Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.5.6.
   1/*
   2 * intel_pt.c: Intel Processor Trace support
   3 * Copyright (c) 2013-2015, Intel Corporation.
   4 *
   5 * This program is free software; you can redistribute it and/or modify it
   6 * under the terms and conditions of the GNU General Public License,
   7 * version 2, as published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope it will be useful, but WITHOUT
  10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12 * more details.
  13 *
  14 */
  15
  16#include <stdio.h>
  17#include <stdbool.h>
  18#include <errno.h>
  19#include <linux/kernel.h>
  20#include <linux/types.h>
  21
  22#include "../perf.h"
  23#include "session.h"
  24#include "machine.h"
  25#include "sort.h"
  26#include "tool.h"
  27#include "event.h"
  28#include "evlist.h"
  29#include "evsel.h"
  30#include "map.h"
  31#include "color.h"
  32#include "util.h"
  33#include "thread.h"
  34#include "thread-stack.h"
  35#include "symbol.h"
  36#include "callchain.h"
  37#include "dso.h"
  38#include "debug.h"
  39#include "auxtrace.h"
  40#include "tsc.h"
  41#include "intel-pt.h"
  42#include "config.h"
  43
  44#include "intel-pt-decoder/intel-pt-log.h"
  45#include "intel-pt-decoder/intel-pt-decoder.h"
  46#include "intel-pt-decoder/intel-pt-insn-decoder.h"
  47#include "intel-pt-decoder/intel-pt-pkt-decoder.h"
  48
  49#define MAX_TIMESTAMP (~0ULL)
  50
  51struct intel_pt {
  52	struct auxtrace auxtrace;
  53	struct auxtrace_queues queues;
  54	struct auxtrace_heap heap;
  55	u32 auxtrace_type;
  56	struct perf_session *session;
  57	struct machine *machine;
  58	struct perf_evsel *switch_evsel;
  59	struct thread *unknown_thread;
  60	bool timeless_decoding;
  61	bool sampling_mode;
  62	bool snapshot_mode;
  63	bool per_cpu_mmaps;
  64	bool have_tsc;
  65	bool data_queued;
  66	bool est_tsc;
  67	bool sync_switch;
  68	bool mispred_all;
  69	int have_sched_switch;
  70	u32 pmu_type;
  71	u64 kernel_start;
  72	u64 switch_ip;
  73	u64 ptss_ip;
  74
  75	struct perf_tsc_conversion tc;
  76	bool cap_user_time_zero;
  77
  78	struct itrace_synth_opts synth_opts;
  79
  80	bool sample_instructions;
  81	u64 instructions_sample_type;
  82	u64 instructions_sample_period;
  83	u64 instructions_id;
  84
  85	bool sample_branches;
  86	u32 branches_filter;
  87	u64 branches_sample_type;
  88	u64 branches_id;
  89
  90	bool sample_transactions;
  91	u64 transactions_sample_type;
  92	u64 transactions_id;
  93
  94	bool synth_needs_swap;
  95
  96	u64 tsc_bit;
  97	u64 mtc_bit;
  98	u64 mtc_freq_bits;
  99	u32 tsc_ctc_ratio_n;
 100	u32 tsc_ctc_ratio_d;
 101	u64 cyc_bit;
 102	u64 noretcomp_bit;
 103	unsigned max_non_turbo_ratio;
 104
 105	unsigned long num_events;
 106
 107	char *filter;
 108	struct addr_filters filts;
 109};
 110
 111enum switch_state {
 112	INTEL_PT_SS_NOT_TRACING,
 113	INTEL_PT_SS_UNKNOWN,
 114	INTEL_PT_SS_TRACING,
 115	INTEL_PT_SS_EXPECTING_SWITCH_EVENT,
 116	INTEL_PT_SS_EXPECTING_SWITCH_IP,
 117};
 118
 119struct intel_pt_queue {
 120	struct intel_pt *pt;
 121	unsigned int queue_nr;
 122	struct auxtrace_buffer *buffer;
 123	void *decoder;
 124	const struct intel_pt_state *state;
 125	struct ip_callchain *chain;
 126	struct branch_stack *last_branch;
 127	struct branch_stack *last_branch_rb;
 128	size_t last_branch_pos;
 129	union perf_event *event_buf;
 130	bool on_heap;
 131	bool stop;
 132	bool step_through_buffers;
 133	bool use_buffer_pid_tid;
 134	pid_t pid, tid;
 135	int cpu;
 136	int switch_state;
 137	pid_t next_tid;
 138	struct thread *thread;
 139	bool exclude_kernel;
 140	bool have_sample;
 141	u64 time;
 142	u64 timestamp;
 143	u32 flags;
 144	u16 insn_len;
 145	u64 last_insn_cnt;
 146	char insn[INTEL_PT_INSN_BUF_SZ];
 147};
 148
 149static void intel_pt_dump(struct intel_pt *pt __maybe_unused,
 150			  unsigned char *buf, size_t len)
 151{
 152	struct intel_pt_pkt packet;
 153	size_t pos = 0;
 154	int ret, pkt_len, i;
 155	char desc[INTEL_PT_PKT_DESC_MAX];
 156	const char *color = PERF_COLOR_BLUE;
 157
 158	color_fprintf(stdout, color,
 159		      ". ... Intel Processor Trace data: size %zu bytes\n",
 160		      len);
 161
 162	while (len) {
 163		ret = intel_pt_get_packet(buf, len, &packet);
 164		if (ret > 0)
 165			pkt_len = ret;
 166		else
 167			pkt_len = 1;
 168		printf(".");
 169		color_fprintf(stdout, color, "  %08x: ", pos);
 170		for (i = 0; i < pkt_len; i++)
 171			color_fprintf(stdout, color, " %02x", buf[i]);
 172		for (; i < 16; i++)
 173			color_fprintf(stdout, color, "   ");
 174		if (ret > 0) {
 175			ret = intel_pt_pkt_desc(&packet, desc,
 176						INTEL_PT_PKT_DESC_MAX);
 177			if (ret > 0)
 178				color_fprintf(stdout, color, " %s\n", desc);
 179		} else {
 180			color_fprintf(stdout, color, " Bad packet!\n");
 181		}
 182		pos += pkt_len;
 183		buf += pkt_len;
 184		len -= pkt_len;
 185	}
 186}
 187
 188static void intel_pt_dump_event(struct intel_pt *pt, unsigned char *buf,
 189				size_t len)
 190{
 191	printf(".\n");
 192	intel_pt_dump(pt, buf, len);
 193}
 194
 195static int intel_pt_do_fix_overlap(struct intel_pt *pt, struct auxtrace_buffer *a,
 196				   struct auxtrace_buffer *b)
 197{
 198	void *start;
 199
 200	start = intel_pt_find_overlap(a->data, a->size, b->data, b->size,
 201				      pt->have_tsc);
 202	if (!start)
 203		return -EINVAL;
 204	b->use_size = b->data + b->size - start;
 205	b->use_data = start;
 206	return 0;
 207}
 208
 209static void intel_pt_use_buffer_pid_tid(struct intel_pt_queue *ptq,
 210					struct auxtrace_queue *queue,
 211					struct auxtrace_buffer *buffer)
 212{
 213	if (queue->cpu == -1 && buffer->cpu != -1)
 214		ptq->cpu = buffer->cpu;
 215
 216	ptq->pid = buffer->pid;
 217	ptq->tid = buffer->tid;
 218
 219	intel_pt_log("queue %u cpu %d pid %d tid %d\n",
 220		     ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
 221
 222	thread__zput(ptq->thread);
 223
 224	if (ptq->tid != -1) {
 225		if (ptq->pid != -1)
 226			ptq->thread = machine__findnew_thread(ptq->pt->machine,
 227							      ptq->pid,
 228							      ptq->tid);
 229		else
 230			ptq->thread = machine__find_thread(ptq->pt->machine, -1,
 231							   ptq->tid);
 232	}
 233}
 234
 235/* This function assumes data is processed sequentially only */
 236static int intel_pt_get_trace(struct intel_pt_buffer *b, void *data)
 237{
 238	struct intel_pt_queue *ptq = data;
 239	struct auxtrace_buffer *buffer = ptq->buffer, *old_buffer = buffer;
 240	struct auxtrace_queue *queue;
 241
 242	if (ptq->stop) {
 243		b->len = 0;
 244		return 0;
 245	}
 246
 247	queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
 248next:
 249	buffer = auxtrace_buffer__next(queue, buffer);
 250	if (!buffer) {
 251		if (old_buffer)
 252			auxtrace_buffer__drop_data(old_buffer);
 253		b->len = 0;
 254		return 0;
 255	}
 256
 257	ptq->buffer = buffer;
 258
 259	if (!buffer->data) {
 260		int fd = perf_data_file__fd(ptq->pt->session->file);
 261
 262		buffer->data = auxtrace_buffer__get_data(buffer, fd);
 263		if (!buffer->data)
 264			return -ENOMEM;
 265	}
 266
 267	if (ptq->pt->snapshot_mode && !buffer->consecutive && old_buffer &&
 268	    intel_pt_do_fix_overlap(ptq->pt, old_buffer, buffer))
 269		return -ENOMEM;
 270
 271	if (buffer->use_data) {
 272		b->len = buffer->use_size;
 273		b->buf = buffer->use_data;
 274	} else {
 275		b->len = buffer->size;
 276		b->buf = buffer->data;
 277	}
 278	b->ref_timestamp = buffer->reference;
 279
 280	/*
 281	 * If in snapshot mode and the buffer has no usable data, get next
 282	 * buffer and again check overlap against old_buffer.
 283	 */
 284	if (ptq->pt->snapshot_mode && !b->len)
 285		goto next;
 286
 287	if (old_buffer)
 288		auxtrace_buffer__drop_data(old_buffer);
 289
 290	if (!old_buffer || ptq->pt->sampling_mode || (ptq->pt->snapshot_mode &&
 291						      !buffer->consecutive)) {
 292		b->consecutive = false;
 293		b->trace_nr = buffer->buffer_nr + 1;
 294	} else {
 295		b->consecutive = true;
 296	}
 297
 298	if (ptq->use_buffer_pid_tid && (ptq->pid != buffer->pid ||
 299					ptq->tid != buffer->tid))
 300		intel_pt_use_buffer_pid_tid(ptq, queue, buffer);
 301
 302	if (ptq->step_through_buffers)
 303		ptq->stop = true;
 304
 305	if (!b->len)
 306		return intel_pt_get_trace(b, data);
 307
 308	return 0;
 309}
 310
 311struct intel_pt_cache_entry {
 312	struct auxtrace_cache_entry	entry;
 313	u64				insn_cnt;
 314	u64				byte_cnt;
 315	enum intel_pt_insn_op		op;
 316	enum intel_pt_insn_branch	branch;
 317	int				length;
 318	int32_t				rel;
 319	char				insn[INTEL_PT_INSN_BUF_SZ];
 320};
 321
 322static int intel_pt_config_div(const char *var, const char *value, void *data)
 323{
 324	int *d = data;
 325	long val;
 326
 327	if (!strcmp(var, "intel-pt.cache-divisor")) {
 328		val = strtol(value, NULL, 0);
 329		if (val > 0 && val <= INT_MAX)
 330			*d = val;
 331	}
 332
 333	return 0;
 334}
 335
 336static int intel_pt_cache_divisor(void)
 337{
 338	static int d;
 339
 340	if (d)
 341		return d;
 342
 343	perf_config(intel_pt_config_div, &d);
 344
 345	if (!d)
 346		d = 64;
 347
 348	return d;
 349}
 350
 351static unsigned int intel_pt_cache_size(struct dso *dso,
 352					struct machine *machine)
 353{
 354	off_t size;
 355
 356	size = dso__data_size(dso, machine);
 357	size /= intel_pt_cache_divisor();
 358	if (size < 1000)
 359		return 10;
 360	if (size > (1 << 21))
 361		return 21;
 362	return 32 - __builtin_clz(size);
 363}
 364
 365static struct auxtrace_cache *intel_pt_cache(struct dso *dso,
 366					     struct machine *machine)
 367{
 368	struct auxtrace_cache *c;
 369	unsigned int bits;
 370
 371	if (dso->auxtrace_cache)
 372		return dso->auxtrace_cache;
 373
 374	bits = intel_pt_cache_size(dso, machine);
 375
 376	/* Ignoring cache creation failure */
 377	c = auxtrace_cache__new(bits, sizeof(struct intel_pt_cache_entry), 200);
 378
 379	dso->auxtrace_cache = c;
 380
 381	return c;
 382}
 383
 384static int intel_pt_cache_add(struct dso *dso, struct machine *machine,
 385			      u64 offset, u64 insn_cnt, u64 byte_cnt,
 386			      struct intel_pt_insn *intel_pt_insn)
 387{
 388	struct auxtrace_cache *c = intel_pt_cache(dso, machine);
 389	struct intel_pt_cache_entry *e;
 390	int err;
 391
 392	if (!c)
 393		return -ENOMEM;
 394
 395	e = auxtrace_cache__alloc_entry(c);
 396	if (!e)
 397		return -ENOMEM;
 398
 399	e->insn_cnt = insn_cnt;
 400	e->byte_cnt = byte_cnt;
 401	e->op = intel_pt_insn->op;
 402	e->branch = intel_pt_insn->branch;
 403	e->length = intel_pt_insn->length;
 404	e->rel = intel_pt_insn->rel;
 405	memcpy(e->insn, intel_pt_insn->buf, INTEL_PT_INSN_BUF_SZ);
 406
 407	err = auxtrace_cache__add(c, offset, &e->entry);
 408	if (err)
 409		auxtrace_cache__free_entry(c, e);
 410
 411	return err;
 412}
 413
 414static struct intel_pt_cache_entry *
 415intel_pt_cache_lookup(struct dso *dso, struct machine *machine, u64 offset)
 416{
 417	struct auxtrace_cache *c = intel_pt_cache(dso, machine);
 418
 419	if (!c)
 420		return NULL;
 421
 422	return auxtrace_cache__lookup(dso->auxtrace_cache, offset);
 423}
 424
 425static int intel_pt_walk_next_insn(struct intel_pt_insn *intel_pt_insn,
 426				   uint64_t *insn_cnt_ptr, uint64_t *ip,
 427				   uint64_t to_ip, uint64_t max_insn_cnt,
 428				   void *data)
 429{
 430	struct intel_pt_queue *ptq = data;
 431	struct machine *machine = ptq->pt->machine;
 432	struct thread *thread;
 433	struct addr_location al;
 434	unsigned char buf[INTEL_PT_INSN_BUF_SZ];
 435	ssize_t len;
 436	int x86_64;
 437	u8 cpumode;
 438	u64 offset, start_offset, start_ip;
 439	u64 insn_cnt = 0;
 440	bool one_map = true;
 441
 442	intel_pt_insn->length = 0;
 443
 444	if (to_ip && *ip == to_ip)
 445		goto out_no_cache;
 446
 447	if (*ip >= ptq->pt->kernel_start)
 448		cpumode = PERF_RECORD_MISC_KERNEL;
 449	else
 450		cpumode = PERF_RECORD_MISC_USER;
 451
 452	thread = ptq->thread;
 453	if (!thread) {
 454		if (cpumode != PERF_RECORD_MISC_KERNEL)
 455			return -EINVAL;
 456		thread = ptq->pt->unknown_thread;
 457	}
 458
 459	while (1) {
 460		thread__find_addr_map(thread, cpumode, MAP__FUNCTION, *ip, &al);
 461		if (!al.map || !al.map->dso)
 462			return -EINVAL;
 463
 464		if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
 465		    dso__data_status_seen(al.map->dso,
 466					  DSO_DATA_STATUS_SEEN_ITRACE))
 467			return -ENOENT;
 468
 469		offset = al.map->map_ip(al.map, *ip);
 470
 471		if (!to_ip && one_map) {
 472			struct intel_pt_cache_entry *e;
 473
 474			e = intel_pt_cache_lookup(al.map->dso, machine, offset);
 475			if (e &&
 476			    (!max_insn_cnt || e->insn_cnt <= max_insn_cnt)) {
 477				*insn_cnt_ptr = e->insn_cnt;
 478				*ip += e->byte_cnt;
 479				intel_pt_insn->op = e->op;
 480				intel_pt_insn->branch = e->branch;
 481				intel_pt_insn->length = e->length;
 482				intel_pt_insn->rel = e->rel;
 483				memcpy(intel_pt_insn->buf, e->insn,
 484				       INTEL_PT_INSN_BUF_SZ);
 485				intel_pt_log_insn_no_data(intel_pt_insn, *ip);
 486				return 0;
 487			}
 488		}
 489
 490		start_offset = offset;
 491		start_ip = *ip;
 492
 493		/* Load maps to ensure dso->is_64_bit has been updated */
 494		map__load(al.map);
 495
 496		x86_64 = al.map->dso->is_64_bit;
 497
 498		while (1) {
 499			len = dso__data_read_offset(al.map->dso, machine,
 500						    offset, buf,
 501						    INTEL_PT_INSN_BUF_SZ);
 502			if (len <= 0)
 503				return -EINVAL;
 504
 505			if (intel_pt_get_insn(buf, len, x86_64, intel_pt_insn))
 506				return -EINVAL;
 507
 508			intel_pt_log_insn(intel_pt_insn, *ip);
 509
 510			insn_cnt += 1;
 511
 512			if (intel_pt_insn->branch != INTEL_PT_BR_NO_BRANCH)
 513				goto out;
 514
 515			if (max_insn_cnt && insn_cnt >= max_insn_cnt)
 516				goto out_no_cache;
 517
 518			*ip += intel_pt_insn->length;
 519
 520			if (to_ip && *ip == to_ip)
 521				goto out_no_cache;
 522
 523			if (*ip >= al.map->end)
 524				break;
 525
 526			offset += intel_pt_insn->length;
 527		}
 528		one_map = false;
 529	}
 530out:
 531	*insn_cnt_ptr = insn_cnt;
 532
 533	if (!one_map)
 534		goto out_no_cache;
 535
 536	/*
 537	 * Didn't lookup in the 'to_ip' case, so do it now to prevent duplicate
 538	 * entries.
 539	 */
 540	if (to_ip) {
 541		struct intel_pt_cache_entry *e;
 542
 543		e = intel_pt_cache_lookup(al.map->dso, machine, start_offset);
 544		if (e)
 545			return 0;
 546	}
 547
 548	/* Ignore cache errors */
 549	intel_pt_cache_add(al.map->dso, machine, start_offset, insn_cnt,
 550			   *ip - start_ip, intel_pt_insn);
 551
 552	return 0;
 553
 554out_no_cache:
 555	*insn_cnt_ptr = insn_cnt;
 556	return 0;
 557}
 558
 559static bool intel_pt_match_pgd_ip(struct intel_pt *pt, uint64_t ip,
 560				  uint64_t offset, const char *filename)
 561{
 562	struct addr_filter *filt;
 563	bool have_filter   = false;
 564	bool hit_tracestop = false;
 565	bool hit_filter    = false;
 566
 567	list_for_each_entry(filt, &pt->filts.head, list) {
 568		if (filt->start)
 569			have_filter = true;
 570
 571		if ((filename && !filt->filename) ||
 572		    (!filename && filt->filename) ||
 573		    (filename && strcmp(filename, filt->filename)))
 574			continue;
 575
 576		if (!(offset >= filt->addr && offset < filt->addr + filt->size))
 577			continue;
 578
 579		intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s hit filter: %s offset %#"PRIx64" size %#"PRIx64"\n",
 580			     ip, offset, filename ? filename : "[kernel]",
 581			     filt->start ? "filter" : "stop",
 582			     filt->addr, filt->size);
 583
 584		if (filt->start)
 585			hit_filter = true;
 586		else
 587			hit_tracestop = true;
 588	}
 589
 590	if (!hit_tracestop && !hit_filter)
 591		intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s is not in a filter region\n",
 592			     ip, offset, filename ? filename : "[kernel]");
 593
 594	return hit_tracestop || (have_filter && !hit_filter);
 595}
 596
 597static int __intel_pt_pgd_ip(uint64_t ip, void *data)
 598{
 599	struct intel_pt_queue *ptq = data;
 600	struct thread *thread;
 601	struct addr_location al;
 602	u8 cpumode;
 603	u64 offset;
 604
 605	if (ip >= ptq->pt->kernel_start)
 606		return intel_pt_match_pgd_ip(ptq->pt, ip, ip, NULL);
 607
 608	cpumode = PERF_RECORD_MISC_USER;
 609
 610	thread = ptq->thread;
 611	if (!thread)
 612		return -EINVAL;
 613
 614	thread__find_addr_map(thread, cpumode, MAP__FUNCTION, ip, &al);
 615	if (!al.map || !al.map->dso)
 616		return -EINVAL;
 617
 618	offset = al.map->map_ip(al.map, ip);
 619
 620	return intel_pt_match_pgd_ip(ptq->pt, ip, offset,
 621				     al.map->dso->long_name);
 622}
 623
 624static bool intel_pt_pgd_ip(uint64_t ip, void *data)
 625{
 626	return __intel_pt_pgd_ip(ip, data) > 0;
 627}
 628
 629static bool intel_pt_get_config(struct intel_pt *pt,
 630				struct perf_event_attr *attr, u64 *config)
 631{
 632	if (attr->type == pt->pmu_type) {
 633		if (config)
 634			*config = attr->config;
 635		return true;
 636	}
 637
 638	return false;
 639}
 640
 641static bool intel_pt_exclude_kernel(struct intel_pt *pt)
 642{
 643	struct perf_evsel *evsel;
 644
 645	evlist__for_each_entry(pt->session->evlist, evsel) {
 646		if (intel_pt_get_config(pt, &evsel->attr, NULL) &&
 647		    !evsel->attr.exclude_kernel)
 648			return false;
 649	}
 650	return true;
 651}
 652
 653static bool intel_pt_return_compression(struct intel_pt *pt)
 654{
 655	struct perf_evsel *evsel;
 656	u64 config;
 657
 658	if (!pt->noretcomp_bit)
 659		return true;
 660
 661	evlist__for_each_entry(pt->session->evlist, evsel) {
 662		if (intel_pt_get_config(pt, &evsel->attr, &config) &&
 663		    (config & pt->noretcomp_bit))
 664			return false;
 665	}
 666	return true;
 667}
 668
 669static unsigned int intel_pt_mtc_period(struct intel_pt *pt)
 670{
 671	struct perf_evsel *evsel;
 672	unsigned int shift;
 673	u64 config;
 674
 675	if (!pt->mtc_freq_bits)
 676		return 0;
 677
 678	for (shift = 0, config = pt->mtc_freq_bits; !(config & 1); shift++)
 679		config >>= 1;
 680
 681	evlist__for_each_entry(pt->session->evlist, evsel) {
 682		if (intel_pt_get_config(pt, &evsel->attr, &config))
 683			return (config & pt->mtc_freq_bits) >> shift;
 684	}
 685	return 0;
 686}
 687
 688static bool intel_pt_timeless_decoding(struct intel_pt *pt)
 689{
 690	struct perf_evsel *evsel;
 691	bool timeless_decoding = true;
 692	u64 config;
 693
 694	if (!pt->tsc_bit || !pt->cap_user_time_zero)
 695		return true;
 696
 697	evlist__for_each_entry(pt->session->evlist, evsel) {
 698		if (!(evsel->attr.sample_type & PERF_SAMPLE_TIME))
 699			return true;
 700		if (intel_pt_get_config(pt, &evsel->attr, &config)) {
 701			if (config & pt->tsc_bit)
 702				timeless_decoding = false;
 703			else
 704				return true;
 705		}
 706	}
 707	return timeless_decoding;
 708}
 709
 710static bool intel_pt_tracing_kernel(struct intel_pt *pt)
 711{
 712	struct perf_evsel *evsel;
 713
 714	evlist__for_each_entry(pt->session->evlist, evsel) {
 715		if (intel_pt_get_config(pt, &evsel->attr, NULL) &&
 716		    !evsel->attr.exclude_kernel)
 717			return true;
 718	}
 719	return false;
 720}
 721
 722static bool intel_pt_have_tsc(struct intel_pt *pt)
 723{
 724	struct perf_evsel *evsel;
 725	bool have_tsc = false;
 726	u64 config;
 727
 728	if (!pt->tsc_bit)
 729		return false;
 730
 731	evlist__for_each_entry(pt->session->evlist, evsel) {
 732		if (intel_pt_get_config(pt, &evsel->attr, &config)) {
 733			if (config & pt->tsc_bit)
 734				have_tsc = true;
 735			else
 736				return false;
 737		}
 738	}
 739	return have_tsc;
 740}
 741
 742static u64 intel_pt_ns_to_ticks(const struct intel_pt *pt, u64 ns)
 743{
 744	u64 quot, rem;
 745
 746	quot = ns / pt->tc.time_mult;
 747	rem  = ns % pt->tc.time_mult;
 748	return (quot << pt->tc.time_shift) + (rem << pt->tc.time_shift) /
 749		pt->tc.time_mult;
 750}
 751
 752static struct intel_pt_queue *intel_pt_alloc_queue(struct intel_pt *pt,
 753						   unsigned int queue_nr)
 754{
 755	struct intel_pt_params params = { .get_trace = 0, };
 756	struct intel_pt_queue *ptq;
 757
 758	ptq = zalloc(sizeof(struct intel_pt_queue));
 759	if (!ptq)
 760		return NULL;
 761
 762	if (pt->synth_opts.callchain) {
 763		size_t sz = sizeof(struct ip_callchain);
 764
 765		sz += pt->synth_opts.callchain_sz * sizeof(u64);
 766		ptq->chain = zalloc(sz);
 767		if (!ptq->chain)
 768			goto out_free;
 769	}
 770
 771	if (pt->synth_opts.last_branch) {
 772		size_t sz = sizeof(struct branch_stack);
 773
 774		sz += pt->synth_opts.last_branch_sz *
 775		      sizeof(struct branch_entry);
 776		ptq->last_branch = zalloc(sz);
 777		if (!ptq->last_branch)
 778			goto out_free;
 779		ptq->last_branch_rb = zalloc(sz);
 780		if (!ptq->last_branch_rb)
 781			goto out_free;
 782	}
 783
 784	ptq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
 785	if (!ptq->event_buf)
 786		goto out_free;
 787
 788	ptq->pt = pt;
 789	ptq->queue_nr = queue_nr;
 790	ptq->exclude_kernel = intel_pt_exclude_kernel(pt);
 791	ptq->pid = -1;
 792	ptq->tid = -1;
 793	ptq->cpu = -1;
 794	ptq->next_tid = -1;
 795
 796	params.get_trace = intel_pt_get_trace;
 797	params.walk_insn = intel_pt_walk_next_insn;
 798	params.data = ptq;
 799	params.return_compression = intel_pt_return_compression(pt);
 800	params.max_non_turbo_ratio = pt->max_non_turbo_ratio;
 801	params.mtc_period = intel_pt_mtc_period(pt);
 802	params.tsc_ctc_ratio_n = pt->tsc_ctc_ratio_n;
 803	params.tsc_ctc_ratio_d = pt->tsc_ctc_ratio_d;
 804
 805	if (pt->filts.cnt > 0)
 806		params.pgd_ip = intel_pt_pgd_ip;
 807
 808	if (pt->synth_opts.instructions) {
 809		if (pt->synth_opts.period) {
 810			switch (pt->synth_opts.period_type) {
 811			case PERF_ITRACE_PERIOD_INSTRUCTIONS:
 812				params.period_type =
 813						INTEL_PT_PERIOD_INSTRUCTIONS;
 814				params.period = pt->synth_opts.period;
 815				break;
 816			case PERF_ITRACE_PERIOD_TICKS:
 817				params.period_type = INTEL_PT_PERIOD_TICKS;
 818				params.period = pt->synth_opts.period;
 819				break;
 820			case PERF_ITRACE_PERIOD_NANOSECS:
 821				params.period_type = INTEL_PT_PERIOD_TICKS;
 822				params.period = intel_pt_ns_to_ticks(pt,
 823							pt->synth_opts.period);
 824				break;
 825			default:
 826				break;
 827			}
 828		}
 829
 830		if (!params.period) {
 831			params.period_type = INTEL_PT_PERIOD_INSTRUCTIONS;
 832			params.period = 1;
 833		}
 834	}
 835
 836	ptq->decoder = intel_pt_decoder_new(&params);
 837	if (!ptq->decoder)
 838		goto out_free;
 839
 840	return ptq;
 841
 842out_free:
 843	zfree(&ptq->event_buf);
 844	zfree(&ptq->last_branch);
 845	zfree(&ptq->last_branch_rb);
 846	zfree(&ptq->chain);
 847	free(ptq);
 848	return NULL;
 849}
 850
 851static void intel_pt_free_queue(void *priv)
 852{
 853	struct intel_pt_queue *ptq = priv;
 854
 855	if (!ptq)
 856		return;
 857	thread__zput(ptq->thread);
 858	intel_pt_decoder_free(ptq->decoder);
 859	zfree(&ptq->event_buf);
 860	zfree(&ptq->last_branch);
 861	zfree(&ptq->last_branch_rb);
 862	zfree(&ptq->chain);
 863	free(ptq);
 864}
 865
 866static void intel_pt_set_pid_tid_cpu(struct intel_pt *pt,
 867				     struct auxtrace_queue *queue)
 868{
 869	struct intel_pt_queue *ptq = queue->priv;
 870
 871	if (queue->tid == -1 || pt->have_sched_switch) {
 872		ptq->tid = machine__get_current_tid(pt->machine, ptq->cpu);
 873		thread__zput(ptq->thread);
 874	}
 875
 876	if (!ptq->thread && ptq->tid != -1)
 877		ptq->thread = machine__find_thread(pt->machine, -1, ptq->tid);
 878
 879	if (ptq->thread) {
 880		ptq->pid = ptq->thread->pid_;
 881		if (queue->cpu == -1)
 882			ptq->cpu = ptq->thread->cpu;
 883	}
 884}
 885
 886static void intel_pt_sample_flags(struct intel_pt_queue *ptq)
 887{
 888	if (ptq->state->flags & INTEL_PT_ABORT_TX) {
 889		ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_TX_ABORT;
 890	} else if (ptq->state->flags & INTEL_PT_ASYNC) {
 891		if (ptq->state->to_ip)
 892			ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_CALL |
 893				     PERF_IP_FLAG_ASYNC |
 894				     PERF_IP_FLAG_INTERRUPT;
 895		else
 896			ptq->flags = PERF_IP_FLAG_BRANCH |
 897				     PERF_IP_FLAG_TRACE_END;
 898		ptq->insn_len = 0;
 899	} else {
 900		if (ptq->state->from_ip)
 901			ptq->flags = intel_pt_insn_type(ptq->state->insn_op);
 902		else
 903			ptq->flags = PERF_IP_FLAG_BRANCH |
 904				     PERF_IP_FLAG_TRACE_BEGIN;
 905		if (ptq->state->flags & INTEL_PT_IN_TX)
 906			ptq->flags |= PERF_IP_FLAG_IN_TX;
 907		ptq->insn_len = ptq->state->insn_len;
 908		memcpy(ptq->insn, ptq->state->insn, INTEL_PT_INSN_BUF_SZ);
 909	}
 910}
 911
 912static int intel_pt_setup_queue(struct intel_pt *pt,
 913				struct auxtrace_queue *queue,
 914				unsigned int queue_nr)
 915{
 916	struct intel_pt_queue *ptq = queue->priv;
 917
 918	if (list_empty(&queue->head))
 919		return 0;
 920
 921	if (!ptq) {
 922		ptq = intel_pt_alloc_queue(pt, queue_nr);
 923		if (!ptq)
 924			return -ENOMEM;
 925		queue->priv = ptq;
 926
 927		if (queue->cpu != -1)
 928			ptq->cpu = queue->cpu;
 929		ptq->tid = queue->tid;
 930
 931		if (pt->sampling_mode) {
 932			if (pt->timeless_decoding)
 933				ptq->step_through_buffers = true;
 934			if (pt->timeless_decoding || !pt->have_sched_switch)
 935				ptq->use_buffer_pid_tid = true;
 936		}
 937	}
 938
 939	if (!ptq->on_heap &&
 940	    (!pt->sync_switch ||
 941	     ptq->switch_state != INTEL_PT_SS_EXPECTING_SWITCH_EVENT)) {
 942		const struct intel_pt_state *state;
 943		int ret;
 944
 945		if (pt->timeless_decoding)
 946			return 0;
 947
 948		intel_pt_log("queue %u getting timestamp\n", queue_nr);
 949		intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
 950			     queue_nr, ptq->cpu, ptq->pid, ptq->tid);
 951		while (1) {
 952			state = intel_pt_decode(ptq->decoder);
 953			if (state->err) {
 954				if (state->err == INTEL_PT_ERR_NODATA) {
 955					intel_pt_log("queue %u has no timestamp\n",
 956						     queue_nr);
 957					return 0;
 958				}
 959				continue;
 960			}
 961			if (state->timestamp)
 962				break;
 963		}
 964
 965		ptq->timestamp = state->timestamp;
 966		intel_pt_log("queue %u timestamp 0x%" PRIx64 "\n",
 967			     queue_nr, ptq->timestamp);
 968		ptq->state = state;
 969		ptq->have_sample = true;
 970		intel_pt_sample_flags(ptq);
 971		ret = auxtrace_heap__add(&pt->heap, queue_nr, ptq->timestamp);
 972		if (ret)
 973			return ret;
 974		ptq->on_heap = true;
 975	}
 976
 977	return 0;
 978}
 979
 980static int intel_pt_setup_queues(struct intel_pt *pt)
 981{
 982	unsigned int i;
 983	int ret;
 984
 985	for (i = 0; i < pt->queues.nr_queues; i++) {
 986		ret = intel_pt_setup_queue(pt, &pt->queues.queue_array[i], i);
 987		if (ret)
 988			return ret;
 989	}
 990	return 0;
 991}
 992
 993static inline void intel_pt_copy_last_branch_rb(struct intel_pt_queue *ptq)
 994{
 995	struct branch_stack *bs_src = ptq->last_branch_rb;
 996	struct branch_stack *bs_dst = ptq->last_branch;
 997	size_t nr = 0;
 998
 999	bs_dst->nr = bs_src->nr;
1000
1001	if (!bs_src->nr)
1002		return;
1003
1004	nr = ptq->pt->synth_opts.last_branch_sz - ptq->last_branch_pos;
1005	memcpy(&bs_dst->entries[0],
1006	       &bs_src->entries[ptq->last_branch_pos],
1007	       sizeof(struct branch_entry) * nr);
1008
1009	if (bs_src->nr >= ptq->pt->synth_opts.last_branch_sz) {
1010		memcpy(&bs_dst->entries[nr],
1011		       &bs_src->entries[0],
1012		       sizeof(struct branch_entry) * ptq->last_branch_pos);
1013	}
1014}
1015
1016static inline void intel_pt_reset_last_branch_rb(struct intel_pt_queue *ptq)
1017{
1018	ptq->last_branch_pos = 0;
1019	ptq->last_branch_rb->nr = 0;
1020}
1021
1022static void intel_pt_update_last_branch_rb(struct intel_pt_queue *ptq)
1023{
1024	const struct intel_pt_state *state = ptq->state;
1025	struct branch_stack *bs = ptq->last_branch_rb;
1026	struct branch_entry *be;
1027
1028	if (!ptq->last_branch_pos)
1029		ptq->last_branch_pos = ptq->pt->synth_opts.last_branch_sz;
1030
1031	ptq->last_branch_pos -= 1;
1032
1033	be              = &bs->entries[ptq->last_branch_pos];
1034	be->from        = state->from_ip;
1035	be->to          = state->to_ip;
1036	be->flags.abort = !!(state->flags & INTEL_PT_ABORT_TX);
1037	be->flags.in_tx = !!(state->flags & INTEL_PT_IN_TX);
1038	/* No support for mispredict */
1039	be->flags.mispred = ptq->pt->mispred_all;
1040
1041	if (bs->nr < ptq->pt->synth_opts.last_branch_sz)
1042		bs->nr += 1;
1043}
1044
1045static int intel_pt_inject_event(union perf_event *event,
1046				 struct perf_sample *sample, u64 type,
1047				 bool swapped)
1048{
1049	event->header.size = perf_event__sample_event_size(sample, type, 0);
1050	return perf_event__synthesize_sample(event, type, 0, sample, swapped);
1051}
1052
1053static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq)
1054{
1055	int ret;
1056	struct intel_pt *pt = ptq->pt;
1057	union perf_event *event = ptq->event_buf;
1058	struct perf_sample sample = { .ip = 0, };
1059	struct dummy_branch_stack {
1060		u64			nr;
1061		struct branch_entry	entries;
1062	} dummy_bs;
1063
1064	if (pt->branches_filter && !(pt->branches_filter & ptq->flags))
1065		return 0;
1066
1067	if (pt->synth_opts.initial_skip &&
1068	    pt->num_events++ < pt->synth_opts.initial_skip)
1069		return 0;
1070
1071	event->sample.header.type = PERF_RECORD_SAMPLE;
1072	event->sample.header.misc = PERF_RECORD_MISC_USER;
1073	event->sample.header.size = sizeof(struct perf_event_header);
1074
1075	if (!pt->timeless_decoding)
1076		sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1077
1078	sample.cpumode = PERF_RECORD_MISC_USER;
1079	sample.ip = ptq->state->from_ip;
1080	sample.pid = ptq->pid;
1081	sample.tid = ptq->tid;
1082	sample.addr = ptq->state->to_ip;
1083	sample.id = ptq->pt->branches_id;
1084	sample.stream_id = ptq->pt->branches_id;
1085	sample.period = 1;
1086	sample.cpu = ptq->cpu;
1087	sample.flags = ptq->flags;
1088	sample.insn_len = ptq->insn_len;
1089	memcpy(sample.insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1090
1091	/*
1092	 * perf report cannot handle events without a branch stack when using
1093	 * SORT_MODE__BRANCH so make a dummy one.
1094	 */
1095	if (pt->synth_opts.last_branch && sort__mode == SORT_MODE__BRANCH) {
1096		dummy_bs = (struct dummy_branch_stack){
1097			.nr = 1,
1098			.entries = {
1099				.from = sample.ip,
1100				.to = sample.addr,
1101			},
1102		};
1103		sample.branch_stack = (struct branch_stack *)&dummy_bs;
1104	}
1105
1106	if (pt->synth_opts.inject) {
1107		ret = intel_pt_inject_event(event, &sample,
1108					    pt->branches_sample_type,
1109					    pt->synth_needs_swap);
1110		if (ret)
1111			return ret;
1112	}
1113
1114	ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1115	if (ret)
1116		pr_err("Intel Processor Trace: failed to deliver branch event, error %d\n",
1117		       ret);
1118
1119	return ret;
1120}
1121
1122static int intel_pt_synth_instruction_sample(struct intel_pt_queue *ptq)
1123{
1124	int ret;
1125	struct intel_pt *pt = ptq->pt;
1126	union perf_event *event = ptq->event_buf;
1127	struct perf_sample sample = { .ip = 0, };
1128
1129	if (pt->synth_opts.initial_skip &&
1130	    pt->num_events++ < pt->synth_opts.initial_skip)
1131		return 0;
1132
1133	event->sample.header.type = PERF_RECORD_SAMPLE;
1134	event->sample.header.misc = PERF_RECORD_MISC_USER;
1135	event->sample.header.size = sizeof(struct perf_event_header);
1136
1137	if (!pt->timeless_decoding)
1138		sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1139
1140	sample.cpumode = PERF_RECORD_MISC_USER;
1141	sample.ip = ptq->state->from_ip;
1142	sample.pid = ptq->pid;
1143	sample.tid = ptq->tid;
1144	sample.addr = ptq->state->to_ip;
1145	sample.id = ptq->pt->instructions_id;
1146	sample.stream_id = ptq->pt->instructions_id;
1147	sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt;
1148	sample.cpu = ptq->cpu;
1149	sample.flags = ptq->flags;
1150	sample.insn_len = ptq->insn_len;
1151	memcpy(sample.insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1152
1153	ptq->last_insn_cnt = ptq->state->tot_insn_cnt;
1154
1155	if (pt->synth_opts.callchain) {
1156		thread_stack__sample(ptq->thread, ptq->chain,
1157				     pt->synth_opts.callchain_sz, sample.ip);
1158		sample.callchain = ptq->chain;
1159	}
1160
1161	if (pt->synth_opts.last_branch) {
1162		intel_pt_copy_last_branch_rb(ptq);
1163		sample.branch_stack = ptq->last_branch;
1164	}
1165
1166	if (pt->synth_opts.inject) {
1167		ret = intel_pt_inject_event(event, &sample,
1168					    pt->instructions_sample_type,
1169					    pt->synth_needs_swap);
1170		if (ret)
1171			return ret;
1172	}
1173
1174	ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1175	if (ret)
1176		pr_err("Intel Processor Trace: failed to deliver instruction event, error %d\n",
1177		       ret);
1178
1179	if (pt->synth_opts.last_branch)
1180		intel_pt_reset_last_branch_rb(ptq);
1181
1182	return ret;
1183}
1184
1185static int intel_pt_synth_transaction_sample(struct intel_pt_queue *ptq)
1186{
1187	int ret;
1188	struct intel_pt *pt = ptq->pt;
1189	union perf_event *event = ptq->event_buf;
1190	struct perf_sample sample = { .ip = 0, };
1191
1192	if (pt->synth_opts.initial_skip &&
1193	    pt->num_events++ < pt->synth_opts.initial_skip)
1194		return 0;
1195
1196	event->sample.header.type = PERF_RECORD_SAMPLE;
1197	event->sample.header.misc = PERF_RECORD_MISC_USER;
1198	event->sample.header.size = sizeof(struct perf_event_header);
1199
1200	if (!pt->timeless_decoding)
1201		sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1202
1203	sample.cpumode = PERF_RECORD_MISC_USER;
1204	sample.ip = ptq->state->from_ip;
1205	sample.pid = ptq->pid;
1206	sample.tid = ptq->tid;
1207	sample.addr = ptq->state->to_ip;
1208	sample.id = ptq->pt->transactions_id;
1209	sample.stream_id = ptq->pt->transactions_id;
1210	sample.period = 1;
1211	sample.cpu = ptq->cpu;
1212	sample.flags = ptq->flags;
1213	sample.insn_len = ptq->insn_len;
1214	memcpy(sample.insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1215
1216	if (pt->synth_opts.callchain) {
1217		thread_stack__sample(ptq->thread, ptq->chain,
1218				     pt->synth_opts.callchain_sz, sample.ip);
1219		sample.callchain = ptq->chain;
1220	}
1221
1222	if (pt->synth_opts.last_branch) {
1223		intel_pt_copy_last_branch_rb(ptq);
1224		sample.branch_stack = ptq->last_branch;
1225	}
1226
1227	if (pt->synth_opts.inject) {
1228		ret = intel_pt_inject_event(event, &sample,
1229					    pt->transactions_sample_type,
1230					    pt->synth_needs_swap);
1231		if (ret)
1232			return ret;
1233	}
1234
1235	ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1236	if (ret)
1237		pr_err("Intel Processor Trace: failed to deliver transaction event, error %d\n",
1238		       ret);
1239
1240	if (pt->synth_opts.last_branch)
1241		intel_pt_reset_last_branch_rb(ptq);
1242
1243	return ret;
1244}
1245
1246static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,
1247				pid_t pid, pid_t tid, u64 ip)
1248{
1249	union perf_event event;
1250	char msg[MAX_AUXTRACE_ERROR_MSG];
1251	int err;
1252
1253	intel_pt__strerror(code, msg, MAX_AUXTRACE_ERROR_MSG);
1254
1255	auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
1256			     code, cpu, pid, tid, ip, msg);
1257
1258	err = perf_session__deliver_synth_event(pt->session, &event, NULL);
1259	if (err)
1260		pr_err("Intel Processor Trace: failed to deliver error event, error %d\n",
1261		       err);
1262
1263	return err;
1264}
1265
1266static int intel_pt_next_tid(struct intel_pt *pt, struct intel_pt_queue *ptq)
1267{
1268	struct auxtrace_queue *queue;
1269	pid_t tid = ptq->next_tid;
1270	int err;
1271
1272	if (tid == -1)
1273		return 0;
1274
1275	intel_pt_log("switch: cpu %d tid %d\n", ptq->cpu, tid);
1276
1277	err = machine__set_current_tid(pt->machine, ptq->cpu, -1, tid);
1278
1279	queue = &pt->queues.queue_array[ptq->queue_nr];
1280	intel_pt_set_pid_tid_cpu(pt, queue);
1281
1282	ptq->next_tid = -1;
1283
1284	return err;
1285}
1286
1287static inline bool intel_pt_is_switch_ip(struct intel_pt_queue *ptq, u64 ip)
1288{
1289	struct intel_pt *pt = ptq->pt;
1290
1291	return ip == pt->switch_ip &&
1292	       (ptq->flags & PERF_IP_FLAG_BRANCH) &&
1293	       !(ptq->flags & (PERF_IP_FLAG_CONDITIONAL | PERF_IP_FLAG_ASYNC |
1294			       PERF_IP_FLAG_INTERRUPT | PERF_IP_FLAG_TX_ABORT));
1295}
1296
1297static int intel_pt_sample(struct intel_pt_queue *ptq)
1298{
1299	const struct intel_pt_state *state = ptq->state;
1300	struct intel_pt *pt = ptq->pt;
1301	int err;
1302
1303	if (!ptq->have_sample)
1304		return 0;
1305
1306	ptq->have_sample = false;
1307
1308	if (pt->sample_instructions &&
1309	    (state->type & INTEL_PT_INSTRUCTION) &&
1310	    (!pt->synth_opts.initial_skip ||
1311	     pt->num_events++ >= pt->synth_opts.initial_skip)) {
1312		err = intel_pt_synth_instruction_sample(ptq);
1313		if (err)
1314			return err;
1315	}
1316
1317	if (pt->sample_transactions &&
1318	    (state->type & INTEL_PT_TRANSACTION) &&
1319	    (!pt->synth_opts.initial_skip ||
1320	     pt->num_events++ >= pt->synth_opts.initial_skip)) {
1321		err = intel_pt_synth_transaction_sample(ptq);
1322		if (err)
1323			return err;
1324	}
1325
1326	if (!(state->type & INTEL_PT_BRANCH))
1327		return 0;
1328
1329	if (pt->synth_opts.callchain || pt->synth_opts.thread_stack)
1330		thread_stack__event(ptq->thread, ptq->flags, state->from_ip,
1331				    state->to_ip, ptq->insn_len,
1332				    state->trace_nr);
1333	else
1334		thread_stack__set_trace_nr(ptq->thread, state->trace_nr);
1335
1336	if (pt->sample_branches) {
1337		err = intel_pt_synth_branch_sample(ptq);
1338		if (err)
1339			return err;
1340	}
1341
1342	if (pt->synth_opts.last_branch)
1343		intel_pt_update_last_branch_rb(ptq);
1344
1345	if (!pt->sync_switch)
1346		return 0;
1347
1348	if (intel_pt_is_switch_ip(ptq, state->to_ip)) {
1349		switch (ptq->switch_state) {
1350		case INTEL_PT_SS_UNKNOWN:
1351		case INTEL_PT_SS_EXPECTING_SWITCH_IP:
1352			err = intel_pt_next_tid(pt, ptq);
1353			if (err)
1354				return err;
1355			ptq->switch_state = INTEL_PT_SS_TRACING;
1356			break;
1357		default:
1358			ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_EVENT;
1359			return 1;
1360		}
1361	} else if (!state->to_ip) {
1362		ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
1363	} else if (ptq->switch_state == INTEL_PT_SS_NOT_TRACING) {
1364		ptq->switch_state = INTEL_PT_SS_UNKNOWN;
1365	} else if (ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
1366		   state->to_ip == pt->ptss_ip &&
1367		   (ptq->flags & PERF_IP_FLAG_CALL)) {
1368		ptq->switch_state = INTEL_PT_SS_TRACING;
1369	}
1370
1371	return 0;
1372}
1373
1374static u64 intel_pt_switch_ip(struct intel_pt *pt, u64 *ptss_ip)
1375{
1376	struct machine *machine = pt->machine;
1377	struct map *map;
1378	struct symbol *sym, *start;
1379	u64 ip, switch_ip = 0;
1380	const char *ptss;
1381
1382	if (ptss_ip)
1383		*ptss_ip = 0;
1384
1385	map = machine__kernel_map(machine);
1386	if (!map)
1387		return 0;
1388
1389	if (map__load(map))
1390		return 0;
1391
1392	start = dso__first_symbol(map->dso, MAP__FUNCTION);
1393
1394	for (sym = start; sym; sym = dso__next_symbol(sym)) {
1395		if (sym->binding == STB_GLOBAL &&
1396		    !strcmp(sym->name, "__switch_to")) {
1397			ip = map->unmap_ip(map, sym->start);
1398			if (ip >= map->start && ip < map->end) {
1399				switch_ip = ip;
1400				break;
1401			}
1402		}
1403	}
1404
1405	if (!switch_ip || !ptss_ip)
1406		return 0;
1407
1408	if (pt->have_sched_switch == 1)
1409		ptss = "perf_trace_sched_switch";
1410	else
1411		ptss = "__perf_event_task_sched_out";
1412
1413	for (sym = start; sym; sym = dso__next_symbol(sym)) {
1414		if (!strcmp(sym->name, ptss)) {
1415			ip = map->unmap_ip(map, sym->start);
1416			if (ip >= map->start && ip < map->end) {
1417				*ptss_ip = ip;
1418				break;
1419			}
1420		}
1421	}
1422
1423	return switch_ip;
1424}
1425
1426static int intel_pt_run_decoder(struct intel_pt_queue *ptq, u64 *timestamp)
1427{
1428	const struct intel_pt_state *state = ptq->state;
1429	struct intel_pt *pt = ptq->pt;
1430	int err;
1431
1432	if (!pt->kernel_start) {
1433		pt->kernel_start = machine__kernel_start(pt->machine);
1434		if (pt->per_cpu_mmaps &&
1435		    (pt->have_sched_switch == 1 || pt->have_sched_switch == 3) &&
1436		    !pt->timeless_decoding && intel_pt_tracing_kernel(pt) &&
1437		    !pt->sampling_mode) {
1438			pt->switch_ip = intel_pt_switch_ip(pt, &pt->ptss_ip);
1439			if (pt->switch_ip) {
1440				intel_pt_log("switch_ip: %"PRIx64" ptss_ip: %"PRIx64"\n",
1441					     pt->switch_ip, pt->ptss_ip);
1442				pt->sync_switch = true;
1443			}
1444		}
1445	}
1446
1447	intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
1448		     ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
1449	while (1) {
1450		err = intel_pt_sample(ptq);
1451		if (err)
1452			return err;
1453
1454		state = intel_pt_decode(ptq->decoder);
1455		if (state->err) {
1456			if (state->err == INTEL_PT_ERR_NODATA)
1457				return 1;
1458			if (pt->sync_switch &&
1459			    state->from_ip >= pt->kernel_start) {
1460				pt->sync_switch = false;
1461				intel_pt_next_tid(pt, ptq);
1462			}
1463			if (pt->synth_opts.errors) {
1464				err = intel_pt_synth_error(pt, state->err,
1465							   ptq->cpu, ptq->pid,
1466							   ptq->tid,
1467							   state->from_ip);
1468				if (err)
1469					return err;
1470			}
1471			continue;
1472		}
1473
1474		ptq->state = state;
1475		ptq->have_sample = true;
1476		intel_pt_sample_flags(ptq);
1477
1478		/* Use estimated TSC upon return to user space */
1479		if (pt->est_tsc &&
1480		    (state->from_ip >= pt->kernel_start || !state->from_ip) &&
1481		    state->to_ip && state->to_ip < pt->kernel_start) {
1482			intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
1483				     state->timestamp, state->est_timestamp);
1484			ptq->timestamp = state->est_timestamp;
1485		/* Use estimated TSC in unknown switch state */
1486		} else if (pt->sync_switch &&
1487			   ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
1488			   intel_pt_is_switch_ip(ptq, state->to_ip) &&
1489			   ptq->next_tid == -1) {
1490			intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
1491				     state->timestamp, state->est_timestamp);
1492			ptq->timestamp = state->est_timestamp;
1493		} else if (state->timestamp > ptq->timestamp) {
1494			ptq->timestamp = state->timestamp;
1495		}
1496
1497		if (!pt->timeless_decoding && ptq->timestamp >= *timestamp) {
1498			*timestamp = ptq->timestamp;
1499			return 0;
1500		}
1501	}
1502	return 0;
1503}
1504
1505static inline int intel_pt_update_queues(struct intel_pt *pt)
1506{
1507	if (pt->queues.new_data) {
1508		pt->queues.new_data = false;
1509		return intel_pt_setup_queues(pt);
1510	}
1511	return 0;
1512}
1513
1514static int intel_pt_process_queues(struct intel_pt *pt, u64 timestamp)
1515{
1516	unsigned int queue_nr;
1517	u64 ts;
1518	int ret;
1519
1520	while (1) {
1521		struct auxtrace_queue *queue;
1522		struct intel_pt_queue *ptq;
1523
1524		if (!pt->heap.heap_cnt)
1525			return 0;
1526
1527		if (pt->heap.heap_array[0].ordinal >= timestamp)
1528			return 0;
1529
1530		queue_nr = pt->heap.heap_array[0].queue_nr;
1531		queue = &pt->queues.queue_array[queue_nr];
1532		ptq = queue->priv;
1533
1534		intel_pt_log("queue %u processing 0x%" PRIx64 " to 0x%" PRIx64 "\n",
1535			     queue_nr, pt->heap.heap_array[0].ordinal,
1536			     timestamp);
1537
1538		auxtrace_heap__pop(&pt->heap);
1539
1540		if (pt->heap.heap_cnt) {
1541			ts = pt->heap.heap_array[0].ordinal + 1;
1542			if (ts > timestamp)
1543				ts = timestamp;
1544		} else {
1545			ts = timestamp;
1546		}
1547
1548		intel_pt_set_pid_tid_cpu(pt, queue);
1549
1550		ret = intel_pt_run_decoder(ptq, &ts);
1551
1552		if (ret < 0) {
1553			auxtrace_heap__add(&pt->heap, queue_nr, ts);
1554			return ret;
1555		}
1556
1557		if (!ret) {
1558			ret = auxtrace_heap__add(&pt->heap, queue_nr, ts);
1559			if (ret < 0)
1560				return ret;
1561		} else {
1562			ptq->on_heap = false;
1563		}
1564	}
1565
1566	return 0;
1567}
1568
1569static int intel_pt_process_timeless_queues(struct intel_pt *pt, pid_t tid,
1570					    u64 time_)
1571{
1572	struct auxtrace_queues *queues = &pt->queues;
1573	unsigned int i;
1574	u64 ts = 0;
1575
1576	for (i = 0; i < queues->nr_queues; i++) {
1577		struct auxtrace_queue *queue = &pt->queues.queue_array[i];
1578		struct intel_pt_queue *ptq = queue->priv;
1579
1580		if (ptq && (tid == -1 || ptq->tid == tid)) {
1581			ptq->time = time_;
1582			intel_pt_set_pid_tid_cpu(pt, queue);
1583			intel_pt_run_decoder(ptq, &ts);
1584		}
1585	}
1586	return 0;
1587}
1588
1589static int intel_pt_lost(struct intel_pt *pt, struct perf_sample *sample)
1590{
1591	return intel_pt_synth_error(pt, INTEL_PT_ERR_LOST, sample->cpu,
1592				    sample->pid, sample->tid, 0);
1593}
1594
1595static struct intel_pt_queue *intel_pt_cpu_to_ptq(struct intel_pt *pt, int cpu)
1596{
1597	unsigned i, j;
1598
1599	if (cpu < 0 || !pt->queues.nr_queues)
1600		return NULL;
1601
1602	if ((unsigned)cpu >= pt->queues.nr_queues)
1603		i = pt->queues.nr_queues - 1;
1604	else
1605		i = cpu;
1606
1607	if (pt->queues.queue_array[i].cpu == cpu)
1608		return pt->queues.queue_array[i].priv;
1609
1610	for (j = 0; i > 0; j++) {
1611		if (pt->queues.queue_array[--i].cpu == cpu)
1612			return pt->queues.queue_array[i].priv;
1613	}
1614
1615	for (; j < pt->queues.nr_queues; j++) {
1616		if (pt->queues.queue_array[j].cpu == cpu)
1617			return pt->queues.queue_array[j].priv;
1618	}
1619
1620	return NULL;
1621}
1622
1623static int intel_pt_sync_switch(struct intel_pt *pt, int cpu, pid_t tid,
1624				u64 timestamp)
1625{
1626	struct intel_pt_queue *ptq;
1627	int err;
1628
1629	if (!pt->sync_switch)
1630		return 1;
1631
1632	ptq = intel_pt_cpu_to_ptq(pt, cpu);
1633	if (!ptq)
1634		return 1;
1635
1636	switch (ptq->switch_state) {
1637	case INTEL_PT_SS_NOT_TRACING:
1638		ptq->next_tid = -1;
1639		break;
1640	case INTEL_PT_SS_UNKNOWN:
1641	case INTEL_PT_SS_TRACING:
1642		ptq->next_tid = tid;
1643		ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_IP;
1644		return 0;
1645	case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
1646		if (!ptq->on_heap) {
1647			ptq->timestamp = perf_time_to_tsc(timestamp,
1648							  &pt->tc);
1649			err = auxtrace_heap__add(&pt->heap, ptq->queue_nr,
1650						 ptq->timestamp);
1651			if (err)
1652				return err;
1653			ptq->on_heap = true;
1654		}
1655		ptq->switch_state = INTEL_PT_SS_TRACING;
1656		break;
1657	case INTEL_PT_SS_EXPECTING_SWITCH_IP:
1658		ptq->next_tid = tid;
1659		intel_pt_log("ERROR: cpu %d expecting switch ip\n", cpu);
1660		break;
1661	default:
1662		break;
1663	}
1664
1665	return 1;
1666}
1667
1668static int intel_pt_process_switch(struct intel_pt *pt,
1669				   struct perf_sample *sample)
1670{
1671	struct perf_evsel *evsel;
1672	pid_t tid;
1673	int cpu, ret;
1674
1675	evsel = perf_evlist__id2evsel(pt->session->evlist, sample->id);
1676	if (evsel != pt->switch_evsel)
1677		return 0;
1678
1679	tid = perf_evsel__intval(evsel, sample, "next_pid");
1680	cpu = sample->cpu;
1681
1682	intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1683		     cpu, tid, sample->time, perf_time_to_tsc(sample->time,
1684		     &pt->tc));
1685
1686	ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
1687	if (ret <= 0)
1688		return ret;
1689
1690	return machine__set_current_tid(pt->machine, cpu, -1, tid);
1691}
1692
1693static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,
1694				   struct perf_sample *sample)
1695{
1696	bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1697	pid_t pid, tid;
1698	int cpu, ret;
1699
1700	cpu = sample->cpu;
1701
1702	if (pt->have_sched_switch == 3) {
1703		if (!out)
1704			return 0;
1705		if (event->header.type != PERF_RECORD_SWITCH_CPU_WIDE) {
1706			pr_err("Expecting CPU-wide context switch event\n");
1707			return -EINVAL;
1708		}
1709		pid = event->context_switch.next_prev_pid;
1710		tid = event->context_switch.next_prev_tid;
1711	} else {
1712		if (out)
1713			return 0;
1714		pid = sample->pid;
1715		tid = sample->tid;
1716	}
1717
1718	if (tid == -1) {
1719		pr_err("context_switch event has no tid\n");
1720		return -EINVAL;
1721	}
1722
1723	intel_pt_log("context_switch: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1724		     cpu, pid, tid, sample->time, perf_time_to_tsc(sample->time,
1725		     &pt->tc));
1726
1727	ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
1728	if (ret <= 0)
1729		return ret;
1730
1731	return machine__set_current_tid(pt->machine, cpu, pid, tid);
1732}
1733
1734static int intel_pt_process_itrace_start(struct intel_pt *pt,
1735					 union perf_event *event,
1736					 struct perf_sample *sample)
1737{
1738	if (!pt->per_cpu_mmaps)
1739		return 0;
1740
1741	intel_pt_log("itrace_start: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1742		     sample->cpu, event->itrace_start.pid,
1743		     event->itrace_start.tid, sample->time,
1744		     perf_time_to_tsc(sample->time, &pt->tc));
1745
1746	return machine__set_current_tid(pt->machine, sample->cpu,
1747					event->itrace_start.pid,
1748					event->itrace_start.tid);
1749}
1750
1751static int intel_pt_process_event(struct perf_session *session,
1752				  union perf_event *event,
1753				  struct perf_sample *sample,
1754				  struct perf_tool *tool)
1755{
1756	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1757					   auxtrace);
1758	u64 timestamp;
1759	int err = 0;
1760
1761	if (dump_trace)
1762		return 0;
1763
1764	if (!tool->ordered_events) {
1765		pr_err("Intel Processor Trace requires ordered events\n");
1766		return -EINVAL;
1767	}
1768
1769	if (sample->time && sample->time != (u64)-1)
1770		timestamp = perf_time_to_tsc(sample->time, &pt->tc);
1771	else
1772		timestamp = 0;
1773
1774	if (timestamp || pt->timeless_decoding) {
1775		err = intel_pt_update_queues(pt);
1776		if (err)
1777			return err;
1778	}
1779
1780	if (pt->timeless_decoding) {
1781		if (event->header.type == PERF_RECORD_EXIT) {
1782			err = intel_pt_process_timeless_queues(pt,
1783							       event->fork.tid,
1784							       sample->time);
1785		}
1786	} else if (timestamp) {
1787		err = intel_pt_process_queues(pt, timestamp);
1788	}
1789	if (err)
1790		return err;
1791
1792	if (event->header.type == PERF_RECORD_AUX &&
1793	    (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
1794	    pt->synth_opts.errors) {
1795		err = intel_pt_lost(pt, sample);
1796		if (err)
1797			return err;
1798	}
1799
1800	if (pt->switch_evsel && event->header.type == PERF_RECORD_SAMPLE)
1801		err = intel_pt_process_switch(pt, sample);
1802	else if (event->header.type == PERF_RECORD_ITRACE_START)
1803		err = intel_pt_process_itrace_start(pt, event, sample);
1804	else if (event->header.type == PERF_RECORD_SWITCH ||
1805		 event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
1806		err = intel_pt_context_switch(pt, event, sample);
1807
1808	intel_pt_log("event %s (%u): cpu %d time %"PRIu64" tsc %#"PRIx64"\n",
1809		     perf_event__name(event->header.type), event->header.type,
1810		     sample->cpu, sample->time, timestamp);
1811
1812	return err;
1813}
1814
1815static int intel_pt_flush(struct perf_session *session, struct perf_tool *tool)
1816{
1817	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1818					   auxtrace);
1819	int ret;
1820
1821	if (dump_trace)
1822		return 0;
1823
1824	if (!tool->ordered_events)
1825		return -EINVAL;
1826
1827	ret = intel_pt_update_queues(pt);
1828	if (ret < 0)
1829		return ret;
1830
1831	if (pt->timeless_decoding)
1832		return intel_pt_process_timeless_queues(pt, -1,
1833							MAX_TIMESTAMP - 1);
1834
1835	return intel_pt_process_queues(pt, MAX_TIMESTAMP);
1836}
1837
1838static void intel_pt_free_events(struct perf_session *session)
1839{
1840	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1841					   auxtrace);
1842	struct auxtrace_queues *queues = &pt->queues;
1843	unsigned int i;
1844
1845	for (i = 0; i < queues->nr_queues; i++) {
1846		intel_pt_free_queue(queues->queue_array[i].priv);
1847		queues->queue_array[i].priv = NULL;
1848	}
1849	intel_pt_log_disable();
1850	auxtrace_queues__free(queues);
1851}
1852
1853static void intel_pt_free(struct perf_session *session)
1854{
1855	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1856					   auxtrace);
1857
1858	auxtrace_heap__free(&pt->heap);
1859	intel_pt_free_events(session);
1860	session->auxtrace = NULL;
1861	thread__put(pt->unknown_thread);
1862	addr_filters__exit(&pt->filts);
1863	zfree(&pt->filter);
1864	free(pt);
1865}
1866
1867static int intel_pt_process_auxtrace_event(struct perf_session *session,
1868					   union perf_event *event,
1869					   struct perf_tool *tool __maybe_unused)
1870{
1871	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1872					   auxtrace);
1873
1874	if (pt->sampling_mode)
1875		return 0;
1876
1877	if (!pt->data_queued) {
1878		struct auxtrace_buffer *buffer;
1879		off_t data_offset;
1880		int fd = perf_data_file__fd(session->file);
1881		int err;
1882
1883		if (perf_data_file__is_pipe(session->file)) {
1884			data_offset = 0;
1885		} else {
1886			data_offset = lseek(fd, 0, SEEK_CUR);
1887			if (data_offset == -1)
1888				return -errno;
1889		}
1890
1891		err = auxtrace_queues__add_event(&pt->queues, session, event,
1892						 data_offset, &buffer);
1893		if (err)
1894			return err;
1895
1896		/* Dump here now we have copied a piped trace out of the pipe */
1897		if (dump_trace) {
1898			if (auxtrace_buffer__get_data(buffer, fd)) {
1899				intel_pt_dump_event(pt, buffer->data,
1900						    buffer->size);
1901				auxtrace_buffer__put_data(buffer);
1902			}
1903		}
1904	}
1905
1906	return 0;
1907}
1908
1909struct intel_pt_synth {
1910	struct perf_tool dummy_tool;
1911	struct perf_session *session;
1912};
1913
1914static int intel_pt_event_synth(struct perf_tool *tool,
1915				union perf_event *event,
1916				struct perf_sample *sample __maybe_unused,
1917				struct machine *machine __maybe_unused)
1918{
1919	struct intel_pt_synth *intel_pt_synth =
1920			container_of(tool, struct intel_pt_synth, dummy_tool);
1921
1922	return perf_session__deliver_synth_event(intel_pt_synth->session, event,
1923						 NULL);
1924}
1925
1926static int intel_pt_synth_event(struct perf_session *session,
1927				struct perf_event_attr *attr, u64 id)
1928{
1929	struct intel_pt_synth intel_pt_synth;
1930
1931	memset(&intel_pt_synth, 0, sizeof(struct intel_pt_synth));
1932	intel_pt_synth.session = session;
1933
1934	return perf_event__synthesize_attr(&intel_pt_synth.dummy_tool, attr, 1,
1935					   &id, intel_pt_event_synth);
1936}
1937
1938static int intel_pt_synth_events(struct intel_pt *pt,
1939				 struct perf_session *session)
1940{
1941	struct perf_evlist *evlist = session->evlist;
1942	struct perf_evsel *evsel;
1943	struct perf_event_attr attr;
1944	bool found = false;
1945	u64 id;
1946	int err;
1947
1948	evlist__for_each_entry(evlist, evsel) {
1949		if (evsel->attr.type == pt->pmu_type && evsel->ids) {
1950			found = true;
1951			break;
1952		}
1953	}
1954
1955	if (!found) {
1956		pr_debug("There are no selected events with Intel Processor Trace data\n");
1957		return 0;
1958	}
1959
1960	memset(&attr, 0, sizeof(struct perf_event_attr));
1961	attr.size = sizeof(struct perf_event_attr);
1962	attr.type = PERF_TYPE_HARDWARE;
1963	attr.sample_type = evsel->attr.sample_type & PERF_SAMPLE_MASK;
1964	attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
1965			    PERF_SAMPLE_PERIOD;
1966	if (pt->timeless_decoding)
1967		attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
1968	else
1969		attr.sample_type |= PERF_SAMPLE_TIME;
1970	if (!pt->per_cpu_mmaps)
1971		attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
1972	attr.exclude_user = evsel->attr.exclude_user;
1973	attr.exclude_kernel = evsel->attr.exclude_kernel;
1974	attr.exclude_hv = evsel->attr.exclude_hv;
1975	attr.exclude_host = evsel->attr.exclude_host;
1976	attr.exclude_guest = evsel->attr.exclude_guest;
1977	attr.sample_id_all = evsel->attr.sample_id_all;
1978	attr.read_format = evsel->attr.read_format;
1979
1980	id = evsel->id[0] + 1000000000;
1981	if (!id)
1982		id = 1;
1983
1984	if (pt->synth_opts.instructions) {
1985		attr.config = PERF_COUNT_HW_INSTRUCTIONS;
1986		if (pt->synth_opts.period_type == PERF_ITRACE_PERIOD_NANOSECS)
1987			attr.sample_period =
1988				intel_pt_ns_to_ticks(pt, pt->synth_opts.period);
1989		else
1990			attr.sample_period = pt->synth_opts.period;
1991		pt->instructions_sample_period = attr.sample_period;
1992		if (pt->synth_opts.callchain)
1993			attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
1994		if (pt->synth_opts.last_branch)
1995			attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
1996		pr_debug("Synthesizing 'instructions' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
1997			 id, (u64)attr.sample_type);
1998		err = intel_pt_synth_event(session, &attr, id);
1999		if (err) {
2000			pr_err("%s: failed to synthesize 'instructions' event type\n",
2001			       __func__);
2002			return err;
2003		}
2004		pt->sample_instructions = true;
2005		pt->instructions_sample_type = attr.sample_type;
2006		pt->instructions_id = id;
2007		id += 1;
2008	}
2009
2010	if (pt->synth_opts.transactions) {
2011		attr.config = PERF_COUNT_HW_INSTRUCTIONS;
2012		attr.sample_period = 1;
2013		if (pt->synth_opts.callchain)
2014			attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
2015		if (pt->synth_opts.last_branch)
2016			attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
2017		pr_debug("Synthesizing 'transactions' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
2018			 id, (u64)attr.sample_type);
2019		err = intel_pt_synth_event(session, &attr, id);
2020		if (err) {
2021			pr_err("%s: failed to synthesize 'transactions' event type\n",
2022			       __func__);
2023			return err;
2024		}
2025		pt->sample_transactions = true;
2026		pt->transactions_id = id;
2027		id += 1;
2028		evlist__for_each_entry(evlist, evsel) {
2029			if (evsel->id && evsel->id[0] == pt->transactions_id) {
2030				if (evsel->name)
2031					zfree(&evsel->name);
2032				evsel->name = strdup("transactions");
2033				break;
2034			}
2035		}
2036	}
2037
2038	if (pt->synth_opts.branches) {
2039		attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
2040		attr.sample_period = 1;
2041		attr.sample_type |= PERF_SAMPLE_ADDR;
2042		attr.sample_type &= ~(u64)PERF_SAMPLE_CALLCHAIN;
2043		attr.sample_type &= ~(u64)PERF_SAMPLE_BRANCH_STACK;
2044		pr_debug("Synthesizing 'branches' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
2045			 id, (u64)attr.sample_type);
2046		err = intel_pt_synth_event(session, &attr, id);
2047		if (err) {
2048			pr_err("%s: failed to synthesize 'branches' event type\n",
2049			       __func__);
2050			return err;
2051		}
2052		pt->sample_branches = true;
2053		pt->branches_sample_type = attr.sample_type;
2054		pt->branches_id = id;
2055	}
2056
2057	pt->synth_needs_swap = evsel->needs_swap;
2058
2059	return 0;
2060}
2061
2062static struct perf_evsel *intel_pt_find_sched_switch(struct perf_evlist *evlist)
2063{
2064	struct perf_evsel *evsel;
2065
2066	evlist__for_each_entry_reverse(evlist, evsel) {
2067		const char *name = perf_evsel__name(evsel);
2068
2069		if (!strcmp(name, "sched:sched_switch"))
2070			return evsel;
2071	}
2072
2073	return NULL;
2074}
2075
2076static bool intel_pt_find_switch(struct perf_evlist *evlist)
2077{
2078	struct perf_evsel *evsel;
2079
2080	evlist__for_each_entry(evlist, evsel) {
2081		if (evsel->attr.context_switch)
2082			return true;
2083	}
2084
2085	return false;
2086}
2087
2088static int intel_pt_perf_config(const char *var, const char *value, void *data)
2089{
2090	struct intel_pt *pt = data;
2091
2092	if (!strcmp(var, "intel-pt.mispred-all"))
2093		pt->mispred_all = perf_config_bool(var, value);
2094
2095	return 0;
2096}
2097
2098static const char * const intel_pt_info_fmts[] = {
2099	[INTEL_PT_PMU_TYPE]		= "  PMU Type            %"PRId64"\n",
2100	[INTEL_PT_TIME_SHIFT]		= "  Time Shift          %"PRIu64"\n",
2101	[INTEL_PT_TIME_MULT]		= "  Time Muliplier      %"PRIu64"\n",
2102	[INTEL_PT_TIME_ZERO]		= "  Time Zero           %"PRIu64"\n",
2103	[INTEL_PT_CAP_USER_TIME_ZERO]	= "  Cap Time Zero       %"PRId64"\n",
2104	[INTEL_PT_TSC_BIT]		= "  TSC bit             %#"PRIx64"\n",
2105	[INTEL_PT_NORETCOMP_BIT]	= "  NoRETComp bit       %#"PRIx64"\n",
2106	[INTEL_PT_HAVE_SCHED_SWITCH]	= "  Have sched_switch   %"PRId64"\n",
2107	[INTEL_PT_SNAPSHOT_MODE]	= "  Snapshot mode       %"PRId64"\n",
2108	[INTEL_PT_PER_CPU_MMAPS]	= "  Per-cpu maps        %"PRId64"\n",
2109	[INTEL_PT_MTC_BIT]		= "  MTC bit             %#"PRIx64"\n",
2110	[INTEL_PT_TSC_CTC_N]		= "  TSC:CTC numerator   %"PRIu64"\n",
2111	[INTEL_PT_TSC_CTC_D]		= "  TSC:CTC denominator %"PRIu64"\n",
2112	[INTEL_PT_CYC_BIT]		= "  CYC bit             %#"PRIx64"\n",
2113	[INTEL_PT_MAX_NONTURBO_RATIO]	= "  Max non-turbo ratio %"PRIu64"\n",
2114	[INTEL_PT_FILTER_STR_LEN]	= "  Filter string len.  %"PRIu64"\n",
2115};
2116
2117static void intel_pt_print_info(u64 *arr, int start, int finish)
2118{
2119	int i;
2120
2121	if (!dump_trace)
2122		return;
2123
2124	for (i = start; i <= finish; i++)
2125		fprintf(stdout, intel_pt_info_fmts[i], arr[i]);
2126}
2127
2128static void intel_pt_print_info_str(const char *name, const char *str)
2129{
2130	if (!dump_trace)
2131		return;
2132
2133	fprintf(stdout, "  %-20s%s\n", name, str ? str : "");
2134}
2135
2136static bool intel_pt_has(struct auxtrace_info_event *auxtrace_info, int pos)
2137{
2138	return auxtrace_info->header.size >=
2139		sizeof(struct auxtrace_info_event) + (sizeof(u64) * (pos + 1));
2140}
2141
2142int intel_pt_process_auxtrace_info(union perf_event *event,
2143				   struct perf_session *session)
2144{
2145	struct auxtrace_info_event *auxtrace_info = &event->auxtrace_info;
2146	size_t min_sz = sizeof(u64) * INTEL_PT_PER_CPU_MMAPS;
2147	struct intel_pt *pt;
2148	void *info_end;
2149	u64 *info;
2150	int err;
2151
2152	if (auxtrace_info->header.size < sizeof(struct auxtrace_info_event) +
2153					min_sz)
2154		return -EINVAL;
2155
2156	pt = zalloc(sizeof(struct intel_pt));
2157	if (!pt)
2158		return -ENOMEM;
2159
2160	addr_filters__init(&pt->filts);
2161
2162	perf_config(intel_pt_perf_config, pt);
2163
2164	err = auxtrace_queues__init(&pt->queues);
2165	if (err)
2166		goto err_free;
2167
2168	intel_pt_log_set_name(INTEL_PT_PMU_NAME);
2169
2170	pt->session = session;
2171	pt->machine = &session->machines.host; /* No kvm support */
2172	pt->auxtrace_type = auxtrace_info->type;
2173	pt->pmu_type = auxtrace_info->priv[INTEL_PT_PMU_TYPE];
2174	pt->tc.time_shift = auxtrace_info->priv[INTEL_PT_TIME_SHIFT];
2175	pt->tc.time_mult = auxtrace_info->priv[INTEL_PT_TIME_MULT];
2176	pt->tc.time_zero = auxtrace_info->priv[INTEL_PT_TIME_ZERO];
2177	pt->cap_user_time_zero = auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO];
2178	pt->tsc_bit = auxtrace_info->priv[INTEL_PT_TSC_BIT];
2179	pt->noretcomp_bit = auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT];
2180	pt->have_sched_switch = auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH];
2181	pt->snapshot_mode = auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE];
2182	pt->per_cpu_mmaps = auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS];
2183	intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_PMU_TYPE,
2184			    INTEL_PT_PER_CPU_MMAPS);
2185
2186	if (intel_pt_has(auxtrace_info, INTEL_PT_CYC_BIT)) {
2187		pt->mtc_bit = auxtrace_info->priv[INTEL_PT_MTC_BIT];
2188		pt->mtc_freq_bits = auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS];
2189		pt->tsc_ctc_ratio_n = auxtrace_info->priv[INTEL_PT_TSC_CTC_N];
2190		pt->tsc_ctc_ratio_d = auxtrace_info->priv[INTEL_PT_TSC_CTC_D];
2191		pt->cyc_bit = auxtrace_info->priv[INTEL_PT_CYC_BIT];
2192		intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_MTC_BIT,
2193				    INTEL_PT_CYC_BIT);
2194	}
2195
2196	if (intel_pt_has(auxtrace_info, INTEL_PT_MAX_NONTURBO_RATIO)) {
2197		pt->max_non_turbo_ratio =
2198			auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO];
2199		intel_pt_print_info(&auxtrace_info->priv[0],
2200				    INTEL_PT_MAX_NONTURBO_RATIO,
2201				    INTEL_PT_MAX_NONTURBO_RATIO);
2202	}
2203
2204	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
2205	info_end = (void *)info + auxtrace_info->header.size;
2206
2207	if (intel_pt_has(auxtrace_info, INTEL_PT_FILTER_STR_LEN)) {
2208		size_t len;
2209
2210		len = auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN];
2211		intel_pt_print_info(&auxtrace_info->priv[0],
2212				    INTEL_PT_FILTER_STR_LEN,
2213				    INTEL_PT_FILTER_STR_LEN);
2214		if (len) {
2215			const char *filter = (const char *)info;
2216
2217			len = roundup(len + 1, 8);
2218			info += len >> 3;
2219			if ((void *)info > info_end) {
2220				pr_err("%s: bad filter string length\n", __func__);
2221				err = -EINVAL;
2222				goto err_free_queues;
2223			}
2224			pt->filter = memdup(filter, len);
2225			if (!pt->filter) {
2226				err = -ENOMEM;
2227				goto err_free_queues;
2228			}
2229			if (session->header.needs_swap)
2230				mem_bswap_64(pt->filter, len);
2231			if (pt->filter[len - 1]) {
2232				pr_err("%s: filter string not null terminated\n", __func__);
2233				err = -EINVAL;
2234				goto err_free_queues;
2235			}
2236			err = addr_filters__parse_bare_filter(&pt->filts,
2237							      filter);
2238			if (err)
2239				goto err_free_queues;
2240		}
2241		intel_pt_print_info_str("Filter string", pt->filter);
2242	}
2243
2244	pt->timeless_decoding = intel_pt_timeless_decoding(pt);
2245	pt->have_tsc = intel_pt_have_tsc(pt);
2246	pt->sampling_mode = false;
2247	pt->est_tsc = !pt->timeless_decoding;
2248
2249	pt->unknown_thread = thread__new(999999999, 999999999);
2250	if (!pt->unknown_thread) {
2251		err = -ENOMEM;
2252		goto err_free_queues;
2253	}
2254
2255	/*
2256	 * Since this thread will not be kept in any rbtree not in a
2257	 * list, initialize its list node so that at thread__put() the
2258	 * current thread lifetime assuption is kept and we don't segfault
2259	 * at list_del_init().
2260	 */
2261	INIT_LIST_HEAD(&pt->unknown_thread->node);
2262
2263	err = thread__set_comm(pt->unknown_thread, "unknown", 0);
2264	if (err)
2265		goto err_delete_thread;
2266	if (thread__init_map_groups(pt->unknown_thread, pt->machine)) {
2267		err = -ENOMEM;
2268		goto err_delete_thread;
2269	}
2270
2271	pt->auxtrace.process_event = intel_pt_process_event;
2272	pt->auxtrace.process_auxtrace_event = intel_pt_process_auxtrace_event;
2273	pt->auxtrace.flush_events = intel_pt_flush;
2274	pt->auxtrace.free_events = intel_pt_free_events;
2275	pt->auxtrace.free = intel_pt_free;
2276	session->auxtrace = &pt->auxtrace;
2277
2278	if (dump_trace)
2279		return 0;
2280
2281	if (pt->have_sched_switch == 1) {
2282		pt->switch_evsel = intel_pt_find_sched_switch(session->evlist);
2283		if (!pt->switch_evsel) {
2284			pr_err("%s: missing sched_switch event\n", __func__);
2285			err = -EINVAL;
2286			goto err_delete_thread;
2287		}
2288	} else if (pt->have_sched_switch == 2 &&
2289		   !intel_pt_find_switch(session->evlist)) {
2290		pr_err("%s: missing context_switch attribute flag\n", __func__);
2291		err = -EINVAL;
2292		goto err_delete_thread;
2293	}
2294
2295	if (session->itrace_synth_opts && session->itrace_synth_opts->set) {
2296		pt->synth_opts = *session->itrace_synth_opts;
2297	} else {
2298		itrace_synth_opts__set_default(&pt->synth_opts);
2299		if (use_browser != -1) {
2300			pt->synth_opts.branches = false;
2301			pt->synth_opts.callchain = true;
2302		}
2303		if (session->itrace_synth_opts)
2304			pt->synth_opts.thread_stack =
2305				session->itrace_synth_opts->thread_stack;
2306	}
2307
2308	if (pt->synth_opts.log)
2309		intel_pt_log_enable();
2310
2311	/* Maximum non-turbo ratio is TSC freq / 100 MHz */
2312	if (pt->tc.time_mult) {
2313		u64 tsc_freq = intel_pt_ns_to_ticks(pt, 1000000000);
2314
2315		if (!pt->max_non_turbo_ratio)
2316			pt->max_non_turbo_ratio =
2317					(tsc_freq + 50000000) / 100000000;
2318		intel_pt_log("TSC frequency %"PRIu64"\n", tsc_freq);
2319		intel_pt_log("Maximum non-turbo ratio %u\n",
2320			     pt->max_non_turbo_ratio);
2321	}
2322
2323	if (pt->synth_opts.calls)
2324		pt->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
2325				       PERF_IP_FLAG_TRACE_END;
2326	if (pt->synth_opts.returns)
2327		pt->branches_filter |= PERF_IP_FLAG_RETURN |
2328				       PERF_IP_FLAG_TRACE_BEGIN;
2329
2330	if (pt->synth_opts.callchain && !symbol_conf.use_callchain) {
2331		symbol_conf.use_callchain = true;
2332		if (callchain_register_param(&callchain_param) < 0) {
2333			symbol_conf.use_callchain = false;
2334			pt->synth_opts.callchain = false;
2335		}
2336	}
2337
2338	err = intel_pt_synth_events(pt, session);
2339	if (err)
2340		goto err_delete_thread;
2341
2342	err = auxtrace_queues__process_index(&pt->queues, session);
2343	if (err)
2344		goto err_delete_thread;
2345
2346	if (pt->queues.populated)
2347		pt->data_queued = true;
2348
2349	if (pt->timeless_decoding)
2350		pr_debug2("Intel PT decoding without timestamps\n");
2351
2352	return 0;
2353
2354err_delete_thread:
2355	thread__zput(pt->unknown_thread);
2356err_free_queues:
2357	intel_pt_log_disable();
2358	auxtrace_queues__free(&pt->queues);
2359	session->auxtrace = NULL;
2360err_free:
2361	addr_filters__exit(&pt->filts);
2362	zfree(&pt->filter);
2363	free(pt);
2364	return err;
2365}