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 <inttypes.h>
   8#include <linux/perf_event.h>
   9#include <stdio.h>
  10#include <stdbool.h>
  11#include <errno.h>
  12#include <linux/kernel.h>
  13#include <linux/string.h>
  14#include <linux/types.h>
  15#include <linux/zalloc.h>
  16
  17#include "session.h"
  18#include "machine.h"
  19#include "memswap.h"
  20#include "sort.h"
  21#include "tool.h"
  22#include "event.h"
  23#include "evlist.h"
  24#include "evsel.h"
  25#include "map.h"
  26#include "color.h"
  27#include "thread.h"
  28#include "thread-stack.h"
  29#include "symbol.h"
  30#include "callchain.h"
  31#include "dso.h"
  32#include "debug.h"
  33#include "auxtrace.h"
  34#include "tsc.h"
  35#include "intel-pt.h"
  36#include "config.h"
  37#include "util/perf_api_probe.h"
  38#include "util/synthetic-events.h"
  39#include "time-utils.h"
  40
  41#include "../arch/x86/include/uapi/asm/perf_regs.h"
  42
  43#include "intel-pt-decoder/intel-pt-log.h"
  44#include "intel-pt-decoder/intel-pt-decoder.h"
  45#include "intel-pt-decoder/intel-pt-insn-decoder.h"
  46#include "intel-pt-decoder/intel-pt-pkt-decoder.h"
  47
  48#define MAX_TIMESTAMP (~0ULL)
  49
  50#define INTEL_PT_CFG_PASS_THRU	BIT_ULL(0)
  51#define INTEL_PT_CFG_PWR_EVT_EN	BIT_ULL(4)
  52#define INTEL_PT_CFG_BRANCH_EN	BIT_ULL(13)
  53#define INTEL_PT_CFG_EVT_EN	BIT_ULL(31)
  54#define INTEL_PT_CFG_TNT_DIS	BIT_ULL(55)
  55
  56struct range {
  57	u64 start;
  58	u64 end;
  59};
  60
  61struct intel_pt {
  62	struct auxtrace auxtrace;
  63	struct auxtrace_queues queues;
  64	struct auxtrace_heap heap;
  65	u32 auxtrace_type;
  66	struct perf_session *session;
  67	struct machine *machine;
  68	struct evsel *switch_evsel;
  69	struct thread *unknown_thread;
  70	bool timeless_decoding;
  71	bool sampling_mode;
  72	bool snapshot_mode;
  73	bool per_cpu_mmaps;
  74	bool have_tsc;
  75	bool data_queued;
  76	bool est_tsc;
  77	bool sync_switch;
  78	bool sync_switch_not_supported;
  79	bool mispred_all;
  80	bool use_thread_stack;
  81	bool callstack;
  82	bool cap_event_trace;
  83	bool have_guest_sideband;
  84	unsigned int br_stack_sz;
  85	unsigned int br_stack_sz_plus;
  86	int have_sched_switch;
  87	u32 pmu_type;
  88	u64 kernel_start;
  89	u64 switch_ip;
  90	u64 ptss_ip;
  91	u64 first_timestamp;
  92
  93	struct perf_tsc_conversion tc;
  94	bool cap_user_time_zero;
  95
  96	struct itrace_synth_opts synth_opts;
  97
  98	bool sample_instructions;
  99	u64 instructions_sample_type;
 100	u64 instructions_id;
 101
 102	bool sample_cycles;
 103	u64 cycles_sample_type;
 104	u64 cycles_id;
 105
 106	bool sample_branches;
 107	u32 branches_filter;
 108	u64 branches_sample_type;
 109	u64 branches_id;
 110
 111	bool sample_transactions;
 112	u64 transactions_sample_type;
 113	u64 transactions_id;
 114
 115	bool sample_ptwrites;
 116	u64 ptwrites_sample_type;
 117	u64 ptwrites_id;
 118
 119	bool sample_pwr_events;
 120	u64 pwr_events_sample_type;
 121	u64 mwait_id;
 122	u64 pwre_id;
 123	u64 exstop_id;
 124	u64 pwrx_id;
 125	u64 cbr_id;
 126	u64 psb_id;
 127
 128	bool single_pebs;
 129	bool sample_pebs;
 130	struct evsel *pebs_evsel;
 131
 132	u64 evt_sample_type;
 133	u64 evt_id;
 134
 135	u64 iflag_chg_sample_type;
 136	u64 iflag_chg_id;
 137
 138	u64 tsc_bit;
 139	u64 mtc_bit;
 140	u64 mtc_freq_bits;
 141	u32 tsc_ctc_ratio_n;
 142	u32 tsc_ctc_ratio_d;
 143	u64 cyc_bit;
 144	u64 noretcomp_bit;
 145	unsigned max_non_turbo_ratio;
 146	unsigned cbr2khz;
 147	int max_loops;
 148
 149	unsigned long num_events;
 150
 151	char *filter;
 152	struct addr_filters filts;
 153
 154	struct range *time_ranges;
 155	unsigned int range_cnt;
 156
 157	struct ip_callchain *chain;
 158	struct branch_stack *br_stack;
 159
 160	u64 dflt_tsc_offset;
 161	struct rb_root vmcs_info;
 162};
 163
 164enum switch_state {
 165	INTEL_PT_SS_NOT_TRACING,
 166	INTEL_PT_SS_UNKNOWN,
 167	INTEL_PT_SS_TRACING,
 168	INTEL_PT_SS_EXPECTING_SWITCH_EVENT,
 169	INTEL_PT_SS_EXPECTING_SWITCH_IP,
 170};
 171
 172/* applicable_counters is 64-bits */
 173#define INTEL_PT_MAX_PEBS 64
 174
 175struct intel_pt_pebs_event {
 176	struct evsel *evsel;
 177	u64 id;
 178};
 179
 180struct intel_pt_queue {
 181	struct intel_pt *pt;
 182	unsigned int queue_nr;
 183	struct auxtrace_buffer *buffer;
 184	struct auxtrace_buffer *old_buffer;
 185	void *decoder;
 186	const struct intel_pt_state *state;
 187	struct ip_callchain *chain;
 188	struct branch_stack *last_branch;
 189	union perf_event *event_buf;
 190	bool on_heap;
 191	bool stop;
 192	bool step_through_buffers;
 193	bool use_buffer_pid_tid;
 194	bool sync_switch;
 195	bool sample_ipc;
 196	pid_t pid, tid;
 197	int cpu;
 198	int switch_state;
 199	pid_t next_tid;
 200	struct thread *thread;
 201	struct machine *guest_machine;
 202	struct thread *guest_thread;
 203	struct thread *unknown_guest_thread;
 204	pid_t guest_machine_pid;
 205	pid_t guest_pid;
 206	pid_t guest_tid;
 207	int vcpu;
 208	bool exclude_kernel;
 209	bool have_sample;
 210	u64 time;
 211	u64 timestamp;
 212	u64 sel_timestamp;
 213	bool sel_start;
 214	unsigned int sel_idx;
 215	u32 flags;
 216	u16 insn_len;
 217	u64 last_insn_cnt;
 218	u64 ipc_insn_cnt;
 219	u64 ipc_cyc_cnt;
 220	u64 last_in_insn_cnt;
 221	u64 last_in_cyc_cnt;
 222	u64 last_cy_insn_cnt;
 223	u64 last_cy_cyc_cnt;
 224	u64 last_br_insn_cnt;
 225	u64 last_br_cyc_cnt;
 226	unsigned int cbr_seen;
 227	char insn[INTEL_PT_INSN_BUF_SZ];
 228	struct intel_pt_pebs_event pebs[INTEL_PT_MAX_PEBS];
 229};
 230
 231static void intel_pt_dump(struct intel_pt *pt __maybe_unused,
 232			  unsigned char *buf, size_t len)
 233{
 234	struct intel_pt_pkt packet;
 235	size_t pos = 0;
 236	int ret, pkt_len, i;
 237	char desc[INTEL_PT_PKT_DESC_MAX];
 238	const char *color = PERF_COLOR_BLUE;
 239	enum intel_pt_pkt_ctx ctx = INTEL_PT_NO_CTX;
 240
 241	color_fprintf(stdout, color,
 242		      ". ... Intel Processor Trace data: size %zu bytes\n",
 243		      len);
 244
 245	while (len) {
 246		ret = intel_pt_get_packet(buf, len, &packet, &ctx);
 247		if (ret > 0)
 248			pkt_len = ret;
 249		else
 250			pkt_len = 1;
 251		printf(".");
 252		color_fprintf(stdout, color, "  %08zx: ", pos);
 253		for (i = 0; i < pkt_len; i++)
 254			color_fprintf(stdout, color, " %02x", buf[i]);
 255		for (; i < 16; i++)
 256			color_fprintf(stdout, color, "   ");
 257		if (ret > 0) {
 258			ret = intel_pt_pkt_desc(&packet, desc,
 259						INTEL_PT_PKT_DESC_MAX);
 260			if (ret > 0)
 261				color_fprintf(stdout, color, " %s\n", desc);
 262		} else {
 263			color_fprintf(stdout, color, " Bad packet!\n");
 264		}
 265		pos += pkt_len;
 266		buf += pkt_len;
 267		len -= pkt_len;
 268	}
 269}
 270
 271static void intel_pt_dump_event(struct intel_pt *pt, unsigned char *buf,
 272				size_t len)
 273{
 274	printf(".\n");
 275	intel_pt_dump(pt, buf, len);
 276}
 277
 278static void intel_pt_log_event(union perf_event *event)
 279{
 280	FILE *f = intel_pt_log_fp();
 281
 282	if (!intel_pt_enable_logging || !f)
 283		return;
 284
 285	perf_event__fprintf(event, NULL, f);
 286}
 287
 288static void intel_pt_dump_sample(struct perf_session *session,
 289				 struct perf_sample *sample)
 290{
 291	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
 292					   auxtrace);
 293
 294	printf("\n");
 295	intel_pt_dump(pt, sample->aux_sample.data, sample->aux_sample.size);
 296}
 297
 298static bool intel_pt_log_events(struct intel_pt *pt, u64 tm)
 299{
 300	struct perf_time_interval *range = pt->synth_opts.ptime_range;
 301	int n = pt->synth_opts.range_num;
 302
 303	if (pt->synth_opts.log_plus_flags & AUXTRACE_LOG_FLG_ALL_PERF_EVTS)
 304		return true;
 305
 306	if (pt->synth_opts.log_minus_flags & AUXTRACE_LOG_FLG_ALL_PERF_EVTS)
 307		return false;
 308
 309	/* perf_time__ranges_skip_sample does not work if time is zero */
 310	if (!tm)
 311		tm = 1;
 312
 313	return !n || !perf_time__ranges_skip_sample(range, n, tm);
 314}
 315
 316static struct intel_pt_vmcs_info *intel_pt_findnew_vmcs(struct rb_root *rb_root,
 317							u64 vmcs,
 318							u64 dflt_tsc_offset)
 319{
 320	struct rb_node **p = &rb_root->rb_node;
 321	struct rb_node *parent = NULL;
 322	struct intel_pt_vmcs_info *v;
 323
 324	while (*p) {
 325		parent = *p;
 326		v = rb_entry(parent, struct intel_pt_vmcs_info, rb_node);
 327
 328		if (v->vmcs == vmcs)
 329			return v;
 330
 331		if (vmcs < v->vmcs)
 332			p = &(*p)->rb_left;
 333		else
 334			p = &(*p)->rb_right;
 335	}
 336
 337	v = zalloc(sizeof(*v));
 338	if (v) {
 339		v->vmcs = vmcs;
 340		v->tsc_offset = dflt_tsc_offset;
 341		v->reliable = dflt_tsc_offset;
 342
 343		rb_link_node(&v->rb_node, parent, p);
 344		rb_insert_color(&v->rb_node, rb_root);
 345	}
 346
 347	return v;
 348}
 349
 350static struct intel_pt_vmcs_info *intel_pt_findnew_vmcs_info(void *data, uint64_t vmcs)
 351{
 352	struct intel_pt_queue *ptq = data;
 353	struct intel_pt *pt = ptq->pt;
 354
 355	if (!vmcs && !pt->dflt_tsc_offset)
 356		return NULL;
 357
 358	return intel_pt_findnew_vmcs(&pt->vmcs_info, vmcs, pt->dflt_tsc_offset);
 359}
 360
 361static void intel_pt_free_vmcs_info(struct intel_pt *pt)
 362{
 363	struct intel_pt_vmcs_info *v;
 364	struct rb_node *n;
 365
 366	n = rb_first(&pt->vmcs_info);
 367	while (n) {
 368		v = rb_entry(n, struct intel_pt_vmcs_info, rb_node);
 369		n = rb_next(n);
 370		rb_erase(&v->rb_node, &pt->vmcs_info);
 371		free(v);
 372	}
 373}
 374
 375static int intel_pt_do_fix_overlap(struct intel_pt *pt, struct auxtrace_buffer *a,
 376				   struct auxtrace_buffer *b)
 377{
 378	bool consecutive = false;
 379	void *start;
 380
 381	start = intel_pt_find_overlap(a->data, a->size, b->data, b->size,
 382				      pt->have_tsc, &consecutive,
 383				      pt->synth_opts.vm_time_correlation);
 384	if (!start)
 385		return -EINVAL;
 386	/*
 387	 * In the case of vm_time_correlation, the overlap might contain TSC
 388	 * packets that will not be fixed, and that will then no longer work for
 389	 * overlap detection. Avoid that by zeroing out the overlap.
 390	 */
 391	if (pt->synth_opts.vm_time_correlation)
 392		memset(b->data, 0, start - b->data);
 393	b->use_size = b->data + b->size - start;
 394	b->use_data = start;
 395	if (b->use_size && consecutive)
 396		b->consecutive = true;
 397	return 0;
 398}
 399
 400static int intel_pt_get_buffer(struct intel_pt_queue *ptq,
 401			       struct auxtrace_buffer *buffer,
 402			       struct auxtrace_buffer *old_buffer,
 403			       struct intel_pt_buffer *b)
 404{
 405	bool might_overlap;
 406
 407	if (!buffer->data) {
 408		int fd = perf_data__fd(ptq->pt->session->data);
 409
 410		buffer->data = auxtrace_buffer__get_data(buffer, fd);
 411		if (!buffer->data)
 412			return -ENOMEM;
 413	}
 414
 415	might_overlap = ptq->pt->snapshot_mode || ptq->pt->sampling_mode;
 416	if (might_overlap && !buffer->consecutive && old_buffer &&
 417	    intel_pt_do_fix_overlap(ptq->pt, old_buffer, buffer))
 418		return -ENOMEM;
 419
 420	if (buffer->use_data) {
 421		b->len = buffer->use_size;
 422		b->buf = buffer->use_data;
 423	} else {
 424		b->len = buffer->size;
 425		b->buf = buffer->data;
 426	}
 427	b->ref_timestamp = buffer->reference;
 428
 429	if (!old_buffer || (might_overlap && !buffer->consecutive)) {
 430		b->consecutive = false;
 431		b->trace_nr = buffer->buffer_nr + 1;
 432	} else {
 433		b->consecutive = true;
 434	}
 435
 436	return 0;
 437}
 438
 439/* Do not drop buffers with references - refer intel_pt_get_trace() */
 440static void intel_pt_lookahead_drop_buffer(struct intel_pt_queue *ptq,
 441					   struct auxtrace_buffer *buffer)
 442{
 443	if (!buffer || buffer == ptq->buffer || buffer == ptq->old_buffer)
 444		return;
 445
 446	auxtrace_buffer__drop_data(buffer);
 447}
 448
 449/* Must be serialized with respect to intel_pt_get_trace() */
 450static int intel_pt_lookahead(void *data, intel_pt_lookahead_cb_t cb,
 451			      void *cb_data)
 452{
 453	struct intel_pt_queue *ptq = data;
 454	struct auxtrace_buffer *buffer = ptq->buffer;
 455	struct auxtrace_buffer *old_buffer = ptq->old_buffer;
 456	struct auxtrace_queue *queue;
 457	int err = 0;
 458
 459	queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
 460
 461	while (1) {
 462		struct intel_pt_buffer b = { .len = 0 };
 463
 464		buffer = auxtrace_buffer__next(queue, buffer);
 465		if (!buffer)
 466			break;
 467
 468		err = intel_pt_get_buffer(ptq, buffer, old_buffer, &b);
 469		if (err)
 470			break;
 471
 472		if (b.len) {
 473			intel_pt_lookahead_drop_buffer(ptq, old_buffer);
 474			old_buffer = buffer;
 475		} else {
 476			intel_pt_lookahead_drop_buffer(ptq, buffer);
 477			continue;
 478		}
 479
 480		err = cb(&b, cb_data);
 481		if (err)
 482			break;
 483	}
 484
 485	if (buffer != old_buffer)
 486		intel_pt_lookahead_drop_buffer(ptq, buffer);
 487	intel_pt_lookahead_drop_buffer(ptq, old_buffer);
 488
 489	return err;
 490}
 491
 492/*
 493 * This function assumes data is processed sequentially only.
 494 * Must be serialized with respect to intel_pt_lookahead()
 495 */
 496static int intel_pt_get_trace(struct intel_pt_buffer *b, void *data)
 497{
 498	struct intel_pt_queue *ptq = data;
 499	struct auxtrace_buffer *buffer = ptq->buffer;
 500	struct auxtrace_buffer *old_buffer = ptq->old_buffer;
 501	struct auxtrace_queue *queue;
 502	int err;
 503
 504	if (ptq->stop) {
 505		b->len = 0;
 506		return 0;
 507	}
 508
 509	queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
 510
 511	buffer = auxtrace_buffer__next(queue, buffer);
 512	if (!buffer) {
 513		if (old_buffer)
 514			auxtrace_buffer__drop_data(old_buffer);
 515		b->len = 0;
 516		return 0;
 517	}
 518
 519	ptq->buffer = buffer;
 520
 521	err = intel_pt_get_buffer(ptq, buffer, old_buffer, b);
 522	if (err)
 523		return err;
 524
 525	if (ptq->step_through_buffers)
 526		ptq->stop = true;
 527
 528	if (b->len) {
 529		if (old_buffer)
 530			auxtrace_buffer__drop_data(old_buffer);
 531		ptq->old_buffer = buffer;
 532	} else {
 533		auxtrace_buffer__drop_data(buffer);
 534		return intel_pt_get_trace(b, data);
 535	}
 536
 537	return 0;
 538}
 539
 540struct intel_pt_cache_entry {
 541	struct auxtrace_cache_entry	entry;
 542	u64				insn_cnt;
 543	u64				byte_cnt;
 544	enum intel_pt_insn_op		op;
 545	enum intel_pt_insn_branch	branch;
 546	bool				emulated_ptwrite;
 547	int				length;
 548	int32_t				rel;
 549	char				insn[INTEL_PT_INSN_BUF_SZ];
 550};
 551
 552static int intel_pt_config_div(const char *var, const char *value, void *data)
 553{
 554	int *d = data;
 555	long val;
 556
 557	if (!strcmp(var, "intel-pt.cache-divisor")) {
 558		val = strtol(value, NULL, 0);
 559		if (val > 0 && val <= INT_MAX)
 560			*d = val;
 561	}
 562
 563	return 0;
 564}
 565
 566static int intel_pt_cache_divisor(void)
 567{
 568	static int d;
 569
 570	if (d)
 571		return d;
 572
 573	perf_config(intel_pt_config_div, &d);
 574
 575	if (!d)
 576		d = 64;
 577
 578	return d;
 579}
 580
 581static unsigned int intel_pt_cache_size(struct dso *dso,
 582					struct machine *machine)
 583{
 584	off_t size;
 585
 586	size = dso__data_size(dso, machine);
 587	size /= intel_pt_cache_divisor();
 588	if (size < 1000)
 589		return 10;
 590	if (size > (1 << 21))
 591		return 21;
 592	return 32 - __builtin_clz(size);
 593}
 594
 595static struct auxtrace_cache *intel_pt_cache(struct dso *dso,
 596					     struct machine *machine)
 597{
 598	struct auxtrace_cache *c;
 599	unsigned int bits;
 600
 601	if (dso__auxtrace_cache(dso))
 602		return dso__auxtrace_cache(dso);
 603
 604	bits = intel_pt_cache_size(dso, machine);
 605
 606	/* Ignoring cache creation failure */
 607	c = auxtrace_cache__new(bits, sizeof(struct intel_pt_cache_entry), 200);
 608
 609	dso__set_auxtrace_cache(dso, c);
 610
 611	return c;
 612}
 613
 614static int intel_pt_cache_add(struct dso *dso, struct machine *machine,
 615			      u64 offset, u64 insn_cnt, u64 byte_cnt,
 616			      struct intel_pt_insn *intel_pt_insn)
 617{
 618	struct auxtrace_cache *c = intel_pt_cache(dso, machine);
 619	struct intel_pt_cache_entry *e;
 620	int err;
 621
 622	if (!c)
 623		return -ENOMEM;
 624
 625	e = auxtrace_cache__alloc_entry(c);
 626	if (!e)
 627		return -ENOMEM;
 628
 629	e->insn_cnt = insn_cnt;
 630	e->byte_cnt = byte_cnt;
 631	e->op = intel_pt_insn->op;
 632	e->branch = intel_pt_insn->branch;
 633	e->emulated_ptwrite = intel_pt_insn->emulated_ptwrite;
 634	e->length = intel_pt_insn->length;
 635	e->rel = intel_pt_insn->rel;
 636	memcpy(e->insn, intel_pt_insn->buf, INTEL_PT_INSN_BUF_SZ);
 637
 638	err = auxtrace_cache__add(c, offset, &e->entry);
 639	if (err)
 640		auxtrace_cache__free_entry(c, e);
 641
 642	return err;
 643}
 644
 645static struct intel_pt_cache_entry *
 646intel_pt_cache_lookup(struct dso *dso, struct machine *machine, u64 offset)
 647{
 648	struct auxtrace_cache *c = intel_pt_cache(dso, machine);
 649
 650	if (!c)
 651		return NULL;
 652
 653	return auxtrace_cache__lookup(dso__auxtrace_cache(dso), offset);
 654}
 655
 656static void intel_pt_cache_invalidate(struct dso *dso, struct machine *machine,
 657				      u64 offset)
 658{
 659	struct auxtrace_cache *c = intel_pt_cache(dso, machine);
 660
 661	if (!c)
 662		return;
 663
 664	auxtrace_cache__remove(dso__auxtrace_cache(dso), offset);
 665}
 666
 667static inline bool intel_pt_guest_kernel_ip(uint64_t ip)
 668{
 669	/* Assumes 64-bit kernel */
 670	return ip & (1ULL << 63);
 671}
 672
 673static inline u8 intel_pt_nr_cpumode(struct intel_pt_queue *ptq, uint64_t ip, bool nr)
 674{
 675	if (nr) {
 676		return intel_pt_guest_kernel_ip(ip) ?
 677		       PERF_RECORD_MISC_GUEST_KERNEL :
 678		       PERF_RECORD_MISC_GUEST_USER;
 679	}
 680
 681	return ip >= ptq->pt->kernel_start ?
 682	       PERF_RECORD_MISC_KERNEL :
 683	       PERF_RECORD_MISC_USER;
 684}
 685
 686static inline u8 intel_pt_cpumode(struct intel_pt_queue *ptq, uint64_t from_ip, uint64_t to_ip)
 687{
 688	/* No support for non-zero CS base */
 689	if (from_ip)
 690		return intel_pt_nr_cpumode(ptq, from_ip, ptq->state->from_nr);
 691	return intel_pt_nr_cpumode(ptq, to_ip, ptq->state->to_nr);
 692}
 693
 694static int intel_pt_get_guest(struct intel_pt_queue *ptq)
 695{
 696	struct machines *machines = &ptq->pt->session->machines;
 697	struct machine *machine;
 698	pid_t pid = ptq->pid <= 0 ? DEFAULT_GUEST_KERNEL_ID : ptq->pid;
 699
 700	if (ptq->guest_machine && pid == ptq->guest_machine->pid)
 701		return 0;
 702
 703	ptq->guest_machine = NULL;
 704	thread__zput(ptq->unknown_guest_thread);
 705
 706	if (symbol_conf.guest_code) {
 707		thread__zput(ptq->guest_thread);
 708		ptq->guest_thread = machines__findnew_guest_code(machines, pid);
 709	}
 710
 711	machine = machines__find_guest(machines, pid);
 712	if (!machine)
 713		return -1;
 714
 715	ptq->unknown_guest_thread = machine__idle_thread(machine);
 716	if (!ptq->unknown_guest_thread)
 717		return -1;
 718
 719	ptq->guest_machine = machine;
 720
 721	return 0;
 722}
 723
 724static inline bool intel_pt_jmp_16(struct intel_pt_insn *intel_pt_insn)
 725{
 726	return intel_pt_insn->rel == 16 && intel_pt_insn->branch == INTEL_PT_BR_UNCONDITIONAL;
 727}
 728
 729#define PTWRITE_MAGIC		"\x0f\x0bperf,ptwrite  "
 730#define PTWRITE_MAGIC_LEN	16
 731
 732static bool intel_pt_emulated_ptwrite(struct dso *dso, struct machine *machine, u64 offset)
 733{
 734	unsigned char buf[PTWRITE_MAGIC_LEN];
 735	ssize_t len;
 736
 737	len = dso__data_read_offset(dso, machine, offset, buf, PTWRITE_MAGIC_LEN);
 738	if (len == PTWRITE_MAGIC_LEN && !memcmp(buf, PTWRITE_MAGIC, PTWRITE_MAGIC_LEN)) {
 739		intel_pt_log("Emulated ptwrite signature found\n");
 740		return true;
 741	}
 742	intel_pt_log("Emulated ptwrite signature not found\n");
 743	return false;
 744}
 745
 746static int intel_pt_walk_next_insn(struct intel_pt_insn *intel_pt_insn,
 747				   uint64_t *insn_cnt_ptr, uint64_t *ip,
 748				   uint64_t to_ip, uint64_t max_insn_cnt,
 749				   void *data)
 750{
 751	struct intel_pt_queue *ptq = data;
 752	struct machine *machine = ptq->pt->machine;
 753	struct thread *thread;
 754	struct addr_location al;
 755	unsigned char buf[INTEL_PT_INSN_BUF_SZ];
 756	ssize_t len;
 757	int x86_64, ret = 0;
 758	u8 cpumode;
 759	u64 offset, start_offset, start_ip;
 760	u64 insn_cnt = 0;
 761	bool one_map = true;
 762	bool nr;
 763
 764
 765	addr_location__init(&al);
 766	intel_pt_insn->length = 0;
 767	intel_pt_insn->op = INTEL_PT_OP_OTHER;
 768
 769	if (to_ip && *ip == to_ip)
 770		goto out_no_cache;
 771
 772	nr = ptq->state->to_nr;
 773	cpumode = intel_pt_nr_cpumode(ptq, *ip, nr);
 774
 775	if (nr) {
 776		if (ptq->pt->have_guest_sideband) {
 777			if (!ptq->guest_machine || ptq->guest_machine_pid != ptq->pid) {
 778				intel_pt_log("ERROR: guest sideband but no guest machine\n");
 779				ret = -EINVAL;
 780				goto out_ret;
 781			}
 782		} else if ((!symbol_conf.guest_code && cpumode != PERF_RECORD_MISC_GUEST_KERNEL) ||
 783			   intel_pt_get_guest(ptq)) {
 784			intel_pt_log("ERROR: no guest machine\n");
 785			ret = -EINVAL;
 786			goto out_ret;
 787		}
 788		machine = ptq->guest_machine;
 789		thread = ptq->guest_thread;
 790		if (!thread) {
 791			if (cpumode != PERF_RECORD_MISC_GUEST_KERNEL) {
 792				intel_pt_log("ERROR: no guest thread\n");
 793				ret = -EINVAL;
 794				goto out_ret;
 795			}
 796			thread = ptq->unknown_guest_thread;
 797		}
 798	} else {
 799		thread = ptq->thread;
 800		if (!thread) {
 801			if (cpumode != PERF_RECORD_MISC_KERNEL) {
 802				intel_pt_log("ERROR: no thread\n");
 803				ret = -EINVAL;
 804				goto out_ret;
 805			}
 806			thread = ptq->pt->unknown_thread;
 807		}
 808	}
 809
 810	while (1) {
 811		struct dso *dso;
 
 812
 813		if (!thread__find_map(thread, cpumode, *ip, &al) || !map__dso(al.map)) {
 814			if (al.map)
 815				intel_pt_log("ERROR: thread has no dso for %#" PRIx64 "\n", *ip);
 816			else
 817				intel_pt_log("ERROR: thread has no map for %#" PRIx64 "\n", *ip);
 818			addr_location__exit(&al);
 819			ret = -EINVAL;
 820			goto out_ret;
 821		}
 822		dso = map__dso(al.map);
 823
 824		if (dso__data(dso)->status == DSO_DATA_STATUS_ERROR &&
 825		    dso__data_status_seen(dso, DSO_DATA_STATUS_SEEN_ITRACE)) {
 826			ret = -ENOENT;
 827			goto out_ret;
 828		}
 829
 830		offset = map__map_ip(al.map, *ip);
 831
 832		if (!to_ip && one_map) {
 833			struct intel_pt_cache_entry *e;
 834
 835			e = intel_pt_cache_lookup(dso, machine, offset);
 836			if (e &&
 837			    (!max_insn_cnt || e->insn_cnt <= max_insn_cnt)) {
 838				*insn_cnt_ptr = e->insn_cnt;
 839				*ip += e->byte_cnt;
 840				intel_pt_insn->op = e->op;
 841				intel_pt_insn->branch = e->branch;
 842				intel_pt_insn->emulated_ptwrite = e->emulated_ptwrite;
 843				intel_pt_insn->length = e->length;
 844				intel_pt_insn->rel = e->rel;
 845				memcpy(intel_pt_insn->buf, e->insn, INTEL_PT_INSN_BUF_SZ);
 
 846				intel_pt_log_insn_no_data(intel_pt_insn, *ip);
 847				ret = 0;
 848				goto out_ret;
 849			}
 850		}
 851
 852		start_offset = offset;
 853		start_ip = *ip;
 854
 855		/* Load maps to ensure dso->is_64_bit has been updated */
 856		map__load(al.map);
 857
 858		x86_64 = dso__is_64_bit(dso);
 859
 860		while (1) {
 861			len = dso__data_read_offset(dso, machine,
 862						    offset, buf,
 863						    INTEL_PT_INSN_BUF_SZ);
 864			if (len <= 0) {
 865				intel_pt_log("ERROR: failed to read at offset %#" PRIx64 " ",
 866					     offset);
 867				if (intel_pt_enable_logging)
 868					dso__fprintf(dso, intel_pt_log_fp());
 869				ret = -EINVAL;
 870				goto out_ret;
 871			}
 872
 873			if (intel_pt_get_insn(buf, len, x86_64, intel_pt_insn)) {
 874				ret = -EINVAL;
 875				goto out_ret;
 876			}
 877
 878			intel_pt_log_insn(intel_pt_insn, *ip);
 879
 880			insn_cnt += 1;
 881
 882			if (intel_pt_insn->branch != INTEL_PT_BR_NO_BRANCH) {
 883				bool eptw;
 884				u64 offs;
 885
 886				if (!intel_pt_jmp_16(intel_pt_insn))
 887					goto out;
 888				/* Check for emulated ptwrite */
 889				offs = offset + intel_pt_insn->length;
 890				eptw = intel_pt_emulated_ptwrite(dso, machine, offs);
 891				intel_pt_insn->emulated_ptwrite = eptw;
 892				goto out;
 893			}
 894
 895			if (max_insn_cnt && insn_cnt >= max_insn_cnt)
 896				goto out_no_cache;
 897
 898			*ip += intel_pt_insn->length;
 899
 900			if (to_ip && *ip == to_ip) {
 901				intel_pt_insn->length = 0;
 902				intel_pt_insn->op = INTEL_PT_OP_OTHER;
 903				goto out_no_cache;
 904			}
 905
 906			if (*ip >= map__end(al.map))
 907				break;
 908
 909			offset += intel_pt_insn->length;
 910		}
 911		one_map = false;
 912	}
 913out:
 914	*insn_cnt_ptr = insn_cnt;
 915
 916	if (!one_map)
 917		goto out_no_cache;
 918
 919	/*
 920	 * Didn't lookup in the 'to_ip' case, so do it now to prevent duplicate
 921	 * entries.
 922	 */
 923	if (to_ip) {
 924		struct intel_pt_cache_entry *e;
 925
 926		e = intel_pt_cache_lookup(map__dso(al.map), machine, start_offset);
 927		if (e)
 928			goto out_ret;
 929	}
 930
 931	/* Ignore cache errors */
 932	intel_pt_cache_add(map__dso(al.map), machine, start_offset, insn_cnt,
 933			   *ip - start_ip, intel_pt_insn);
 934
 935out_ret:
 936	addr_location__exit(&al);
 937	return ret;
 938
 939out_no_cache:
 940	*insn_cnt_ptr = insn_cnt;
 941	addr_location__exit(&al);
 942	return 0;
 943}
 944
 945static bool intel_pt_match_pgd_ip(struct intel_pt *pt, uint64_t ip,
 946				  uint64_t offset, const char *filename)
 947{
 948	struct addr_filter *filt;
 949	bool have_filter   = false;
 950	bool hit_tracestop = false;
 951	bool hit_filter    = false;
 952
 953	list_for_each_entry(filt, &pt->filts.head, list) {
 954		if (filt->start)
 955			have_filter = true;
 956
 957		if ((filename && !filt->filename) ||
 958		    (!filename && filt->filename) ||
 959		    (filename && strcmp(filename, filt->filename)))
 960			continue;
 961
 962		if (!(offset >= filt->addr && offset < filt->addr + filt->size))
 963			continue;
 964
 965		intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s hit filter: %s offset %#"PRIx64" size %#"PRIx64"\n",
 966			     ip, offset, filename ? filename : "[kernel]",
 967			     filt->start ? "filter" : "stop",
 968			     filt->addr, filt->size);
 969
 970		if (filt->start)
 971			hit_filter = true;
 972		else
 973			hit_tracestop = true;
 974	}
 975
 976	if (!hit_tracestop && !hit_filter)
 977		intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s is not in a filter region\n",
 978			     ip, offset, filename ? filename : "[kernel]");
 979
 980	return hit_tracestop || (have_filter && !hit_filter);
 981}
 982
 983static int __intel_pt_pgd_ip(uint64_t ip, void *data)
 984{
 985	struct intel_pt_queue *ptq = data;
 986	struct thread *thread;
 987	struct addr_location al;
 988	u8 cpumode;
 989	u64 offset;
 990	int res;
 991
 992	if (ptq->state->to_nr) {
 993		if (intel_pt_guest_kernel_ip(ip))
 994			return intel_pt_match_pgd_ip(ptq->pt, ip, ip, NULL);
 995		/* No support for decoding guest user space */
 996		return -EINVAL;
 997	} else if (ip >= ptq->pt->kernel_start) {
 998		return intel_pt_match_pgd_ip(ptq->pt, ip, ip, NULL);
 999	}
1000
1001	cpumode = PERF_RECORD_MISC_USER;
1002
1003	thread = ptq->thread;
1004	if (!thread)
1005		return -EINVAL;
1006
1007	addr_location__init(&al);
1008	if (!thread__find_map(thread, cpumode, ip, &al) || !map__dso(al.map))
1009		return -EINVAL;
1010
1011	offset = map__map_ip(al.map, ip);
1012
1013	res = intel_pt_match_pgd_ip(ptq->pt, ip, offset, dso__long_name(map__dso(al.map)));
1014	addr_location__exit(&al);
1015	return res;
1016}
1017
1018static bool intel_pt_pgd_ip(uint64_t ip, void *data)
1019{
1020	return __intel_pt_pgd_ip(ip, data) > 0;
1021}
1022
1023static bool intel_pt_get_config(struct intel_pt *pt,
1024				struct perf_event_attr *attr, u64 *config)
1025{
1026	if (attr->type == pt->pmu_type) {
1027		if (config)
1028			*config = attr->config;
1029		return true;
1030	}
1031
1032	return false;
1033}
1034
1035static bool intel_pt_exclude_kernel(struct intel_pt *pt)
1036{
1037	struct evsel *evsel;
1038
1039	evlist__for_each_entry(pt->session->evlist, evsel) {
1040		if (intel_pt_get_config(pt, &evsel->core.attr, NULL) &&
1041		    !evsel->core.attr.exclude_kernel)
1042			return false;
1043	}
1044	return true;
1045}
1046
1047static bool intel_pt_return_compression(struct intel_pt *pt)
1048{
1049	struct evsel *evsel;
1050	u64 config;
1051
1052	if (!pt->noretcomp_bit)
1053		return true;
1054
1055	evlist__for_each_entry(pt->session->evlist, evsel) {
1056		if (intel_pt_get_config(pt, &evsel->core.attr, &config) &&
1057		    (config & pt->noretcomp_bit))
1058			return false;
1059	}
1060	return true;
1061}
1062
1063static bool intel_pt_branch_enable(struct intel_pt *pt)
1064{
1065	struct evsel *evsel;
1066	u64 config;
1067
1068	evlist__for_each_entry(pt->session->evlist, evsel) {
1069		if (intel_pt_get_config(pt, &evsel->core.attr, &config) &&
1070		    (config & INTEL_PT_CFG_PASS_THRU) &&
1071		    !(config & INTEL_PT_CFG_BRANCH_EN))
1072			return false;
1073	}
1074	return true;
1075}
1076
1077static bool intel_pt_disabled_tnt(struct intel_pt *pt)
1078{
1079	struct evsel *evsel;
1080	u64 config;
1081
1082	evlist__for_each_entry(pt->session->evlist, evsel) {
1083		if (intel_pt_get_config(pt, &evsel->core.attr, &config) &&
1084		    config & INTEL_PT_CFG_TNT_DIS)
1085			return true;
1086	}
1087	return false;
1088}
1089
1090static unsigned int intel_pt_mtc_period(struct intel_pt *pt)
1091{
1092	struct evsel *evsel;
1093	unsigned int shift;
1094	u64 config;
1095
1096	if (!pt->mtc_freq_bits)
1097		return 0;
1098
1099	for (shift = 0, config = pt->mtc_freq_bits; !(config & 1); shift++)
1100		config >>= 1;
1101
1102	evlist__for_each_entry(pt->session->evlist, evsel) {
1103		if (intel_pt_get_config(pt, &evsel->core.attr, &config))
1104			return (config & pt->mtc_freq_bits) >> shift;
1105	}
1106	return 0;
1107}
1108
1109static bool intel_pt_timeless_decoding(struct intel_pt *pt)
1110{
1111	struct evsel *evsel;
1112	bool timeless_decoding = true;
1113	u64 config;
1114
1115	if (!pt->tsc_bit || !pt->cap_user_time_zero || pt->synth_opts.timeless_decoding)
1116		return true;
1117
1118	evlist__for_each_entry(pt->session->evlist, evsel) {
1119		if (!(evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
1120			return true;
1121		if (intel_pt_get_config(pt, &evsel->core.attr, &config)) {
1122			if (config & pt->tsc_bit)
1123				timeless_decoding = false;
1124			else
1125				return true;
1126		}
1127	}
1128	return timeless_decoding;
1129}
1130
1131static bool intel_pt_tracing_kernel(struct intel_pt *pt)
1132{
1133	struct evsel *evsel;
1134
1135	evlist__for_each_entry(pt->session->evlist, evsel) {
1136		if (intel_pt_get_config(pt, &evsel->core.attr, NULL) &&
1137		    !evsel->core.attr.exclude_kernel)
1138			return true;
1139	}
1140	return false;
1141}
1142
1143static bool intel_pt_have_tsc(struct intel_pt *pt)
1144{
1145	struct evsel *evsel;
1146	bool have_tsc = false;
1147	u64 config;
1148
1149	if (!pt->tsc_bit)
1150		return false;
1151
1152	evlist__for_each_entry(pt->session->evlist, evsel) {
1153		if (intel_pt_get_config(pt, &evsel->core.attr, &config)) {
1154			if (config & pt->tsc_bit)
1155				have_tsc = true;
1156			else
1157				return false;
1158		}
1159	}
1160	return have_tsc;
1161}
1162
1163static bool intel_pt_have_mtc(struct intel_pt *pt)
1164{
1165	struct evsel *evsel;
1166	u64 config;
1167
1168	evlist__for_each_entry(pt->session->evlist, evsel) {
1169		if (intel_pt_get_config(pt, &evsel->core.attr, &config) &&
1170		    (config & pt->mtc_bit))
1171			return true;
1172	}
1173	return false;
1174}
1175
1176static bool intel_pt_sampling_mode(struct intel_pt *pt)
1177{
1178	struct evsel *evsel;
1179
1180	evlist__for_each_entry(pt->session->evlist, evsel) {
1181		if ((evsel->core.attr.sample_type & PERF_SAMPLE_AUX) &&
1182		    evsel->core.attr.aux_sample_size)
1183			return true;
1184	}
1185	return false;
1186}
1187
1188static u64 intel_pt_ctl(struct intel_pt *pt)
1189{
1190	struct evsel *evsel;
1191	u64 config;
1192
1193	evlist__for_each_entry(pt->session->evlist, evsel) {
1194		if (intel_pt_get_config(pt, &evsel->core.attr, &config))
1195			return config;
1196	}
1197	return 0;
1198}
1199
1200static u64 intel_pt_ns_to_ticks(const struct intel_pt *pt, u64 ns)
1201{
1202	u64 quot, rem;
1203
1204	quot = ns / pt->tc.time_mult;
1205	rem  = ns % pt->tc.time_mult;
1206	return (quot << pt->tc.time_shift) + (rem << pt->tc.time_shift) /
1207		pt->tc.time_mult;
1208}
1209
1210static struct ip_callchain *intel_pt_alloc_chain(struct intel_pt *pt)
1211{
1212	size_t sz = sizeof(struct ip_callchain);
1213
1214	/* Add 1 to callchain_sz for callchain context */
1215	sz += (pt->synth_opts.callchain_sz + 1) * sizeof(u64);
1216	return zalloc(sz);
1217}
1218
1219static int intel_pt_callchain_init(struct intel_pt *pt)
1220{
1221	struct evsel *evsel;
1222
1223	evlist__for_each_entry(pt->session->evlist, evsel) {
1224		if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CALLCHAIN))
1225			evsel->synth_sample_type |= PERF_SAMPLE_CALLCHAIN;
1226	}
1227
1228	pt->chain = intel_pt_alloc_chain(pt);
1229	if (!pt->chain)
1230		return -ENOMEM;
1231
1232	return 0;
1233}
1234
1235static void intel_pt_add_callchain(struct intel_pt *pt,
1236				   struct perf_sample *sample)
1237{
1238	struct thread *thread = machine__findnew_thread(pt->machine,
1239							sample->pid,
1240							sample->tid);
1241
1242	thread_stack__sample_late(thread, sample->cpu, pt->chain,
1243				  pt->synth_opts.callchain_sz + 1, sample->ip,
1244				  pt->kernel_start);
1245
1246	sample->callchain = pt->chain;
1247}
1248
1249static struct branch_stack *intel_pt_alloc_br_stack(unsigned int entry_cnt)
1250{
1251	size_t sz = sizeof(struct branch_stack);
1252
1253	sz += entry_cnt * sizeof(struct branch_entry);
1254	return zalloc(sz);
1255}
1256
1257static int intel_pt_br_stack_init(struct intel_pt *pt)
1258{
1259	struct evsel *evsel;
1260
1261	evlist__for_each_entry(pt->session->evlist, evsel) {
1262		if (!(evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK))
1263			evsel->synth_sample_type |= PERF_SAMPLE_BRANCH_STACK;
1264	}
1265
1266	pt->br_stack = intel_pt_alloc_br_stack(pt->br_stack_sz);
1267	if (!pt->br_stack)
1268		return -ENOMEM;
1269
1270	return 0;
1271}
1272
1273static void intel_pt_add_br_stack(struct intel_pt *pt,
1274				  struct perf_sample *sample)
1275{
1276	struct thread *thread = machine__findnew_thread(pt->machine,
1277							sample->pid,
1278							sample->tid);
1279
1280	thread_stack__br_sample_late(thread, sample->cpu, pt->br_stack,
1281				     pt->br_stack_sz, sample->ip,
1282				     pt->kernel_start);
1283
1284	sample->branch_stack = pt->br_stack;
1285	thread__put(thread);
1286}
1287
1288/* INTEL_PT_LBR_0, INTEL_PT_LBR_1 and INTEL_PT_LBR_2 */
1289#define LBRS_MAX (INTEL_PT_BLK_ITEM_ID_CNT * 3U)
1290
1291static struct intel_pt_queue *intel_pt_alloc_queue(struct intel_pt *pt,
1292						   unsigned int queue_nr)
1293{
1294	struct intel_pt_params params = { .get_trace = 0, };
1295	struct perf_env *env = pt->machine->env;
1296	struct intel_pt_queue *ptq;
1297
1298	ptq = zalloc(sizeof(struct intel_pt_queue));
1299	if (!ptq)
1300		return NULL;
1301
1302	if (pt->synth_opts.callchain) {
1303		ptq->chain = intel_pt_alloc_chain(pt);
1304		if (!ptq->chain)
1305			goto out_free;
1306	}
1307
1308	if (pt->synth_opts.last_branch || pt->synth_opts.other_events) {
1309		unsigned int entry_cnt = max(LBRS_MAX, pt->br_stack_sz);
1310
1311		ptq->last_branch = intel_pt_alloc_br_stack(entry_cnt);
1312		if (!ptq->last_branch)
1313			goto out_free;
1314	}
1315
1316	ptq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
1317	if (!ptq->event_buf)
1318		goto out_free;
1319
1320	ptq->pt = pt;
1321	ptq->queue_nr = queue_nr;
1322	ptq->exclude_kernel = intel_pt_exclude_kernel(pt);
1323	ptq->pid = -1;
1324	ptq->tid = -1;
1325	ptq->cpu = -1;
1326	ptq->next_tid = -1;
1327
1328	params.get_trace = intel_pt_get_trace;
1329	params.walk_insn = intel_pt_walk_next_insn;
1330	params.lookahead = intel_pt_lookahead;
1331	params.findnew_vmcs_info = intel_pt_findnew_vmcs_info;
1332	params.data = ptq;
1333	params.return_compression = intel_pt_return_compression(pt);
1334	params.branch_enable = intel_pt_branch_enable(pt);
1335	params.ctl = intel_pt_ctl(pt);
1336	params.max_non_turbo_ratio = pt->max_non_turbo_ratio;
1337	params.mtc_period = intel_pt_mtc_period(pt);
1338	params.tsc_ctc_ratio_n = pt->tsc_ctc_ratio_n;
1339	params.tsc_ctc_ratio_d = pt->tsc_ctc_ratio_d;
1340	params.quick = pt->synth_opts.quick;
1341	params.vm_time_correlation = pt->synth_opts.vm_time_correlation;
1342	params.vm_tm_corr_dry_run = pt->synth_opts.vm_tm_corr_dry_run;
1343	params.first_timestamp = pt->first_timestamp;
1344	params.max_loops = pt->max_loops;
1345
1346	/* Cannot walk code without TNT, so force 'quick' mode */
1347	if (params.branch_enable && intel_pt_disabled_tnt(pt) && !params.quick)
1348		params.quick = 1;
1349
1350	if (pt->filts.cnt > 0)
1351		params.pgd_ip = intel_pt_pgd_ip;
1352
1353	if (pt->synth_opts.instructions || pt->synth_opts.cycles) {
1354		if (pt->synth_opts.period) {
1355			switch (pt->synth_opts.period_type) {
1356			case PERF_ITRACE_PERIOD_INSTRUCTIONS:
1357				params.period_type =
1358						INTEL_PT_PERIOD_INSTRUCTIONS;
1359				params.period = pt->synth_opts.period;
1360				break;
1361			case PERF_ITRACE_PERIOD_TICKS:
1362				params.period_type = INTEL_PT_PERIOD_TICKS;
1363				params.period = pt->synth_opts.period;
1364				break;
1365			case PERF_ITRACE_PERIOD_NANOSECS:
1366				params.period_type = INTEL_PT_PERIOD_TICKS;
1367				params.period = intel_pt_ns_to_ticks(pt,
1368							pt->synth_opts.period);
1369				break;
1370			default:
1371				break;
1372			}
1373		}
1374
1375		if (!params.period) {
1376			params.period_type = INTEL_PT_PERIOD_INSTRUCTIONS;
1377			params.period = 1;
1378		}
1379	}
1380
1381	if (env->cpuid && !strncmp(env->cpuid, "GenuineIntel,6,92,", 18))
1382		params.flags |= INTEL_PT_FUP_WITH_NLIP;
1383
1384	ptq->decoder = intel_pt_decoder_new(&params);
1385	if (!ptq->decoder)
1386		goto out_free;
1387
1388	return ptq;
1389
1390out_free:
1391	zfree(&ptq->event_buf);
1392	zfree(&ptq->last_branch);
1393	zfree(&ptq->chain);
1394	free(ptq);
1395	return NULL;
1396}
1397
1398static void intel_pt_free_queue(void *priv)
1399{
1400	struct intel_pt_queue *ptq = priv;
1401
1402	if (!ptq)
1403		return;
1404	thread__zput(ptq->thread);
1405	thread__zput(ptq->guest_thread);
1406	thread__zput(ptq->unknown_guest_thread);
1407	intel_pt_decoder_free(ptq->decoder);
1408	zfree(&ptq->event_buf);
1409	zfree(&ptq->last_branch);
1410	zfree(&ptq->chain);
1411	free(ptq);
1412}
1413
1414static void intel_pt_first_timestamp(struct intel_pt *pt, u64 timestamp)
1415{
1416	unsigned int i;
1417
1418	pt->first_timestamp = timestamp;
1419
1420	for (i = 0; i < pt->queues.nr_queues; i++) {
1421		struct auxtrace_queue *queue = &pt->queues.queue_array[i];
1422		struct intel_pt_queue *ptq = queue->priv;
1423
1424		if (ptq && ptq->decoder)
1425			intel_pt_set_first_timestamp(ptq->decoder, timestamp);
1426	}
1427}
1428
1429static int intel_pt_get_guest_from_sideband(struct intel_pt_queue *ptq)
1430{
1431	struct machines *machines = &ptq->pt->session->machines;
1432	struct machine *machine;
1433	pid_t machine_pid = ptq->pid;
1434	pid_t tid;
1435	int vcpu;
1436
1437	if (machine_pid <= 0)
1438		return 0; /* Not a guest machine */
1439
1440	machine = machines__find(machines, machine_pid);
1441	if (!machine)
1442		return 0; /* Not a guest machine */
1443
1444	if (ptq->guest_machine != machine) {
1445		ptq->guest_machine = NULL;
1446		thread__zput(ptq->guest_thread);
1447		thread__zput(ptq->unknown_guest_thread);
1448
1449		ptq->unknown_guest_thread = machine__find_thread(machine, 0, 0);
1450		if (!ptq->unknown_guest_thread)
1451			return -1;
1452		ptq->guest_machine = machine;
1453	}
1454
1455	vcpu = ptq->thread ? thread__guest_cpu(ptq->thread) : -1;
1456	if (vcpu < 0)
1457		return -1;
1458
1459	tid = machine__get_current_tid(machine, vcpu);
1460
1461	if (ptq->guest_thread && thread__tid(ptq->guest_thread) != tid)
1462		thread__zput(ptq->guest_thread);
1463
1464	if (!ptq->guest_thread) {
1465		ptq->guest_thread = machine__find_thread(machine, -1, tid);
1466		if (!ptq->guest_thread)
1467			return -1;
1468	}
1469
1470	ptq->guest_machine_pid = machine_pid;
1471	ptq->guest_pid = thread__pid(ptq->guest_thread);
1472	ptq->guest_tid = tid;
1473	ptq->vcpu = vcpu;
1474
1475	return 0;
1476}
1477
1478static void intel_pt_set_pid_tid_cpu(struct intel_pt *pt,
1479				     struct auxtrace_queue *queue)
1480{
1481	struct intel_pt_queue *ptq = queue->priv;
1482
1483	if (queue->tid == -1 || pt->have_sched_switch) {
1484		ptq->tid = machine__get_current_tid(pt->machine, ptq->cpu);
1485		if (ptq->tid == -1)
1486			ptq->pid = -1;
1487		thread__zput(ptq->thread);
1488	}
1489
1490	if (!ptq->thread && ptq->tid != -1)
1491		ptq->thread = machine__find_thread(pt->machine, -1, ptq->tid);
1492
1493	if (ptq->thread) {
1494		ptq->pid = thread__pid(ptq->thread);
1495		if (queue->cpu == -1)
1496			ptq->cpu = thread__cpu(ptq->thread);
1497	}
1498
1499	if (pt->have_guest_sideband && intel_pt_get_guest_from_sideband(ptq)) {
1500		ptq->guest_machine_pid = 0;
1501		ptq->guest_pid = -1;
1502		ptq->guest_tid = -1;
1503		ptq->vcpu = -1;
1504	}
1505}
1506
1507static void intel_pt_sample_flags(struct intel_pt_queue *ptq)
1508{
1509	struct intel_pt *pt = ptq->pt;
1510
1511	ptq->insn_len = 0;
1512	if (ptq->state->flags & INTEL_PT_ABORT_TX) {
1513		ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_TX_ABORT;
1514	} else if (ptq->state->flags & INTEL_PT_ASYNC) {
1515		if (!ptq->state->to_ip)
1516			ptq->flags = PERF_IP_FLAG_BRANCH |
1517				     PERF_IP_FLAG_ASYNC |
1518				     PERF_IP_FLAG_TRACE_END;
1519		else if (ptq->state->from_nr && !ptq->state->to_nr)
1520			ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_CALL |
1521				     PERF_IP_FLAG_ASYNC |
1522				     PERF_IP_FLAG_VMEXIT;
1523		else
1524			ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_CALL |
1525				     PERF_IP_FLAG_ASYNC |
1526				     PERF_IP_FLAG_INTERRUPT;
 
 
 
 
1527	} else {
1528		if (ptq->state->from_ip)
1529			ptq->flags = intel_pt_insn_type(ptq->state->insn_op);
1530		else
1531			ptq->flags = PERF_IP_FLAG_BRANCH |
1532				     PERF_IP_FLAG_TRACE_BEGIN;
1533		if (ptq->state->flags & INTEL_PT_IN_TX)
1534			ptq->flags |= PERF_IP_FLAG_IN_TX;
1535		ptq->insn_len = ptq->state->insn_len;
1536		memcpy(ptq->insn, ptq->state->insn, INTEL_PT_INSN_BUF_SZ);
1537	}
1538
1539	if (ptq->state->type & INTEL_PT_TRACE_BEGIN)
1540		ptq->flags |= PERF_IP_FLAG_TRACE_BEGIN;
1541	if (ptq->state->type & INTEL_PT_TRACE_END)
1542		ptq->flags |= PERF_IP_FLAG_TRACE_END;
1543
1544	if (pt->cap_event_trace) {
1545		if (ptq->state->type & INTEL_PT_IFLAG_CHG) {
1546			if (!ptq->state->from_iflag)
1547				ptq->flags |= PERF_IP_FLAG_INTR_DISABLE;
1548			if (ptq->state->from_iflag != ptq->state->to_iflag)
1549				ptq->flags |= PERF_IP_FLAG_INTR_TOGGLE;
1550		} else if (!ptq->state->to_iflag) {
1551			ptq->flags |= PERF_IP_FLAG_INTR_DISABLE;
1552		}
1553	}
1554}
1555
1556static void intel_pt_setup_time_range(struct intel_pt *pt,
1557				      struct intel_pt_queue *ptq)
1558{
1559	if (!pt->range_cnt)
1560		return;
1561
1562	ptq->sel_timestamp = pt->time_ranges[0].start;
1563	ptq->sel_idx = 0;
1564
1565	if (ptq->sel_timestamp) {
1566		ptq->sel_start = true;
1567	} else {
1568		ptq->sel_timestamp = pt->time_ranges[0].end;
1569		ptq->sel_start = false;
1570	}
1571}
1572
1573static int intel_pt_setup_queue(struct intel_pt *pt,
1574				struct auxtrace_queue *queue,
1575				unsigned int queue_nr)
1576{
1577	struct intel_pt_queue *ptq = queue->priv;
1578
1579	if (list_empty(&queue->head))
1580		return 0;
1581
1582	if (!ptq) {
1583		ptq = intel_pt_alloc_queue(pt, queue_nr);
1584		if (!ptq)
1585			return -ENOMEM;
1586		queue->priv = ptq;
1587
1588		if (queue->cpu != -1)
1589			ptq->cpu = queue->cpu;
1590		ptq->tid = queue->tid;
1591
1592		ptq->cbr_seen = UINT_MAX;
1593
1594		if (pt->sampling_mode && !pt->snapshot_mode &&
1595		    pt->timeless_decoding)
1596			ptq->step_through_buffers = true;
1597
1598		ptq->sync_switch = pt->sync_switch;
1599
1600		intel_pt_setup_time_range(pt, ptq);
1601	}
1602
1603	if (!ptq->on_heap &&
1604	    (!ptq->sync_switch ||
1605	     ptq->switch_state != INTEL_PT_SS_EXPECTING_SWITCH_EVENT)) {
1606		const struct intel_pt_state *state;
1607		int ret;
1608
1609		if (pt->timeless_decoding)
1610			return 0;
1611
1612		intel_pt_log("queue %u getting timestamp\n", queue_nr);
1613		intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
1614			     queue_nr, ptq->cpu, ptq->pid, ptq->tid);
1615
1616		if (ptq->sel_start && ptq->sel_timestamp) {
1617			ret = intel_pt_fast_forward(ptq->decoder,
1618						    ptq->sel_timestamp);
1619			if (ret)
1620				return ret;
1621		}
1622
1623		while (1) {
1624			state = intel_pt_decode(ptq->decoder);
1625			if (state->err) {
1626				if (state->err == INTEL_PT_ERR_NODATA) {
1627					intel_pt_log("queue %u has no timestamp\n",
1628						     queue_nr);
1629					return 0;
1630				}
1631				continue;
1632			}
1633			if (state->timestamp)
1634				break;
1635		}
1636
1637		ptq->timestamp = state->timestamp;
1638		intel_pt_log("queue %u timestamp 0x%" PRIx64 "\n",
1639			     queue_nr, ptq->timestamp);
1640		ptq->state = state;
1641		ptq->have_sample = true;
1642		if (ptq->sel_start && ptq->sel_timestamp &&
1643		    ptq->timestamp < ptq->sel_timestamp)
1644			ptq->have_sample = false;
1645		intel_pt_sample_flags(ptq);
1646		ret = auxtrace_heap__add(&pt->heap, queue_nr, ptq->timestamp);
1647		if (ret)
1648			return ret;
1649		ptq->on_heap = true;
1650	}
1651
1652	return 0;
1653}
1654
1655static int intel_pt_setup_queues(struct intel_pt *pt)
1656{
1657	unsigned int i;
1658	int ret;
1659
1660	for (i = 0; i < pt->queues.nr_queues; i++) {
1661		ret = intel_pt_setup_queue(pt, &pt->queues.queue_array[i], i);
1662		if (ret)
1663			return ret;
1664	}
1665	return 0;
1666}
1667
1668static inline bool intel_pt_skip_event(struct intel_pt *pt)
1669{
1670	return pt->synth_opts.initial_skip &&
1671	       pt->num_events++ < pt->synth_opts.initial_skip;
1672}
1673
1674/*
1675 * Cannot count CBR as skipped because it won't go away until cbr == cbr_seen.
1676 * Also ensure CBR is first non-skipped event by allowing for 4 more samples
1677 * from this decoder state.
1678 */
1679static inline bool intel_pt_skip_cbr_event(struct intel_pt *pt)
1680{
1681	return pt->synth_opts.initial_skip &&
1682	       pt->num_events + 4 < pt->synth_opts.initial_skip;
1683}
1684
1685static void intel_pt_prep_a_sample(struct intel_pt_queue *ptq,
1686				   union perf_event *event,
1687				   struct perf_sample *sample)
1688{
1689	event->sample.header.type = PERF_RECORD_SAMPLE;
1690	event->sample.header.size = sizeof(struct perf_event_header);
1691
1692	sample->pid = ptq->pid;
1693	sample->tid = ptq->tid;
1694
1695	if (ptq->pt->have_guest_sideband) {
1696		if ((ptq->state->from_ip && ptq->state->from_nr) ||
1697		    (ptq->state->to_ip && ptq->state->to_nr)) {
1698			sample->pid = ptq->guest_pid;
1699			sample->tid = ptq->guest_tid;
1700			sample->machine_pid = ptq->guest_machine_pid;
1701			sample->vcpu = ptq->vcpu;
1702		}
1703	}
1704
1705	sample->cpu = ptq->cpu;
1706	sample->insn_len = ptq->insn_len;
1707	memcpy(sample->insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1708}
1709
1710static void intel_pt_prep_b_sample(struct intel_pt *pt,
1711				   struct intel_pt_queue *ptq,
1712				   union perf_event *event,
1713				   struct perf_sample *sample)
1714{
1715	intel_pt_prep_a_sample(ptq, event, sample);
1716
1717	if (!pt->timeless_decoding)
1718		sample->time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1719
1720	sample->ip = ptq->state->from_ip;
 
1721	sample->addr = ptq->state->to_ip;
1722	sample->cpumode = intel_pt_cpumode(ptq, sample->ip, sample->addr);
1723	sample->period = 1;
1724	sample->flags = ptq->flags;
1725
1726	event->sample.header.misc = sample->cpumode;
1727}
1728
1729static int intel_pt_inject_event(union perf_event *event,
1730				 struct perf_sample *sample, u64 type)
1731{
1732	event->header.size = perf_event__sample_event_size(sample, type, 0);
1733	return perf_event__synthesize_sample(event, type, 0, sample);
1734}
1735
1736static inline int intel_pt_opt_inject(struct intel_pt *pt,
1737				      union perf_event *event,
1738				      struct perf_sample *sample, u64 type)
1739{
1740	if (!pt->synth_opts.inject)
1741		return 0;
1742
1743	return intel_pt_inject_event(event, sample, type);
1744}
1745
1746static int intel_pt_deliver_synth_event(struct intel_pt *pt,
1747					union perf_event *event,
1748					struct perf_sample *sample, u64 type)
1749{
1750	int ret;
1751
1752	ret = intel_pt_opt_inject(pt, event, sample, type);
1753	if (ret)
1754		return ret;
1755
1756	ret = perf_session__deliver_synth_event(pt->session, event, sample);
1757	if (ret)
1758		pr_err("Intel PT: failed to deliver event, error %d\n", ret);
1759
1760	return ret;
1761}
1762
1763static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq)
1764{
1765	struct intel_pt *pt = ptq->pt;
1766	union perf_event *event = ptq->event_buf;
1767	struct perf_sample sample = { .ip = 0, };
1768	struct dummy_branch_stack {
1769		u64			nr;
1770		u64			hw_idx;
1771		struct branch_entry	entries;
1772	} dummy_bs;
1773
1774	if (pt->branches_filter && !(pt->branches_filter & ptq->flags))
1775		return 0;
1776
1777	if (intel_pt_skip_event(pt))
1778		return 0;
1779
1780	intel_pt_prep_b_sample(pt, ptq, event, &sample);
1781
1782	sample.id = ptq->pt->branches_id;
1783	sample.stream_id = ptq->pt->branches_id;
1784
1785	/*
1786	 * perf report cannot handle events without a branch stack when using
1787	 * SORT_MODE__BRANCH so make a dummy one.
1788	 */
1789	if (pt->synth_opts.last_branch && sort__mode == SORT_MODE__BRANCH) {
1790		dummy_bs = (struct dummy_branch_stack){
1791			.nr = 1,
1792			.hw_idx = -1ULL,
1793			.entries = {
1794				.from = sample.ip,
1795				.to = sample.addr,
1796			},
1797		};
1798		sample.branch_stack = (struct branch_stack *)&dummy_bs;
1799	}
1800
1801	if (ptq->sample_ipc)
1802		sample.cyc_cnt = ptq->ipc_cyc_cnt - ptq->last_br_cyc_cnt;
1803	if (sample.cyc_cnt) {
1804		sample.insn_cnt = ptq->ipc_insn_cnt - ptq->last_br_insn_cnt;
1805		ptq->last_br_insn_cnt = ptq->ipc_insn_cnt;
1806		ptq->last_br_cyc_cnt = ptq->ipc_cyc_cnt;
1807	}
1808
1809	return intel_pt_deliver_synth_event(pt, event, &sample,
1810					    pt->branches_sample_type);
1811}
1812
1813static void intel_pt_prep_sample(struct intel_pt *pt,
1814				 struct intel_pt_queue *ptq,
1815				 union perf_event *event,
1816				 struct perf_sample *sample)
1817{
1818	intel_pt_prep_b_sample(pt, ptq, event, sample);
1819
1820	if (pt->synth_opts.callchain) {
1821		thread_stack__sample(ptq->thread, ptq->cpu, ptq->chain,
1822				     pt->synth_opts.callchain_sz + 1,
1823				     sample->ip, pt->kernel_start);
1824		sample->callchain = ptq->chain;
1825	}
1826
1827	if (pt->synth_opts.last_branch) {
1828		thread_stack__br_sample(ptq->thread, ptq->cpu, ptq->last_branch,
1829					pt->br_stack_sz);
1830		sample->branch_stack = ptq->last_branch;
1831	}
1832}
1833
1834static int intel_pt_synth_instruction_sample(struct intel_pt_queue *ptq)
1835{
1836	struct intel_pt *pt = ptq->pt;
1837	union perf_event *event = ptq->event_buf;
1838	struct perf_sample sample = { .ip = 0, };
1839
1840	if (intel_pt_skip_event(pt))
1841		return 0;
1842
1843	intel_pt_prep_sample(pt, ptq, event, &sample);
1844
1845	sample.id = ptq->pt->instructions_id;
1846	sample.stream_id = ptq->pt->instructions_id;
1847	if (pt->synth_opts.quick)
1848		sample.period = 1;
1849	else
1850		sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt;
1851
1852	if (ptq->sample_ipc)
1853		sample.cyc_cnt = ptq->ipc_cyc_cnt - ptq->last_in_cyc_cnt;
1854	if (sample.cyc_cnt) {
1855		sample.insn_cnt = ptq->ipc_insn_cnt - ptq->last_in_insn_cnt;
1856		ptq->last_in_insn_cnt = ptq->ipc_insn_cnt;
1857		ptq->last_in_cyc_cnt = ptq->ipc_cyc_cnt;
1858	}
1859
1860	ptq->last_insn_cnt = ptq->state->tot_insn_cnt;
1861
1862	return intel_pt_deliver_synth_event(pt, event, &sample,
1863					    pt->instructions_sample_type);
1864}
1865
1866static int intel_pt_synth_cycle_sample(struct intel_pt_queue *ptq)
1867{
1868	struct intel_pt *pt = ptq->pt;
1869	union perf_event *event = ptq->event_buf;
1870	struct perf_sample sample = { .ip = 0, };
1871	u64 period = 0;
1872
1873	if (ptq->sample_ipc)
1874		period = ptq->ipc_cyc_cnt - ptq->last_cy_cyc_cnt;
1875
1876	if (!period || intel_pt_skip_event(pt))
1877		return 0;
1878
1879	intel_pt_prep_sample(pt, ptq, event, &sample);
1880
1881	sample.id = ptq->pt->cycles_id;
1882	sample.stream_id = ptq->pt->cycles_id;
1883	sample.period = period;
1884
1885	sample.cyc_cnt = period;
1886	sample.insn_cnt = ptq->ipc_insn_cnt - ptq->last_cy_insn_cnt;
1887	ptq->last_cy_insn_cnt = ptq->ipc_insn_cnt;
1888	ptq->last_cy_cyc_cnt = ptq->ipc_cyc_cnt;
1889
1890	return intel_pt_deliver_synth_event(pt, event, &sample, pt->cycles_sample_type);
1891}
1892
1893static int intel_pt_synth_transaction_sample(struct intel_pt_queue *ptq)
1894{
1895	struct intel_pt *pt = ptq->pt;
1896	union perf_event *event = ptq->event_buf;
1897	struct perf_sample sample = { .ip = 0, };
1898
1899	if (intel_pt_skip_event(pt))
1900		return 0;
1901
1902	intel_pt_prep_sample(pt, ptq, event, &sample);
1903
1904	sample.id = ptq->pt->transactions_id;
1905	sample.stream_id = ptq->pt->transactions_id;
1906
1907	return intel_pt_deliver_synth_event(pt, event, &sample,
1908					    pt->transactions_sample_type);
1909}
1910
1911static void intel_pt_prep_p_sample(struct intel_pt *pt,
1912				   struct intel_pt_queue *ptq,
1913				   union perf_event *event,
1914				   struct perf_sample *sample)
1915{
1916	intel_pt_prep_sample(pt, ptq, event, sample);
1917
1918	/*
1919	 * Zero IP is used to mean "trace start" but that is not the case for
1920	 * power or PTWRITE events with no IP, so clear the flags.
1921	 */
1922	if (!sample->ip)
1923		sample->flags = 0;
1924}
1925
1926static int intel_pt_synth_ptwrite_sample(struct intel_pt_queue *ptq)
1927{
1928	struct intel_pt *pt = ptq->pt;
1929	union perf_event *event = ptq->event_buf;
1930	struct perf_sample sample = { .ip = 0, };
1931	struct perf_synth_intel_ptwrite raw;
1932
1933	if (intel_pt_skip_event(pt))
1934		return 0;
1935
1936	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1937
1938	sample.id = ptq->pt->ptwrites_id;
1939	sample.stream_id = ptq->pt->ptwrites_id;
1940
1941	raw.flags = 0;
1942	raw.ip = !!(ptq->state->flags & INTEL_PT_FUP_IP);
1943	raw.payload = cpu_to_le64(ptq->state->ptw_payload);
1944
1945	sample.raw_size = perf_synth__raw_size(raw);
1946	sample.raw_data = perf_synth__raw_data(&raw);
1947
1948	return intel_pt_deliver_synth_event(pt, event, &sample,
1949					    pt->ptwrites_sample_type);
1950}
1951
1952static int intel_pt_synth_cbr_sample(struct intel_pt_queue *ptq)
1953{
1954	struct intel_pt *pt = ptq->pt;
1955	union perf_event *event = ptq->event_buf;
1956	struct perf_sample sample = { .ip = 0, };
1957	struct perf_synth_intel_cbr raw;
1958	u32 flags;
1959
1960	if (intel_pt_skip_cbr_event(pt))
1961		return 0;
1962
1963	ptq->cbr_seen = ptq->state->cbr;
1964
1965	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1966
1967	sample.id = ptq->pt->cbr_id;
1968	sample.stream_id = ptq->pt->cbr_id;
1969
1970	flags = (u16)ptq->state->cbr_payload | (pt->max_non_turbo_ratio << 16);
1971	raw.flags = cpu_to_le32(flags);
1972	raw.freq = cpu_to_le32(raw.cbr * pt->cbr2khz);
1973	raw.reserved3 = 0;
1974
1975	sample.raw_size = perf_synth__raw_size(raw);
1976	sample.raw_data = perf_synth__raw_data(&raw);
1977
1978	return intel_pt_deliver_synth_event(pt, event, &sample,
1979					    pt->pwr_events_sample_type);
1980}
1981
1982static int intel_pt_synth_psb_sample(struct intel_pt_queue *ptq)
1983{
1984	struct intel_pt *pt = ptq->pt;
1985	union perf_event *event = ptq->event_buf;
1986	struct perf_sample sample = { .ip = 0, };
1987	struct perf_synth_intel_psb raw;
1988
1989	if (intel_pt_skip_event(pt))
1990		return 0;
1991
1992	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1993
1994	sample.id = ptq->pt->psb_id;
1995	sample.stream_id = ptq->pt->psb_id;
1996	sample.flags = 0;
1997
1998	raw.reserved = 0;
1999	raw.offset = ptq->state->psb_offset;
2000
2001	sample.raw_size = perf_synth__raw_size(raw);
2002	sample.raw_data = perf_synth__raw_data(&raw);
2003
2004	return intel_pt_deliver_synth_event(pt, event, &sample,
2005					    pt->pwr_events_sample_type);
2006}
2007
2008static int intel_pt_synth_mwait_sample(struct intel_pt_queue *ptq)
2009{
2010	struct intel_pt *pt = ptq->pt;
2011	union perf_event *event = ptq->event_buf;
2012	struct perf_sample sample = { .ip = 0, };
2013	struct perf_synth_intel_mwait raw;
2014
2015	if (intel_pt_skip_event(pt))
2016		return 0;
2017
2018	intel_pt_prep_p_sample(pt, ptq, event, &sample);
2019
2020	sample.id = ptq->pt->mwait_id;
2021	sample.stream_id = ptq->pt->mwait_id;
2022
2023	raw.reserved = 0;
2024	raw.payload = cpu_to_le64(ptq->state->mwait_payload);
2025
2026	sample.raw_size = perf_synth__raw_size(raw);
2027	sample.raw_data = perf_synth__raw_data(&raw);
2028
2029	return intel_pt_deliver_synth_event(pt, event, &sample,
2030					    pt->pwr_events_sample_type);
2031}
2032
2033static int intel_pt_synth_pwre_sample(struct intel_pt_queue *ptq)
2034{
2035	struct intel_pt *pt = ptq->pt;
2036	union perf_event *event = ptq->event_buf;
2037	struct perf_sample sample = { .ip = 0, };
2038	struct perf_synth_intel_pwre raw;
2039
2040	if (intel_pt_skip_event(pt))
2041		return 0;
2042
2043	intel_pt_prep_p_sample(pt, ptq, event, &sample);
2044
2045	sample.id = ptq->pt->pwre_id;
2046	sample.stream_id = ptq->pt->pwre_id;
2047
2048	raw.reserved = 0;
2049	raw.payload = cpu_to_le64(ptq->state->pwre_payload);
2050
2051	sample.raw_size = perf_synth__raw_size(raw);
2052	sample.raw_data = perf_synth__raw_data(&raw);
2053
2054	return intel_pt_deliver_synth_event(pt, event, &sample,
2055					    pt->pwr_events_sample_type);
2056}
2057
2058static int intel_pt_synth_exstop_sample(struct intel_pt_queue *ptq)
2059{
2060	struct intel_pt *pt = ptq->pt;
2061	union perf_event *event = ptq->event_buf;
2062	struct perf_sample sample = { .ip = 0, };
2063	struct perf_synth_intel_exstop raw;
2064
2065	if (intel_pt_skip_event(pt))
2066		return 0;
2067
2068	intel_pt_prep_p_sample(pt, ptq, event, &sample);
2069
2070	sample.id = ptq->pt->exstop_id;
2071	sample.stream_id = ptq->pt->exstop_id;
2072
2073	raw.flags = 0;
2074	raw.ip = !!(ptq->state->flags & INTEL_PT_FUP_IP);
2075
2076	sample.raw_size = perf_synth__raw_size(raw);
2077	sample.raw_data = perf_synth__raw_data(&raw);
2078
2079	return intel_pt_deliver_synth_event(pt, event, &sample,
2080					    pt->pwr_events_sample_type);
2081}
2082
2083static int intel_pt_synth_pwrx_sample(struct intel_pt_queue *ptq)
2084{
2085	struct intel_pt *pt = ptq->pt;
2086	union perf_event *event = ptq->event_buf;
2087	struct perf_sample sample = { .ip = 0, };
2088	struct perf_synth_intel_pwrx raw;
2089
2090	if (intel_pt_skip_event(pt))
2091		return 0;
2092
2093	intel_pt_prep_p_sample(pt, ptq, event, &sample);
2094
2095	sample.id = ptq->pt->pwrx_id;
2096	sample.stream_id = ptq->pt->pwrx_id;
2097
2098	raw.reserved = 0;
2099	raw.payload = cpu_to_le64(ptq->state->pwrx_payload);
2100
2101	sample.raw_size = perf_synth__raw_size(raw);
2102	sample.raw_data = perf_synth__raw_data(&raw);
2103
2104	return intel_pt_deliver_synth_event(pt, event, &sample,
2105					    pt->pwr_events_sample_type);
2106}
2107
2108/*
2109 * PEBS gp_regs array indexes plus 1 so that 0 means not present. Refer
2110 * intel_pt_add_gp_regs().
2111 */
2112static const int pebs_gp_regs[] = {
2113	[PERF_REG_X86_FLAGS]	= 1,
2114	[PERF_REG_X86_IP]	= 2,
2115	[PERF_REG_X86_AX]	= 3,
2116	[PERF_REG_X86_CX]	= 4,
2117	[PERF_REG_X86_DX]	= 5,
2118	[PERF_REG_X86_BX]	= 6,
2119	[PERF_REG_X86_SP]	= 7,
2120	[PERF_REG_X86_BP]	= 8,
2121	[PERF_REG_X86_SI]	= 9,
2122	[PERF_REG_X86_DI]	= 10,
2123	[PERF_REG_X86_R8]	= 11,
2124	[PERF_REG_X86_R9]	= 12,
2125	[PERF_REG_X86_R10]	= 13,
2126	[PERF_REG_X86_R11]	= 14,
2127	[PERF_REG_X86_R12]	= 15,
2128	[PERF_REG_X86_R13]	= 16,
2129	[PERF_REG_X86_R14]	= 17,
2130	[PERF_REG_X86_R15]	= 18,
2131};
2132
2133static u64 *intel_pt_add_gp_regs(struct regs_dump *intr_regs, u64 *pos,
2134				 const struct intel_pt_blk_items *items,
2135				 u64 regs_mask)
2136{
2137	const u64 *gp_regs = items->val[INTEL_PT_GP_REGS_POS];
2138	u32 mask = items->mask[INTEL_PT_GP_REGS_POS];
2139	u32 bit;
2140	int i;
2141
2142	for (i = 0, bit = 1; i < PERF_REG_X86_64_MAX; i++, bit <<= 1) {
2143		/* Get the PEBS gp_regs array index */
2144		int n = pebs_gp_regs[i] - 1;
2145
2146		if (n < 0)
2147			continue;
2148		/*
2149		 * Add only registers that were requested (i.e. 'regs_mask') and
2150		 * that were provided (i.e. 'mask'), and update the resulting
2151		 * mask (i.e. 'intr_regs->mask') accordingly.
2152		 */
2153		if (mask & 1 << n && regs_mask & bit) {
2154			intr_regs->mask |= bit;
2155			*pos++ = gp_regs[n];
2156		}
2157	}
2158
2159	return pos;
2160}
2161
2162#ifndef PERF_REG_X86_XMM0
2163#define PERF_REG_X86_XMM0 32
2164#endif
2165
2166static void intel_pt_add_xmm(struct regs_dump *intr_regs, u64 *pos,
2167			     const struct intel_pt_blk_items *items,
2168			     u64 regs_mask)
2169{
2170	u32 mask = items->has_xmm & (regs_mask >> PERF_REG_X86_XMM0);
2171	const u64 *xmm = items->xmm;
2172
2173	/*
2174	 * If there are any XMM registers, then there should be all of them.
2175	 * Nevertheless, follow the logic to add only registers that were
2176	 * requested (i.e. 'regs_mask') and that were provided (i.e. 'mask'),
2177	 * and update the resulting mask (i.e. 'intr_regs->mask') accordingly.
2178	 */
2179	intr_regs->mask |= (u64)mask << PERF_REG_X86_XMM0;
2180
2181	for (; mask; mask >>= 1, xmm++) {
2182		if (mask & 1)
2183			*pos++ = *xmm;
2184	}
2185}
2186
2187#define LBR_INFO_MISPRED	(1ULL << 63)
2188#define LBR_INFO_IN_TX		(1ULL << 62)
2189#define LBR_INFO_ABORT		(1ULL << 61)
2190#define LBR_INFO_CYCLES		0xffff
2191
2192/* Refer kernel's intel_pmu_store_pebs_lbrs() */
2193static u64 intel_pt_lbr_flags(u64 info)
2194{
2195	union {
2196		struct branch_flags flags;
2197		u64 result;
2198	} u;
2199
2200	u.result	  = 0;
2201	u.flags.mispred	  = !!(info & LBR_INFO_MISPRED);
2202	u.flags.predicted = !(info & LBR_INFO_MISPRED);
2203	u.flags.in_tx	  = !!(info & LBR_INFO_IN_TX);
2204	u.flags.abort	  = !!(info & LBR_INFO_ABORT);
2205	u.flags.cycles	  = info & LBR_INFO_CYCLES;
2206
2207	return u.result;
2208}
2209
2210static void intel_pt_add_lbrs(struct branch_stack *br_stack,
2211			      const struct intel_pt_blk_items *items)
2212{
2213	u64 *to;
2214	int i;
2215
2216	br_stack->nr = 0;
2217
2218	to = &br_stack->entries[0].from;
2219
2220	for (i = INTEL_PT_LBR_0_POS; i <= INTEL_PT_LBR_2_POS; i++) {
2221		u32 mask = items->mask[i];
2222		const u64 *from = items->val[i];
2223
2224		for (; mask; mask >>= 3, from += 3) {
2225			if ((mask & 7) == 7) {
2226				*to++ = from[0];
2227				*to++ = from[1];
2228				*to++ = intel_pt_lbr_flags(from[2]);
2229				br_stack->nr += 1;
2230			}
2231		}
2232	}
2233}
2234
2235static int intel_pt_do_synth_pebs_sample(struct intel_pt_queue *ptq, struct evsel *evsel, u64 id)
2236{
2237	const struct intel_pt_blk_items *items = &ptq->state->items;
2238	struct perf_sample sample = { .ip = 0, };
2239	union perf_event *event = ptq->event_buf;
2240	struct intel_pt *pt = ptq->pt;
 
2241	u64 sample_type = evsel->core.attr.sample_type;
 
2242	u8 cpumode;
2243	u64 regs[8 * sizeof(sample.intr_regs.mask)];
2244
2245	if (intel_pt_skip_event(pt))
2246		return 0;
2247
2248	intel_pt_prep_a_sample(ptq, event, &sample);
2249
2250	sample.id = id;
2251	sample.stream_id = id;
2252
2253	if (!evsel->core.attr.freq)
2254		sample.period = evsel->core.attr.sample_period;
2255
2256	/* No support for non-zero CS base */
2257	if (items->has_ip)
2258		sample.ip = items->ip;
2259	else if (items->has_rip)
2260		sample.ip = items->rip;
2261	else
2262		sample.ip = ptq->state->from_ip;
2263
2264	cpumode = intel_pt_cpumode(ptq, sample.ip, 0);
 
 
 
2265
2266	event->sample.header.misc = cpumode | PERF_RECORD_MISC_EXACT_IP;
2267
2268	sample.cpumode = cpumode;
2269
2270	if (sample_type & PERF_SAMPLE_TIME) {
2271		u64 timestamp = 0;
2272
2273		if (items->has_timestamp)
2274			timestamp = items->timestamp;
2275		else if (!pt->timeless_decoding)
2276			timestamp = ptq->timestamp;
2277		if (timestamp)
2278			sample.time = tsc_to_perf_time(timestamp, &pt->tc);
2279	}
2280
2281	if (sample_type & PERF_SAMPLE_CALLCHAIN &&
2282	    pt->synth_opts.callchain) {
2283		thread_stack__sample(ptq->thread, ptq->cpu, ptq->chain,
2284				     pt->synth_opts.callchain_sz, sample.ip,
2285				     pt->kernel_start);
2286		sample.callchain = ptq->chain;
2287	}
2288
2289	if (sample_type & PERF_SAMPLE_REGS_INTR &&
2290	    (items->mask[INTEL_PT_GP_REGS_POS] ||
2291	     items->mask[INTEL_PT_XMM_POS])) {
2292		u64 regs_mask = evsel->core.attr.sample_regs_intr;
2293		u64 *pos;
2294
2295		sample.intr_regs.abi = items->is_32_bit ?
2296				       PERF_SAMPLE_REGS_ABI_32 :
2297				       PERF_SAMPLE_REGS_ABI_64;
2298		sample.intr_regs.regs = regs;
2299
2300		pos = intel_pt_add_gp_regs(&sample.intr_regs, regs, items, regs_mask);
2301
2302		intel_pt_add_xmm(&sample.intr_regs, pos, items, regs_mask);
2303	}
2304
2305	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
2306		if (items->mask[INTEL_PT_LBR_0_POS] ||
2307		    items->mask[INTEL_PT_LBR_1_POS] ||
2308		    items->mask[INTEL_PT_LBR_2_POS]) {
2309			intel_pt_add_lbrs(ptq->last_branch, items);
2310		} else if (pt->synth_opts.last_branch) {
2311			thread_stack__br_sample(ptq->thread, ptq->cpu,
2312						ptq->last_branch,
2313						pt->br_stack_sz);
2314		} else {
2315			ptq->last_branch->nr = 0;
2316		}
2317		sample.branch_stack = ptq->last_branch;
2318	}
2319
2320	if (sample_type & PERF_SAMPLE_ADDR && items->has_mem_access_address)
2321		sample.addr = items->mem_access_address;
2322
2323	if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
2324		/*
2325		 * Refer kernel's setup_pebs_adaptive_sample_data() and
2326		 * intel_hsw_weight().
2327		 */
2328		if (items->has_mem_access_latency) {
2329			u64 weight = items->mem_access_latency >> 32;
2330
2331			/*
2332			 * Starts from SPR, the mem access latency field
2333			 * contains both cache latency [47:32] and instruction
2334			 * latency [15:0]. The cache latency is the same as the
2335			 * mem access latency on previous platforms.
2336			 *
2337			 * In practice, no memory access could last than 4G
2338			 * cycles. Use latency >> 32 to distinguish the
2339			 * different format of the mem access latency field.
2340			 */
2341			if (weight > 0) {
2342				sample.weight = weight & 0xffff;
2343				sample.ins_lat = items->mem_access_latency & 0xffff;
2344			} else
2345				sample.weight = items->mem_access_latency;
2346		}
2347		if (!sample.weight && items->has_tsx_aux_info) {
2348			/* Cycles last block */
2349			sample.weight = (u32)items->tsx_aux_info;
2350		}
2351	}
2352
2353	if (sample_type & PERF_SAMPLE_TRANSACTION && items->has_tsx_aux_info) {
2354		u64 ax = items->has_rax ? items->rax : 0;
2355		/* Refer kernel's intel_hsw_transaction() */
2356		u64 txn = (u8)(items->tsx_aux_info >> 32);
2357
2358		/* For RTM XABORTs also log the abort code from AX */
2359		if (txn & PERF_TXN_TRANSACTION && ax & 1)
2360			txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
2361		sample.transaction = txn;
2362	}
2363
2364	return intel_pt_deliver_synth_event(pt, event, &sample, sample_type);
2365}
2366
2367static int intel_pt_synth_single_pebs_sample(struct intel_pt_queue *ptq)
2368{
2369	struct intel_pt *pt = ptq->pt;
2370	struct evsel *evsel = pt->pebs_evsel;
2371	u64 id = evsel->core.id[0];
2372
2373	return intel_pt_do_synth_pebs_sample(ptq, evsel, id);
2374}
2375
2376static int intel_pt_synth_pebs_sample(struct intel_pt_queue *ptq)
2377{
2378	const struct intel_pt_blk_items *items = &ptq->state->items;
2379	struct intel_pt_pebs_event *pe;
2380	struct intel_pt *pt = ptq->pt;
2381	int err = -EINVAL;
2382	int hw_id;
2383
2384	if (!items->has_applicable_counters || !items->applicable_counters) {
2385		if (!pt->single_pebs)
2386			pr_err("PEBS-via-PT record with no applicable_counters\n");
2387		return intel_pt_synth_single_pebs_sample(ptq);
2388	}
2389
2390	for_each_set_bit(hw_id, (unsigned long *)&items->applicable_counters, INTEL_PT_MAX_PEBS) {
2391		pe = &ptq->pebs[hw_id];
2392		if (!pe->evsel) {
2393			if (!pt->single_pebs)
2394				pr_err("PEBS-via-PT record with no matching event, hw_id %d\n",
2395				       hw_id);
2396			return intel_pt_synth_single_pebs_sample(ptq);
2397		}
2398		err = intel_pt_do_synth_pebs_sample(ptq, pe->evsel, pe->id);
2399		if (err)
2400			return err;
2401	}
2402
2403	return err;
2404}
2405
2406static int intel_pt_synth_events_sample(struct intel_pt_queue *ptq)
2407{
2408	struct intel_pt *pt = ptq->pt;
2409	union perf_event *event = ptq->event_buf;
2410	struct perf_sample sample = { .ip = 0, };
2411	struct {
2412		struct perf_synth_intel_evt cfe;
2413		struct perf_synth_intel_evd evd[INTEL_PT_MAX_EVDS];
2414	} raw;
2415	int i;
2416
2417	if (intel_pt_skip_event(pt))
2418		return 0;
2419
2420	intel_pt_prep_p_sample(pt, ptq, event, &sample);
2421
2422	sample.id        = ptq->pt->evt_id;
2423	sample.stream_id = ptq->pt->evt_id;
2424
2425	raw.cfe.type     = ptq->state->cfe_type;
2426	raw.cfe.reserved = 0;
2427	raw.cfe.ip       = !!(ptq->state->flags & INTEL_PT_FUP_IP);
2428	raw.cfe.vector   = ptq->state->cfe_vector;
2429	raw.cfe.evd_cnt  = ptq->state->evd_cnt;
2430
2431	for (i = 0; i < ptq->state->evd_cnt; i++) {
2432		raw.evd[i].et       = 0;
2433		raw.evd[i].evd_type = ptq->state->evd[i].type;
2434		raw.evd[i].payload  = ptq->state->evd[i].payload;
2435	}
2436
2437	sample.raw_size = perf_synth__raw_size(raw) +
2438			  ptq->state->evd_cnt * sizeof(struct perf_synth_intel_evd);
2439	sample.raw_data = perf_synth__raw_data(&raw);
2440
2441	return intel_pt_deliver_synth_event(pt, event, &sample,
2442					    pt->evt_sample_type);
2443}
2444
2445static int intel_pt_synth_iflag_chg_sample(struct intel_pt_queue *ptq)
2446{
2447	struct intel_pt *pt = ptq->pt;
2448	union perf_event *event = ptq->event_buf;
2449	struct perf_sample sample = { .ip = 0, };
2450	struct perf_synth_intel_iflag_chg raw;
2451
2452	if (intel_pt_skip_event(pt))
2453		return 0;
2454
2455	intel_pt_prep_p_sample(pt, ptq, event, &sample);
2456
2457	sample.id = ptq->pt->iflag_chg_id;
2458	sample.stream_id = ptq->pt->iflag_chg_id;
2459
2460	raw.flags = 0;
2461	raw.iflag = ptq->state->to_iflag;
2462
2463	if (ptq->state->type & INTEL_PT_BRANCH) {
2464		raw.via_branch = 1;
2465		raw.branch_ip = ptq->state->to_ip;
2466	} else {
2467		sample.addr = 0;
2468	}
2469	sample.flags = ptq->flags;
2470
2471	sample.raw_size = perf_synth__raw_size(raw);
2472	sample.raw_data = perf_synth__raw_data(&raw);
2473
2474	return intel_pt_deliver_synth_event(pt, event, &sample,
2475					    pt->iflag_chg_sample_type);
2476}
2477
2478static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,
2479				pid_t pid, pid_t tid, u64 ip, u64 timestamp,
2480				pid_t machine_pid, int vcpu)
2481{
2482	bool dump_log_on_error = pt->synth_opts.log_plus_flags & AUXTRACE_LOG_FLG_ON_ERROR;
2483	bool log_on_stdout = pt->synth_opts.log_plus_flags & AUXTRACE_LOG_FLG_USE_STDOUT;
2484	union perf_event event;
2485	char msg[MAX_AUXTRACE_ERROR_MSG];
2486	int err;
2487
2488	if (pt->synth_opts.error_minus_flags) {
2489		if (code == INTEL_PT_ERR_OVR &&
2490		    pt->synth_opts.error_minus_flags & AUXTRACE_ERR_FLG_OVERFLOW)
2491			return 0;
2492		if (code == INTEL_PT_ERR_LOST &&
2493		    pt->synth_opts.error_minus_flags & AUXTRACE_ERR_FLG_DATA_LOST)
2494			return 0;
2495	}
2496
2497	intel_pt__strerror(code, msg, MAX_AUXTRACE_ERROR_MSG);
2498
2499	auxtrace_synth_guest_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
2500				   code, cpu, pid, tid, ip, msg, timestamp,
2501				   machine_pid, vcpu);
2502
2503	if (intel_pt_enable_logging && !log_on_stdout) {
2504		FILE *fp = intel_pt_log_fp();
2505
2506		if (fp)
2507			perf_event__fprintf_auxtrace_error(&event, fp);
2508	}
2509
2510	if (code != INTEL_PT_ERR_LOST && dump_log_on_error)
2511		intel_pt_log_dump_buf();
2512
2513	err = perf_session__deliver_synth_event(pt->session, &event, NULL);
2514	if (err)
2515		pr_err("Intel Processor Trace: failed to deliver error event, error %d\n",
2516		       err);
2517
2518	return err;
2519}
2520
2521static int intel_ptq_synth_error(struct intel_pt_queue *ptq,
2522				 const struct intel_pt_state *state)
2523{
2524	struct intel_pt *pt = ptq->pt;
2525	u64 tm = ptq->timestamp;
2526	pid_t machine_pid = 0;
2527	pid_t pid = ptq->pid;
2528	pid_t tid = ptq->tid;
2529	int vcpu = -1;
2530
2531	tm = pt->timeless_decoding ? 0 : tsc_to_perf_time(tm, &pt->tc);
2532
2533	if (pt->have_guest_sideband && state->from_nr) {
2534		machine_pid = ptq->guest_machine_pid;
2535		vcpu = ptq->vcpu;
2536		pid = ptq->guest_pid;
2537		tid = ptq->guest_tid;
2538	}
2539
2540	return intel_pt_synth_error(pt, state->err, ptq->cpu, pid, tid,
2541				    state->from_ip, tm, machine_pid, vcpu);
2542}
2543
2544static int intel_pt_next_tid(struct intel_pt *pt, struct intel_pt_queue *ptq)
2545{
2546	struct auxtrace_queue *queue;
2547	pid_t tid = ptq->next_tid;
2548	int err;
2549
2550	if (tid == -1)
2551		return 0;
2552
2553	intel_pt_log("switch: cpu %d tid %d\n", ptq->cpu, tid);
2554
2555	err = machine__set_current_tid(pt->machine, ptq->cpu, -1, tid);
2556
2557	queue = &pt->queues.queue_array[ptq->queue_nr];
2558	intel_pt_set_pid_tid_cpu(pt, queue);
2559
2560	ptq->next_tid = -1;
2561
2562	return err;
2563}
2564
2565static inline bool intel_pt_is_switch_ip(struct intel_pt_queue *ptq, u64 ip)
2566{
2567	struct intel_pt *pt = ptq->pt;
2568
2569	return ip == pt->switch_ip &&
2570	       (ptq->flags & PERF_IP_FLAG_BRANCH) &&
2571	       !(ptq->flags & (PERF_IP_FLAG_CONDITIONAL | PERF_IP_FLAG_ASYNC |
2572			       PERF_IP_FLAG_INTERRUPT | PERF_IP_FLAG_TX_ABORT));
2573}
2574
2575#define INTEL_PT_PWR_EVT (INTEL_PT_MWAIT_OP | INTEL_PT_PWR_ENTRY | \
2576			  INTEL_PT_EX_STOP | INTEL_PT_PWR_EXIT)
2577
2578static int intel_pt_sample(struct intel_pt_queue *ptq)
2579{
2580	const struct intel_pt_state *state = ptq->state;
2581	struct intel_pt *pt = ptq->pt;
2582	int err;
2583
2584	if (!ptq->have_sample)
2585		return 0;
2586
2587	ptq->have_sample = false;
2588
2589	if (pt->synth_opts.approx_ipc) {
2590		ptq->ipc_insn_cnt = ptq->state->tot_insn_cnt;
2591		ptq->ipc_cyc_cnt = ptq->state->cycles;
2592		ptq->sample_ipc = true;
2593	} else {
2594		ptq->ipc_insn_cnt = ptq->state->tot_insn_cnt;
2595		ptq->ipc_cyc_cnt = ptq->state->tot_cyc_cnt;
2596		ptq->sample_ipc = ptq->state->flags & INTEL_PT_SAMPLE_IPC;
2597	}
2598
2599	/* Ensure guest code maps are set up */
2600	if (symbol_conf.guest_code && (state->from_nr || state->to_nr))
2601		intel_pt_get_guest(ptq);
2602
2603	/*
2604	 * Do PEBS first to allow for the possibility that the PEBS timestamp
2605	 * precedes the current timestamp.
2606	 */
2607	if (pt->sample_pebs && state->type & INTEL_PT_BLK_ITEMS) {
2608		err = intel_pt_synth_pebs_sample(ptq);
2609		if (err)
2610			return err;
2611	}
2612
2613	if (pt->synth_opts.intr_events) {
2614		if (state->type & INTEL_PT_EVT) {
2615			err = intel_pt_synth_events_sample(ptq);
2616			if (err)
2617				return err;
2618		}
2619		if (state->type & INTEL_PT_IFLAG_CHG) {
2620			err = intel_pt_synth_iflag_chg_sample(ptq);
2621			if (err)
2622				return err;
2623		}
2624	}
2625
2626	if (pt->sample_pwr_events) {
2627		if (state->type & INTEL_PT_PSB_EVT) {
2628			err = intel_pt_synth_psb_sample(ptq);
2629			if (err)
2630				return err;
2631		}
2632		if (ptq->state->cbr != ptq->cbr_seen) {
2633			err = intel_pt_synth_cbr_sample(ptq);
2634			if (err)
2635				return err;
2636		}
2637		if (state->type & INTEL_PT_PWR_EVT) {
2638			if (state->type & INTEL_PT_MWAIT_OP) {
2639				err = intel_pt_synth_mwait_sample(ptq);
2640				if (err)
2641					return err;
2642			}
2643			if (state->type & INTEL_PT_PWR_ENTRY) {
2644				err = intel_pt_synth_pwre_sample(ptq);
2645				if (err)
2646					return err;
2647			}
2648			if (state->type & INTEL_PT_EX_STOP) {
2649				err = intel_pt_synth_exstop_sample(ptq);
2650				if (err)
2651					return err;
2652			}
2653			if (state->type & INTEL_PT_PWR_EXIT) {
2654				err = intel_pt_synth_pwrx_sample(ptq);
2655				if (err)
2656					return err;
2657			}
2658		}
2659	}
2660
2661	if (state->type & INTEL_PT_INSTRUCTION) {
2662		if (pt->sample_instructions) {
2663			err = intel_pt_synth_instruction_sample(ptq);
2664			if (err)
2665				return err;
2666		}
2667		if (pt->sample_cycles) {
2668			err = intel_pt_synth_cycle_sample(ptq);
2669			if (err)
2670				return err;
2671		}
2672	}
2673
2674	if (pt->sample_transactions && (state->type & INTEL_PT_TRANSACTION)) {
2675		err = intel_pt_synth_transaction_sample(ptq);
2676		if (err)
2677			return err;
2678	}
2679
2680	if (pt->sample_ptwrites && (state->type & INTEL_PT_PTW)) {
2681		err = intel_pt_synth_ptwrite_sample(ptq);
2682		if (err)
2683			return err;
2684	}
2685
2686	if (!(state->type & INTEL_PT_BRANCH))
2687		return 0;
2688
2689	if (pt->use_thread_stack) {
2690		thread_stack__event(ptq->thread, ptq->cpu, ptq->flags,
2691				    state->from_ip, state->to_ip, ptq->insn_len,
2692				    state->trace_nr, pt->callstack,
2693				    pt->br_stack_sz_plus,
2694				    pt->mispred_all);
2695	} else {
2696		thread_stack__set_trace_nr(ptq->thread, ptq->cpu, state->trace_nr);
2697	}
2698
2699	if (pt->sample_branches) {
2700		if (state->from_nr != state->to_nr &&
2701		    state->from_ip && state->to_ip) {
2702			struct intel_pt_state *st = (struct intel_pt_state *)state;
2703			u64 to_ip = st->to_ip;
2704			u64 from_ip = st->from_ip;
2705
2706			/*
2707			 * perf cannot handle having different machines for ip
2708			 * and addr, so create 2 branches.
2709			 */
2710			st->to_ip = 0;
2711			err = intel_pt_synth_branch_sample(ptq);
2712			if (err)
2713				return err;
2714			st->from_ip = 0;
2715			st->to_ip = to_ip;
2716			err = intel_pt_synth_branch_sample(ptq);
2717			st->from_ip = from_ip;
2718		} else {
2719			err = intel_pt_synth_branch_sample(ptq);
2720		}
2721		if (err)
2722			return err;
2723	}
2724
2725	if (!ptq->sync_switch)
2726		return 0;
2727
2728	if (intel_pt_is_switch_ip(ptq, state->to_ip)) {
2729		switch (ptq->switch_state) {
2730		case INTEL_PT_SS_NOT_TRACING:
2731		case INTEL_PT_SS_UNKNOWN:
2732		case INTEL_PT_SS_EXPECTING_SWITCH_IP:
2733			err = intel_pt_next_tid(pt, ptq);
2734			if (err)
2735				return err;
2736			ptq->switch_state = INTEL_PT_SS_TRACING;
2737			break;
2738		default:
2739			ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_EVENT;
2740			return 1;
2741		}
2742	} else if (!state->to_ip) {
2743		ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
2744	} else if (ptq->switch_state == INTEL_PT_SS_NOT_TRACING) {
2745		ptq->switch_state = INTEL_PT_SS_UNKNOWN;
2746	} else if (ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
2747		   state->to_ip == pt->ptss_ip &&
2748		   (ptq->flags & PERF_IP_FLAG_CALL)) {
2749		ptq->switch_state = INTEL_PT_SS_TRACING;
2750	}
2751
2752	return 0;
2753}
2754
2755static u64 intel_pt_switch_ip(struct intel_pt *pt, u64 *ptss_ip)
2756{
2757	struct machine *machine = pt->machine;
2758	struct map *map;
2759	struct symbol *sym, *start;
2760	u64 ip, switch_ip = 0;
2761	const char *ptss;
2762
2763	if (ptss_ip)
2764		*ptss_ip = 0;
2765
2766	map = machine__kernel_map(machine);
2767	if (!map)
2768		return 0;
2769
2770	if (map__load(map))
2771		return 0;
2772
2773	start = dso__first_symbol(map__dso(map));
2774
2775	for (sym = start; sym; sym = dso__next_symbol(sym)) {
2776		if (sym->binding == STB_GLOBAL &&
2777		    !strcmp(sym->name, "__switch_to")) {
2778			ip = map__unmap_ip(map, sym->start);
2779			if (ip >= map__start(map) && ip < map__end(map)) {
2780				switch_ip = ip;
2781				break;
2782			}
2783		}
2784	}
2785
2786	if (!switch_ip || !ptss_ip)
2787		return 0;
2788
2789	if (pt->have_sched_switch == 1)
2790		ptss = "perf_trace_sched_switch";
2791	else
2792		ptss = "__perf_event_task_sched_out";
2793
2794	for (sym = start; sym; sym = dso__next_symbol(sym)) {
2795		if (!strcmp(sym->name, ptss)) {
2796			ip = map__unmap_ip(map, sym->start);
2797			if (ip >= map__start(map) && ip < map__end(map)) {
2798				*ptss_ip = ip;
2799				break;
2800			}
2801		}
2802	}
2803
2804	return switch_ip;
2805}
2806
2807static void intel_pt_enable_sync_switch(struct intel_pt *pt)
2808{
2809	unsigned int i;
2810
2811	if (pt->sync_switch_not_supported)
2812		return;
2813
2814	pt->sync_switch = true;
2815
2816	for (i = 0; i < pt->queues.nr_queues; i++) {
2817		struct auxtrace_queue *queue = &pt->queues.queue_array[i];
2818		struct intel_pt_queue *ptq = queue->priv;
2819
2820		if (ptq)
2821			ptq->sync_switch = true;
2822	}
2823}
2824
2825static void intel_pt_disable_sync_switch(struct intel_pt *pt)
2826{
2827	unsigned int i;
2828
2829	pt->sync_switch = false;
2830
2831	for (i = 0; i < pt->queues.nr_queues; i++) {
2832		struct auxtrace_queue *queue = &pt->queues.queue_array[i];
2833		struct intel_pt_queue *ptq = queue->priv;
2834
2835		if (ptq) {
2836			ptq->sync_switch = false;
2837			intel_pt_next_tid(pt, ptq);
2838		}
2839	}
2840}
2841
2842/*
2843 * To filter against time ranges, it is only necessary to look at the next start
2844 * or end time.
2845 */
2846static bool intel_pt_next_time(struct intel_pt_queue *ptq)
2847{
2848	struct intel_pt *pt = ptq->pt;
2849
2850	if (ptq->sel_start) {
2851		/* Next time is an end time */
2852		ptq->sel_start = false;
2853		ptq->sel_timestamp = pt->time_ranges[ptq->sel_idx].end;
2854		return true;
2855	} else if (ptq->sel_idx + 1 < pt->range_cnt) {
2856		/* Next time is a start time */
2857		ptq->sel_start = true;
2858		ptq->sel_idx += 1;
2859		ptq->sel_timestamp = pt->time_ranges[ptq->sel_idx].start;
2860		return true;
2861	}
2862
2863	/* No next time */
2864	return false;
2865}
2866
2867static int intel_pt_time_filter(struct intel_pt_queue *ptq, u64 *ff_timestamp)
2868{
2869	int err;
2870
2871	while (1) {
2872		if (ptq->sel_start) {
2873			if (ptq->timestamp >= ptq->sel_timestamp) {
2874				/* After start time, so consider next time */
2875				intel_pt_next_time(ptq);
2876				if (!ptq->sel_timestamp) {
2877					/* No end time */
2878					return 0;
2879				}
2880				/* Check against end time */
2881				continue;
2882			}
2883			/* Before start time, so fast forward */
2884			ptq->have_sample = false;
2885			if (ptq->sel_timestamp > *ff_timestamp) {
2886				if (ptq->sync_switch) {
2887					intel_pt_next_tid(ptq->pt, ptq);
2888					ptq->switch_state = INTEL_PT_SS_UNKNOWN;
2889				}
2890				*ff_timestamp = ptq->sel_timestamp;
2891				err = intel_pt_fast_forward(ptq->decoder,
2892							    ptq->sel_timestamp);
2893				if (err)
2894					return err;
2895			}
2896			return 0;
2897		} else if (ptq->timestamp > ptq->sel_timestamp) {
2898			/* After end time, so consider next time */
2899			if (!intel_pt_next_time(ptq)) {
2900				/* No next time range, so stop decoding */
2901				ptq->have_sample = false;
2902				ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
2903				return 1;
2904			}
2905			/* Check against next start time */
2906			continue;
2907		} else {
2908			/* Before end time */
2909			return 0;
2910		}
2911	}
2912}
2913
2914static int intel_pt_run_decoder(struct intel_pt_queue *ptq, u64 *timestamp)
2915{
2916	const struct intel_pt_state *state = ptq->state;
2917	struct intel_pt *pt = ptq->pt;
2918	u64 ff_timestamp = 0;
2919	int err;
2920
2921	if (!pt->kernel_start) {
2922		pt->kernel_start = machine__kernel_start(pt->machine);
2923		if (pt->per_cpu_mmaps &&
2924		    (pt->have_sched_switch == 1 || pt->have_sched_switch == 3) &&
2925		    !pt->timeless_decoding && intel_pt_tracing_kernel(pt) &&
2926		    !pt->sampling_mode && !pt->synth_opts.vm_time_correlation) {
2927			pt->switch_ip = intel_pt_switch_ip(pt, &pt->ptss_ip);
2928			if (pt->switch_ip) {
2929				intel_pt_log("switch_ip: %"PRIx64" ptss_ip: %"PRIx64"\n",
2930					     pt->switch_ip, pt->ptss_ip);
2931				intel_pt_enable_sync_switch(pt);
2932			}
2933		}
2934	}
2935
2936	intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
2937		     ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
2938	while (1) {
2939		err = intel_pt_sample(ptq);
2940		if (err)
2941			return err;
2942
2943		state = intel_pt_decode(ptq->decoder);
2944		if (state->err) {
2945			if (state->err == INTEL_PT_ERR_NODATA)
2946				return 1;
2947			if (ptq->sync_switch &&
2948			    state->from_ip >= pt->kernel_start) {
2949				ptq->sync_switch = false;
2950				intel_pt_next_tid(pt, ptq);
2951			}
2952			ptq->timestamp = state->est_timestamp;
2953			if (pt->synth_opts.errors) {
2954				err = intel_ptq_synth_error(ptq, state);
2955				if (err)
2956					return err;
2957			}
2958			continue;
2959		}
2960
2961		ptq->state = state;
2962		ptq->have_sample = true;
2963		intel_pt_sample_flags(ptq);
2964
2965		/* Use estimated TSC upon return to user space */
2966		if (pt->est_tsc &&
2967		    (state->from_ip >= pt->kernel_start || !state->from_ip) &&
2968		    state->to_ip && state->to_ip < pt->kernel_start) {
2969			intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
2970				     state->timestamp, state->est_timestamp);
2971			ptq->timestamp = state->est_timestamp;
2972		/* Use estimated TSC in unknown switch state */
2973		} else if (ptq->sync_switch &&
2974			   ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
2975			   intel_pt_is_switch_ip(ptq, state->to_ip) &&
2976			   ptq->next_tid == -1) {
2977			intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
2978				     state->timestamp, state->est_timestamp);
2979			ptq->timestamp = state->est_timestamp;
2980		} else if (state->timestamp > ptq->timestamp) {
2981			ptq->timestamp = state->timestamp;
2982		}
2983
2984		if (ptq->sel_timestamp) {
2985			err = intel_pt_time_filter(ptq, &ff_timestamp);
2986			if (err)
2987				return err;
2988		}
2989
2990		if (!pt->timeless_decoding && ptq->timestamp >= *timestamp) {
2991			*timestamp = ptq->timestamp;
2992			return 0;
2993		}
2994	}
2995	return 0;
2996}
2997
2998static inline int intel_pt_update_queues(struct intel_pt *pt)
2999{
3000	if (pt->queues.new_data) {
3001		pt->queues.new_data = false;
3002		return intel_pt_setup_queues(pt);
3003	}
3004	return 0;
3005}
3006
3007static int intel_pt_process_queues(struct intel_pt *pt, u64 timestamp)
3008{
3009	unsigned int queue_nr;
3010	u64 ts;
3011	int ret;
3012
3013	while (1) {
3014		struct auxtrace_queue *queue;
3015		struct intel_pt_queue *ptq;
3016
3017		if (!pt->heap.heap_cnt)
3018			return 0;
3019
3020		if (pt->heap.heap_array[0].ordinal >= timestamp)
3021			return 0;
3022
3023		queue_nr = pt->heap.heap_array[0].queue_nr;
3024		queue = &pt->queues.queue_array[queue_nr];
3025		ptq = queue->priv;
3026
3027		intel_pt_log("queue %u processing 0x%" PRIx64 " to 0x%" PRIx64 "\n",
3028			     queue_nr, pt->heap.heap_array[0].ordinal,
3029			     timestamp);
3030
3031		auxtrace_heap__pop(&pt->heap);
3032
3033		if (pt->heap.heap_cnt) {
3034			ts = pt->heap.heap_array[0].ordinal + 1;
3035			if (ts > timestamp)
3036				ts = timestamp;
3037		} else {
3038			ts = timestamp;
3039		}
3040
3041		intel_pt_set_pid_tid_cpu(pt, queue);
3042
3043		ret = intel_pt_run_decoder(ptq, &ts);
3044
3045		if (ret < 0) {
3046			auxtrace_heap__add(&pt->heap, queue_nr, ts);
3047			return ret;
3048		}
3049
3050		if (!ret) {
3051			ret = auxtrace_heap__add(&pt->heap, queue_nr, ts);
3052			if (ret < 0)
3053				return ret;
3054		} else {
3055			ptq->on_heap = false;
3056		}
3057	}
3058
3059	return 0;
3060}
3061
3062static int intel_pt_process_timeless_queues(struct intel_pt *pt, pid_t tid,
3063					    u64 time_)
3064{
3065	struct auxtrace_queues *queues = &pt->queues;
3066	unsigned int i;
3067	u64 ts = 0;
3068
3069	for (i = 0; i < queues->nr_queues; i++) {
3070		struct auxtrace_queue *queue = &pt->queues.queue_array[i];
3071		struct intel_pt_queue *ptq = queue->priv;
3072
3073		if (ptq && (tid == -1 || ptq->tid == tid)) {
3074			ptq->time = time_;
3075			intel_pt_set_pid_tid_cpu(pt, queue);
3076			intel_pt_run_decoder(ptq, &ts);
3077		}
3078	}
3079	return 0;
3080}
3081
3082static void intel_pt_sample_set_pid_tid_cpu(struct intel_pt_queue *ptq,
3083					    struct auxtrace_queue *queue,
3084					    struct perf_sample *sample)
3085{
3086	struct machine *m = ptq->pt->machine;
3087
3088	ptq->pid = sample->pid;
3089	ptq->tid = sample->tid;
3090	ptq->cpu = queue->cpu;
3091
3092	intel_pt_log("queue %u cpu %d pid %d tid %d\n",
3093		     ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
3094
3095	thread__zput(ptq->thread);
3096
3097	if (ptq->tid == -1)
3098		return;
3099
3100	if (ptq->pid == -1) {
3101		ptq->thread = machine__find_thread(m, -1, ptq->tid);
3102		if (ptq->thread)
3103			ptq->pid = thread__pid(ptq->thread);
3104		return;
3105	}
3106
3107	ptq->thread = machine__findnew_thread(m, ptq->pid, ptq->tid);
3108}
3109
3110static int intel_pt_process_timeless_sample(struct intel_pt *pt,
3111					    struct perf_sample *sample)
3112{
3113	struct auxtrace_queue *queue;
3114	struct intel_pt_queue *ptq;
3115	u64 ts = 0;
3116
3117	queue = auxtrace_queues__sample_queue(&pt->queues, sample, pt->session);
3118	if (!queue)
3119		return -EINVAL;
3120
3121	ptq = queue->priv;
3122	if (!ptq)
3123		return 0;
3124
3125	ptq->stop = false;
3126	ptq->time = sample->time;
3127	intel_pt_sample_set_pid_tid_cpu(ptq, queue, sample);
3128	intel_pt_run_decoder(ptq, &ts);
3129	return 0;
3130}
3131
3132static int intel_pt_lost(struct intel_pt *pt, struct perf_sample *sample)
3133{
3134	return intel_pt_synth_error(pt, INTEL_PT_ERR_LOST, sample->cpu,
3135				    sample->pid, sample->tid, 0, sample->time,
3136				    sample->machine_pid, sample->vcpu);
3137}
3138
3139static struct intel_pt_queue *intel_pt_cpu_to_ptq(struct intel_pt *pt, int cpu)
3140{
3141	unsigned i, j;
3142
3143	if (cpu < 0 || !pt->queues.nr_queues)
3144		return NULL;
3145
3146	if ((unsigned)cpu >= pt->queues.nr_queues)
3147		i = pt->queues.nr_queues - 1;
3148	else
3149		i = cpu;
3150
3151	if (pt->queues.queue_array[i].cpu == cpu)
3152		return pt->queues.queue_array[i].priv;
3153
3154	for (j = 0; i > 0; j++) {
3155		if (pt->queues.queue_array[--i].cpu == cpu)
3156			return pt->queues.queue_array[i].priv;
3157	}
3158
3159	for (; j < pt->queues.nr_queues; j++) {
3160		if (pt->queues.queue_array[j].cpu == cpu)
3161			return pt->queues.queue_array[j].priv;
3162	}
3163
3164	return NULL;
3165}
3166
3167static int intel_pt_sync_switch(struct intel_pt *pt, int cpu, pid_t tid,
3168				u64 timestamp)
3169{
3170	struct intel_pt_queue *ptq;
3171	int err;
3172
3173	if (!pt->sync_switch)
3174		return 1;
3175
3176	ptq = intel_pt_cpu_to_ptq(pt, cpu);
3177	if (!ptq || !ptq->sync_switch)
3178		return 1;
3179
3180	switch (ptq->switch_state) {
3181	case INTEL_PT_SS_NOT_TRACING:
3182		break;
3183	case INTEL_PT_SS_UNKNOWN:
3184	case INTEL_PT_SS_TRACING:
3185		ptq->next_tid = tid;
3186		ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_IP;
3187		return 0;
3188	case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
3189		if (!ptq->on_heap) {
3190			ptq->timestamp = perf_time_to_tsc(timestamp,
3191							  &pt->tc);
3192			err = auxtrace_heap__add(&pt->heap, ptq->queue_nr,
3193						 ptq->timestamp);
3194			if (err)
3195				return err;
3196			ptq->on_heap = true;
3197		}
3198		ptq->switch_state = INTEL_PT_SS_TRACING;
3199		break;
3200	case INTEL_PT_SS_EXPECTING_SWITCH_IP:
3201		intel_pt_log("ERROR: cpu %d expecting switch ip\n", cpu);
3202		break;
3203	default:
3204		break;
3205	}
3206
3207	ptq->next_tid = -1;
3208
3209	return 1;
3210}
3211
3212#ifdef HAVE_LIBTRACEEVENT
3213static int intel_pt_process_switch(struct intel_pt *pt,
3214				   struct perf_sample *sample)
3215{
 
3216	pid_t tid;
3217	int cpu, ret;
3218	struct evsel *evsel = evlist__id2evsel(pt->session->evlist, sample->id);
3219
 
3220	if (evsel != pt->switch_evsel)
3221		return 0;
3222
3223	tid = evsel__intval(evsel, sample, "next_pid");
3224	cpu = sample->cpu;
3225
3226	intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
3227		     cpu, tid, sample->time, perf_time_to_tsc(sample->time,
3228		     &pt->tc));
3229
3230	ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
3231	if (ret <= 0)
3232		return ret;
3233
3234	return machine__set_current_tid(pt->machine, cpu, -1, tid);
3235}
3236#endif /* HAVE_LIBTRACEEVENT */
3237
3238static int intel_pt_context_switch_in(struct intel_pt *pt,
3239				      struct perf_sample *sample)
3240{
3241	pid_t pid = sample->pid;
3242	pid_t tid = sample->tid;
3243	int cpu = sample->cpu;
3244
3245	if (pt->sync_switch) {
3246		struct intel_pt_queue *ptq;
3247
3248		ptq = intel_pt_cpu_to_ptq(pt, cpu);
3249		if (ptq && ptq->sync_switch) {
3250			ptq->next_tid = -1;
3251			switch (ptq->switch_state) {
3252			case INTEL_PT_SS_NOT_TRACING:
3253			case INTEL_PT_SS_UNKNOWN:
3254			case INTEL_PT_SS_TRACING:
3255				break;
3256			case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
3257			case INTEL_PT_SS_EXPECTING_SWITCH_IP:
3258				ptq->switch_state = INTEL_PT_SS_TRACING;
3259				break;
3260			default:
3261				break;
3262			}
3263		}
3264	}
3265
3266	/*
3267	 * If the current tid has not been updated yet, ensure it is now that
3268	 * a "switch in" event has occurred.
3269	 */
3270	if (machine__get_current_tid(pt->machine, cpu) == tid)
3271		return 0;
3272
3273	return machine__set_current_tid(pt->machine, cpu, pid, tid);
3274}
3275
3276static int intel_pt_guest_context_switch(struct intel_pt *pt,
3277					 union perf_event *event,
3278					 struct perf_sample *sample)
3279{
3280	bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
3281	struct machines *machines = &pt->session->machines;
3282	struct machine *machine = machines__find(machines, sample->machine_pid);
3283
3284	pt->have_guest_sideband = true;
3285
3286	/*
3287	 * sync_switch cannot handle guest machines at present, so just disable
3288	 * it.
3289	 */
3290	pt->sync_switch_not_supported = true;
3291	if (pt->sync_switch)
3292		intel_pt_disable_sync_switch(pt);
3293
3294	if (out)
3295		return 0;
3296
3297	if (!machine)
3298		return -EINVAL;
3299
3300	return machine__set_current_tid(machine, sample->vcpu, sample->pid, sample->tid);
3301}
3302
3303static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,
3304				   struct perf_sample *sample)
3305{
3306	bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
3307	pid_t pid, tid;
3308	int cpu, ret;
3309
3310	if (perf_event__is_guest(event))
3311		return intel_pt_guest_context_switch(pt, event, sample);
3312
3313	cpu = sample->cpu;
3314
3315	if (pt->have_sched_switch == 3) {
3316		if (!out)
3317			return intel_pt_context_switch_in(pt, sample);
3318		if (event->header.type != PERF_RECORD_SWITCH_CPU_WIDE) {
3319			pr_err("Expecting CPU-wide context switch event\n");
3320			return -EINVAL;
3321		}
3322		pid = event->context_switch.next_prev_pid;
3323		tid = event->context_switch.next_prev_tid;
3324	} else {
3325		if (out)
3326			return 0;
3327		pid = sample->pid;
3328		tid = sample->tid;
3329	}
3330
3331	if (tid == -1)
3332		intel_pt_log("context_switch event has no tid\n");
 
 
3333
3334	ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
3335	if (ret <= 0)
3336		return ret;
3337
3338	return machine__set_current_tid(pt->machine, cpu, pid, tid);
3339}
3340
3341static int intel_pt_process_itrace_start(struct intel_pt *pt,
3342					 union perf_event *event,
3343					 struct perf_sample *sample)
3344{
3345	if (!pt->per_cpu_mmaps)
3346		return 0;
3347
3348	intel_pt_log("itrace_start: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
3349		     sample->cpu, event->itrace_start.pid,
3350		     event->itrace_start.tid, sample->time,
3351		     perf_time_to_tsc(sample->time, &pt->tc));
3352
3353	return machine__set_current_tid(pt->machine, sample->cpu,
3354					event->itrace_start.pid,
3355					event->itrace_start.tid);
3356}
3357
3358static int intel_pt_process_aux_output_hw_id(struct intel_pt *pt,
3359					     union perf_event *event,
3360					     struct perf_sample *sample)
3361{
3362	u64 hw_id = event->aux_output_hw_id.hw_id;
3363	struct auxtrace_queue *queue;
3364	struct intel_pt_queue *ptq;
3365	struct evsel *evsel;
3366
3367	queue = auxtrace_queues__sample_queue(&pt->queues, sample, pt->session);
3368	evsel = evlist__id2evsel_strict(pt->session->evlist, sample->id);
3369	if (!queue || !queue->priv || !evsel || hw_id > INTEL_PT_MAX_PEBS) {
3370		pr_err("Bad AUX output hardware ID\n");
3371		return -EINVAL;
3372	}
3373
3374	ptq = queue->priv;
3375
3376	ptq->pebs[hw_id].evsel = evsel;
3377	ptq->pebs[hw_id].id = sample->id;
3378
3379	return 0;
3380}
3381
3382static int intel_pt_find_map(struct thread *thread, u8 cpumode, u64 addr,
3383			     struct addr_location *al)
3384{
3385	if (!al->map || addr < map__start(al->map) || addr >= map__end(al->map)) {
3386		if (!thread__find_map(thread, cpumode, addr, al))
3387			return -1;
3388	}
3389
3390	return 0;
3391}
3392
3393/* Invalidate all instruction cache entries that overlap the text poke */
3394static int intel_pt_text_poke(struct intel_pt *pt, union perf_event *event)
3395{
3396	u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
3397	u64 addr = event->text_poke.addr + event->text_poke.new_len - 1;
3398	/* Assume text poke begins in a basic block no more than 4096 bytes */
3399	int cnt = 4096 + event->text_poke.new_len;
3400	struct thread *thread = pt->unknown_thread;
3401	struct addr_location al;
3402	struct machine *machine = pt->machine;
3403	struct intel_pt_cache_entry *e;
3404	u64 offset;
3405	int ret = 0;
3406
3407	addr_location__init(&al);
3408	if (!event->text_poke.new_len)
3409		goto out;
3410
3411	for (; cnt; cnt--, addr--) {
3412		struct dso *dso;
3413
3414		if (intel_pt_find_map(thread, cpumode, addr, &al)) {
3415			if (addr < event->text_poke.addr)
3416				goto out;
3417			continue;
3418		}
3419
3420		dso = map__dso(al.map);
3421		if (!dso || !dso__auxtrace_cache(dso))
3422			continue;
3423
3424		offset = map__map_ip(al.map, addr);
3425
3426		e = intel_pt_cache_lookup(dso, machine, offset);
3427		if (!e)
3428			continue;
3429
3430		if (addr + e->byte_cnt + e->length <= event->text_poke.addr) {
3431			/*
3432			 * No overlap. Working backwards there cannot be another
3433			 * basic block that overlaps the text poke if there is a
3434			 * branch instruction before the text poke address.
3435			 */
3436			if (e->branch != INTEL_PT_BR_NO_BRANCH)
3437				goto out;
3438		} else {
3439			intel_pt_cache_invalidate(dso, machine, offset);
3440			intel_pt_log("Invalidated instruction cache for %s at %#"PRIx64"\n",
3441				     dso__long_name(dso), addr);
3442		}
3443	}
3444out:
3445	addr_location__exit(&al);
3446	return ret;
3447}
3448
3449static int intel_pt_process_event(struct perf_session *session,
3450				  union perf_event *event,
3451				  struct perf_sample *sample,
3452				  const struct perf_tool *tool)
3453{
3454	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
3455					   auxtrace);
3456	u64 timestamp;
3457	int err = 0;
3458
3459	if (dump_trace)
3460		return 0;
3461
3462	if (!tool->ordered_events) {
3463		pr_err("Intel Processor Trace requires ordered events\n");
3464		return -EINVAL;
3465	}
3466
3467	if (sample->time && sample->time != (u64)-1)
3468		timestamp = perf_time_to_tsc(sample->time, &pt->tc);
3469	else
3470		timestamp = 0;
3471
3472	if (timestamp || pt->timeless_decoding) {
3473		err = intel_pt_update_queues(pt);
3474		if (err)
3475			return err;
3476	}
3477
3478	if (pt->timeless_decoding) {
3479		if (pt->sampling_mode) {
3480			if (sample->aux_sample.size)
3481				err = intel_pt_process_timeless_sample(pt,
3482								       sample);
3483		} else if (event->header.type == PERF_RECORD_EXIT) {
3484			err = intel_pt_process_timeless_queues(pt,
3485							       event->fork.tid,
3486							       sample->time);
3487		}
3488	} else if (timestamp) {
3489		if (!pt->first_timestamp)
3490			intel_pt_first_timestamp(pt, timestamp);
3491		err = intel_pt_process_queues(pt, timestamp);
3492	}
3493	if (err)
3494		return err;
3495
3496	if (event->header.type == PERF_RECORD_SAMPLE) {
3497		if (pt->synth_opts.add_callchain && !sample->callchain)
3498			intel_pt_add_callchain(pt, sample);
3499		if (pt->synth_opts.add_last_branch && !sample->branch_stack)
3500			intel_pt_add_br_stack(pt, sample);
3501	}
3502
3503	if (event->header.type == PERF_RECORD_AUX &&
3504	    (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
3505	    pt->synth_opts.errors) {
3506		err = intel_pt_lost(pt, sample);
3507		if (err)
3508			return err;
3509	}
3510
3511#ifdef HAVE_LIBTRACEEVENT
3512	if (pt->switch_evsel && event->header.type == PERF_RECORD_SAMPLE)
3513		err = intel_pt_process_switch(pt, sample);
3514	else
3515#endif
3516	if (event->header.type == PERF_RECORD_ITRACE_START)
3517		err = intel_pt_process_itrace_start(pt, event, sample);
3518	else if (event->header.type == PERF_RECORD_AUX_OUTPUT_HW_ID)
3519		err = intel_pt_process_aux_output_hw_id(pt, event, sample);
3520	else if (event->header.type == PERF_RECORD_SWITCH ||
3521		 event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
3522		err = intel_pt_context_switch(pt, event, sample);
3523
3524	if (!err && event->header.type == PERF_RECORD_TEXT_POKE)
3525		err = intel_pt_text_poke(pt, event);
3526
3527	if (intel_pt_enable_logging && intel_pt_log_events(pt, sample->time)) {
3528		intel_pt_log("event %u: cpu %d time %"PRIu64" tsc %#"PRIx64" ",
3529			     event->header.type, sample->cpu, sample->time, timestamp);
3530		intel_pt_log_event(event);
3531	}
3532
3533	return err;
3534}
3535
3536static int intel_pt_flush(struct perf_session *session, const struct perf_tool *tool)
3537{
3538	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
3539					   auxtrace);
3540	int ret;
3541
3542	if (dump_trace)
3543		return 0;
3544
3545	if (!tool->ordered_events)
3546		return -EINVAL;
3547
3548	ret = intel_pt_update_queues(pt);
3549	if (ret < 0)
3550		return ret;
3551
3552	if (pt->timeless_decoding)
3553		return intel_pt_process_timeless_queues(pt, -1,
3554							MAX_TIMESTAMP - 1);
3555
3556	return intel_pt_process_queues(pt, MAX_TIMESTAMP);
3557}
3558
3559static void intel_pt_free_events(struct perf_session *session)
3560{
3561	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
3562					   auxtrace);
3563	struct auxtrace_queues *queues = &pt->queues;
3564	unsigned int i;
3565
3566	for (i = 0; i < queues->nr_queues; i++) {
3567		intel_pt_free_queue(queues->queue_array[i].priv);
3568		queues->queue_array[i].priv = NULL;
3569	}
3570	intel_pt_log_disable();
3571	auxtrace_queues__free(queues);
3572}
3573
3574static void intel_pt_free(struct perf_session *session)
3575{
3576	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
3577					   auxtrace);
3578
3579	auxtrace_heap__free(&pt->heap);
3580	intel_pt_free_events(session);
3581	session->auxtrace = NULL;
3582	intel_pt_free_vmcs_info(pt);
3583	thread__put(pt->unknown_thread);
3584	addr_filters__exit(&pt->filts);
3585	zfree(&pt->chain);
3586	zfree(&pt->filter);
3587	zfree(&pt->time_ranges);
3588	zfree(&pt->br_stack);
3589	free(pt);
3590}
3591
3592static bool intel_pt_evsel_is_auxtrace(struct perf_session *session,
3593				       struct evsel *evsel)
3594{
3595	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
3596					   auxtrace);
3597
3598	return evsel->core.attr.type == pt->pmu_type;
3599}
3600
3601static int intel_pt_process_auxtrace_event(struct perf_session *session,
3602					   union perf_event *event,
3603					   const struct perf_tool *tool __maybe_unused)
3604{
3605	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
3606					   auxtrace);
3607
3608	if (!pt->data_queued) {
3609		struct auxtrace_buffer *buffer;
3610		off_t data_offset;
3611		int fd = perf_data__fd(session->data);
3612		int err;
3613
3614		if (perf_data__is_pipe(session->data)) {
3615			data_offset = 0;
3616		} else {
3617			data_offset = lseek(fd, 0, SEEK_CUR);
3618			if (data_offset == -1)
3619				return -errno;
3620		}
3621
3622		err = auxtrace_queues__add_event(&pt->queues, session, event,
3623						 data_offset, &buffer);
3624		if (err)
3625			return err;
3626
3627		/* Dump here now we have copied a piped trace out of the pipe */
3628		if (dump_trace) {
3629			if (auxtrace_buffer__get_data(buffer, fd)) {
3630				intel_pt_dump_event(pt, buffer->data,
3631						    buffer->size);
3632				auxtrace_buffer__put_data(buffer);
3633			}
3634		}
3635	}
3636
3637	return 0;
3638}
3639
3640static int intel_pt_queue_data(struct perf_session *session,
3641			       struct perf_sample *sample,
3642			       union perf_event *event, u64 data_offset)
3643{
3644	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
3645					   auxtrace);
3646	u64 timestamp;
3647
3648	if (event) {
3649		return auxtrace_queues__add_event(&pt->queues, session, event,
3650						  data_offset, NULL);
3651	}
3652
3653	if (sample->time && sample->time != (u64)-1)
3654		timestamp = perf_time_to_tsc(sample->time, &pt->tc);
3655	else
3656		timestamp = 0;
3657
3658	return auxtrace_queues__add_sample(&pt->queues, session, sample,
3659					   data_offset, timestamp);
3660}
3661
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3662static int intel_pt_synth_event(struct perf_session *session, const char *name,
3663				struct perf_event_attr *attr, u64 id)
3664{
 
3665	int err;
3666
3667	pr_debug("Synthesizing '%s' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
3668		 name, id, (u64)attr->sample_type);
3669
3670	err = perf_session__deliver_synth_attr_event(session, attr, id);
 
 
 
 
3671	if (err)
3672		pr_err("%s: failed to synthesize '%s' event type\n",
3673		       __func__, name);
3674
3675	return err;
3676}
3677
3678static void intel_pt_set_event_name(struct evlist *evlist, u64 id,
3679				    const char *name)
3680{
3681	struct evsel *evsel;
3682
3683	evlist__for_each_entry(evlist, evsel) {
3684		if (evsel->core.id && evsel->core.id[0] == id) {
3685			if (evsel->name)
3686				zfree(&evsel->name);
3687			evsel->name = strdup(name);
3688			break;
3689		}
3690	}
3691}
3692
3693static struct evsel *intel_pt_evsel(struct intel_pt *pt,
3694					 struct evlist *evlist)
3695{
3696	struct evsel *evsel;
3697
3698	evlist__for_each_entry(evlist, evsel) {
3699		if (evsel->core.attr.type == pt->pmu_type && evsel->core.ids)
3700			return evsel;
3701	}
3702
3703	return NULL;
3704}
3705
3706static int intel_pt_synth_events(struct intel_pt *pt,
3707				 struct perf_session *session)
3708{
3709	struct evlist *evlist = session->evlist;
3710	struct evsel *evsel = intel_pt_evsel(pt, evlist);
3711	struct perf_event_attr attr;
3712	u64 id;
3713	int err;
3714
3715	if (!evsel) {
3716		pr_debug("There are no selected events with Intel Processor Trace data\n");
3717		return 0;
3718	}
3719
3720	memset(&attr, 0, sizeof(struct perf_event_attr));
3721	attr.size = sizeof(struct perf_event_attr);
3722	attr.type = PERF_TYPE_HARDWARE;
3723	attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK;
3724	attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
3725			    PERF_SAMPLE_PERIOD;
3726	if (pt->timeless_decoding)
3727		attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
3728	else
3729		attr.sample_type |= PERF_SAMPLE_TIME;
3730	if (!pt->per_cpu_mmaps)
3731		attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
3732	attr.exclude_user = evsel->core.attr.exclude_user;
3733	attr.exclude_kernel = evsel->core.attr.exclude_kernel;
3734	attr.exclude_hv = evsel->core.attr.exclude_hv;
3735	attr.exclude_host = evsel->core.attr.exclude_host;
3736	attr.exclude_guest = evsel->core.attr.exclude_guest;
3737	attr.sample_id_all = evsel->core.attr.sample_id_all;
3738	attr.read_format = evsel->core.attr.read_format;
3739
3740	id = evsel->core.id[0] + 1000000000;
3741	if (!id)
3742		id = 1;
3743
3744	if (pt->synth_opts.branches) {
3745		attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
3746		attr.sample_period = 1;
3747		attr.sample_type |= PERF_SAMPLE_ADDR;
3748		err = intel_pt_synth_event(session, "branches", &attr, id);
3749		if (err)
3750			return err;
3751		pt->sample_branches = true;
3752		pt->branches_sample_type = attr.sample_type;
3753		pt->branches_id = id;
3754		id += 1;
3755		attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR;
3756	}
3757
3758	if (pt->synth_opts.callchain)
3759		attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
3760	if (pt->synth_opts.last_branch) {
3761		attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
3762		/*
3763		 * We don't use the hardware index, but the sample generation
3764		 * code uses the new format branch_stack with this field,
3765		 * so the event attributes must indicate that it's present.
3766		 */
3767		attr.branch_sample_type |= PERF_SAMPLE_BRANCH_HW_INDEX;
3768	}
3769
3770	if (pt->synth_opts.instructions) {
3771		attr.config = PERF_COUNT_HW_INSTRUCTIONS;
3772		if (pt->synth_opts.period_type == PERF_ITRACE_PERIOD_NANOSECS)
3773			attr.sample_period =
3774				intel_pt_ns_to_ticks(pt, pt->synth_opts.period);
3775		else
3776			attr.sample_period = pt->synth_opts.period;
3777		err = intel_pt_synth_event(session, "instructions", &attr, id);
3778		if (err)
3779			return err;
3780		pt->sample_instructions = true;
3781		pt->instructions_sample_type = attr.sample_type;
3782		pt->instructions_id = id;
3783		id += 1;
3784	}
3785
3786	if (pt->synth_opts.cycles) {
3787		attr.config = PERF_COUNT_HW_CPU_CYCLES;
3788		if (pt->synth_opts.period_type == PERF_ITRACE_PERIOD_NANOSECS)
3789			attr.sample_period =
3790				intel_pt_ns_to_ticks(pt, pt->synth_opts.period);
3791		else
3792			attr.sample_period = pt->synth_opts.period;
3793		err = intel_pt_synth_event(session, "cycles", &attr, id);
3794		if (err)
3795			return err;
3796		pt->sample_cycles = true;
3797		pt->cycles_sample_type = attr.sample_type;
3798		pt->cycles_id = id;
3799		id += 1;
3800	}
3801
3802	attr.sample_type &= ~(u64)PERF_SAMPLE_PERIOD;
3803	attr.sample_period = 1;
3804
3805	if (pt->synth_opts.transactions) {
3806		attr.config = PERF_COUNT_HW_INSTRUCTIONS;
3807		err = intel_pt_synth_event(session, "transactions", &attr, id);
3808		if (err)
3809			return err;
3810		pt->sample_transactions = true;
3811		pt->transactions_sample_type = attr.sample_type;
3812		pt->transactions_id = id;
3813		intel_pt_set_event_name(evlist, id, "transactions");
3814		id += 1;
3815	}
3816
3817	attr.type = PERF_TYPE_SYNTH;
3818	attr.sample_type |= PERF_SAMPLE_RAW;
3819
3820	if (pt->synth_opts.ptwrites) {
3821		attr.config = PERF_SYNTH_INTEL_PTWRITE;
3822		err = intel_pt_synth_event(session, "ptwrite", &attr, id);
3823		if (err)
3824			return err;
3825		pt->sample_ptwrites = true;
3826		pt->ptwrites_sample_type = attr.sample_type;
3827		pt->ptwrites_id = id;
3828		intel_pt_set_event_name(evlist, id, "ptwrite");
3829		id += 1;
3830	}
3831
3832	if (pt->synth_opts.pwr_events) {
3833		pt->sample_pwr_events = true;
3834		pt->pwr_events_sample_type = attr.sample_type;
3835
3836		attr.config = PERF_SYNTH_INTEL_CBR;
3837		err = intel_pt_synth_event(session, "cbr", &attr, id);
3838		if (err)
3839			return err;
3840		pt->cbr_id = id;
3841		intel_pt_set_event_name(evlist, id, "cbr");
3842		id += 1;
3843
3844		attr.config = PERF_SYNTH_INTEL_PSB;
3845		err = intel_pt_synth_event(session, "psb", &attr, id);
3846		if (err)
3847			return err;
3848		pt->psb_id = id;
3849		intel_pt_set_event_name(evlist, id, "psb");
3850		id += 1;
3851	}
3852
3853	if (pt->synth_opts.pwr_events && (evsel->core.attr.config & INTEL_PT_CFG_PWR_EVT_EN)) {
3854		attr.config = PERF_SYNTH_INTEL_MWAIT;
3855		err = intel_pt_synth_event(session, "mwait", &attr, id);
3856		if (err)
3857			return err;
3858		pt->mwait_id = id;
3859		intel_pt_set_event_name(evlist, id, "mwait");
3860		id += 1;
3861
3862		attr.config = PERF_SYNTH_INTEL_PWRE;
3863		err = intel_pt_synth_event(session, "pwre", &attr, id);
3864		if (err)
3865			return err;
3866		pt->pwre_id = id;
3867		intel_pt_set_event_name(evlist, id, "pwre");
3868		id += 1;
3869
3870		attr.config = PERF_SYNTH_INTEL_EXSTOP;
3871		err = intel_pt_synth_event(session, "exstop", &attr, id);
3872		if (err)
3873			return err;
3874		pt->exstop_id = id;
3875		intel_pt_set_event_name(evlist, id, "exstop");
3876		id += 1;
3877
3878		attr.config = PERF_SYNTH_INTEL_PWRX;
3879		err = intel_pt_synth_event(session, "pwrx", &attr, id);
3880		if (err)
3881			return err;
3882		pt->pwrx_id = id;
3883		intel_pt_set_event_name(evlist, id, "pwrx");
3884		id += 1;
3885	}
3886
3887	if (pt->synth_opts.intr_events && (evsel->core.attr.config & INTEL_PT_CFG_EVT_EN)) {
3888		attr.config = PERF_SYNTH_INTEL_EVT;
3889		err = intel_pt_synth_event(session, "evt", &attr, id);
3890		if (err)
3891			return err;
3892		pt->evt_sample_type = attr.sample_type;
3893		pt->evt_id = id;
3894		intel_pt_set_event_name(evlist, id, "evt");
3895		id += 1;
3896	}
3897
3898	if (pt->synth_opts.intr_events && pt->cap_event_trace) {
3899		attr.config = PERF_SYNTH_INTEL_IFLAG_CHG;
3900		err = intel_pt_synth_event(session, "iflag", &attr, id);
3901		if (err)
3902			return err;
3903		pt->iflag_chg_sample_type = attr.sample_type;
3904		pt->iflag_chg_id = id;
3905		intel_pt_set_event_name(evlist, id, "iflag");
3906		id += 1;
3907	}
3908
3909	return 0;
3910}
3911
3912static void intel_pt_setup_pebs_events(struct intel_pt *pt)
3913{
3914	struct evsel *evsel;
3915
3916	if (!pt->synth_opts.other_events)
3917		return;
3918
3919	evlist__for_each_entry(pt->session->evlist, evsel) {
3920		if (evsel->core.attr.aux_output && evsel->core.id) {
3921			if (pt->single_pebs) {
3922				pt->single_pebs = false;
3923				return;
3924			}
3925			pt->single_pebs = true;
3926			pt->sample_pebs = true;
3927			pt->pebs_evsel = evsel;
 
3928		}
3929	}
3930}
3931
3932static struct evsel *intel_pt_find_sched_switch(struct evlist *evlist)
3933{
3934	struct evsel *evsel;
3935
3936	evlist__for_each_entry_reverse(evlist, evsel) {
3937		const char *name = evsel__name(evsel);
3938
3939		if (!strcmp(name, "sched:sched_switch"))
3940			return evsel;
3941	}
3942
3943	return NULL;
3944}
3945
3946static bool intel_pt_find_switch(struct evlist *evlist)
3947{
3948	struct evsel *evsel;
3949
3950	evlist__for_each_entry(evlist, evsel) {
3951		if (evsel->core.attr.context_switch)
3952			return true;
3953	}
3954
3955	return false;
3956}
3957
3958static int intel_pt_perf_config(const char *var, const char *value, void *data)
3959{
3960	struct intel_pt *pt = data;
3961
3962	if (!strcmp(var, "intel-pt.mispred-all"))
3963		pt->mispred_all = perf_config_bool(var, value);
3964
3965	if (!strcmp(var, "intel-pt.max-loops"))
3966		perf_config_int(&pt->max_loops, var, value);
3967
3968	return 0;
3969}
3970
3971/* Find least TSC which converts to ns or later */
3972static u64 intel_pt_tsc_start(u64 ns, struct intel_pt *pt)
3973{
3974	u64 tsc, tm;
3975
3976	tsc = perf_time_to_tsc(ns, &pt->tc);
3977
3978	while (1) {
3979		tm = tsc_to_perf_time(tsc, &pt->tc);
3980		if (tm < ns)
3981			break;
3982		tsc -= 1;
3983	}
3984
3985	while (tm < ns)
3986		tm = tsc_to_perf_time(++tsc, &pt->tc);
3987
3988	return tsc;
3989}
3990
3991/* Find greatest TSC which converts to ns or earlier */
3992static u64 intel_pt_tsc_end(u64 ns, struct intel_pt *pt)
3993{
3994	u64 tsc, tm;
3995
3996	tsc = perf_time_to_tsc(ns, &pt->tc);
3997
3998	while (1) {
3999		tm = tsc_to_perf_time(tsc, &pt->tc);
4000		if (tm > ns)
4001			break;
4002		tsc += 1;
4003	}
4004
4005	while (tm > ns)
4006		tm = tsc_to_perf_time(--tsc, &pt->tc);
4007
4008	return tsc;
4009}
4010
4011static int intel_pt_setup_time_ranges(struct intel_pt *pt,
4012				      struct itrace_synth_opts *opts)
4013{
4014	struct perf_time_interval *p = opts->ptime_range;
4015	int n = opts->range_num;
4016	int i;
4017
4018	if (!n || !p || pt->timeless_decoding)
4019		return 0;
4020
4021	pt->time_ranges = calloc(n, sizeof(struct range));
4022	if (!pt->time_ranges)
4023		return -ENOMEM;
4024
4025	pt->range_cnt = n;
4026
4027	intel_pt_log("%s: %u range(s)\n", __func__, n);
4028
4029	for (i = 0; i < n; i++) {
4030		struct range *r = &pt->time_ranges[i];
4031		u64 ts = p[i].start;
4032		u64 te = p[i].end;
4033
4034		/*
4035		 * Take care to ensure the TSC range matches the perf-time range
4036		 * when converted back to perf-time.
4037		 */
4038		r->start = ts ? intel_pt_tsc_start(ts, pt) : 0;
4039		r->end   = te ? intel_pt_tsc_end(te, pt) : 0;
4040
4041		intel_pt_log("range %d: perf time interval: %"PRIu64" to %"PRIu64"\n",
4042			     i, ts, te);
4043		intel_pt_log("range %d: TSC time interval: %#"PRIx64" to %#"PRIx64"\n",
4044			     i, r->start, r->end);
4045	}
4046
4047	return 0;
4048}
4049
4050static int intel_pt_parse_vm_tm_corr_arg(struct intel_pt *pt, char **args)
4051{
4052	struct intel_pt_vmcs_info *vmcs_info;
4053	u64 tsc_offset, vmcs;
4054	char *p = *args;
4055
4056	errno = 0;
4057
4058	p = skip_spaces(p);
4059	if (!*p)
4060		return 1;
4061
4062	tsc_offset = strtoull(p, &p, 0);
4063	if (errno)
4064		return -errno;
4065	p = skip_spaces(p);
4066	if (*p != ':') {
4067		pt->dflt_tsc_offset = tsc_offset;
4068		*args = p;
4069		return 0;
4070	}
4071	p += 1;
4072	while (1) {
4073		vmcs = strtoull(p, &p, 0);
4074		if (errno)
4075			return -errno;
4076		if (!vmcs)
4077			return -EINVAL;
4078		vmcs_info = intel_pt_findnew_vmcs(&pt->vmcs_info, vmcs, tsc_offset);
4079		if (!vmcs_info)
4080			return -ENOMEM;
4081		p = skip_spaces(p);
4082		if (*p != ',')
4083			break;
4084		p += 1;
4085	}
4086	*args = p;
4087	return 0;
4088}
4089
4090static int intel_pt_parse_vm_tm_corr_args(struct intel_pt *pt)
4091{
4092	char *args = pt->synth_opts.vm_tm_corr_args;
4093	int ret;
4094
4095	if (!args)
4096		return 0;
4097
4098	do {
4099		ret = intel_pt_parse_vm_tm_corr_arg(pt, &args);
4100	} while (!ret);
4101
4102	if (ret < 0) {
4103		pr_err("Failed to parse VM Time Correlation options\n");
4104		return ret;
4105	}
4106
4107	return 0;
4108}
4109
4110static const char * const intel_pt_info_fmts[] = {
4111	[INTEL_PT_PMU_TYPE]		= "  PMU Type            %"PRId64"\n",
4112	[INTEL_PT_TIME_SHIFT]		= "  Time Shift          %"PRIu64"\n",
4113	[INTEL_PT_TIME_MULT]		= "  Time Multiplier     %"PRIu64"\n",
4114	[INTEL_PT_TIME_ZERO]		= "  Time Zero           %"PRIu64"\n",
4115	[INTEL_PT_CAP_USER_TIME_ZERO]	= "  Cap Time Zero       %"PRId64"\n",
4116	[INTEL_PT_TSC_BIT]		= "  TSC bit             %#"PRIx64"\n",
4117	[INTEL_PT_NORETCOMP_BIT]	= "  NoRETComp bit       %#"PRIx64"\n",
4118	[INTEL_PT_HAVE_SCHED_SWITCH]	= "  Have sched_switch   %"PRId64"\n",
4119	[INTEL_PT_SNAPSHOT_MODE]	= "  Snapshot mode       %"PRId64"\n",
4120	[INTEL_PT_PER_CPU_MMAPS]	= "  Per-cpu maps        %"PRId64"\n",
4121	[INTEL_PT_MTC_BIT]		= "  MTC bit             %#"PRIx64"\n",
4122	[INTEL_PT_MTC_FREQ_BITS]	= "  MTC freq bits       %#"PRIx64"\n",
4123	[INTEL_PT_TSC_CTC_N]		= "  TSC:CTC numerator   %"PRIu64"\n",
4124	[INTEL_PT_TSC_CTC_D]		= "  TSC:CTC denominator %"PRIu64"\n",
4125	[INTEL_PT_CYC_BIT]		= "  CYC bit             %#"PRIx64"\n",
4126	[INTEL_PT_MAX_NONTURBO_RATIO]	= "  Max non-turbo ratio %"PRIu64"\n",
4127	[INTEL_PT_FILTER_STR_LEN]	= "  Filter string len.  %"PRIu64"\n",
4128};
4129
4130static void intel_pt_print_info(__u64 *arr, int start, int finish)
4131{
4132	int i;
4133
4134	if (!dump_trace)
4135		return;
4136
4137	for (i = start; i <= finish; i++) {
4138		const char *fmt = intel_pt_info_fmts[i];
4139
4140		if (fmt)
4141			fprintf(stdout, fmt, arr[i]);
4142	}
4143}
4144
4145static void intel_pt_print_info_str(const char *name, const char *str)
4146{
4147	if (!dump_trace)
4148		return;
4149
4150	fprintf(stdout, "  %-20s%s\n", name, str ? str : "");
4151}
4152
4153static bool intel_pt_has(struct perf_record_auxtrace_info *auxtrace_info, int pos)
4154{
4155	return auxtrace_info->header.size >=
4156		sizeof(struct perf_record_auxtrace_info) + (sizeof(u64) * (pos + 1));
4157}
4158
4159int intel_pt_process_auxtrace_info(union perf_event *event,
4160				   struct perf_session *session)
4161{
4162	struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info;
4163	size_t min_sz = sizeof(u64) * INTEL_PT_PER_CPU_MMAPS;
4164	struct intel_pt *pt;
4165	void *info_end;
4166	__u64 *info;
4167	int err;
4168
4169	if (auxtrace_info->header.size < sizeof(struct perf_record_auxtrace_info) +
4170					min_sz)
4171		return -EINVAL;
4172
4173	pt = zalloc(sizeof(struct intel_pt));
4174	if (!pt)
4175		return -ENOMEM;
4176
4177	pt->vmcs_info = RB_ROOT;
4178
4179	addr_filters__init(&pt->filts);
4180
4181	err = perf_config(intel_pt_perf_config, pt);
4182	if (err)
4183		goto err_free;
4184
4185	err = auxtrace_queues__init(&pt->queues);
4186	if (err)
4187		goto err_free;
4188
4189	if (session->itrace_synth_opts->set) {
4190		pt->synth_opts = *session->itrace_synth_opts;
4191	} else {
4192		struct itrace_synth_opts *opts = session->itrace_synth_opts;
4193
4194		itrace_synth_opts__set_default(&pt->synth_opts, opts->default_no_sample);
4195		if (!opts->default_no_sample && !opts->inject) {
4196			pt->synth_opts.branches = false;
4197			pt->synth_opts.callchain = true;
4198			pt->synth_opts.add_callchain = true;
4199		}
4200		pt->synth_opts.thread_stack = opts->thread_stack;
4201	}
4202
4203	if (!(pt->synth_opts.log_plus_flags & AUXTRACE_LOG_FLG_USE_STDOUT))
4204		intel_pt_log_set_name(INTEL_PT_PMU_NAME);
4205
4206	pt->session = session;
4207	pt->machine = &session->machines.host; /* No kvm support */
4208	pt->auxtrace_type = auxtrace_info->type;
4209	pt->pmu_type = auxtrace_info->priv[INTEL_PT_PMU_TYPE];
4210	pt->tc.time_shift = auxtrace_info->priv[INTEL_PT_TIME_SHIFT];
4211	pt->tc.time_mult = auxtrace_info->priv[INTEL_PT_TIME_MULT];
4212	pt->tc.time_zero = auxtrace_info->priv[INTEL_PT_TIME_ZERO];
4213	pt->cap_user_time_zero = auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO];
4214	pt->tsc_bit = auxtrace_info->priv[INTEL_PT_TSC_BIT];
4215	pt->noretcomp_bit = auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT];
4216	pt->have_sched_switch = auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH];
4217	pt->snapshot_mode = auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE];
4218	pt->per_cpu_mmaps = auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS];
4219	intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_PMU_TYPE,
4220			    INTEL_PT_PER_CPU_MMAPS);
4221
4222	if (intel_pt_has(auxtrace_info, INTEL_PT_CYC_BIT)) {
4223		pt->mtc_bit = auxtrace_info->priv[INTEL_PT_MTC_BIT];
4224		pt->mtc_freq_bits = auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS];
4225		pt->tsc_ctc_ratio_n = auxtrace_info->priv[INTEL_PT_TSC_CTC_N];
4226		pt->tsc_ctc_ratio_d = auxtrace_info->priv[INTEL_PT_TSC_CTC_D];
4227		pt->cyc_bit = auxtrace_info->priv[INTEL_PT_CYC_BIT];
4228		intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_MTC_BIT,
4229				    INTEL_PT_CYC_BIT);
4230	}
4231
4232	if (intel_pt_has(auxtrace_info, INTEL_PT_MAX_NONTURBO_RATIO)) {
4233		pt->max_non_turbo_ratio =
4234			auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO];
4235		intel_pt_print_info(&auxtrace_info->priv[0],
4236				    INTEL_PT_MAX_NONTURBO_RATIO,
4237				    INTEL_PT_MAX_NONTURBO_RATIO);
4238	}
4239
4240	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
4241	info_end = (void *)auxtrace_info + auxtrace_info->header.size;
4242
4243	if (intel_pt_has(auxtrace_info, INTEL_PT_FILTER_STR_LEN)) {
4244		size_t len;
4245
4246		len = auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN];
4247		intel_pt_print_info(&auxtrace_info->priv[0],
4248				    INTEL_PT_FILTER_STR_LEN,
4249				    INTEL_PT_FILTER_STR_LEN);
4250		if (len) {
4251			const char *filter = (const char *)info;
4252
4253			len = roundup(len + 1, 8);
4254			info += len >> 3;
4255			if ((void *)info > info_end) {
4256				pr_err("%s: bad filter string length\n", __func__);
4257				err = -EINVAL;
4258				goto err_free_queues;
4259			}
4260			pt->filter = memdup(filter, len);
4261			if (!pt->filter) {
4262				err = -ENOMEM;
4263				goto err_free_queues;
4264			}
4265			if (session->header.needs_swap)
4266				mem_bswap_64(pt->filter, len);
4267			if (pt->filter[len - 1]) {
4268				pr_err("%s: filter string not null terminated\n", __func__);
4269				err = -EINVAL;
4270				goto err_free_queues;
4271			}
4272			err = addr_filters__parse_bare_filter(&pt->filts,
4273							      filter);
4274			if (err)
4275				goto err_free_queues;
4276		}
4277		intel_pt_print_info_str("Filter string", pt->filter);
4278	}
4279
4280	if ((void *)info < info_end) {
4281		pt->cap_event_trace = *info++;
4282		if (dump_trace)
4283			fprintf(stdout, "  Cap Event Trace     %d\n",
4284				pt->cap_event_trace);
4285	}
4286
4287	pt->timeless_decoding = intel_pt_timeless_decoding(pt);
4288	if (pt->timeless_decoding && !pt->tc.time_mult)
4289		pt->tc.time_mult = 1;
4290	pt->have_tsc = intel_pt_have_tsc(pt);
4291	pt->sampling_mode = intel_pt_sampling_mode(pt);
4292	pt->est_tsc = !pt->timeless_decoding;
4293
4294	if (pt->synth_opts.vm_time_correlation) {
4295		if (pt->timeless_decoding) {
4296			pr_err("Intel PT has no time information for VM Time Correlation\n");
4297			err = -EINVAL;
4298			goto err_free_queues;
4299		}
4300		if (session->itrace_synth_opts->ptime_range) {
4301			pr_err("Time ranges cannot be specified with VM Time Correlation\n");
4302			err = -EINVAL;
4303			goto err_free_queues;
4304		}
4305		/* Currently TSC Offset is calculated using MTC packets */
4306		if (!intel_pt_have_mtc(pt)) {
4307			pr_err("MTC packets must have been enabled for VM Time Correlation\n");
4308			err = -EINVAL;
4309			goto err_free_queues;
4310		}
4311		err = intel_pt_parse_vm_tm_corr_args(pt);
4312		if (err)
4313			goto err_free_queues;
4314	}
4315
4316	pt->unknown_thread = thread__new(999999999, 999999999);
4317	if (!pt->unknown_thread) {
4318		err = -ENOMEM;
4319		goto err_free_queues;
4320	}
4321
 
 
 
 
 
 
 
 
4322	err = thread__set_comm(pt->unknown_thread, "unknown", 0);
4323	if (err)
4324		goto err_delete_thread;
4325	if (thread__init_maps(pt->unknown_thread, pt->machine)) {
4326		err = -ENOMEM;
4327		goto err_delete_thread;
4328	}
4329
4330	pt->auxtrace.process_event = intel_pt_process_event;
4331	pt->auxtrace.process_auxtrace_event = intel_pt_process_auxtrace_event;
4332	pt->auxtrace.queue_data = intel_pt_queue_data;
4333	pt->auxtrace.dump_auxtrace_sample = intel_pt_dump_sample;
4334	pt->auxtrace.flush_events = intel_pt_flush;
4335	pt->auxtrace.free_events = intel_pt_free_events;
4336	pt->auxtrace.free = intel_pt_free;
4337	pt->auxtrace.evsel_is_auxtrace = intel_pt_evsel_is_auxtrace;
4338	session->auxtrace = &pt->auxtrace;
4339
4340	if (dump_trace)
4341		return 0;
4342
4343	if (pt->have_sched_switch == 1) {
4344		pt->switch_evsel = intel_pt_find_sched_switch(session->evlist);
4345		if (!pt->switch_evsel) {
4346			pr_err("%s: missing sched_switch event\n", __func__);
4347			err = -EINVAL;
4348			goto err_delete_thread;
4349		}
4350	} else if (pt->have_sched_switch == 2 &&
4351		   !intel_pt_find_switch(session->evlist)) {
4352		pr_err("%s: missing context_switch attribute flag\n", __func__);
4353		err = -EINVAL;
4354		goto err_delete_thread;
4355	}
4356
4357	if (pt->synth_opts.log) {
4358		bool log_on_error = pt->synth_opts.log_plus_flags & AUXTRACE_LOG_FLG_ON_ERROR;
4359		unsigned int log_on_error_size = pt->synth_opts.log_on_error_size;
4360
4361		intel_pt_log_enable(log_on_error, log_on_error_size);
 
 
 
 
 
 
 
 
4362	}
4363
 
 
 
4364	/* Maximum non-turbo ratio is TSC freq / 100 MHz */
4365	if (pt->tc.time_mult) {
4366		u64 tsc_freq = intel_pt_ns_to_ticks(pt, 1000000000);
4367
4368		if (!pt->max_non_turbo_ratio)
4369			pt->max_non_turbo_ratio =
4370					(tsc_freq + 50000000) / 100000000;
4371		intel_pt_log("TSC frequency %"PRIu64"\n", tsc_freq);
4372		intel_pt_log("Maximum non-turbo ratio %u\n",
4373			     pt->max_non_turbo_ratio);
4374		pt->cbr2khz = tsc_freq / pt->max_non_turbo_ratio / 1000;
4375	}
4376
4377	err = intel_pt_setup_time_ranges(pt, session->itrace_synth_opts);
4378	if (err)
4379		goto err_delete_thread;
4380
4381	if (pt->synth_opts.calls)
4382		pt->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
4383				       PERF_IP_FLAG_TRACE_END;
4384	if (pt->synth_opts.returns)
4385		pt->branches_filter |= PERF_IP_FLAG_RETURN |
4386				       PERF_IP_FLAG_TRACE_BEGIN;
4387
4388	if ((pt->synth_opts.callchain || pt->synth_opts.add_callchain) &&
4389	    !symbol_conf.use_callchain) {
4390		symbol_conf.use_callchain = true;
4391		if (callchain_register_param(&callchain_param) < 0) {
4392			symbol_conf.use_callchain = false;
4393			pt->synth_opts.callchain = false;
4394			pt->synth_opts.add_callchain = false;
4395		}
4396	}
4397
4398	if (pt->synth_opts.add_callchain) {
4399		err = intel_pt_callchain_init(pt);
4400		if (err)
4401			goto err_delete_thread;
4402	}
4403
4404	if (pt->synth_opts.last_branch || pt->synth_opts.add_last_branch) {
4405		pt->br_stack_sz = pt->synth_opts.last_branch_sz;
4406		pt->br_stack_sz_plus = pt->br_stack_sz;
4407	}
4408
4409	if (pt->synth_opts.add_last_branch) {
4410		err = intel_pt_br_stack_init(pt);
4411		if (err)
4412			goto err_delete_thread;
4413		/*
4414		 * Additional branch stack size to cater for tracing from the
4415		 * actual sample ip to where the sample time is recorded.
4416		 * Measured at about 200 branches, but generously set to 1024.
4417		 * If kernel space is not being traced, then add just 1 for the
4418		 * branch to kernel space.
4419		 */
4420		if (intel_pt_tracing_kernel(pt))
4421			pt->br_stack_sz_plus += 1024;
4422		else
4423			pt->br_stack_sz_plus += 1;
4424	}
4425
4426	pt->use_thread_stack = pt->synth_opts.callchain ||
4427			       pt->synth_opts.add_callchain ||
4428			       pt->synth_opts.thread_stack ||
4429			       pt->synth_opts.last_branch ||
4430			       pt->synth_opts.add_last_branch;
4431
4432	pt->callstack = pt->synth_opts.callchain ||
4433			pt->synth_opts.add_callchain ||
4434			pt->synth_opts.thread_stack;
4435
4436	err = intel_pt_synth_events(pt, session);
4437	if (err)
4438		goto err_delete_thread;
4439
4440	intel_pt_setup_pebs_events(pt);
4441
4442	if (perf_data__is_pipe(session->data)) {
4443		pr_warning("WARNING: Intel PT with pipe mode is not recommended.\n"
4444			   "         The output cannot relied upon.  In particular,\n"
4445			   "         timestamps and the order of events may be incorrect.\n");
4446	}
4447
4448	if (pt->sampling_mode || list_empty(&session->auxtrace_index))
4449		err = auxtrace_queue_data(session, true, true);
4450	else
4451		err = auxtrace_queues__process_index(&pt->queues, session);
4452	if (err)
4453		goto err_delete_thread;
4454
4455	if (pt->queues.populated)
4456		pt->data_queued = true;
4457
4458	if (pt->timeless_decoding)
4459		pr_debug2("Intel PT decoding without timestamps\n");
4460
4461	return 0;
4462
4463err_delete_thread:
4464	zfree(&pt->chain);
4465	thread__zput(pt->unknown_thread);
4466err_free_queues:
4467	intel_pt_log_disable();
4468	auxtrace_queues__free(&pt->queues);
4469	session->auxtrace = NULL;
4470err_free:
4471	addr_filters__exit(&pt->filts);
4472	zfree(&pt->filter);
4473	zfree(&pt->time_ranges);
4474	free(pt);
4475	return err;
4476}