trace reading thread to NUMA node cpu mask or cpu of processed mmap buffer",
3507		     record__parse_affinity),
3508#ifdef HAVE_ZSTD_SUPPORT
3509	OPT_CALLBACK_OPTARG('z', "compression-level", &record.opts, &comp_level_default, "n",
3510			    "Compress records using specified level (default: 1 - fastest compression, 22 - greatest compression)",
3511			    record__parse_comp_level),
3512#endif
3513	OPT_CALLBACK(0, "max-size", &record.output_max_size,
3514		     "size", "Limit the maximum size of the output file", parse_output_max_size),
3515	OPT_UINTEGER(0, "num-thread-synthesize",
3516		     &record.opts.nr_threads_synthesize,
3517		     "number of threads to run for event synthesis"),
3518#ifdef HAVE_LIBPFM
3519	OPT_CALLBACK(0, "pfm-events", &record.evlist, "event",
3520		"libpfm4 event selector. use 'perf list' to list available events",
3521		parse_libpfm_events_option),
3522#endif
3523	OPT_CALLBACK(0, "control", &record.opts, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
3524		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events,\n"
3525		     "\t\t\t  'snapshot': AUX area tracing snapshot).\n"
3526		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
3527		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
3528		      parse_control_option),
3529	OPT_CALLBACK(0, "synth", &record.opts, "no|all|task|mmap|cgroup",
3530		     "Fine-tune event synthesis: default=all", parse_record_synth_option),
3531	OPT_STRING_OPTARG_SET(0, "debuginfod", &record.debuginfod.urls,
3532			  &record.debuginfod.set, "debuginfod urls",
3533			  "Enable debuginfod data retrieval from DEBUGINFOD_URLS or specified urls",
3534			  "system"),
3535	OPT_CALLBACK_OPTARG(0, "threads", &record.opts, NULL, "spec",
3536			    "write collected trace data into several data files using parallel threads",
3537			    record__parse_threads),
3538	OPT_BOOLEAN(0, "off-cpu", &record.off_cpu, "Enable off-cpu analysis"),
3539	OPT_END()
3540};
3541
3542struct option *record_options = __record_options;
3543
3544static int record__mmap_cpu_mask_init(struct mmap_cpu_mask *mask, struct perf_cpu_map *cpus)
3545{
3546	struct perf_cpu cpu;
3547	int idx;
3548
3549	if (cpu_map__is_dummy(cpus))
3550		return 0;
3551
3552	perf_cpu_map__for_each_cpu(cpu, idx, cpus) {
3553		if (cpu.cpu == -1)
3554			continue;
3555		/* Return ENODEV is input cpu is greater than max cpu */
3556		if ((unsigned long)cpu.cpu > mask->nbits)
3557			return -ENODEV;
3558		__set_bit(cpu.cpu, mask->bits);
3559	}
3560
3561	return 0;
3562}
3563
3564static int record__mmap_cpu_mask_init_spec(struct mmap_cpu_mask *mask, const char *mask_spec)
3565{
3566	struct perf_cpu_map *cpus;
3567
3568	cpus = perf_cpu_map__new(mask_spec);
3569	if (!cpus)
3570		return -ENOMEM;
3571
3572	bitmap_zero(mask->bits, mask->nbits);
3573	if (record__mmap_cpu_mask_init(mask, cpus))
3574		return -ENODEV;
3575
3576	perf_cpu_map__put(cpus);
3577
3578	return 0;
3579}
3580
3581static void record__free_thread_masks(struct record *rec, int nr_threads)
3582{
3583	int t;
3584
3585	if (rec->thread_masks)
3586		for (t = 0; t < nr_threads; t++)
3587			record__thread_mask_free(&rec->thread_masks[t]);
3588
3589	zfree(&rec->thread_masks);
3590}
3591
3592static int record__alloc_thread_masks(struct record *rec, int nr_threads, int nr_bits)
3593{
3594	int t, ret;
3595
3596	rec->thread_masks = zalloc(nr_threads * sizeof(*(rec->thread_masks)));
3597	if (!rec->thread_masks) {
3598		pr_err("Failed to allocate thread masks\n");
3599		return -ENOMEM;
3600	}
3601
3602	for (t = 0; t < nr_threads; t++) {
3603		ret = record__thread_mask_alloc(&rec->thread_masks[t], nr_bits);
3604		if (ret) {
3605			pr_err("Failed to allocate thread masks[%d]\n", t);
3606			goto out_free;
3607		}
3608	}
3609
3610	return 0;
3611
3612out_free:
3613	record__free_thread_masks(rec, nr_threads);
3614
3615	return ret;
3616}
3617
3618static int record__init_thread_cpu_masks(struct record *rec, struct perf_cpu_map *cpus)
3619{
3620	int t, ret, nr_cpus = perf_cpu_map__nr(cpus);
3621
3622	ret = record__alloc_thread_masks(rec, nr_cpus, cpu__max_cpu().cpu);
3623	if (ret)
3624		return ret;
3625
3626	rec->nr_threads = nr_cpus;
3627	pr_debug("nr_threads: %d\n", rec->nr_threads);
3628
3629	for (t = 0; t < rec->nr_threads; t++) {
3630		__set_bit(perf_cpu_map__cpu(cpus, t).cpu, rec->thread_masks[t].maps.bits);
3631		__set_bit(perf_cpu_map__cpu(cpus, t).cpu, rec->thread_masks[t].affinity.bits);
3632		if (verbose > 0) {
3633			pr_debug("thread_masks[%d]: ", t);
3634			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].maps, "maps");
3635			pr_debug("thread_masks[%d]: ", t);
3636			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].affinity, "affinity");
3637		}
3638	}
3639
3640	return 0;
3641}
3642
3643static int record__init_thread_masks_spec(struct record *rec, struct perf_cpu_map *cpus,
3644					  const char **maps_spec, const char **affinity_spec,
3645					  u32 nr_spec)
3646{
3647	u32 s;
3648	int ret = 0, t = 0;
3649	struct mmap_cpu_mask cpus_mask;
3650	struct thread_mask thread_mask, full_mask, *thread_masks;
3651
3652	ret = record__mmap_cpu_mask_alloc(&cpus_mask, cpu__max_cpu().cpu);
3653	if (ret) {
3654		pr_err("Failed to allocate CPUs mask\n");
3655		return ret;
3656	}
3657
3658	ret = record__mmap_cpu_mask_init(&cpus_mask, cpus);
3659	if (ret) {
3660		pr_err("Failed to init cpu mask\n");
3661		goto out_free_cpu_mask;
3662	}
3663
3664	ret = record__thread_mask_alloc(&full_mask, cpu__max_cpu().cpu);
3665	if (ret) {
3666		pr_err("Failed to allocate full mask\n");
3667		goto out_free_cpu_mask;
3668	}
3669
3670	ret = record__thread_mask_alloc(&thread_mask, cpu__max_cpu().cpu);
3671	if (ret) {
3672		pr_err("Failed to allocate thread mask\n");
3673		goto out_free_full_and_cpu_masks;
3674	}
3675
3676	for (s = 0; s < nr_spec; s++) {
3677		ret = record__mmap_cpu_mask_init_spec(&thread_mask.maps, maps_spec[s]);
3678		if (ret) {
3679			pr_err("Failed to initialize maps thread mask\n");
3680			goto out_free;
3681		}
3682		ret = record__mmap_cpu_mask_init_spec(&thread_mask.affinity, affinity_spec[s]);
3683		if (ret) {
3684			pr_err("Failed to initialize affinity thread mask\n");
3685			goto out_free;
3686		}
3687
3688		/* ignore invalid CPUs but do not allow empty masks */
3689		if (!bitmap_and(thread_mask.maps.bits, thread_mask.maps.bits,
3690				cpus_mask.bits, thread_mask.maps.nbits)) {
3691			pr_err("Empty maps mask: %s\n", maps_spec[s]);
3692			ret = -EINVAL;
3693			goto out_free;
3694		}
3695		if (!bitmap_and(thread_mask.affinity.bits, thread_mask.affinity.bits,
3696				cpus_mask.bits, thread_mask.affinity.nbits)) {
3697			pr_err("Empty affinity mask: %s\n", affinity_spec[s]);
3698			ret = -EINVAL;
3699			goto out_free;
3700		}
3701
3702		/* do not allow intersection with other masks (full_mask) */
3703		if (bitmap_intersects(thread_mask.maps.bits, full_mask.maps.bits,
3704				      thread_mask.maps.nbits)) {
3705			pr_err("Intersecting maps mask: %s\n", maps_spec[s]);
3706			ret = -EINVAL;
3707			goto out_free;
3708		}
3709		if (bitmap_intersects(thread_mask.affinity.bits, full_mask.affinity.bits,
3710				      thread_mask.affinity.nbits)) {
3711			pr_err("Intersecting affinity mask: %s\n", affinity_spec[s]);
3712			ret = -EINVAL;
3713			goto out_free;
3714		}
3715
3716		bitmap_or(full_mask.maps.bits, full_mask.maps.bits,
3717			  thread_mask.maps.bits, full_mask.maps.nbits);
3718		bitmap_or(full_mask.affinity.bits, full_mask.affinity.bits,
3719			  thread_mask.affinity.bits, full_mask.maps.nbits);
3720
3721		thread_masks = realloc(rec->thread_masks, (t + 1) * sizeof(struct thread_mask));
3722		if (!thread_masks) {
3723			pr_err("Failed to reallocate thread masks\n");
3724			ret = -ENOMEM;
3725			goto out_free;
3726		}
3727		rec->thread_masks = thread_masks;
3728		rec->thread_masks[t] = thread_mask;
3729		if (verbose > 0) {
3730			pr_debug("thread_masks[%d]: ", t);
3731			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].maps, "maps");
3732			pr_debug("thread_masks[%d]: ", t);
3733			mmap_cpu_mask__scnprintf(&rec->thread_masks[t].affinity, "affinity");
3734		}
3735		t++;
3736		ret = record__thread_mask_alloc(&thread_mask, cpu__max_cpu().cpu);
3737		if (ret) {
3738			pr_err("Failed to allocate thread mask\n");
3739			goto out_free_full_and_cpu_masks;
3740		}
3741	}
3742	rec->nr_threads = t;
3743	pr_debug("nr_threads: %d\n", rec->nr_threads);
3744	if (!rec->nr_threads)
3745		ret = -EINVAL;
3746
3747out_free:
3748	record__thread_mask_free(&thread_mask);
3749out_free_full_and_cpu_masks:
3750	record__thread_mask_free(&full_mask);
3751out_free_cpu_mask:
3752	record__mmap_cpu_mask_free(&cpus_mask);
3753
3754	return ret;
3755}
3756
3757static int record__init_thread_core_masks(struct record *rec, struct perf_cpu_map *cpus)
3758{
3759	int ret;
3760	struct cpu_topology *topo;
3761
3762	topo = cpu_topology__new();
3763	if (!topo) {
3764		pr_err("Failed to allocate CPU topology\n");
3765		return -ENOMEM;
3766	}
3767
3768	ret = record__init_thread_masks_spec(rec, cpus, topo->core_cpus_list,
3769					     topo->core_cpus_list, topo->core_cpus_lists);
3770	cpu_topology__delete(topo);
3771
3772	return ret;
3773}
3774
3775static int record__init_thread_package_masks(struct record *rec, struct perf_cpu_map *cpus)
3776{
3777	int ret;
3778	struct cpu_topology *topo;
3779
3780	topo = cpu_topology__new();
3781	if (!topo) {
3782		pr_err("Failed to allocate CPU topology\n");
3783		return -ENOMEM;
3784	}
3785
3786	ret = record__init_thread_masks_spec(rec, cpus, topo->package_cpus_list,
3787					     topo->package_cpus_list, topo->package_cpus_lists);
3788	cpu_topology__delete(topo);
3789
3790	return ret;
3791}
3792
3793static int record__init_thread_numa_masks(struct record *rec, struct perf_cpu_map *cpus)
3794{
3795	u32 s;
3796	int ret;
3797	const char **spec;
3798	struct numa_topology *topo;
3799
3800	topo = numa_topology__new();
3801	if (!topo) {
3802		pr_err("Failed to allocate NUMA topology\n");
3803		return -ENOMEM;
3804	}
3805
3806	spec = zalloc(topo->nr * sizeof(char *));
3807	if (!spec) {
3808		pr_err("Failed to allocate NUMA spec\n");
3809		ret = -ENOMEM;
3810		goto out_delete_topo;
3811	}
3812	for (s = 0; s < topo->nr; s++)
3813		spec[s] = topo->nodes[s].cpus;
3814
3815	ret = record__init_thread_masks_spec(rec, cpus, spec, spec, topo->nr);
3816
3817	zfree(&spec);
3818
3819out_delete_topo:
3820	numa_topology__delete(topo);
3821
3822	return ret;
3823}
3824
3825static int record__init_thread_user_masks(struct record *rec, struct perf_cpu_map *cpus)
3826{
3827	int t, ret;
3828	u32 s, nr_spec = 0;
3829	char **maps_spec = NULL, **affinity_spec = NULL, **tmp_spec;
3830	char *user_spec, *spec, *spec_ptr, *mask, *mask_ptr, *dup_mask = NULL;
3831
3832	for (t = 0, user_spec = (char *)rec->opts.threads_user_spec; ; t++, user_spec = NULL) {
3833		spec = strtok_r(user_spec, ":", &spec_ptr);
3834		if (spec == NULL)
3835			break;
3836		pr_debug2("threads_spec[%d]: %s\n", t, spec);
3837		mask = strtok_r(spec, "/", &mask_ptr);
3838		if (mask == NULL)
3839			break;
3840		pr_debug2("  maps mask: %s\n", mask);
3841		tmp_spec = realloc(maps_spec, (nr_spec + 1) * sizeof(char *));
3842		if (!tmp_spec) {
3843			pr_err("Failed to reallocate maps spec\n");
3844			ret = -ENOMEM;
3845			goto out_free;
3846		}
3847		maps_spec = tmp_spec;
3848		maps_spec[nr_spec] = dup_mask = strdup(mask);
3849		if (!maps_spec[nr_spec]) {
3850			pr_err("Failed to allocate maps spec[%d]\n", nr_spec);
3851			ret = -ENOMEM;
3852			goto out_free;
3853		}
3854		mask = strtok_r(NULL, "/", &mask_ptr);
3855		if (mask == NULL) {
3856			pr_err("Invalid thread maps or affinity specs\n");
3857			ret = -EINVAL;
3858			goto out_free;
3859		}
3860		pr_debug2("  affinity mask: %s\n", mask);
3861		tmp_spec = realloc(affinity_spec, (nr_spec + 1) * sizeof(char *));
3862		if (!tmp_spec) {
3863			pr_err("Failed to reallocate affinity spec\n");
3864			ret = -ENOMEM;
3865			goto out_free;
3866		}
3867		affinity_spec = tmp_spec;
3868		affinity_spec[nr_spec] = strdup(mask);
3869		if (!affinity_spec[nr_spec]) {
3870			pr_err("Failed to allocate affinity spec[%d]\n", nr_spec);
3871			ret = -ENOMEM;
3872			goto out_free;
3873		}
3874		dup_mask = NULL;
3875		nr_spec++;
3876	}
3877
3878	ret = record__init_thread_masks_spec(rec, cpus, (const char **)maps_spec,
3879					     (const char **)affinity_spec, nr_spec);
3880
3881out_free:
3882	free(dup_mask);
3883	for (s = 0; s < nr_spec; s++) {
3884		if (maps_spec)
3885			free(maps_spec[s]);
3886		if (affinity_spec)
3887			free(affinity_spec[s]);
3888	}
3889	free(affinity_spec);
3890	free(maps_spec);
3891
3892	return ret;
3893}
3894
3895static int record__init_thread_default_masks(struct record *rec, struct perf_cpu_map *cpus)
3896{
3897	int ret;
3898
3899	ret = record__alloc_thread_masks(rec, 1, cpu__max_cpu().cpu);
3900	if (ret)
3901		return ret;
3902
3903	if (record__mmap_cpu_mask_init(&rec->thread_masks->maps, cpus))
3904		return -ENODEV;
3905
3906	rec->nr_threads = 1;
3907
3908	return 0;
3909}
3910
3911static int record__init_thread_masks(struct record *rec)
3912{
3913	int ret = 0;
3914	struct perf_cpu_map *cpus = rec->evlist->core.all_cpus;
3915
3916	if (!record__threads_enabled(rec))
3917		return record__init_thread_default_masks(rec, cpus);
3918
3919	if (evlist__per_thread(rec->evlist)) {
3920		pr_err("--per-thread option is mutually exclusive to parallel streaming mode.\n");
3921		return -EINVAL;
3922	}
3923
3924	switch (rec->opts.threads_spec) {
3925	case THREAD_SPEC__CPU:
3926		ret = record__init_thread_cpu_masks(rec, cpus);
3927		break;
3928	case THREAD_SPEC__CORE:
3929		ret = record__init_thread_core_masks(rec, cpus);
3930		break;
3931	case THREAD_SPEC__PACKAGE:
3932		ret = record__init_thread_package_masks(rec, cpus);
3933		break;
3934	case THREAD_SPEC__NUMA:
3935		ret = record__init_thread_numa_masks(rec, cpus);
3936		break;
3937	case THREAD_SPEC__USER:
3938		ret = record__init_thread_user_masks(rec, cpus);
3939		break;
3940	default:
3941		break;
3942	}
3943
3944	return ret;
3945}
3946
3947int cmd_record(int argc, const char **argv)
3948{
3949	int err;
3950	struct record *rec = &record;
3951	char errbuf[BUFSIZ];
3952
3953	setlocale(LC_ALL, "");
3954
3955#ifndef HAVE_LIBBPF_SUPPORT
3956# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
3957	set_nobuild('\0', "clang-path", true);
3958	set_nobuild('\0', "clang-opt", true);
3959# undef set_nobuild
3960#endif
3961
3962#ifndef HAVE_BPF_PROLOGUE
3963# if !defined (HAVE_DWARF_SUPPORT)
3964#  define REASON  "NO_DWARF=1"
3965# elif !defined (HAVE_LIBBPF_SUPPORT)
3966#  define REASON  "NO_LIBBPF=1"
3967# else
3968#  define REASON  "this architecture doesn't support BPF prologue"
3969# endif
3970# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
3971	set_nobuild('\0', "vmlinux", true);
3972# undef set_nobuild
3973# undef REASON
3974#endif
3975
3976#ifndef HAVE_BPF_SKEL
3977# define set_nobuild(s, l, m, c) set_option_nobuild(record_options, s, l, m, c)
3978	set_nobuild('\0', "off-cpu", "no BUILD_BPF_SKEL=1", true);
3979# undef set_nobuild
3980#endif
3981
3982	rec->opts.affinity = PERF_AFFINITY_SYS;
3983
3984	rec->evlist = evlist__new();
3985	if (rec->evlist == NULL)
3986		return -ENOMEM;
3987
3988	err = perf_config(perf_record_config, rec);
3989	if (err)
3990		return err;
3991
3992	argc = parse_options(argc, argv, record_options, record_usage,
3993			    PARSE_OPT_STOP_AT_NON_OPTION);
3994	if (quiet)
3995		perf_quiet_option();
3996
3997	err = symbol__validate_sym_arguments();
3998	if (err)
3999		return err;
4000
4001	perf_debuginfod_setup(&record.debuginfod);
4002
4003	/* Make system wide (-a) the default target. */
4004	if (!argc && target__none(&rec->opts.target))
4005		rec->opts.target.system_wide = true;
4006
4007	if (nr_cgroups && !rec->opts.target.system_wide) {
4008		usage_with_options_msg(record_usage, record_options,
4009			"cgroup monitoring only available in system-wide mode");
4010
4011	}
4012
4013	if (rec->buildid_mmap) {
4014		if (!perf_can_record_build_id()) {
4015			pr_err("Failed: no support to record build id in mmap events, update your kernel.\n");
4016			err = -EINVAL;
4017			goto out_opts;
4018		}
4019		pr_debug("Enabling build id in mmap2 events.\n");
4020		/* Enable mmap build id synthesizing. */
4021		symbol_conf.buildid_mmap2 = true;
4022		/* Enable perf_event_attr::build_id bit. */
4023		rec->opts.build_id = true;
4024		/* Disable build id cache. */
4025		rec->no_buildid = true;
4026	}
4027
4028	if (rec->opts.record_cgroup && !perf_can_record_cgroup()) {
4029		pr_err("Kernel has no cgroup sampling support.\n");
4030		err = -EINVAL;
4031		goto out_opts;
4032	}
4033
4034	if (rec->opts.kcore)
4035		rec->opts.text_poke = true;
4036
4037	if (rec->opts.kcore || record__threads_enabled(rec))
4038		rec->data.is_dir = true;
4039
4040	if (record__threads_enabled(rec)) {
4041		if (rec->opts.affinity != PERF_AFFINITY_SYS) {
4042			pr_err("--affinity option is mutually exclusive to parallel streaming mode.\n");
4043			goto out_opts;
4044		}
4045		if (record__aio_enabled(rec)) {
4046			pr_err("Asynchronous streaming mode (--aio) is mutually exclusive to parallel streaming mode.\n");
4047			goto out_opts;
4048		}
4049	}
4050
4051	if (rec->opts.comp_level != 0) {
4052		pr_debug("Compression enabled, disabling build id collection at the end of the session.\n");
4053		rec->no_buildid = true;
4054	}
4055
4056	if (rec->opts.record_switch_events &&
4057	    !perf_can_record_switch_events()) {
4058		ui__error("kernel does not support recording context switch events\n");
4059		parse_options_usage(record_usage, record_options, "switch-events", 0);
4060		err = -EINVAL;
4061		goto out_opts;
4062	}
4063
4064	if (switch_output_setup(rec)) {
4065		parse_options_usage(record_usage, record_options, "switch-output", 0);
4066		err = -EINVAL;
4067		goto out_opts;
4068	}
4069
4070	if (rec->switch_output.time) {
4071		signal(SIGALRM, alarm_sig_handler);
4072		alarm(rec->switch_output.time);
4073	}
4074
4075	if (rec->switch_output.num_files) {
4076		rec->switch_output.filenames = calloc(sizeof(char *),
4077						      rec->switch_output.num_files);
4078		if (!rec->switch_output.filenames) {
4079			err = -EINVAL;
4080			goto out_opts;
4081		}
4082	}
4083
4084	if (rec->timestamp_filename && record__threads_enabled(rec)) {
4085		rec->timestamp_filename = false;
4086		pr_warning("WARNING: --timestamp-filename option is not available in parallel streaming mode.\n");
4087	}
4088
4089	/*
4090	 * Allow aliases to facilitate the lookup of symbols for address
4091	 * filters. Refer to auxtrace_parse_filters().
4092	 */
4093	symbol_conf.allow_aliases = true;
4094
4095	symbol__init(NULL);
4096
4097	err = record__auxtrace_init(rec);
4098	if (err)
4099		goto out;
4100
4101	if (dry_run)
4102		goto out;
4103
4104	err = bpf__setup_stdout(rec->evlist);
4105	if (err) {
4106		bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
4107		pr_err("ERROR: Setup BPF stdout failed: %s\n",
4108			 errbuf);
4109		goto out;
4110	}
4111
4112	err = -ENOMEM;
4113
4114	if (rec->no_buildid_cache || rec->no_buildid) {
4115		disable_buildid_cache();
4116	} else if (rec->switch_output.enabled) {
4117		/*
4118		 * In 'perf record --switch-output', disable buildid
4119		 * generation by default to reduce data file switching
4120		 * overhead. Still generate buildid if they are required
4121		 * explicitly using
4122		 *
4123		 *  perf record --switch-output --no-no-buildid \
4124		 *              --no-no-buildid-cache
4125		 *
4126		 * Following code equals to:
4127		 *
4128		 * if ((rec->no_buildid || !rec->no_buildid_set) &&
4129		 *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
4130		 *         disable_buildid_cache();
4131		 */
4132		bool disable = true;
4133
4134		if (rec->no_buildid_set && !rec->no_buildid)
4135			disable = false;
4136		if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
4137			disable = false;
4138		if (disable) {
4139			rec->no_buildid = true;
4140			rec->no_buildid_cache = true;
4141			disable_buildid_cache();
4142		}
4143	}
4144
4145	if (record.opts.overwrite)
4146		record.opts.tail_synthesize = true;
4147
4148	if (rec->evlist->core.nr_entries == 0) {
4149		if (perf_pmu__has_hybrid()) {
4150			err = evlist__add_default_hybrid(rec->evlist,
4151							 !record.opts.no_samples);
4152		} else {
4153			err = __evlist__add_default(rec->evlist,
4154						    !record.opts.no_samples);
4155		}
4156
4157		if (err < 0) {
4158			pr_err("Not enough memory for event selector list\n");
4159			goto out;
4160		}
4161	}
4162
4163	if (rec->opts.target.tid && !rec->opts.no_inherit_set)
4164		rec->opts.no_inherit = true;
4165
4166	err = target__validate(&rec->opts.target);
4167	if (err) {
4168		target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
4169		ui__warning("%s\n", errbuf);
4170	}
4171
4172	err = target__parse_uid(&rec->opts.target);
4173	if (err) {
4174		int saved_errno = errno;
4175
4176		target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
4177		ui__error("%s", errbuf);
4178
4179		err = -saved_errno;
4180		goto out;
4181	}
4182
4183	/* Enable ignoring missing threads when -u/-p option is defined. */
4184	rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
4185
4186	if (evlist__fix_hybrid_cpus(rec->evlist, rec->opts.target.cpu_list)) {
4187		pr_err("failed to use cpu list %s\n",
4188		       rec->opts.target.cpu_list);
4189		goto out;
4190	}
4191
4192	rec->opts.target.hybrid = perf_pmu__has_hybrid();
4193
4194	if (callchain_param.enabled && callchain_param.record_mode == CALLCHAIN_FP)
4195		arch__add_leaf_frame_record_opts(&rec->opts);
4196
4197	err = -ENOMEM;
4198	if (evlist__create_maps(rec->evlist, &rec->opts.target) < 0) {
4199		if (rec->opts.target.pid != NULL) {
4200			pr_err("Couldn't create thread/CPU maps: %s\n",
4201				errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
4202			goto out;
4203		}
4204		else
4205			usage_with_options(record_usage, record_options);
4206	}
4207
4208	err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
4209	if (err)
4210		goto out;
4211
4212	/*
4213	 * We take all buildids when the file contains
4214	 * AUX area tracing data because we do not decode the
4215	 * trace because it would take too long.
4216	 */
4217	if (rec->opts.full_auxtrace)
4218		rec->buildid_all = true;
4219
4220	if (rec->opts.text_poke) {
4221		err = record__config_text_poke(rec->evlist);
4222		if (err) {
4223			pr_err("record__config_text_poke failed, error %d\n", err);
4224			goto out;
4225		}
4226	}
4227
4228	if (rec->off_cpu) {
4229		err = record__config_off_cpu(rec);
4230		if (err) {
4231			pr_err("record__config_off_cpu failed, error %d\n", err);
4232			goto out;
4233		}
4234	}
4235
4236	if (record_opts__config(&rec->opts)) {
4237		err = -EINVAL;
4238		goto out;
4239	}
4240
4241	err = record__init_thread_masks(rec);
4242	if (err) {
4243		pr_err("Failed to initialize parallel data streaming masks\n");
4244		goto out;
4245	}
4246
4247	if (rec->opts.nr_cblocks > nr_cblocks_max)
4248		rec->opts.nr_cblocks = nr_cblocks_max;
4249	pr_debug("nr_cblocks: %d\n", rec->opts.nr_cblocks);
4250
4251	pr_debug("affinity: %s\n", affinity_tags[rec->opts.affinity]);
4252	pr_debug("mmap flush: %d\n", rec->opts.mmap_flush);
4253
4254	if (rec->opts.comp_level > comp_level_max)
4255		rec->opts.comp_level = comp_level_max;
4256	pr_debug("comp level: %d\n", rec->opts.comp_level);
4257
4258	err = __cmd_record(&record, argc, argv);
4259out:
4260	evlist__delete(rec->evlist);
4261	symbol__exit();
4262	auxtrace_record__free(rec->itr);
4263out_opts:
4264	record__free_thread_masks(rec, rec->nr_threads);
4265	rec->nr_threads = 0;
4266	evlist__close_control(rec->opts.ctl_fd, rec->opts.ctl_fd_ack, &rec->opts.ctl_fd_close);
4267	return err;
4268}
4269
4270static void snapshot_sig_handler(int sig __maybe_unused)
4271{
4272	struct record *rec = &record;
4273
4274	hit_auxtrace_snapshot_trigger(rec);
4275
4276	if (switch_output_signal(rec))
4277		trigger_hit(&switch_output_trigger);
4278}
4279
4280static void alarm_sig_handler(int sig __maybe_unused)
4281{
4282	struct record *rec = &record;
4283
4284	if (switch_output_time(rec))
4285		trigger_hit(&switch_output_trigger);
4286}