intel-pt.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * intel_pt.c: Intel Processor Trace support
   4 * Copyright (c) 2013-2015, Intel Corporation.
   5 */
   6
   7#include <errno.h>
   8#include <stdbool.h>
   9#include <linux/kernel.h>
  10#include <linux/types.h>
  11#include <linux/bitops.h>
  12#include <linux/log2.h>
  13#include <linux/zalloc.h>
  14#include <linux/err.h>
  15
  16#include "../../../util/session.h"
  17#include "../../../util/event.h"
  18#include "../../../util/evlist.h"
  19#include "../../../util/evsel.h"
  20#include "../../../util/evsel_config.h"
  21#include "../../../util/config.h"
  22#include "../../../util/cpumap.h"
  23#include "../../../util/mmap.h"
  24#include <subcmd/parse-options.h>
  25#include "../../../util/parse-events.h"
  26#include "../../../util/pmus.h"
  27#include "../../../util/debug.h"
  28#include "../../../util/auxtrace.h"
  29#include "../../../util/perf_api_probe.h"
  30#include "../../../util/record.h"
  31#include "../../../util/target.h"
  32#include "../../../util/tsc.h"
  33#include <internal/lib.h> // page_size
  34#include "../../../util/intel-pt.h"
  35#include <api/fs/fs.h>
  36#include "cpuid.h"
  37
  38#define KiB(x) ((x) * 1024)
  39#define MiB(x) ((x) * 1024 * 1024)
  40#define KiB_MASK(x) (KiB(x) - 1)
  41#define MiB_MASK(x) (MiB(x) - 1)
  42
  43#define INTEL_PT_PSB_PERIOD_NEAR	256
  44
  45struct intel_pt_snapshot_ref {
  46	void *ref_buf;
  47	size_t ref_offset;
  48	bool wrapped;
  49};
  50
  51struct intel_pt_recording {
  52	struct auxtrace_record		itr;
  53	struct perf_pmu			*intel_pt_pmu;
  54	int				have_sched_switch;
  55	struct evlist		*evlist;
  56	bool				all_switch_events;
  57	bool				snapshot_mode;
  58	bool				snapshot_init_done;
  59	size_t				snapshot_size;
  60	size_t				snapshot_ref_buf_size;
  61	int				snapshot_ref_cnt;
  62	struct intel_pt_snapshot_ref	*snapshot_refs;
  63	size_t				priv_size;
  64};
  65
  66static int intel_pt_parse_terms_with_default(const struct perf_pmu *pmu,
  67					     const char *str,
  68					     u64 *config)
  69{
  70	struct parse_events_terms terms;
  71	struct perf_event_attr attr = { .size = 0, };
  72	int err;
  73
  74	parse_events_terms__init(&terms);
  75	err = parse_events_terms(&terms, str);
  76	if (err)
  77		goto out_free;
  78
  79	attr.config = *config;
  80	err = perf_pmu__config_terms(pmu, &attr, &terms, /*zero=*/true, /*apply_hardcoded=*/false,
  81				     /*err=*/NULL);
  82	if (err)
  83		goto out_free;
  84
  85	*config = attr.config;
  86out_free:
  87	parse_events_terms__exit(&terms);
  88	return err;
  89}
  90
  91static int intel_pt_parse_terms(const struct perf_pmu *pmu, const char *str, u64 *config)
  92{
  93	*config = 0;
  94	return intel_pt_parse_terms_with_default(pmu, str, config);
  95}
  96
  97static u64 intel_pt_masked_bits(u64 mask, u64 bits)
  98{
  99	const u64 top_bit = 1ULL << 63;
 100	u64 res = 0;
 101	int i;
 102
 103	for (i = 0; i < 64; i++) {
 104		if (mask & top_bit) {
 105			res <<= 1;
 106			if (bits & top_bit)
 107				res |= 1;
 108		}
 109		mask <<= 1;
 110		bits <<= 1;
 111	}
 112
 113	return res;
 114}
 115
 116static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
 117				struct evlist *evlist, u64 *res)
 118{
 119	struct evsel *evsel;
 120	u64 mask;
 121
 122	*res = 0;
 123
 124	mask = perf_pmu__format_bits(intel_pt_pmu, str);
 125	if (!mask)
 126		return -EINVAL;
 127
 128	evlist__for_each_entry(evlist, evsel) {
 129		if (evsel->core.attr.type == intel_pt_pmu->type) {
 130			*res = intel_pt_masked_bits(mask, evsel->core.attr.config);
 131			return 0;
 132		}
 133	}
 134
 135	return -EINVAL;
 136}
 137
 138static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
 139				  struct evlist *evlist)
 140{
 141	u64 val;
 142	int err, topa_multiple_entries;
 143	size_t psb_period;
 144
 145	if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
 146				"%d", &topa_multiple_entries) != 1)
 147		topa_multiple_entries = 0;
 148
 149	/*
 150	 * Use caps/topa_multiple_entries to indicate early hardware that had
 151	 * extra frequent PSBs.
 152	 */
 153	if (!topa_multiple_entries) {
 154		psb_period = 256;
 155		goto out;
 156	}
 157
 158	err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
 159	if (err)
 160		val = 0;
 161
 162	psb_period = 1 << (val + 11);
 163out:
 164	pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
 165	return psb_period;
 166}
 167
 168static int intel_pt_pick_bit(int bits, int target)
 169{
 170	int pos, pick = -1;
 171
 172	for (pos = 0; bits; bits >>= 1, pos++) {
 173		if (bits & 1) {
 174			if (pos <= target || pick < 0)
 175				pick = pos;
 176			if (pos >= target)
 177				break;
 178		}
 179	}
 180
 181	return pick;
 182}
 183
 184static u64 intel_pt_default_config(const struct perf_pmu *intel_pt_pmu)
 185{
 186	char buf[256];
 187	int mtc, mtc_periods = 0, mtc_period;
 188	int psb_cyc, psb_periods, psb_period;
 189	int pos = 0;
 190	u64 config;
 191	char c;
 192	int dirfd;
 193
 194	dirfd = perf_pmu__event_source_devices_fd();
 195
 196	pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
 197
 198	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/mtc", "%d",
 199				   &mtc) != 1)
 200		mtc = 1;
 201
 202	if (mtc) {
 203		if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/mtc_periods", "%x",
 204					   &mtc_periods) != 1)
 205			mtc_periods = 0;
 206		if (mtc_periods) {
 207			mtc_period = intel_pt_pick_bit(mtc_periods, 3);
 208			pos += scnprintf(buf + pos, sizeof(buf) - pos,
 209					 ",mtc,mtc_period=%d", mtc_period);
 210		}
 211	}
 212
 213	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/psb_cyc", "%d",
 214				   &psb_cyc) != 1)
 215		psb_cyc = 1;
 216
 217	if (psb_cyc && mtc_periods) {
 218		if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/psb_periods", "%x",
 219					   &psb_periods) != 1)
 220			psb_periods = 0;
 221		if (psb_periods) {
 222			psb_period = intel_pt_pick_bit(psb_periods, 3);
 223			pos += scnprintf(buf + pos, sizeof(buf) - pos,
 224					 ",psb_period=%d", psb_period);
 225		}
 226	}
 227
 228	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/pt", "%c", &c) == 1 &&
 229	    perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/branch", "%c", &c) == 1)
 230		pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
 231
 232	pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
 233
 234	intel_pt_parse_terms(intel_pt_pmu, buf, &config);
 235
 236	close(dirfd);
 237	return config;
 238}
 239
 240static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
 241					   struct record_opts *opts,
 242					   const char *str)
 243{
 244	struct intel_pt_recording *ptr =
 245			container_of(itr, struct intel_pt_recording, itr);
 246	unsigned long long snapshot_size = 0;
 247	char *endptr;
 248
 249	if (str) {
 250		snapshot_size = strtoull(str, &endptr, 0);
 251		if (*endptr || snapshot_size > SIZE_MAX)
 252			return -1;
 253	}
 254
 255	opts->auxtrace_snapshot_mode = true;
 256	opts->auxtrace_snapshot_size = snapshot_size;
 257
 258	ptr->snapshot_size = snapshot_size;
 259
 260	return 0;
 261}
 262
 263void intel_pt_pmu_default_config(const struct perf_pmu *intel_pt_pmu,
 264				 struct perf_event_attr *attr)
 265{
 266	static u64 config;
 267	static bool initialized;
 268
 269	if (!initialized) {
 270		config = intel_pt_default_config(intel_pt_pmu);
 271		initialized = true;
 272	}
 273	attr->config = config;
 274}
 275
 276static const char *intel_pt_find_filter(struct evlist *evlist,
 277					struct perf_pmu *intel_pt_pmu)
 278{
 279	struct evsel *evsel;
 280
 281	evlist__for_each_entry(evlist, evsel) {
 282		if (evsel->core.attr.type == intel_pt_pmu->type)
 283			return evsel->filter;
 284	}
 285
 286	return NULL;
 287}
 288
 289static size_t intel_pt_filter_bytes(const char *filter)
 290{
 291	size_t len = filter ? strlen(filter) : 0;
 292
 293	return len ? roundup(len + 1, 8) : 0;
 294}
 295
 296static size_t
 297intel_pt_info_priv_size(struct auxtrace_record *itr, struct evlist *evlist)
 298{
 299	struct intel_pt_recording *ptr =
 300			container_of(itr, struct intel_pt_recording, itr);
 301	const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
 302
 303	ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
 304			 intel_pt_filter_bytes(filter);
 305	ptr->priv_size += sizeof(u64); /* Cap Event Trace */
 306
 307	return ptr->priv_size;
 308}
 309
 310static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
 311{
 312	unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
 313
 314	cpuid(0x15, 0, &eax, &ebx, &ecx, &edx);
 315	*n = ebx;
 316	*d = eax;
 317}
 318
 319static int intel_pt_info_fill(struct auxtrace_record *itr,
 320			      struct perf_session *session,
 321			      struct perf_record_auxtrace_info *auxtrace_info,
 322			      size_t priv_size)
 323{
 324	struct intel_pt_recording *ptr =
 325			container_of(itr, struct intel_pt_recording, itr);
 326	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 327	struct perf_event_mmap_page *pc;
 328	struct perf_tsc_conversion tc = { .time_mult = 0, };
 329	bool cap_user_time_zero = false, per_cpu_mmaps;
 330	u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
 331	u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
 332	unsigned long max_non_turbo_ratio;
 333	size_t filter_str_len;
 334	const char *filter;
 335	int event_trace;
 336	__u64 *info;
 337	int err;
 338
 339	if (priv_size != ptr->priv_size)
 340		return -EINVAL;
 341
 342	intel_pt_parse_terms(intel_pt_pmu, "tsc", &tsc_bit);
 343	intel_pt_parse_terms(intel_pt_pmu, "noretcomp", &noretcomp_bit);
 344	intel_pt_parse_terms(intel_pt_pmu, "mtc", &mtc_bit);
 345	mtc_freq_bits = perf_pmu__format_bits(intel_pt_pmu, "mtc_period");
 346	intel_pt_parse_terms(intel_pt_pmu, "cyc", &cyc_bit);
 347
 348	intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
 349
 350	if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
 351				"%lu", &max_non_turbo_ratio) != 1)
 352		max_non_turbo_ratio = 0;
 353	if (perf_pmu__scan_file(intel_pt_pmu, "caps/event_trace",
 354				"%d", &event_trace) != 1)
 355		event_trace = 0;
 356
 357	filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
 358	filter_str_len = filter ? strlen(filter) : 0;
 359
 360	if (!session->evlist->core.nr_mmaps)
 361		return -EINVAL;
 362
 363	pc = session->evlist->mmap[0].core.base;
 364	if (pc) {
 365		err = perf_read_tsc_conversion(pc, &tc);
 366		if (err) {
 367			if (err != -EOPNOTSUPP)
 368				return err;
 369		} else {
 370			cap_user_time_zero = tc.time_mult != 0;
 371		}
 372		if (!cap_user_time_zero)
 373			ui__warning("Intel Processor Trace: TSC not available\n");
 374	}
 375
 376	per_cpu_mmaps = !perf_cpu_map__is_any_cpu_or_is_empty(session->evlist->core.user_requested_cpus);
 377
 378	auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
 379	auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
 380	auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
 381	auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
 382	auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
 383	auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
 384	auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
 385	auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
 386	auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
 387	auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
 388	auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
 389	auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
 390	auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
 391	auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
 392	auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
 393	auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
 394	auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
 395	auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
 396
 397	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
 398
 399	if (filter_str_len) {
 400		size_t len = intel_pt_filter_bytes(filter);
 401
 402		strncpy((char *)info, filter, len);
 403		info += len >> 3;
 404	}
 405
 406	*info++ = event_trace;
 407
 408	return 0;
 409}
 410
 411#ifdef HAVE_LIBTRACEEVENT
 412static int intel_pt_track_switches(struct evlist *evlist)
 413{
 414	const char *sched_switch = "sched:sched_switch";
 415	struct evsel *evsel;
 416	int err;
 417
 418	if (!evlist__can_select_event(evlist, sched_switch))
 419		return -EPERM;
 420
 421	evsel = evlist__add_sched_switch(evlist, true);
 422	if (IS_ERR(evsel)) {
 423		err = PTR_ERR(evsel);
 424		pr_debug2("%s: failed to create %s, error = %d\n",
 425			  __func__, sched_switch, err);
 426		return err;
 427	}
 428
 429	evsel->immediate = true;
 430
 431	return 0;
 432}
 433#endif
 434
 435static bool intel_pt_exclude_guest(void)
 436{
 437	int pt_mode;
 438
 439	if (sysfs__read_int("module/kvm_intel/parameters/pt_mode", &pt_mode))
 440		pt_mode = 0;
 441
 442	return pt_mode == 1;
 443}
 444
 445static void intel_pt_valid_str(char *str, size_t len, u64 valid)
 446{
 447	unsigned int val, last = 0, state = 1;
 448	int p = 0;
 449
 450	str[0] = '\0';
 451
 452	for (val = 0; val <= 64; val++, valid >>= 1) {
 453		if (valid & 1) {
 454			last = val;
 455			switch (state) {
 456			case 0:
 457				p += scnprintf(str + p, len - p, ",");
 458				/* Fall through */
 459			case 1:
 460				p += scnprintf(str + p, len - p, "%u", val);
 461				state = 2;
 462				break;
 463			case 2:
 464				state = 3;
 465				break;
 466			case 3:
 467				state = 4;
 468				break;
 469			default:
 470				break;
 471			}
 472		} else {
 473			switch (state) {
 474			case 3:
 475				p += scnprintf(str + p, len - p, ",%u", last);
 476				state = 0;
 477				break;
 478			case 4:
 479				p += scnprintf(str + p, len - p, "-%u", last);
 480				state = 0;
 481				break;
 482			default:
 483				break;
 484			}
 485			if (state != 1)
 486				state = 0;
 487		}
 488	}
 489}
 490
 491static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu, int dirfd,
 492				    const char *caps, const char *name,
 493				    const char *supported, u64 config)
 494{
 495	char valid_str[256];
 496	unsigned int shift;
 497	unsigned long long valid;
 498	u64 bits;
 499	int ok;
 500
 501	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, caps, "%llx", &valid) != 1)
 502		valid = 0;
 503
 504	if (supported &&
 505	    perf_pmu__scan_file_at(intel_pt_pmu, dirfd, supported, "%d", &ok) == 1 && !ok)
 506		valid = 0;
 507
 508	valid |= 1;
 509
 510	bits = perf_pmu__format_bits(intel_pt_pmu, name);
 511
 512	config &= bits;
 513
 514	for (shift = 0; bits && !(bits & 1); shift++)
 515		bits >>= 1;
 516
 517	config >>= shift;
 518
 519	if (config > 63)
 520		goto out_err;
 521
 522	if (valid & (1 << config))
 523		return 0;
 524out_err:
 525	intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
 526	pr_err("Invalid %s for %s. Valid values are: %s\n",
 527	       name, INTEL_PT_PMU_NAME, valid_str);
 528	return -EINVAL;
 529}
 530
 531static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
 532				    struct evsel *evsel)
 533{
 534	int err, dirfd;
 535	char c;
 536
 537	if (!evsel)
 538		return 0;
 539
 540	dirfd = perf_pmu__event_source_devices_fd();
 541	if (dirfd < 0)
 542		return dirfd;
 543
 544	/*
 545	 * If supported, force pass-through config term (pt=1) even if user
 546	 * sets pt=0, which avoids senseless kernel errors.
 547	 */
 548	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/pt", "%c", &c) == 1 &&
 549	    !(evsel->core.attr.config & 1)) {
 550		pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
 551		evsel->core.attr.config |= 1;
 552	}
 553
 554	err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/cycle_thresholds",
 555				       "cyc_thresh", "caps/psb_cyc",
 556				       evsel->core.attr.config);
 557	if (err)
 558		goto out;
 559
 560	err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/mtc_periods",
 561				       "mtc_period", "caps/mtc",
 562				       evsel->core.attr.config);
 563	if (err)
 564		goto out;
 565
 566	err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/psb_periods",
 567					"psb_period", "caps/psb_cyc",
 568					evsel->core.attr.config);
 569
 570out:
 571	close(dirfd);
 572	return err;
 573}
 574
 575static void intel_pt_min_max_sample_sz(struct evlist *evlist,
 576				       size_t *min_sz, size_t *max_sz)
 577{
 578	struct evsel *evsel;
 579
 580	evlist__for_each_entry(evlist, evsel) {
 581		size_t sz = evsel->core.attr.aux_sample_size;
 582
 583		if (!sz)
 584			continue;
 585		if (min_sz && (sz < *min_sz || !*min_sz))
 586			*min_sz = sz;
 587		if (max_sz && sz > *max_sz)
 588			*max_sz = sz;
 589	}
 590}
 591
 592/*
 593 * Currently, there is not enough information to disambiguate different PEBS
 594 * events, so only allow one.
 595 */
 596static bool intel_pt_too_many_aux_output(struct evlist *evlist)
 597{
 598	struct evsel *evsel;
 599	int aux_output_cnt = 0;
 600
 601	evlist__for_each_entry(evlist, evsel)
 602		aux_output_cnt += !!evsel->core.attr.aux_output;
 603
 604	if (aux_output_cnt > 1) {
 605		pr_err(INTEL_PT_PMU_NAME " supports at most one event with aux-output\n");
 606		return true;
 607	}
 608
 609	return false;
 610}
 611
 612static int intel_pt_recording_options(struct auxtrace_record *itr,
 613				      struct evlist *evlist,
 614				      struct record_opts *opts)
 615{
 616	struct intel_pt_recording *ptr =
 617			container_of(itr, struct intel_pt_recording, itr);
 618	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 619	bool have_timing_info, need_immediate = false;
 620	struct evsel *evsel, *intel_pt_evsel = NULL;
 621	const struct perf_cpu_map *cpus = evlist->core.user_requested_cpus;
 622	bool privileged = perf_event_paranoid_check(-1);
 623	u64 tsc_bit;
 624	int err;
 625
 626	ptr->evlist = evlist;
 627	ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
 628
 629	evlist__for_each_entry(evlist, evsel) {
 630		if (evsel->core.attr.type == intel_pt_pmu->type) {
 631			if (intel_pt_evsel) {
 632				pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
 633				return -EINVAL;
 634			}
 635			evsel->core.attr.freq = 0;
 636			evsel->core.attr.sample_period = 1;
 637			evsel->core.attr.exclude_guest = intel_pt_exclude_guest();
 638			evsel->no_aux_samples = true;
 639			evsel->needs_auxtrace_mmap = true;
 640			intel_pt_evsel = evsel;
 641			opts->full_auxtrace = true;
 642		}
 643	}
 644
 645	if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
 646		pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
 647		return -EINVAL;
 648	}
 649
 650	if (opts->auxtrace_snapshot_mode && opts->auxtrace_sample_mode) {
 651		pr_err("Snapshot mode (" INTEL_PT_PMU_NAME " PMU) and sample trace cannot be used together\n");
 652		return -EINVAL;
 653	}
 654
 655	if (opts->use_clockid) {
 656		pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
 657		return -EINVAL;
 658	}
 659
 660	if (intel_pt_too_many_aux_output(evlist))
 661		return -EINVAL;
 662
 663	if (!opts->full_auxtrace)
 664		return 0;
 665
 666	if (opts->auxtrace_sample_mode)
 667		evsel__set_config_if_unset(intel_pt_pmu, intel_pt_evsel,
 668					   "psb_period", 0);
 669
 670	err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
 671	if (err)
 672		return err;
 673
 674	/* Set default sizes for snapshot mode */
 675	if (opts->auxtrace_snapshot_mode) {
 676		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 677
 678		if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
 679			if (privileged) {
 680				opts->auxtrace_mmap_pages = MiB(4) / page_size;
 681			} else {
 682				opts->auxtrace_mmap_pages = KiB(128) / page_size;
 683				if (opts->mmap_pages == UINT_MAX)
 684					opts->mmap_pages = KiB(256) / page_size;
 685			}
 686		} else if (!opts->auxtrace_mmap_pages && !privileged &&
 687			   opts->mmap_pages == UINT_MAX) {
 688			opts->mmap_pages = KiB(256) / page_size;
 689		}
 690		if (!opts->auxtrace_snapshot_size)
 691			opts->auxtrace_snapshot_size =
 692				opts->auxtrace_mmap_pages * (size_t)page_size;
 693		if (!opts->auxtrace_mmap_pages) {
 694			size_t sz = opts->auxtrace_snapshot_size;
 695
 696			sz = round_up(sz, page_size) / page_size;
 697			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 698		}
 699		if (opts->auxtrace_snapshot_size >
 700				opts->auxtrace_mmap_pages * (size_t)page_size) {
 701			pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
 702			       opts->auxtrace_snapshot_size,
 703			       opts->auxtrace_mmap_pages * (size_t)page_size);
 704			return -EINVAL;
 705		}
 706		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
 707			pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
 708			return -EINVAL;
 709		}
 710		pr_debug2("Intel PT snapshot size: %zu\n",
 711			  opts->auxtrace_snapshot_size);
 712		if (psb_period &&
 713		    opts->auxtrace_snapshot_size <= psb_period +
 714						  INTEL_PT_PSB_PERIOD_NEAR)
 715			ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
 716				    opts->auxtrace_snapshot_size, psb_period);
 717	}
 718
 719	/* Set default sizes for sample mode */
 720	if (opts->auxtrace_sample_mode) {
 721		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 722		size_t min_sz = 0, max_sz = 0;
 723
 724		intel_pt_min_max_sample_sz(evlist, &min_sz, &max_sz);
 725		if (!opts->auxtrace_mmap_pages && !privileged &&
 726		    opts->mmap_pages == UINT_MAX)
 727			opts->mmap_pages = KiB(256) / page_size;
 728		if (!opts->auxtrace_mmap_pages) {
 729			size_t sz = round_up(max_sz, page_size) / page_size;
 730
 731			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 732		}
 733		if (max_sz > opts->auxtrace_mmap_pages * (size_t)page_size) {
 734			pr_err("Sample size %zu must not be greater than AUX area tracing mmap size %zu\n",
 735			       max_sz,
 736			       opts->auxtrace_mmap_pages * (size_t)page_size);
 737			return -EINVAL;
 738		}
 739		pr_debug2("Intel PT min. sample size: %zu max. sample size: %zu\n",
 740			  min_sz, max_sz);
 741		if (psb_period &&
 742		    min_sz <= psb_period + INTEL_PT_PSB_PERIOD_NEAR)
 743			ui__warning("Intel PT sample size (%zu) may be too small for PSB period (%zu)\n",
 744				    min_sz, psb_period);
 745	}
 746
 747	/* Set default sizes for full trace mode */
 748	if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
 749		if (privileged) {
 750			opts->auxtrace_mmap_pages = MiB(4) / page_size;
 751		} else {
 752			opts->auxtrace_mmap_pages = KiB(128) / page_size;
 753			if (opts->mmap_pages == UINT_MAX)
 754				opts->mmap_pages = KiB(256) / page_size;
 755		}
 756	}
 757
 758	/* Validate auxtrace_mmap_pages */
 759	if (opts->auxtrace_mmap_pages) {
 760		size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 761		size_t min_sz;
 762
 763		if (opts->auxtrace_snapshot_mode || opts->auxtrace_sample_mode)
 764			min_sz = KiB(4);
 765		else
 766			min_sz = KiB(8);
 767
 768		if (sz < min_sz || !is_power_of_2(sz)) {
 769			pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
 770			       min_sz / 1024);
 771			return -EINVAL;
 772		}
 773	}
 774
 775	if (!opts->auxtrace_snapshot_mode && !opts->auxtrace_sample_mode) {
 776		size_t aw = opts->auxtrace_mmap_pages * (size_t)page_size / 4;
 777		u32 aux_watermark = aw > UINT_MAX ? UINT_MAX : aw;
 778
 779		intel_pt_evsel->core.attr.aux_watermark = aux_watermark;
 780	}
 781
 782	intel_pt_parse_terms(intel_pt_pmu, "tsc", &tsc_bit);
 783
 784	if (opts->full_auxtrace && (intel_pt_evsel->core.attr.config & tsc_bit))
 785		have_timing_info = true;
 786	else
 787		have_timing_info = false;
 788
 789	/*
 790	 * Per-cpu recording needs sched_switch events to distinguish different
 791	 * threads.
 792	 */
 793	if (have_timing_info && !perf_cpu_map__is_any_cpu_or_is_empty(cpus) &&
 794	    !record_opts__no_switch_events(opts)) {
 795		if (perf_can_record_switch_events()) {
 796			bool cpu_wide = !target__none(&opts->target) &&
 797					!target__has_task(&opts->target);
 798
 799			if (ptr->all_switch_events && !cpu_wide && perf_can_record_cpu_wide()) {
 800				struct evsel *switch_evsel;
 801
 802				switch_evsel = evlist__add_dummy_on_all_cpus(evlist);
 803				if (!switch_evsel)
 804					return -ENOMEM;
 805
 806				switch_evsel->core.attr.context_switch = 1;
 807				switch_evsel->immediate = true;
 808
 809				evsel__set_sample_bit(switch_evsel, TID);
 810				evsel__set_sample_bit(switch_evsel, TIME);
 811				evsel__set_sample_bit(switch_evsel, CPU);
 812				evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
 813
 814				opts->record_switch_events = false;
 815				ptr->have_sched_switch = 3;
 816			} else {
 817				opts->record_switch_events = true;
 818				need_immediate = true;
 819				if (cpu_wide)
 820					ptr->have_sched_switch = 3;
 821				else
 822					ptr->have_sched_switch = 2;
 823			}
 824		} else {
 825#ifdef HAVE_LIBTRACEEVENT
 826			err = intel_pt_track_switches(evlist);
 827			if (err == -EPERM)
 828				pr_debug2("Unable to select sched:sched_switch\n");
 829			else if (err)
 830				return err;
 831			else
 832				ptr->have_sched_switch = 1;
 833#endif
 834		}
 835	}
 836
 837	if (have_timing_info && !intel_pt_evsel->core.attr.exclude_kernel &&
 838	    perf_can_record_text_poke_events() && perf_can_record_cpu_wide())
 839		opts->text_poke = true;
 840
 841	if (intel_pt_evsel) {
 842		/*
 843		 * To obtain the auxtrace buffer file descriptor, the auxtrace
 844		 * event must come first.
 845		 */
 846		evlist__to_front(evlist, intel_pt_evsel);
 847		/*
 848		 * In the case of per-cpu mmaps, we need the CPU on the
 849		 * AUX event.
 850		 */
 851		if (!perf_cpu_map__is_any_cpu_or_is_empty(cpus))
 852			evsel__set_sample_bit(intel_pt_evsel, CPU);
 853	}
 854
 855	/* Add dummy event to keep tracking */
 856	if (opts->full_auxtrace) {
 857		bool need_system_wide_tracking;
 858		struct evsel *tracking_evsel;
 859
 860		/*
 861		 * User space tasks can migrate between CPUs, so when tracing
 862		 * selected CPUs, sideband for all CPUs is still needed.
 863		 */
 864		need_system_wide_tracking = opts->target.cpu_list &&
 865					    !intel_pt_evsel->core.attr.exclude_user;
 866
 867		tracking_evsel = evlist__add_aux_dummy(evlist, need_system_wide_tracking);
 868		if (!tracking_evsel)
 869			return -ENOMEM;
 870
 871		evlist__set_tracking_event(evlist, tracking_evsel);
 872
 873		if (need_immediate)
 874			tracking_evsel->immediate = true;
 875
 876		/* In per-cpu case, always need the time of mmap events etc */
 877		if (!perf_cpu_map__is_any_cpu_or_is_empty(cpus)) {
 878			evsel__set_sample_bit(tracking_evsel, TIME);
 879			/* And the CPU for switch events */
 880			evsel__set_sample_bit(tracking_evsel, CPU);
 881		}
 882		evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
 883	}
 884
 885	/*
 886	 * Warn the user when we do not have enough information to decode i.e.
 887	 * per-cpu with no sched_switch (except workload-only).
 888	 */
 889	if (!ptr->have_sched_switch && !perf_cpu_map__is_any_cpu_or_is_empty(cpus) &&
 890	    !target__none(&opts->target) &&
 891	    !intel_pt_evsel->core.attr.exclude_user)
 892		ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
 893
 894	return 0;
 895}
 896
 897static int intel_pt_snapshot_start(struct auxtrace_record *itr)
 898{
 899	struct intel_pt_recording *ptr =
 900			container_of(itr, struct intel_pt_recording, itr);
 901	struct evsel *evsel;
 902
 903	evlist__for_each_entry(ptr->evlist, evsel) {
 904		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
 905			return evsel__disable(evsel);
 906	}
 907	return -EINVAL;
 908}
 909
 910static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
 911{
 912	struct intel_pt_recording *ptr =
 913			container_of(itr, struct intel_pt_recording, itr);
 914	struct evsel *evsel;
 915
 916	evlist__for_each_entry(ptr->evlist, evsel) {
 917		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
 918			return evsel__enable(evsel);
 919	}
 920	return -EINVAL;
 921}
 922
 923static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
 924{
 925	const size_t sz = sizeof(struct intel_pt_snapshot_ref);
 926	int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
 927	struct intel_pt_snapshot_ref *refs;
 928
 929	if (!new_cnt)
 930		new_cnt = 16;
 931
 932	while (new_cnt <= idx)
 933		new_cnt *= 2;
 934
 935	refs = calloc(new_cnt, sz);
 936	if (!refs)
 937		return -ENOMEM;
 938
 939	memcpy(refs, ptr->snapshot_refs, cnt * sz);
 940
 941	ptr->snapshot_refs = refs;
 942	ptr->snapshot_ref_cnt = new_cnt;
 943
 944	return 0;
 945}
 946
 947static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
 948{
 949	int i;
 950
 951	for (i = 0; i < ptr->snapshot_ref_cnt; i++)
 952		zfree(&ptr->snapshot_refs[i].ref_buf);
 953	zfree(&ptr->snapshot_refs);
 954}
 955
 956static void intel_pt_recording_free(struct auxtrace_record *itr)
 957{
 958	struct intel_pt_recording *ptr =
 959			container_of(itr, struct intel_pt_recording, itr);
 960
 961	intel_pt_free_snapshot_refs(ptr);
 962	free(ptr);
 963}
 964
 965static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
 966				       size_t snapshot_buf_size)
 967{
 968	size_t ref_buf_size = ptr->snapshot_ref_buf_size;
 969	void *ref_buf;
 970
 971	ref_buf = zalloc(ref_buf_size);
 972	if (!ref_buf)
 973		return -ENOMEM;
 974
 975	ptr->snapshot_refs[idx].ref_buf = ref_buf;
 976	ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
 977
 978	return 0;
 979}
 980
 981static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
 982					     size_t snapshot_buf_size)
 983{
 984	const size_t max_size = 256 * 1024;
 985	size_t buf_size = 0, psb_period;
 986
 987	if (ptr->snapshot_size <= 64 * 1024)
 988		return 0;
 989
 990	psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
 991	if (psb_period)
 992		buf_size = psb_period * 2;
 993
 994	if (!buf_size || buf_size > max_size)
 995		buf_size = max_size;
 996
 997	if (buf_size >= snapshot_buf_size)
 998		return 0;
 999
1000	if (buf_size >= ptr->snapshot_size / 2)
1001		return 0;
1002
1003	return buf_size;
1004}
1005
1006static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
1007				  size_t snapshot_buf_size)
1008{
1009	if (ptr->snapshot_init_done)
1010		return 0;
1011
1012	ptr->snapshot_init_done = true;
1013
1014	ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
1015							snapshot_buf_size);
1016
1017	return 0;
1018}
1019
1020/**
1021 * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
1022 * @buf1: first buffer
1023 * @compare_size: number of bytes to compare
1024 * @buf2: second buffer (a circular buffer)
1025 * @offs2: offset in second buffer
1026 * @buf2_size: size of second buffer
1027 *
1028 * The comparison allows for the possibility that the bytes to compare in the
1029 * circular buffer are not contiguous.  It is assumed that @compare_size <=
1030 * @buf2_size.  This function returns %false if the bytes are identical, %true
1031 * otherwise.
1032 */
1033static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
1034				     void *buf2, size_t offs2, size_t buf2_size)
1035{
1036	size_t end2 = offs2 + compare_size, part_size;
1037
1038	if (end2 <= buf2_size)
1039		return memcmp(buf1, buf2 + offs2, compare_size);
1040
1041	part_size = end2 - buf2_size;
1042	if (memcmp(buf1, buf2 + offs2, part_size))
1043		return true;
1044
1045	compare_size -= part_size;
1046
1047	return memcmp(buf1 + part_size, buf2, compare_size);
1048}
1049
1050static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
1051				 size_t ref_size, size_t buf_size,
1052				 void *data, size_t head)
1053{
1054	size_t ref_end = ref_offset + ref_size;
1055
1056	if (ref_end > buf_size) {
1057		if (head > ref_offset || head < ref_end - buf_size)
1058			return true;
1059	} else if (head > ref_offset && head < ref_end) {
1060		return true;
1061	}
1062
1063	return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
1064					buf_size);
1065}
1066
1067static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
1068			      void *data, size_t head)
1069{
1070	if (head >= ref_size) {
1071		memcpy(ref_buf, data + head - ref_size, ref_size);
1072	} else {
1073		memcpy(ref_buf, data, head);
1074		ref_size -= head;
1075		memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
1076	}
1077}
1078
1079static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
1080			     struct auxtrace_mmap *mm, unsigned char *data,
1081			     u64 head)
1082{
1083	struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
1084	bool wrapped;
1085
1086	wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
1087				       ptr->snapshot_ref_buf_size, mm->len,
1088				       data, head);
1089
1090	intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
1091			  data, head);
1092
1093	return wrapped;
1094}
1095
1096static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
1097{
1098	int i, a, b;
1099
1100	b = buf_size >> 3;
1101	a = b - 512;
1102	if (a < 0)
1103		a = 0;
1104
1105	for (i = a; i < b; i++) {
1106		if (data[i])
1107			return true;
1108	}
1109
1110	return false;
1111}
1112
1113static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1114				  struct auxtrace_mmap *mm, unsigned char *data,
1115				  u64 *head, u64 *old)
1116{
1117	struct intel_pt_recording *ptr =
1118			container_of(itr, struct intel_pt_recording, itr);
1119	bool wrapped;
1120	int err;
1121
1122	pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1123		  __func__, idx, (size_t)*old, (size_t)*head);
1124
1125	err = intel_pt_snapshot_init(ptr, mm->len);
1126	if (err)
1127		goto out_err;
1128
1129	if (idx >= ptr->snapshot_ref_cnt) {
1130		err = intel_pt_alloc_snapshot_refs(ptr, idx);
1131		if (err)
1132			goto out_err;
1133	}
1134
1135	if (ptr->snapshot_ref_buf_size) {
1136		if (!ptr->snapshot_refs[idx].ref_buf) {
1137			err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1138			if (err)
1139				goto out_err;
1140		}
1141		wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1142	} else {
1143		wrapped = ptr->snapshot_refs[idx].wrapped;
1144		if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1145			ptr->snapshot_refs[idx].wrapped = true;
1146			wrapped = true;
1147		}
1148	}
1149
1150	/*
1151	 * In full trace mode 'head' continually increases.  However in snapshot
1152	 * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1153	 * are adjusted to match the full trace case which expects that 'old' is
1154	 * always less than 'head'.
1155	 */
1156	if (wrapped) {
1157		*old = *head;
1158		*head += mm->len;
1159	} else {
1160		if (mm->mask)
1161			*old &= mm->mask;
1162		else
1163			*old %= mm->len;
1164		if (*old > *head)
1165			*head += mm->len;
1166	}
1167
1168	pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1169		  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1170
1171	return 0;
1172
1173out_err:
1174	pr_err("%s: failed, error %d\n", __func__, err);
1175	return err;
1176}
1177
1178static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1179{
1180	return rdtsc();
1181}
1182
1183static int intel_pt_perf_config(const char *var, const char *value, void *data)
1184{
1185	struct intel_pt_recording *ptr = data;
1186
1187	if (!strcmp(var, "intel-pt.all-switch-events"))
1188		ptr->all_switch_events = perf_config_bool(var, value);
1189
1190	return 0;
1191}
1192
1193struct auxtrace_record *intel_pt_recording_init(int *err)
1194{
1195	struct perf_pmu *intel_pt_pmu = perf_pmus__find(INTEL_PT_PMU_NAME);
1196	struct intel_pt_recording *ptr;
1197
1198	if (!intel_pt_pmu)
1199		return NULL;
1200
1201	if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1202		*err = -errno;
1203		return NULL;
1204	}
1205
1206	ptr = zalloc(sizeof(struct intel_pt_recording));
1207	if (!ptr) {
1208		*err = -ENOMEM;
1209		return NULL;
1210	}
1211
1212	perf_config(intel_pt_perf_config, ptr);
1213
1214	ptr->intel_pt_pmu = intel_pt_pmu;
1215	ptr->itr.recording_options = intel_pt_recording_options;
1216	ptr->itr.info_priv_size = intel_pt_info_priv_size;
1217	ptr->itr.info_fill = intel_pt_info_fill;
1218	ptr->itr.free = intel_pt_recording_free;
1219	ptr->itr.snapshot_start = intel_pt_snapshot_start;
1220	ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1221	ptr->itr.find_snapshot = intel_pt_find_snapshot;
1222	ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1223	ptr->itr.reference = intel_pt_reference;
1224	ptr->itr.read_finish = auxtrace_record__read_finish;
1225	/*
1226	 * Decoding starts at a PSB packet. Minimum PSB period is 2K so 4K
1227	 * should give at least 1 PSB per sample.
1228	 */
1229	ptr->itr.default_aux_sample_size = 4096;
1230	return &ptr->itr;
1231}