xref: /linux/tools/perf/util/intel-pt.c (revision 7a309195d11cde854eb75559fbd6b48f9e518f25)
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 <stdio.h>
9 #include <stdbool.h>
10 #include <errno.h>
11 #include <linux/kernel.h>
12 #include <linux/string.h>
13 #include <linux/types.h>
14 #include <linux/zalloc.h>
15 
16 #include "session.h"
17 #include "machine.h"
18 #include "memswap.h"
19 #include "sort.h"
20 #include "tool.h"
21 #include "event.h"
22 #include "evlist.h"
23 #include "evsel.h"
24 #include "map.h"
25 #include "color.h"
26 #include "thread.h"
27 #include "thread-stack.h"
28 #include "symbol.h"
29 #include "callchain.h"
30 #include "dso.h"
31 #include "debug.h"
32 #include "auxtrace.h"
33 #include "tsc.h"
34 #include "intel-pt.h"
35 #include "config.h"
36 #include "util/perf_api_probe.h"
37 #include "util/synthetic-events.h"
38 #include "time-utils.h"
39 
40 #include "../arch/x86/include/uapi/asm/perf_regs.h"
41 
42 #include "intel-pt-decoder/intel-pt-log.h"
43 #include "intel-pt-decoder/intel-pt-decoder.h"
44 #include "intel-pt-decoder/intel-pt-insn-decoder.h"
45 #include "intel-pt-decoder/intel-pt-pkt-decoder.h"
46 
47 #define MAX_TIMESTAMP (~0ULL)
48 
49 struct range {
50 	u64 start;
51 	u64 end;
52 };
53 
54 struct intel_pt {
55 	struct auxtrace auxtrace;
56 	struct auxtrace_queues queues;
57 	struct auxtrace_heap heap;
58 	u32 auxtrace_type;
59 	struct perf_session *session;
60 	struct machine *machine;
61 	struct evsel *switch_evsel;
62 	struct thread *unknown_thread;
63 	bool timeless_decoding;
64 	bool sampling_mode;
65 	bool snapshot_mode;
66 	bool per_cpu_mmaps;
67 	bool have_tsc;
68 	bool data_queued;
69 	bool est_tsc;
70 	bool sync_switch;
71 	bool mispred_all;
72 	bool use_thread_stack;
73 	bool callstack;
74 	unsigned int br_stack_sz;
75 	unsigned int br_stack_sz_plus;
76 	int have_sched_switch;
77 	u32 pmu_type;
78 	u64 kernel_start;
79 	u64 switch_ip;
80 	u64 ptss_ip;
81 
82 	struct perf_tsc_conversion tc;
83 	bool cap_user_time_zero;
84 
85 	struct itrace_synth_opts synth_opts;
86 
87 	bool sample_instructions;
88 	u64 instructions_sample_type;
89 	u64 instructions_id;
90 
91 	bool sample_branches;
92 	u32 branches_filter;
93 	u64 branches_sample_type;
94 	u64 branches_id;
95 
96 	bool sample_transactions;
97 	u64 transactions_sample_type;
98 	u64 transactions_id;
99 
100 	bool sample_ptwrites;
101 	u64 ptwrites_sample_type;
102 	u64 ptwrites_id;
103 
104 	bool sample_pwr_events;
105 	u64 pwr_events_sample_type;
106 	u64 mwait_id;
107 	u64 pwre_id;
108 	u64 exstop_id;
109 	u64 pwrx_id;
110 	u64 cbr_id;
111 
112 	bool sample_pebs;
113 	struct evsel *pebs_evsel;
114 
115 	u64 tsc_bit;
116 	u64 mtc_bit;
117 	u64 mtc_freq_bits;
118 	u32 tsc_ctc_ratio_n;
119 	u32 tsc_ctc_ratio_d;
120 	u64 cyc_bit;
121 	u64 noretcomp_bit;
122 	unsigned max_non_turbo_ratio;
123 	unsigned cbr2khz;
124 
125 	unsigned long num_events;
126 
127 	char *filter;
128 	struct addr_filters filts;
129 
130 	struct range *time_ranges;
131 	unsigned int range_cnt;
132 
133 	struct ip_callchain *chain;
134 	struct branch_stack *br_stack;
135 };
136 
137 enum switch_state {
138 	INTEL_PT_SS_NOT_TRACING,
139 	INTEL_PT_SS_UNKNOWN,
140 	INTEL_PT_SS_TRACING,
141 	INTEL_PT_SS_EXPECTING_SWITCH_EVENT,
142 	INTEL_PT_SS_EXPECTING_SWITCH_IP,
143 };
144 
145 struct intel_pt_queue {
146 	struct intel_pt *pt;
147 	unsigned int queue_nr;
148 	struct auxtrace_buffer *buffer;
149 	struct auxtrace_buffer *old_buffer;
150 	void *decoder;
151 	const struct intel_pt_state *state;
152 	struct ip_callchain *chain;
153 	struct branch_stack *last_branch;
154 	union perf_event *event_buf;
155 	bool on_heap;
156 	bool stop;
157 	bool step_through_buffers;
158 	bool use_buffer_pid_tid;
159 	bool sync_switch;
160 	pid_t pid, tid;
161 	int cpu;
162 	int switch_state;
163 	pid_t next_tid;
164 	struct thread *thread;
165 	bool exclude_kernel;
166 	bool have_sample;
167 	u64 time;
168 	u64 timestamp;
169 	u64 sel_timestamp;
170 	bool sel_start;
171 	unsigned int sel_idx;
172 	u32 flags;
173 	u16 insn_len;
174 	u64 last_insn_cnt;
175 	u64 ipc_insn_cnt;
176 	u64 ipc_cyc_cnt;
177 	u64 last_in_insn_cnt;
178 	u64 last_in_cyc_cnt;
179 	u64 last_br_insn_cnt;
180 	u64 last_br_cyc_cnt;
181 	unsigned int cbr_seen;
182 	char insn[INTEL_PT_INSN_BUF_SZ];
183 };
184 
185 static void intel_pt_dump(struct intel_pt *pt __maybe_unused,
186 			  unsigned char *buf, size_t len)
187 {
188 	struct intel_pt_pkt packet;
189 	size_t pos = 0;
190 	int ret, pkt_len, i;
191 	char desc[INTEL_PT_PKT_DESC_MAX];
192 	const char *color = PERF_COLOR_BLUE;
193 	enum intel_pt_pkt_ctx ctx = INTEL_PT_NO_CTX;
194 
195 	color_fprintf(stdout, color,
196 		      ". ... Intel Processor Trace data: size %zu bytes\n",
197 		      len);
198 
199 	while (len) {
200 		ret = intel_pt_get_packet(buf, len, &packet, &ctx);
201 		if (ret > 0)
202 			pkt_len = ret;
203 		else
204 			pkt_len = 1;
205 		printf(".");
206 		color_fprintf(stdout, color, "  %08x: ", pos);
207 		for (i = 0; i < pkt_len; i++)
208 			color_fprintf(stdout, color, " %02x", buf[i]);
209 		for (; i < 16; i++)
210 			color_fprintf(stdout, color, "   ");
211 		if (ret > 0) {
212 			ret = intel_pt_pkt_desc(&packet, desc,
213 						INTEL_PT_PKT_DESC_MAX);
214 			if (ret > 0)
215 				color_fprintf(stdout, color, " %s\n", desc);
216 		} else {
217 			color_fprintf(stdout, color, " Bad packet!\n");
218 		}
219 		pos += pkt_len;
220 		buf += pkt_len;
221 		len -= pkt_len;
222 	}
223 }
224 
225 static void intel_pt_dump_event(struct intel_pt *pt, unsigned char *buf,
226 				size_t len)
227 {
228 	printf(".\n");
229 	intel_pt_dump(pt, buf, len);
230 }
231 
232 static void intel_pt_log_event(union perf_event *event)
233 {
234 	FILE *f = intel_pt_log_fp();
235 
236 	if (!intel_pt_enable_logging || !f)
237 		return;
238 
239 	perf_event__fprintf(event, f);
240 }
241 
242 static void intel_pt_dump_sample(struct perf_session *session,
243 				 struct perf_sample *sample)
244 {
245 	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
246 					   auxtrace);
247 
248 	printf("\n");
249 	intel_pt_dump(pt, sample->aux_sample.data, sample->aux_sample.size);
250 }
251 
252 static int intel_pt_do_fix_overlap(struct intel_pt *pt, struct auxtrace_buffer *a,
253 				   struct auxtrace_buffer *b)
254 {
255 	bool consecutive = false;
256 	void *start;
257 
258 	start = intel_pt_find_overlap(a->data, a->size, b->data, b->size,
259 				      pt->have_tsc, &consecutive);
260 	if (!start)
261 		return -EINVAL;
262 	b->use_size = b->data + b->size - start;
263 	b->use_data = start;
264 	if (b->use_size && consecutive)
265 		b->consecutive = true;
266 	return 0;
267 }
268 
269 static int intel_pt_get_buffer(struct intel_pt_queue *ptq,
270 			       struct auxtrace_buffer *buffer,
271 			       struct auxtrace_buffer *old_buffer,
272 			       struct intel_pt_buffer *b)
273 {
274 	bool might_overlap;
275 
276 	if (!buffer->data) {
277 		int fd = perf_data__fd(ptq->pt->session->data);
278 
279 		buffer->data = auxtrace_buffer__get_data(buffer, fd);
280 		if (!buffer->data)
281 			return -ENOMEM;
282 	}
283 
284 	might_overlap = ptq->pt->snapshot_mode || ptq->pt->sampling_mode;
285 	if (might_overlap && !buffer->consecutive && old_buffer &&
286 	    intel_pt_do_fix_overlap(ptq->pt, old_buffer, buffer))
287 		return -ENOMEM;
288 
289 	if (buffer->use_data) {
290 		b->len = buffer->use_size;
291 		b->buf = buffer->use_data;
292 	} else {
293 		b->len = buffer->size;
294 		b->buf = buffer->data;
295 	}
296 	b->ref_timestamp = buffer->reference;
297 
298 	if (!old_buffer || (might_overlap && !buffer->consecutive)) {
299 		b->consecutive = false;
300 		b->trace_nr = buffer->buffer_nr + 1;
301 	} else {
302 		b->consecutive = true;
303 	}
304 
305 	return 0;
306 }
307 
308 /* Do not drop buffers with references - refer intel_pt_get_trace() */
309 static void intel_pt_lookahead_drop_buffer(struct intel_pt_queue *ptq,
310 					   struct auxtrace_buffer *buffer)
311 {
312 	if (!buffer || buffer == ptq->buffer || buffer == ptq->old_buffer)
313 		return;
314 
315 	auxtrace_buffer__drop_data(buffer);
316 }
317 
318 /* Must be serialized with respect to intel_pt_get_trace() */
319 static int intel_pt_lookahead(void *data, intel_pt_lookahead_cb_t cb,
320 			      void *cb_data)
321 {
322 	struct intel_pt_queue *ptq = data;
323 	struct auxtrace_buffer *buffer = ptq->buffer;
324 	struct auxtrace_buffer *old_buffer = ptq->old_buffer;
325 	struct auxtrace_queue *queue;
326 	int err = 0;
327 
328 	queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
329 
330 	while (1) {
331 		struct intel_pt_buffer b = { .len = 0 };
332 
333 		buffer = auxtrace_buffer__next(queue, buffer);
334 		if (!buffer)
335 			break;
336 
337 		err = intel_pt_get_buffer(ptq, buffer, old_buffer, &b);
338 		if (err)
339 			break;
340 
341 		if (b.len) {
342 			intel_pt_lookahead_drop_buffer(ptq, old_buffer);
343 			old_buffer = buffer;
344 		} else {
345 			intel_pt_lookahead_drop_buffer(ptq, buffer);
346 			continue;
347 		}
348 
349 		err = cb(&b, cb_data);
350 		if (err)
351 			break;
352 	}
353 
354 	if (buffer != old_buffer)
355 		intel_pt_lookahead_drop_buffer(ptq, buffer);
356 	intel_pt_lookahead_drop_buffer(ptq, old_buffer);
357 
358 	return err;
359 }
360 
361 /*
362  * This function assumes data is processed sequentially only.
363  * Must be serialized with respect to intel_pt_lookahead()
364  */
365 static int intel_pt_get_trace(struct intel_pt_buffer *b, void *data)
366 {
367 	struct intel_pt_queue *ptq = data;
368 	struct auxtrace_buffer *buffer = ptq->buffer;
369 	struct auxtrace_buffer *old_buffer = ptq->old_buffer;
370 	struct auxtrace_queue *queue;
371 	int err;
372 
373 	if (ptq->stop) {
374 		b->len = 0;
375 		return 0;
376 	}
377 
378 	queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
379 
380 	buffer = auxtrace_buffer__next(queue, buffer);
381 	if (!buffer) {
382 		if (old_buffer)
383 			auxtrace_buffer__drop_data(old_buffer);
384 		b->len = 0;
385 		return 0;
386 	}
387 
388 	ptq->buffer = buffer;
389 
390 	err = intel_pt_get_buffer(ptq, buffer, old_buffer, b);
391 	if (err)
392 		return err;
393 
394 	if (ptq->step_through_buffers)
395 		ptq->stop = true;
396 
397 	if (b->len) {
398 		if (old_buffer)
399 			auxtrace_buffer__drop_data(old_buffer);
400 		ptq->old_buffer = buffer;
401 	} else {
402 		auxtrace_buffer__drop_data(buffer);
403 		return intel_pt_get_trace(b, data);
404 	}
405 
406 	return 0;
407 }
408 
409 struct intel_pt_cache_entry {
410 	struct auxtrace_cache_entry	entry;
411 	u64				insn_cnt;
412 	u64				byte_cnt;
413 	enum intel_pt_insn_op		op;
414 	enum intel_pt_insn_branch	branch;
415 	int				length;
416 	int32_t				rel;
417 	char				insn[INTEL_PT_INSN_BUF_SZ];
418 };
419 
420 static int intel_pt_config_div(const char *var, const char *value, void *data)
421 {
422 	int *d = data;
423 	long val;
424 
425 	if (!strcmp(var, "intel-pt.cache-divisor")) {
426 		val = strtol(value, NULL, 0);
427 		if (val > 0 && val <= INT_MAX)
428 			*d = val;
429 	}
430 
431 	return 0;
432 }
433 
434 static int intel_pt_cache_divisor(void)
435 {
436 	static int d;
437 
438 	if (d)
439 		return d;
440 
441 	perf_config(intel_pt_config_div, &d);
442 
443 	if (!d)
444 		d = 64;
445 
446 	return d;
447 }
448 
449 static unsigned int intel_pt_cache_size(struct dso *dso,
450 					struct machine *machine)
451 {
452 	off_t size;
453 
454 	size = dso__data_size(dso, machine);
455 	size /= intel_pt_cache_divisor();
456 	if (size < 1000)
457 		return 10;
458 	if (size > (1 << 21))
459 		return 21;
460 	return 32 - __builtin_clz(size);
461 }
462 
463 static struct auxtrace_cache *intel_pt_cache(struct dso *dso,
464 					     struct machine *machine)
465 {
466 	struct auxtrace_cache *c;
467 	unsigned int bits;
468 
469 	if (dso->auxtrace_cache)
470 		return dso->auxtrace_cache;
471 
472 	bits = intel_pt_cache_size(dso, machine);
473 
474 	/* Ignoring cache creation failure */
475 	c = auxtrace_cache__new(bits, sizeof(struct intel_pt_cache_entry), 200);
476 
477 	dso->auxtrace_cache = c;
478 
479 	return c;
480 }
481 
482 static int intel_pt_cache_add(struct dso *dso, struct machine *machine,
483 			      u64 offset, u64 insn_cnt, u64 byte_cnt,
484 			      struct intel_pt_insn *intel_pt_insn)
485 {
486 	struct auxtrace_cache *c = intel_pt_cache(dso, machine);
487 	struct intel_pt_cache_entry *e;
488 	int err;
489 
490 	if (!c)
491 		return -ENOMEM;
492 
493 	e = auxtrace_cache__alloc_entry(c);
494 	if (!e)
495 		return -ENOMEM;
496 
497 	e->insn_cnt = insn_cnt;
498 	e->byte_cnt = byte_cnt;
499 	e->op = intel_pt_insn->op;
500 	e->branch = intel_pt_insn->branch;
501 	e->length = intel_pt_insn->length;
502 	e->rel = intel_pt_insn->rel;
503 	memcpy(e->insn, intel_pt_insn->buf, INTEL_PT_INSN_BUF_SZ);
504 
505 	err = auxtrace_cache__add(c, offset, &e->entry);
506 	if (err)
507 		auxtrace_cache__free_entry(c, e);
508 
509 	return err;
510 }
511 
512 static struct intel_pt_cache_entry *
513 intel_pt_cache_lookup(struct dso *dso, struct machine *machine, u64 offset)
514 {
515 	struct auxtrace_cache *c = intel_pt_cache(dso, machine);
516 
517 	if (!c)
518 		return NULL;
519 
520 	return auxtrace_cache__lookup(dso->auxtrace_cache, offset);
521 }
522 
523 static inline u8 intel_pt_cpumode(struct intel_pt *pt, uint64_t ip)
524 {
525 	return ip >= pt->kernel_start ?
526 	       PERF_RECORD_MISC_KERNEL :
527 	       PERF_RECORD_MISC_USER;
528 }
529 
530 static int intel_pt_walk_next_insn(struct intel_pt_insn *intel_pt_insn,
531 				   uint64_t *insn_cnt_ptr, uint64_t *ip,
532 				   uint64_t to_ip, uint64_t max_insn_cnt,
533 				   void *data)
534 {
535 	struct intel_pt_queue *ptq = data;
536 	struct machine *machine = ptq->pt->machine;
537 	struct thread *thread;
538 	struct addr_location al;
539 	unsigned char buf[INTEL_PT_INSN_BUF_SZ];
540 	ssize_t len;
541 	int x86_64;
542 	u8 cpumode;
543 	u64 offset, start_offset, start_ip;
544 	u64 insn_cnt = 0;
545 	bool one_map = true;
546 
547 	intel_pt_insn->length = 0;
548 
549 	if (to_ip && *ip == to_ip)
550 		goto out_no_cache;
551 
552 	cpumode = intel_pt_cpumode(ptq->pt, *ip);
553 
554 	thread = ptq->thread;
555 	if (!thread) {
556 		if (cpumode != PERF_RECORD_MISC_KERNEL)
557 			return -EINVAL;
558 		thread = ptq->pt->unknown_thread;
559 	}
560 
561 	while (1) {
562 		if (!thread__find_map(thread, cpumode, *ip, &al) || !al.map->dso)
563 			return -EINVAL;
564 
565 		if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
566 		    dso__data_status_seen(al.map->dso,
567 					  DSO_DATA_STATUS_SEEN_ITRACE))
568 			return -ENOENT;
569 
570 		offset = al.map->map_ip(al.map, *ip);
571 
572 		if (!to_ip && one_map) {
573 			struct intel_pt_cache_entry *e;
574 
575 			e = intel_pt_cache_lookup(al.map->dso, machine, offset);
576 			if (e &&
577 			    (!max_insn_cnt || e->insn_cnt <= max_insn_cnt)) {
578 				*insn_cnt_ptr = e->insn_cnt;
579 				*ip += e->byte_cnt;
580 				intel_pt_insn->op = e->op;
581 				intel_pt_insn->branch = e->branch;
582 				intel_pt_insn->length = e->length;
583 				intel_pt_insn->rel = e->rel;
584 				memcpy(intel_pt_insn->buf, e->insn,
585 				       INTEL_PT_INSN_BUF_SZ);
586 				intel_pt_log_insn_no_data(intel_pt_insn, *ip);
587 				return 0;
588 			}
589 		}
590 
591 		start_offset = offset;
592 		start_ip = *ip;
593 
594 		/* Load maps to ensure dso->is_64_bit has been updated */
595 		map__load(al.map);
596 
597 		x86_64 = al.map->dso->is_64_bit;
598 
599 		while (1) {
600 			len = dso__data_read_offset(al.map->dso, machine,
601 						    offset, buf,
602 						    INTEL_PT_INSN_BUF_SZ);
603 			if (len <= 0)
604 				return -EINVAL;
605 
606 			if (intel_pt_get_insn(buf, len, x86_64, intel_pt_insn))
607 				return -EINVAL;
608 
609 			intel_pt_log_insn(intel_pt_insn, *ip);
610 
611 			insn_cnt += 1;
612 
613 			if (intel_pt_insn->branch != INTEL_PT_BR_NO_BRANCH)
614 				goto out;
615 
616 			if (max_insn_cnt && insn_cnt >= max_insn_cnt)
617 				goto out_no_cache;
618 
619 			*ip += intel_pt_insn->length;
620 
621 			if (to_ip && *ip == to_ip)
622 				goto out_no_cache;
623 
624 			if (*ip >= al.map->end)
625 				break;
626 
627 			offset += intel_pt_insn->length;
628 		}
629 		one_map = false;
630 	}
631 out:
632 	*insn_cnt_ptr = insn_cnt;
633 
634 	if (!one_map)
635 		goto out_no_cache;
636 
637 	/*
638 	 * Didn't lookup in the 'to_ip' case, so do it now to prevent duplicate
639 	 * entries.
640 	 */
641 	if (to_ip) {
642 		struct intel_pt_cache_entry *e;
643 
644 		e = intel_pt_cache_lookup(al.map->dso, machine, start_offset);
645 		if (e)
646 			return 0;
647 	}
648 
649 	/* Ignore cache errors */
650 	intel_pt_cache_add(al.map->dso, machine, start_offset, insn_cnt,
651 			   *ip - start_ip, intel_pt_insn);
652 
653 	return 0;
654 
655 out_no_cache:
656 	*insn_cnt_ptr = insn_cnt;
657 	return 0;
658 }
659 
660 static bool intel_pt_match_pgd_ip(struct intel_pt *pt, uint64_t ip,
661 				  uint64_t offset, const char *filename)
662 {
663 	struct addr_filter *filt;
664 	bool have_filter   = false;
665 	bool hit_tracestop = false;
666 	bool hit_filter    = false;
667 
668 	list_for_each_entry(filt, &pt->filts.head, list) {
669 		if (filt->start)
670 			have_filter = true;
671 
672 		if ((filename && !filt->filename) ||
673 		    (!filename && filt->filename) ||
674 		    (filename && strcmp(filename, filt->filename)))
675 			continue;
676 
677 		if (!(offset >= filt->addr && offset < filt->addr + filt->size))
678 			continue;
679 
680 		intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s hit filter: %s offset %#"PRIx64" size %#"PRIx64"\n",
681 			     ip, offset, filename ? filename : "[kernel]",
682 			     filt->start ? "filter" : "stop",
683 			     filt->addr, filt->size);
684 
685 		if (filt->start)
686 			hit_filter = true;
687 		else
688 			hit_tracestop = true;
689 	}
690 
691 	if (!hit_tracestop && !hit_filter)
692 		intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s is not in a filter region\n",
693 			     ip, offset, filename ? filename : "[kernel]");
694 
695 	return hit_tracestop || (have_filter && !hit_filter);
696 }
697 
698 static int __intel_pt_pgd_ip(uint64_t ip, void *data)
699 {
700 	struct intel_pt_queue *ptq = data;
701 	struct thread *thread;
702 	struct addr_location al;
703 	u8 cpumode;
704 	u64 offset;
705 
706 	if (ip >= ptq->pt->kernel_start)
707 		return intel_pt_match_pgd_ip(ptq->pt, ip, ip, NULL);
708 
709 	cpumode = PERF_RECORD_MISC_USER;
710 
711 	thread = ptq->thread;
712 	if (!thread)
713 		return -EINVAL;
714 
715 	if (!thread__find_map(thread, cpumode, ip, &al) || !al.map->dso)
716 		return -EINVAL;
717 
718 	offset = al.map->map_ip(al.map, ip);
719 
720 	return intel_pt_match_pgd_ip(ptq->pt, ip, offset,
721 				     al.map->dso->long_name);
722 }
723 
724 static bool intel_pt_pgd_ip(uint64_t ip, void *data)
725 {
726 	return __intel_pt_pgd_ip(ip, data) > 0;
727 }
728 
729 static bool intel_pt_get_config(struct intel_pt *pt,
730 				struct perf_event_attr *attr, u64 *config)
731 {
732 	if (attr->type == pt->pmu_type) {
733 		if (config)
734 			*config = attr->config;
735 		return true;
736 	}
737 
738 	return false;
739 }
740 
741 static bool intel_pt_exclude_kernel(struct intel_pt *pt)
742 {
743 	struct evsel *evsel;
744 
745 	evlist__for_each_entry(pt->session->evlist, evsel) {
746 		if (intel_pt_get_config(pt, &evsel->core.attr, NULL) &&
747 		    !evsel->core.attr.exclude_kernel)
748 			return false;
749 	}
750 	return true;
751 }
752 
753 static bool intel_pt_return_compression(struct intel_pt *pt)
754 {
755 	struct evsel *evsel;
756 	u64 config;
757 
758 	if (!pt->noretcomp_bit)
759 		return true;
760 
761 	evlist__for_each_entry(pt->session->evlist, evsel) {
762 		if (intel_pt_get_config(pt, &evsel->core.attr, &config) &&
763 		    (config & pt->noretcomp_bit))
764 			return false;
765 	}
766 	return true;
767 }
768 
769 static bool intel_pt_branch_enable(struct intel_pt *pt)
770 {
771 	struct evsel *evsel;
772 	u64 config;
773 
774 	evlist__for_each_entry(pt->session->evlist, evsel) {
775 		if (intel_pt_get_config(pt, &evsel->core.attr, &config) &&
776 		    (config & 1) && !(config & 0x2000))
777 			return false;
778 	}
779 	return true;
780 }
781 
782 static unsigned int intel_pt_mtc_period(struct intel_pt *pt)
783 {
784 	struct evsel *evsel;
785 	unsigned int shift;
786 	u64 config;
787 
788 	if (!pt->mtc_freq_bits)
789 		return 0;
790 
791 	for (shift = 0, config = pt->mtc_freq_bits; !(config & 1); shift++)
792 		config >>= 1;
793 
794 	evlist__for_each_entry(pt->session->evlist, evsel) {
795 		if (intel_pt_get_config(pt, &evsel->core.attr, &config))
796 			return (config & pt->mtc_freq_bits) >> shift;
797 	}
798 	return 0;
799 }
800 
801 static bool intel_pt_timeless_decoding(struct intel_pt *pt)
802 {
803 	struct evsel *evsel;
804 	bool timeless_decoding = true;
805 	u64 config;
806 
807 	if (!pt->tsc_bit || !pt->cap_user_time_zero)
808 		return true;
809 
810 	evlist__for_each_entry(pt->session->evlist, evsel) {
811 		if (!(evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
812 			return true;
813 		if (intel_pt_get_config(pt, &evsel->core.attr, &config)) {
814 			if (config & pt->tsc_bit)
815 				timeless_decoding = false;
816 			else
817 				return true;
818 		}
819 	}
820 	return timeless_decoding;
821 }
822 
823 static bool intel_pt_tracing_kernel(struct intel_pt *pt)
824 {
825 	struct evsel *evsel;
826 
827 	evlist__for_each_entry(pt->session->evlist, evsel) {
828 		if (intel_pt_get_config(pt, &evsel->core.attr, NULL) &&
829 		    !evsel->core.attr.exclude_kernel)
830 			return true;
831 	}
832 	return false;
833 }
834 
835 static bool intel_pt_have_tsc(struct intel_pt *pt)
836 {
837 	struct evsel *evsel;
838 	bool have_tsc = false;
839 	u64 config;
840 
841 	if (!pt->tsc_bit)
842 		return false;
843 
844 	evlist__for_each_entry(pt->session->evlist, evsel) {
845 		if (intel_pt_get_config(pt, &evsel->core.attr, &config)) {
846 			if (config & pt->tsc_bit)
847 				have_tsc = true;
848 			else
849 				return false;
850 		}
851 	}
852 	return have_tsc;
853 }
854 
855 static bool intel_pt_sampling_mode(struct intel_pt *pt)
856 {
857 	struct evsel *evsel;
858 
859 	evlist__for_each_entry(pt->session->evlist, evsel) {
860 		if ((evsel->core.attr.sample_type & PERF_SAMPLE_AUX) &&
861 		    evsel->core.attr.aux_sample_size)
862 			return true;
863 	}
864 	return false;
865 }
866 
867 static u64 intel_pt_ns_to_ticks(const struct intel_pt *pt, u64 ns)
868 {
869 	u64 quot, rem;
870 
871 	quot = ns / pt->tc.time_mult;
872 	rem  = ns % pt->tc.time_mult;
873 	return (quot << pt->tc.time_shift) + (rem << pt->tc.time_shift) /
874 		pt->tc.time_mult;
875 }
876 
877 static struct ip_callchain *intel_pt_alloc_chain(struct intel_pt *pt)
878 {
879 	size_t sz = sizeof(struct ip_callchain);
880 
881 	/* Add 1 to callchain_sz for callchain context */
882 	sz += (pt->synth_opts.callchain_sz + 1) * sizeof(u64);
883 	return zalloc(sz);
884 }
885 
886 static int intel_pt_callchain_init(struct intel_pt *pt)
887 {
888 	struct evsel *evsel;
889 
890 	evlist__for_each_entry(pt->session->evlist, evsel) {
891 		if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CALLCHAIN))
892 			evsel->synth_sample_type |= PERF_SAMPLE_CALLCHAIN;
893 	}
894 
895 	pt->chain = intel_pt_alloc_chain(pt);
896 	if (!pt->chain)
897 		return -ENOMEM;
898 
899 	return 0;
900 }
901 
902 static void intel_pt_add_callchain(struct intel_pt *pt,
903 				   struct perf_sample *sample)
904 {
905 	struct thread *thread = machine__findnew_thread(pt->machine,
906 							sample->pid,
907 							sample->tid);
908 
909 	thread_stack__sample_late(thread, sample->cpu, pt->chain,
910 				  pt->synth_opts.callchain_sz + 1, sample->ip,
911 				  pt->kernel_start);
912 
913 	sample->callchain = pt->chain;
914 }
915 
916 static struct branch_stack *intel_pt_alloc_br_stack(unsigned int entry_cnt)
917 {
918 	size_t sz = sizeof(struct branch_stack);
919 
920 	sz += entry_cnt * sizeof(struct branch_entry);
921 	return zalloc(sz);
922 }
923 
924 static int intel_pt_br_stack_init(struct intel_pt *pt)
925 {
926 	struct evsel *evsel;
927 
928 	evlist__for_each_entry(pt->session->evlist, evsel) {
929 		if (!(evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK))
930 			evsel->synth_sample_type |= PERF_SAMPLE_BRANCH_STACK;
931 	}
932 
933 	pt->br_stack = intel_pt_alloc_br_stack(pt->br_stack_sz);
934 	if (!pt->br_stack)
935 		return -ENOMEM;
936 
937 	return 0;
938 }
939 
940 static void intel_pt_add_br_stack(struct intel_pt *pt,
941 				  struct perf_sample *sample)
942 {
943 	struct thread *thread = machine__findnew_thread(pt->machine,
944 							sample->pid,
945 							sample->tid);
946 
947 	thread_stack__br_sample_late(thread, sample->cpu, pt->br_stack,
948 				     pt->br_stack_sz, sample->ip,
949 				     pt->kernel_start);
950 
951 	sample->branch_stack = pt->br_stack;
952 }
953 
954 /* INTEL_PT_LBR_0, INTEL_PT_LBR_1 and INTEL_PT_LBR_2 */
955 #define LBRS_MAX (INTEL_PT_BLK_ITEM_ID_CNT * 3U)
956 
957 static struct intel_pt_queue *intel_pt_alloc_queue(struct intel_pt *pt,
958 						   unsigned int queue_nr)
959 {
960 	struct intel_pt_params params = { .get_trace = 0, };
961 	struct perf_env *env = pt->machine->env;
962 	struct intel_pt_queue *ptq;
963 
964 	ptq = zalloc(sizeof(struct intel_pt_queue));
965 	if (!ptq)
966 		return NULL;
967 
968 	if (pt->synth_opts.callchain) {
969 		ptq->chain = intel_pt_alloc_chain(pt);
970 		if (!ptq->chain)
971 			goto out_free;
972 	}
973 
974 	if (pt->synth_opts.last_branch || pt->synth_opts.other_events) {
975 		unsigned int entry_cnt = max(LBRS_MAX, pt->br_stack_sz);
976 
977 		ptq->last_branch = intel_pt_alloc_br_stack(entry_cnt);
978 		if (!ptq->last_branch)
979 			goto out_free;
980 	}
981 
982 	ptq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
983 	if (!ptq->event_buf)
984 		goto out_free;
985 
986 	ptq->pt = pt;
987 	ptq->queue_nr = queue_nr;
988 	ptq->exclude_kernel = intel_pt_exclude_kernel(pt);
989 	ptq->pid = -1;
990 	ptq->tid = -1;
991 	ptq->cpu = -1;
992 	ptq->next_tid = -1;
993 
994 	params.get_trace = intel_pt_get_trace;
995 	params.walk_insn = intel_pt_walk_next_insn;
996 	params.lookahead = intel_pt_lookahead;
997 	params.data = ptq;
998 	params.return_compression = intel_pt_return_compression(pt);
999 	params.branch_enable = intel_pt_branch_enable(pt);
1000 	params.max_non_turbo_ratio = pt->max_non_turbo_ratio;
1001 	params.mtc_period = intel_pt_mtc_period(pt);
1002 	params.tsc_ctc_ratio_n = pt->tsc_ctc_ratio_n;
1003 	params.tsc_ctc_ratio_d = pt->tsc_ctc_ratio_d;
1004 
1005 	if (pt->filts.cnt > 0)
1006 		params.pgd_ip = intel_pt_pgd_ip;
1007 
1008 	if (pt->synth_opts.instructions) {
1009 		if (pt->synth_opts.period) {
1010 			switch (pt->synth_opts.period_type) {
1011 			case PERF_ITRACE_PERIOD_INSTRUCTIONS:
1012 				params.period_type =
1013 						INTEL_PT_PERIOD_INSTRUCTIONS;
1014 				params.period = pt->synth_opts.period;
1015 				break;
1016 			case PERF_ITRACE_PERIOD_TICKS:
1017 				params.period_type = INTEL_PT_PERIOD_TICKS;
1018 				params.period = pt->synth_opts.period;
1019 				break;
1020 			case PERF_ITRACE_PERIOD_NANOSECS:
1021 				params.period_type = INTEL_PT_PERIOD_TICKS;
1022 				params.period = intel_pt_ns_to_ticks(pt,
1023 							pt->synth_opts.period);
1024 				break;
1025 			default:
1026 				break;
1027 			}
1028 		}
1029 
1030 		if (!params.period) {
1031 			params.period_type = INTEL_PT_PERIOD_INSTRUCTIONS;
1032 			params.period = 1;
1033 		}
1034 	}
1035 
1036 	if (env->cpuid && !strncmp(env->cpuid, "GenuineIntel,6,92,", 18))
1037 		params.flags |= INTEL_PT_FUP_WITH_NLIP;
1038 
1039 	ptq->decoder = intel_pt_decoder_new(&params);
1040 	if (!ptq->decoder)
1041 		goto out_free;
1042 
1043 	return ptq;
1044 
1045 out_free:
1046 	zfree(&ptq->event_buf);
1047 	zfree(&ptq->last_branch);
1048 	zfree(&ptq->chain);
1049 	free(ptq);
1050 	return NULL;
1051 }
1052 
1053 static void intel_pt_free_queue(void *priv)
1054 {
1055 	struct intel_pt_queue *ptq = priv;
1056 
1057 	if (!ptq)
1058 		return;
1059 	thread__zput(ptq->thread);
1060 	intel_pt_decoder_free(ptq->decoder);
1061 	zfree(&ptq->event_buf);
1062 	zfree(&ptq->last_branch);
1063 	zfree(&ptq->chain);
1064 	free(ptq);
1065 }
1066 
1067 static void intel_pt_set_pid_tid_cpu(struct intel_pt *pt,
1068 				     struct auxtrace_queue *queue)
1069 {
1070 	struct intel_pt_queue *ptq = queue->priv;
1071 
1072 	if (queue->tid == -1 || pt->have_sched_switch) {
1073 		ptq->tid = machine__get_current_tid(pt->machine, ptq->cpu);
1074 		thread__zput(ptq->thread);
1075 	}
1076 
1077 	if (!ptq->thread && ptq->tid != -1)
1078 		ptq->thread = machine__find_thread(pt->machine, -1, ptq->tid);
1079 
1080 	if (ptq->thread) {
1081 		ptq->pid = ptq->thread->pid_;
1082 		if (queue->cpu == -1)
1083 			ptq->cpu = ptq->thread->cpu;
1084 	}
1085 }
1086 
1087 static void intel_pt_sample_flags(struct intel_pt_queue *ptq)
1088 {
1089 	if (ptq->state->flags & INTEL_PT_ABORT_TX) {
1090 		ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_TX_ABORT;
1091 	} else if (ptq->state->flags & INTEL_PT_ASYNC) {
1092 		if (ptq->state->to_ip)
1093 			ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_CALL |
1094 				     PERF_IP_FLAG_ASYNC |
1095 				     PERF_IP_FLAG_INTERRUPT;
1096 		else
1097 			ptq->flags = PERF_IP_FLAG_BRANCH |
1098 				     PERF_IP_FLAG_TRACE_END;
1099 		ptq->insn_len = 0;
1100 	} else {
1101 		if (ptq->state->from_ip)
1102 			ptq->flags = intel_pt_insn_type(ptq->state->insn_op);
1103 		else
1104 			ptq->flags = PERF_IP_FLAG_BRANCH |
1105 				     PERF_IP_FLAG_TRACE_BEGIN;
1106 		if (ptq->state->flags & INTEL_PT_IN_TX)
1107 			ptq->flags |= PERF_IP_FLAG_IN_TX;
1108 		ptq->insn_len = ptq->state->insn_len;
1109 		memcpy(ptq->insn, ptq->state->insn, INTEL_PT_INSN_BUF_SZ);
1110 	}
1111 
1112 	if (ptq->state->type & INTEL_PT_TRACE_BEGIN)
1113 		ptq->flags |= PERF_IP_FLAG_TRACE_BEGIN;
1114 	if (ptq->state->type & INTEL_PT_TRACE_END)
1115 		ptq->flags |= PERF_IP_FLAG_TRACE_END;
1116 }
1117 
1118 static void intel_pt_setup_time_range(struct intel_pt *pt,
1119 				      struct intel_pt_queue *ptq)
1120 {
1121 	if (!pt->range_cnt)
1122 		return;
1123 
1124 	ptq->sel_timestamp = pt->time_ranges[0].start;
1125 	ptq->sel_idx = 0;
1126 
1127 	if (ptq->sel_timestamp) {
1128 		ptq->sel_start = true;
1129 	} else {
1130 		ptq->sel_timestamp = pt->time_ranges[0].end;
1131 		ptq->sel_start = false;
1132 	}
1133 }
1134 
1135 static int intel_pt_setup_queue(struct intel_pt *pt,
1136 				struct auxtrace_queue *queue,
1137 				unsigned int queue_nr)
1138 {
1139 	struct intel_pt_queue *ptq = queue->priv;
1140 
1141 	if (list_empty(&queue->head))
1142 		return 0;
1143 
1144 	if (!ptq) {
1145 		ptq = intel_pt_alloc_queue(pt, queue_nr);
1146 		if (!ptq)
1147 			return -ENOMEM;
1148 		queue->priv = ptq;
1149 
1150 		if (queue->cpu != -1)
1151 			ptq->cpu = queue->cpu;
1152 		ptq->tid = queue->tid;
1153 
1154 		ptq->cbr_seen = UINT_MAX;
1155 
1156 		if (pt->sampling_mode && !pt->snapshot_mode &&
1157 		    pt->timeless_decoding)
1158 			ptq->step_through_buffers = true;
1159 
1160 		ptq->sync_switch = pt->sync_switch;
1161 
1162 		intel_pt_setup_time_range(pt, ptq);
1163 	}
1164 
1165 	if (!ptq->on_heap &&
1166 	    (!ptq->sync_switch ||
1167 	     ptq->switch_state != INTEL_PT_SS_EXPECTING_SWITCH_EVENT)) {
1168 		const struct intel_pt_state *state;
1169 		int ret;
1170 
1171 		if (pt->timeless_decoding)
1172 			return 0;
1173 
1174 		intel_pt_log("queue %u getting timestamp\n", queue_nr);
1175 		intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
1176 			     queue_nr, ptq->cpu, ptq->pid, ptq->tid);
1177 
1178 		if (ptq->sel_start && ptq->sel_timestamp) {
1179 			ret = intel_pt_fast_forward(ptq->decoder,
1180 						    ptq->sel_timestamp);
1181 			if (ret)
1182 				return ret;
1183 		}
1184 
1185 		while (1) {
1186 			state = intel_pt_decode(ptq->decoder);
1187 			if (state->err) {
1188 				if (state->err == INTEL_PT_ERR_NODATA) {
1189 					intel_pt_log("queue %u has no timestamp\n",
1190 						     queue_nr);
1191 					return 0;
1192 				}
1193 				continue;
1194 			}
1195 			if (state->timestamp)
1196 				break;
1197 		}
1198 
1199 		ptq->timestamp = state->timestamp;
1200 		intel_pt_log("queue %u timestamp 0x%" PRIx64 "\n",
1201 			     queue_nr, ptq->timestamp);
1202 		ptq->state = state;
1203 		ptq->have_sample = true;
1204 		if (ptq->sel_start && ptq->sel_timestamp &&
1205 		    ptq->timestamp < ptq->sel_timestamp)
1206 			ptq->have_sample = false;
1207 		intel_pt_sample_flags(ptq);
1208 		ret = auxtrace_heap__add(&pt->heap, queue_nr, ptq->timestamp);
1209 		if (ret)
1210 			return ret;
1211 		ptq->on_heap = true;
1212 	}
1213 
1214 	return 0;
1215 }
1216 
1217 static int intel_pt_setup_queues(struct intel_pt *pt)
1218 {
1219 	unsigned int i;
1220 	int ret;
1221 
1222 	for (i = 0; i < pt->queues.nr_queues; i++) {
1223 		ret = intel_pt_setup_queue(pt, &pt->queues.queue_array[i], i);
1224 		if (ret)
1225 			return ret;
1226 	}
1227 	return 0;
1228 }
1229 
1230 static inline bool intel_pt_skip_event(struct intel_pt *pt)
1231 {
1232 	return pt->synth_opts.initial_skip &&
1233 	       pt->num_events++ < pt->synth_opts.initial_skip;
1234 }
1235 
1236 /*
1237  * Cannot count CBR as skipped because it won't go away until cbr == cbr_seen.
1238  * Also ensure CBR is first non-skipped event by allowing for 4 more samples
1239  * from this decoder state.
1240  */
1241 static inline bool intel_pt_skip_cbr_event(struct intel_pt *pt)
1242 {
1243 	return pt->synth_opts.initial_skip &&
1244 	       pt->num_events + 4 < pt->synth_opts.initial_skip;
1245 }
1246 
1247 static void intel_pt_prep_a_sample(struct intel_pt_queue *ptq,
1248 				   union perf_event *event,
1249 				   struct perf_sample *sample)
1250 {
1251 	event->sample.header.type = PERF_RECORD_SAMPLE;
1252 	event->sample.header.size = sizeof(struct perf_event_header);
1253 
1254 	sample->pid = ptq->pid;
1255 	sample->tid = ptq->tid;
1256 	sample->cpu = ptq->cpu;
1257 	sample->insn_len = ptq->insn_len;
1258 	memcpy(sample->insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1259 }
1260 
1261 static void intel_pt_prep_b_sample(struct intel_pt *pt,
1262 				   struct intel_pt_queue *ptq,
1263 				   union perf_event *event,
1264 				   struct perf_sample *sample)
1265 {
1266 	intel_pt_prep_a_sample(ptq, event, sample);
1267 
1268 	if (!pt->timeless_decoding)
1269 		sample->time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1270 
1271 	sample->ip = ptq->state->from_ip;
1272 	sample->cpumode = intel_pt_cpumode(pt, sample->ip);
1273 	sample->addr = ptq->state->to_ip;
1274 	sample->period = 1;
1275 	sample->flags = ptq->flags;
1276 
1277 	event->sample.header.misc = sample->cpumode;
1278 }
1279 
1280 static int intel_pt_inject_event(union perf_event *event,
1281 				 struct perf_sample *sample, u64 type)
1282 {
1283 	event->header.size = perf_event__sample_event_size(sample, type, 0);
1284 	return perf_event__synthesize_sample(event, type, 0, sample);
1285 }
1286 
1287 static inline int intel_pt_opt_inject(struct intel_pt *pt,
1288 				      union perf_event *event,
1289 				      struct perf_sample *sample, u64 type)
1290 {
1291 	if (!pt->synth_opts.inject)
1292 		return 0;
1293 
1294 	return intel_pt_inject_event(event, sample, type);
1295 }
1296 
1297 static int intel_pt_deliver_synth_event(struct intel_pt *pt,
1298 					union perf_event *event,
1299 					struct perf_sample *sample, u64 type)
1300 {
1301 	int ret;
1302 
1303 	ret = intel_pt_opt_inject(pt, event, sample, type);
1304 	if (ret)
1305 		return ret;
1306 
1307 	ret = perf_session__deliver_synth_event(pt->session, event, sample);
1308 	if (ret)
1309 		pr_err("Intel PT: failed to deliver event, error %d\n", ret);
1310 
1311 	return ret;
1312 }
1313 
1314 static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq)
1315 {
1316 	struct intel_pt *pt = ptq->pt;
1317 	union perf_event *event = ptq->event_buf;
1318 	struct perf_sample sample = { .ip = 0, };
1319 	struct dummy_branch_stack {
1320 		u64			nr;
1321 		u64			hw_idx;
1322 		struct branch_entry	entries;
1323 	} dummy_bs;
1324 
1325 	if (pt->branches_filter && !(pt->branches_filter & ptq->flags))
1326 		return 0;
1327 
1328 	if (intel_pt_skip_event(pt))
1329 		return 0;
1330 
1331 	intel_pt_prep_b_sample(pt, ptq, event, &sample);
1332 
1333 	sample.id = ptq->pt->branches_id;
1334 	sample.stream_id = ptq->pt->branches_id;
1335 
1336 	/*
1337 	 * perf report cannot handle events without a branch stack when using
1338 	 * SORT_MODE__BRANCH so make a dummy one.
1339 	 */
1340 	if (pt->synth_opts.last_branch && sort__mode == SORT_MODE__BRANCH) {
1341 		dummy_bs = (struct dummy_branch_stack){
1342 			.nr = 1,
1343 			.hw_idx = -1ULL,
1344 			.entries = {
1345 				.from = sample.ip,
1346 				.to = sample.addr,
1347 			},
1348 		};
1349 		sample.branch_stack = (struct branch_stack *)&dummy_bs;
1350 	}
1351 
1352 	sample.cyc_cnt = ptq->ipc_cyc_cnt - ptq->last_br_cyc_cnt;
1353 	if (sample.cyc_cnt) {
1354 		sample.insn_cnt = ptq->ipc_insn_cnt - ptq->last_br_insn_cnt;
1355 		ptq->last_br_insn_cnt = ptq->ipc_insn_cnt;
1356 		ptq->last_br_cyc_cnt = ptq->ipc_cyc_cnt;
1357 	}
1358 
1359 	return intel_pt_deliver_synth_event(pt, event, &sample,
1360 					    pt->branches_sample_type);
1361 }
1362 
1363 static void intel_pt_prep_sample(struct intel_pt *pt,
1364 				 struct intel_pt_queue *ptq,
1365 				 union perf_event *event,
1366 				 struct perf_sample *sample)
1367 {
1368 	intel_pt_prep_b_sample(pt, ptq, event, sample);
1369 
1370 	if (pt->synth_opts.callchain) {
1371 		thread_stack__sample(ptq->thread, ptq->cpu, ptq->chain,
1372 				     pt->synth_opts.callchain_sz + 1,
1373 				     sample->ip, pt->kernel_start);
1374 		sample->callchain = ptq->chain;
1375 	}
1376 
1377 	if (pt->synth_opts.last_branch) {
1378 		thread_stack__br_sample(ptq->thread, ptq->cpu, ptq->last_branch,
1379 					pt->br_stack_sz);
1380 		sample->branch_stack = ptq->last_branch;
1381 	}
1382 }
1383 
1384 static int intel_pt_synth_instruction_sample(struct intel_pt_queue *ptq)
1385 {
1386 	struct intel_pt *pt = ptq->pt;
1387 	union perf_event *event = ptq->event_buf;
1388 	struct perf_sample sample = { .ip = 0, };
1389 
1390 	if (intel_pt_skip_event(pt))
1391 		return 0;
1392 
1393 	intel_pt_prep_sample(pt, ptq, event, &sample);
1394 
1395 	sample.id = ptq->pt->instructions_id;
1396 	sample.stream_id = ptq->pt->instructions_id;
1397 	sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt;
1398 
1399 	sample.cyc_cnt = ptq->ipc_cyc_cnt - ptq->last_in_cyc_cnt;
1400 	if (sample.cyc_cnt) {
1401 		sample.insn_cnt = ptq->ipc_insn_cnt - ptq->last_in_insn_cnt;
1402 		ptq->last_in_insn_cnt = ptq->ipc_insn_cnt;
1403 		ptq->last_in_cyc_cnt = ptq->ipc_cyc_cnt;
1404 	}
1405 
1406 	ptq->last_insn_cnt = ptq->state->tot_insn_cnt;
1407 
1408 	return intel_pt_deliver_synth_event(pt, event, &sample,
1409 					    pt->instructions_sample_type);
1410 }
1411 
1412 static int intel_pt_synth_transaction_sample(struct intel_pt_queue *ptq)
1413 {
1414 	struct intel_pt *pt = ptq->pt;
1415 	union perf_event *event = ptq->event_buf;
1416 	struct perf_sample sample = { .ip = 0, };
1417 
1418 	if (intel_pt_skip_event(pt))
1419 		return 0;
1420 
1421 	intel_pt_prep_sample(pt, ptq, event, &sample);
1422 
1423 	sample.id = ptq->pt->transactions_id;
1424 	sample.stream_id = ptq->pt->transactions_id;
1425 
1426 	return intel_pt_deliver_synth_event(pt, event, &sample,
1427 					    pt->transactions_sample_type);
1428 }
1429 
1430 static void intel_pt_prep_p_sample(struct intel_pt *pt,
1431 				   struct intel_pt_queue *ptq,
1432 				   union perf_event *event,
1433 				   struct perf_sample *sample)
1434 {
1435 	intel_pt_prep_sample(pt, ptq, event, sample);
1436 
1437 	/*
1438 	 * Zero IP is used to mean "trace start" but that is not the case for
1439 	 * power or PTWRITE events with no IP, so clear the flags.
1440 	 */
1441 	if (!sample->ip)
1442 		sample->flags = 0;
1443 }
1444 
1445 static int intel_pt_synth_ptwrite_sample(struct intel_pt_queue *ptq)
1446 {
1447 	struct intel_pt *pt = ptq->pt;
1448 	union perf_event *event = ptq->event_buf;
1449 	struct perf_sample sample = { .ip = 0, };
1450 	struct perf_synth_intel_ptwrite raw;
1451 
1452 	if (intel_pt_skip_event(pt))
1453 		return 0;
1454 
1455 	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1456 
1457 	sample.id = ptq->pt->ptwrites_id;
1458 	sample.stream_id = ptq->pt->ptwrites_id;
1459 
1460 	raw.flags = 0;
1461 	raw.ip = !!(ptq->state->flags & INTEL_PT_FUP_IP);
1462 	raw.payload = cpu_to_le64(ptq->state->ptw_payload);
1463 
1464 	sample.raw_size = perf_synth__raw_size(raw);
1465 	sample.raw_data = perf_synth__raw_data(&raw);
1466 
1467 	return intel_pt_deliver_synth_event(pt, event, &sample,
1468 					    pt->ptwrites_sample_type);
1469 }
1470 
1471 static int intel_pt_synth_cbr_sample(struct intel_pt_queue *ptq)
1472 {
1473 	struct intel_pt *pt = ptq->pt;
1474 	union perf_event *event = ptq->event_buf;
1475 	struct perf_sample sample = { .ip = 0, };
1476 	struct perf_synth_intel_cbr raw;
1477 	u32 flags;
1478 
1479 	if (intel_pt_skip_cbr_event(pt))
1480 		return 0;
1481 
1482 	ptq->cbr_seen = ptq->state->cbr;
1483 
1484 	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1485 
1486 	sample.id = ptq->pt->cbr_id;
1487 	sample.stream_id = ptq->pt->cbr_id;
1488 
1489 	flags = (u16)ptq->state->cbr_payload | (pt->max_non_turbo_ratio << 16);
1490 	raw.flags = cpu_to_le32(flags);
1491 	raw.freq = cpu_to_le32(raw.cbr * pt->cbr2khz);
1492 	raw.reserved3 = 0;
1493 
1494 	sample.raw_size = perf_synth__raw_size(raw);
1495 	sample.raw_data = perf_synth__raw_data(&raw);
1496 
1497 	return intel_pt_deliver_synth_event(pt, event, &sample,
1498 					    pt->pwr_events_sample_type);
1499 }
1500 
1501 static int intel_pt_synth_mwait_sample(struct intel_pt_queue *ptq)
1502 {
1503 	struct intel_pt *pt = ptq->pt;
1504 	union perf_event *event = ptq->event_buf;
1505 	struct perf_sample sample = { .ip = 0, };
1506 	struct perf_synth_intel_mwait raw;
1507 
1508 	if (intel_pt_skip_event(pt))
1509 		return 0;
1510 
1511 	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1512 
1513 	sample.id = ptq->pt->mwait_id;
1514 	sample.stream_id = ptq->pt->mwait_id;
1515 
1516 	raw.reserved = 0;
1517 	raw.payload = cpu_to_le64(ptq->state->mwait_payload);
1518 
1519 	sample.raw_size = perf_synth__raw_size(raw);
1520 	sample.raw_data = perf_synth__raw_data(&raw);
1521 
1522 	return intel_pt_deliver_synth_event(pt, event, &sample,
1523 					    pt->pwr_events_sample_type);
1524 }
1525 
1526 static int intel_pt_synth_pwre_sample(struct intel_pt_queue *ptq)
1527 {
1528 	struct intel_pt *pt = ptq->pt;
1529 	union perf_event *event = ptq->event_buf;
1530 	struct perf_sample sample = { .ip = 0, };
1531 	struct perf_synth_intel_pwre raw;
1532 
1533 	if (intel_pt_skip_event(pt))
1534 		return 0;
1535 
1536 	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1537 
1538 	sample.id = ptq->pt->pwre_id;
1539 	sample.stream_id = ptq->pt->pwre_id;
1540 
1541 	raw.reserved = 0;
1542 	raw.payload = cpu_to_le64(ptq->state->pwre_payload);
1543 
1544 	sample.raw_size = perf_synth__raw_size(raw);
1545 	sample.raw_data = perf_synth__raw_data(&raw);
1546 
1547 	return intel_pt_deliver_synth_event(pt, event, &sample,
1548 					    pt->pwr_events_sample_type);
1549 }
1550 
1551 static int intel_pt_synth_exstop_sample(struct intel_pt_queue *ptq)
1552 {
1553 	struct intel_pt *pt = ptq->pt;
1554 	union perf_event *event = ptq->event_buf;
1555 	struct perf_sample sample = { .ip = 0, };
1556 	struct perf_synth_intel_exstop raw;
1557 
1558 	if (intel_pt_skip_event(pt))
1559 		return 0;
1560 
1561 	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1562 
1563 	sample.id = ptq->pt->exstop_id;
1564 	sample.stream_id = ptq->pt->exstop_id;
1565 
1566 	raw.flags = 0;
1567 	raw.ip = !!(ptq->state->flags & INTEL_PT_FUP_IP);
1568 
1569 	sample.raw_size = perf_synth__raw_size(raw);
1570 	sample.raw_data = perf_synth__raw_data(&raw);
1571 
1572 	return intel_pt_deliver_synth_event(pt, event, &sample,
1573 					    pt->pwr_events_sample_type);
1574 }
1575 
1576 static int intel_pt_synth_pwrx_sample(struct intel_pt_queue *ptq)
1577 {
1578 	struct intel_pt *pt = ptq->pt;
1579 	union perf_event *event = ptq->event_buf;
1580 	struct perf_sample sample = { .ip = 0, };
1581 	struct perf_synth_intel_pwrx raw;
1582 
1583 	if (intel_pt_skip_event(pt))
1584 		return 0;
1585 
1586 	intel_pt_prep_p_sample(pt, ptq, event, &sample);
1587 
1588 	sample.id = ptq->pt->pwrx_id;
1589 	sample.stream_id = ptq->pt->pwrx_id;
1590 
1591 	raw.reserved = 0;
1592 	raw.payload = cpu_to_le64(ptq->state->pwrx_payload);
1593 
1594 	sample.raw_size = perf_synth__raw_size(raw);
1595 	sample.raw_data = perf_synth__raw_data(&raw);
1596 
1597 	return intel_pt_deliver_synth_event(pt, event, &sample,
1598 					    pt->pwr_events_sample_type);
1599 }
1600 
1601 /*
1602  * PEBS gp_regs array indexes plus 1 so that 0 means not present. Refer
1603  * intel_pt_add_gp_regs().
1604  */
1605 static const int pebs_gp_regs[] = {
1606 	[PERF_REG_X86_FLAGS]	= 1,
1607 	[PERF_REG_X86_IP]	= 2,
1608 	[PERF_REG_X86_AX]	= 3,
1609 	[PERF_REG_X86_CX]	= 4,
1610 	[PERF_REG_X86_DX]	= 5,
1611 	[PERF_REG_X86_BX]	= 6,
1612 	[PERF_REG_X86_SP]	= 7,
1613 	[PERF_REG_X86_BP]	= 8,
1614 	[PERF_REG_X86_SI]	= 9,
1615 	[PERF_REG_X86_DI]	= 10,
1616 	[PERF_REG_X86_R8]	= 11,
1617 	[PERF_REG_X86_R9]	= 12,
1618 	[PERF_REG_X86_R10]	= 13,
1619 	[PERF_REG_X86_R11]	= 14,
1620 	[PERF_REG_X86_R12]	= 15,
1621 	[PERF_REG_X86_R13]	= 16,
1622 	[PERF_REG_X86_R14]	= 17,
1623 	[PERF_REG_X86_R15]	= 18,
1624 };
1625 
1626 static u64 *intel_pt_add_gp_regs(struct regs_dump *intr_regs, u64 *pos,
1627 				 const struct intel_pt_blk_items *items,
1628 				 u64 regs_mask)
1629 {
1630 	const u64 *gp_regs = items->val[INTEL_PT_GP_REGS_POS];
1631 	u32 mask = items->mask[INTEL_PT_GP_REGS_POS];
1632 	u32 bit;
1633 	int i;
1634 
1635 	for (i = 0, bit = 1; i < PERF_REG_X86_64_MAX; i++, bit <<= 1) {
1636 		/* Get the PEBS gp_regs array index */
1637 		int n = pebs_gp_regs[i] - 1;
1638 
1639 		if (n < 0)
1640 			continue;
1641 		/*
1642 		 * Add only registers that were requested (i.e. 'regs_mask') and
1643 		 * that were provided (i.e. 'mask'), and update the resulting
1644 		 * mask (i.e. 'intr_regs->mask') accordingly.
1645 		 */
1646 		if (mask & 1 << n && regs_mask & bit) {
1647 			intr_regs->mask |= bit;
1648 			*pos++ = gp_regs[n];
1649 		}
1650 	}
1651 
1652 	return pos;
1653 }
1654 
1655 #ifndef PERF_REG_X86_XMM0
1656 #define PERF_REG_X86_XMM0 32
1657 #endif
1658 
1659 static void intel_pt_add_xmm(struct regs_dump *intr_regs, u64 *pos,
1660 			     const struct intel_pt_blk_items *items,
1661 			     u64 regs_mask)
1662 {
1663 	u32 mask = items->has_xmm & (regs_mask >> PERF_REG_X86_XMM0);
1664 	const u64 *xmm = items->xmm;
1665 
1666 	/*
1667 	 * If there are any XMM registers, then there should be all of them.
1668 	 * Nevertheless, follow the logic to add only registers that were
1669 	 * requested (i.e. 'regs_mask') and that were provided (i.e. 'mask'),
1670 	 * and update the resulting mask (i.e. 'intr_regs->mask') accordingly.
1671 	 */
1672 	intr_regs->mask |= (u64)mask << PERF_REG_X86_XMM0;
1673 
1674 	for (; mask; mask >>= 1, xmm++) {
1675 		if (mask & 1)
1676 			*pos++ = *xmm;
1677 	}
1678 }
1679 
1680 #define LBR_INFO_MISPRED	(1ULL << 63)
1681 #define LBR_INFO_IN_TX		(1ULL << 62)
1682 #define LBR_INFO_ABORT		(1ULL << 61)
1683 #define LBR_INFO_CYCLES		0xffff
1684 
1685 /* Refer kernel's intel_pmu_store_pebs_lbrs() */
1686 static u64 intel_pt_lbr_flags(u64 info)
1687 {
1688 	union {
1689 		struct branch_flags flags;
1690 		u64 result;
1691 	} u;
1692 
1693 	u.result	  = 0;
1694 	u.flags.mispred	  = !!(info & LBR_INFO_MISPRED);
1695 	u.flags.predicted = !(info & LBR_INFO_MISPRED);
1696 	u.flags.in_tx	  = !!(info & LBR_INFO_IN_TX);
1697 	u.flags.abort	  = !!(info & LBR_INFO_ABORT);
1698 	u.flags.cycles	  = info & LBR_INFO_CYCLES;
1699 
1700 	return u.result;
1701 }
1702 
1703 static void intel_pt_add_lbrs(struct branch_stack *br_stack,
1704 			      const struct intel_pt_blk_items *items)
1705 {
1706 	u64 *to;
1707 	int i;
1708 
1709 	br_stack->nr = 0;
1710 
1711 	to = &br_stack->entries[0].from;
1712 
1713 	for (i = INTEL_PT_LBR_0_POS; i <= INTEL_PT_LBR_2_POS; i++) {
1714 		u32 mask = items->mask[i];
1715 		const u64 *from = items->val[i];
1716 
1717 		for (; mask; mask >>= 3, from += 3) {
1718 			if ((mask & 7) == 7) {
1719 				*to++ = from[0];
1720 				*to++ = from[1];
1721 				*to++ = intel_pt_lbr_flags(from[2]);
1722 				br_stack->nr += 1;
1723 			}
1724 		}
1725 	}
1726 }
1727 
1728 static int intel_pt_synth_pebs_sample(struct intel_pt_queue *ptq)
1729 {
1730 	const struct intel_pt_blk_items *items = &ptq->state->items;
1731 	struct perf_sample sample = { .ip = 0, };
1732 	union perf_event *event = ptq->event_buf;
1733 	struct intel_pt *pt = ptq->pt;
1734 	struct evsel *evsel = pt->pebs_evsel;
1735 	u64 sample_type = evsel->core.attr.sample_type;
1736 	u64 id = evsel->core.id[0];
1737 	u8 cpumode;
1738 	u64 regs[8 * sizeof(sample.intr_regs.mask)];
1739 
1740 	if (intel_pt_skip_event(pt))
1741 		return 0;
1742 
1743 	intel_pt_prep_a_sample(ptq, event, &sample);
1744 
1745 	sample.id = id;
1746 	sample.stream_id = id;
1747 
1748 	if (!evsel->core.attr.freq)
1749 		sample.period = evsel->core.attr.sample_period;
1750 
1751 	/* No support for non-zero CS base */
1752 	if (items->has_ip)
1753 		sample.ip = items->ip;
1754 	else if (items->has_rip)
1755 		sample.ip = items->rip;
1756 	else
1757 		sample.ip = ptq->state->from_ip;
1758 
1759 	/* No support for guest mode at this time */
1760 	cpumode = sample.ip < ptq->pt->kernel_start ?
1761 		  PERF_RECORD_MISC_USER :
1762 		  PERF_RECORD_MISC_KERNEL;
1763 
1764 	event->sample.header.misc = cpumode | PERF_RECORD_MISC_EXACT_IP;
1765 
1766 	sample.cpumode = cpumode;
1767 
1768 	if (sample_type & PERF_SAMPLE_TIME) {
1769 		u64 timestamp = 0;
1770 
1771 		if (items->has_timestamp)
1772 			timestamp = items->timestamp;
1773 		else if (!pt->timeless_decoding)
1774 			timestamp = ptq->timestamp;
1775 		if (timestamp)
1776 			sample.time = tsc_to_perf_time(timestamp, &pt->tc);
1777 	}
1778 
1779 	if (sample_type & PERF_SAMPLE_CALLCHAIN &&
1780 	    pt->synth_opts.callchain) {
1781 		thread_stack__sample(ptq->thread, ptq->cpu, ptq->chain,
1782 				     pt->synth_opts.callchain_sz, sample.ip,
1783 				     pt->kernel_start);
1784 		sample.callchain = ptq->chain;
1785 	}
1786 
1787 	if (sample_type & PERF_SAMPLE_REGS_INTR &&
1788 	    (items->mask[INTEL_PT_GP_REGS_POS] ||
1789 	     items->mask[INTEL_PT_XMM_POS])) {
1790 		u64 regs_mask = evsel->core.attr.sample_regs_intr;
1791 		u64 *pos;
1792 
1793 		sample.intr_regs.abi = items->is_32_bit ?
1794 				       PERF_SAMPLE_REGS_ABI_32 :
1795 				       PERF_SAMPLE_REGS_ABI_64;
1796 		sample.intr_regs.regs = regs;
1797 
1798 		pos = intel_pt_add_gp_regs(&sample.intr_regs, regs, items, regs_mask);
1799 
1800 		intel_pt_add_xmm(&sample.intr_regs, pos, items, regs_mask);
1801 	}
1802 
1803 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1804 		if (items->mask[INTEL_PT_LBR_0_POS] ||
1805 		    items->mask[INTEL_PT_LBR_1_POS] ||
1806 		    items->mask[INTEL_PT_LBR_2_POS]) {
1807 			intel_pt_add_lbrs(ptq->last_branch, items);
1808 		} else if (pt->synth_opts.last_branch) {
1809 			thread_stack__br_sample(ptq->thread, ptq->cpu,
1810 						ptq->last_branch,
1811 						pt->br_stack_sz);
1812 		} else {
1813 			ptq->last_branch->nr = 0;
1814 		}
1815 		sample.branch_stack = ptq->last_branch;
1816 	}
1817 
1818 	if (sample_type & PERF_SAMPLE_ADDR && items->has_mem_access_address)
1819 		sample.addr = items->mem_access_address;
1820 
1821 	if (sample_type & PERF_SAMPLE_WEIGHT) {
1822 		/*
1823 		 * Refer kernel's setup_pebs_adaptive_sample_data() and
1824 		 * intel_hsw_weight().
1825 		 */
1826 		if (items->has_mem_access_latency)
1827 			sample.weight = items->mem_access_latency;
1828 		if (!sample.weight && items->has_tsx_aux_info) {
1829 			/* Cycles last block */
1830 			sample.weight = (u32)items->tsx_aux_info;
1831 		}
1832 	}
1833 
1834 	if (sample_type & PERF_SAMPLE_TRANSACTION && items->has_tsx_aux_info) {
1835 		u64 ax = items->has_rax ? items->rax : 0;
1836 		/* Refer kernel's intel_hsw_transaction() */
1837 		u64 txn = (u8)(items->tsx_aux_info >> 32);
1838 
1839 		/* For RTM XABORTs also log the abort code from AX */
1840 		if (txn & PERF_TXN_TRANSACTION && ax & 1)
1841 			txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1842 		sample.transaction = txn;
1843 	}
1844 
1845 	return intel_pt_deliver_synth_event(pt, event, &sample, sample_type);
1846 }
1847 
1848 static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,
1849 				pid_t pid, pid_t tid, u64 ip, u64 timestamp)
1850 {
1851 	union perf_event event;
1852 	char msg[MAX_AUXTRACE_ERROR_MSG];
1853 	int err;
1854 
1855 	intel_pt__strerror(code, msg, MAX_AUXTRACE_ERROR_MSG);
1856 
1857 	auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
1858 			     code, cpu, pid, tid, ip, msg, timestamp);
1859 
1860 	err = perf_session__deliver_synth_event(pt->session, &event, NULL);
1861 	if (err)
1862 		pr_err("Intel Processor Trace: failed to deliver error event, error %d\n",
1863 		       err);
1864 
1865 	return err;
1866 }
1867 
1868 static int intel_ptq_synth_error(struct intel_pt_queue *ptq,
1869 				 const struct intel_pt_state *state)
1870 {
1871 	struct intel_pt *pt = ptq->pt;
1872 	u64 tm = ptq->timestamp;
1873 
1874 	tm = pt->timeless_decoding ? 0 : tsc_to_perf_time(tm, &pt->tc);
1875 
1876 	return intel_pt_synth_error(pt, state->err, ptq->cpu, ptq->pid,
1877 				    ptq->tid, state->from_ip, tm);
1878 }
1879 
1880 static int intel_pt_next_tid(struct intel_pt *pt, struct intel_pt_queue *ptq)
1881 {
1882 	struct auxtrace_queue *queue;
1883 	pid_t tid = ptq->next_tid;
1884 	int err;
1885 
1886 	if (tid == -1)
1887 		return 0;
1888 
1889 	intel_pt_log("switch: cpu %d tid %d\n", ptq->cpu, tid);
1890 
1891 	err = machine__set_current_tid(pt->machine, ptq->cpu, -1, tid);
1892 
1893 	queue = &pt->queues.queue_array[ptq->queue_nr];
1894 	intel_pt_set_pid_tid_cpu(pt, queue);
1895 
1896 	ptq->next_tid = -1;
1897 
1898 	return err;
1899 }
1900 
1901 static inline bool intel_pt_is_switch_ip(struct intel_pt_queue *ptq, u64 ip)
1902 {
1903 	struct intel_pt *pt = ptq->pt;
1904 
1905 	return ip == pt->switch_ip &&
1906 	       (ptq->flags & PERF_IP_FLAG_BRANCH) &&
1907 	       !(ptq->flags & (PERF_IP_FLAG_CONDITIONAL | PERF_IP_FLAG_ASYNC |
1908 			       PERF_IP_FLAG_INTERRUPT | PERF_IP_FLAG_TX_ABORT));
1909 }
1910 
1911 #define INTEL_PT_PWR_EVT (INTEL_PT_MWAIT_OP | INTEL_PT_PWR_ENTRY | \
1912 			  INTEL_PT_EX_STOP | INTEL_PT_PWR_EXIT)
1913 
1914 static int intel_pt_sample(struct intel_pt_queue *ptq)
1915 {
1916 	const struct intel_pt_state *state = ptq->state;
1917 	struct intel_pt *pt = ptq->pt;
1918 	int err;
1919 
1920 	if (!ptq->have_sample)
1921 		return 0;
1922 
1923 	ptq->have_sample = false;
1924 
1925 	if (ptq->state->tot_cyc_cnt > ptq->ipc_cyc_cnt) {
1926 		/*
1927 		 * Cycle count and instruction count only go together to create
1928 		 * a valid IPC ratio when the cycle count changes.
1929 		 */
1930 		ptq->ipc_insn_cnt = ptq->state->tot_insn_cnt;
1931 		ptq->ipc_cyc_cnt = ptq->state->tot_cyc_cnt;
1932 	}
1933 
1934 	/*
1935 	 * Do PEBS first to allow for the possibility that the PEBS timestamp
1936 	 * precedes the current timestamp.
1937 	 */
1938 	if (pt->sample_pebs && state->type & INTEL_PT_BLK_ITEMS) {
1939 		err = intel_pt_synth_pebs_sample(ptq);
1940 		if (err)
1941 			return err;
1942 	}
1943 
1944 	if (pt->sample_pwr_events) {
1945 		if (ptq->state->cbr != ptq->cbr_seen) {
1946 			err = intel_pt_synth_cbr_sample(ptq);
1947 			if (err)
1948 				return err;
1949 		}
1950 		if (state->type & INTEL_PT_PWR_EVT) {
1951 			if (state->type & INTEL_PT_MWAIT_OP) {
1952 				err = intel_pt_synth_mwait_sample(ptq);
1953 				if (err)
1954 					return err;
1955 			}
1956 			if (state->type & INTEL_PT_PWR_ENTRY) {
1957 				err = intel_pt_synth_pwre_sample(ptq);
1958 				if (err)
1959 					return err;
1960 			}
1961 			if (state->type & INTEL_PT_EX_STOP) {
1962 				err = intel_pt_synth_exstop_sample(ptq);
1963 				if (err)
1964 					return err;
1965 			}
1966 			if (state->type & INTEL_PT_PWR_EXIT) {
1967 				err = intel_pt_synth_pwrx_sample(ptq);
1968 				if (err)
1969 					return err;
1970 			}
1971 		}
1972 	}
1973 
1974 	if (pt->sample_instructions && (state->type & INTEL_PT_INSTRUCTION)) {
1975 		err = intel_pt_synth_instruction_sample(ptq);
1976 		if (err)
1977 			return err;
1978 	}
1979 
1980 	if (pt->sample_transactions && (state->type & INTEL_PT_TRANSACTION)) {
1981 		err = intel_pt_synth_transaction_sample(ptq);
1982 		if (err)
1983 			return err;
1984 	}
1985 
1986 	if (pt->sample_ptwrites && (state->type & INTEL_PT_PTW)) {
1987 		err = intel_pt_synth_ptwrite_sample(ptq);
1988 		if (err)
1989 			return err;
1990 	}
1991 
1992 	if (!(state->type & INTEL_PT_BRANCH))
1993 		return 0;
1994 
1995 	if (pt->use_thread_stack) {
1996 		thread_stack__event(ptq->thread, ptq->cpu, ptq->flags,
1997 				    state->from_ip, state->to_ip, ptq->insn_len,
1998 				    state->trace_nr, pt->callstack,
1999 				    pt->br_stack_sz_plus,
2000 				    pt->mispred_all);
2001 	} else {
2002 		thread_stack__set_trace_nr(ptq->thread, ptq->cpu, state->trace_nr);
2003 	}
2004 
2005 	if (pt->sample_branches) {
2006 		err = intel_pt_synth_branch_sample(ptq);
2007 		if (err)
2008 			return err;
2009 	}
2010 
2011 	if (!ptq->sync_switch)
2012 		return 0;
2013 
2014 	if (intel_pt_is_switch_ip(ptq, state->to_ip)) {
2015 		switch (ptq->switch_state) {
2016 		case INTEL_PT_SS_NOT_TRACING:
2017 		case INTEL_PT_SS_UNKNOWN:
2018 		case INTEL_PT_SS_EXPECTING_SWITCH_IP:
2019 			err = intel_pt_next_tid(pt, ptq);
2020 			if (err)
2021 				return err;
2022 			ptq->switch_state = INTEL_PT_SS_TRACING;
2023 			break;
2024 		default:
2025 			ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_EVENT;
2026 			return 1;
2027 		}
2028 	} else if (!state->to_ip) {
2029 		ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
2030 	} else if (ptq->switch_state == INTEL_PT_SS_NOT_TRACING) {
2031 		ptq->switch_state = INTEL_PT_SS_UNKNOWN;
2032 	} else if (ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
2033 		   state->to_ip == pt->ptss_ip &&
2034 		   (ptq->flags & PERF_IP_FLAG_CALL)) {
2035 		ptq->switch_state = INTEL_PT_SS_TRACING;
2036 	}
2037 
2038 	return 0;
2039 }
2040 
2041 static u64 intel_pt_switch_ip(struct intel_pt *pt, u64 *ptss_ip)
2042 {
2043 	struct machine *machine = pt->machine;
2044 	struct map *map;
2045 	struct symbol *sym, *start;
2046 	u64 ip, switch_ip = 0;
2047 	const char *ptss;
2048 
2049 	if (ptss_ip)
2050 		*ptss_ip = 0;
2051 
2052 	map = machine__kernel_map(machine);
2053 	if (!map)
2054 		return 0;
2055 
2056 	if (map__load(map))
2057 		return 0;
2058 
2059 	start = dso__first_symbol(map->dso);
2060 
2061 	for (sym = start; sym; sym = dso__next_symbol(sym)) {
2062 		if (sym->binding == STB_GLOBAL &&
2063 		    !strcmp(sym->name, "__switch_to")) {
2064 			ip = map->unmap_ip(map, sym->start);
2065 			if (ip >= map->start && ip < map->end) {
2066 				switch_ip = ip;
2067 				break;
2068 			}
2069 		}
2070 	}
2071 
2072 	if (!switch_ip || !ptss_ip)
2073 		return 0;
2074 
2075 	if (pt->have_sched_switch == 1)
2076 		ptss = "perf_trace_sched_switch";
2077 	else
2078 		ptss = "__perf_event_task_sched_out";
2079 
2080 	for (sym = start; sym; sym = dso__next_symbol(sym)) {
2081 		if (!strcmp(sym->name, ptss)) {
2082 			ip = map->unmap_ip(map, sym->start);
2083 			if (ip >= map->start && ip < map->end) {
2084 				*ptss_ip = ip;
2085 				break;
2086 			}
2087 		}
2088 	}
2089 
2090 	return switch_ip;
2091 }
2092 
2093 static void intel_pt_enable_sync_switch(struct intel_pt *pt)
2094 {
2095 	unsigned int i;
2096 
2097 	pt->sync_switch = true;
2098 
2099 	for (i = 0; i < pt->queues.nr_queues; i++) {
2100 		struct auxtrace_queue *queue = &pt->queues.queue_array[i];
2101 		struct intel_pt_queue *ptq = queue->priv;
2102 
2103 		if (ptq)
2104 			ptq->sync_switch = true;
2105 	}
2106 }
2107 
2108 /*
2109  * To filter against time ranges, it is only necessary to look at the next start
2110  * or end time.
2111  */
2112 static bool intel_pt_next_time(struct intel_pt_queue *ptq)
2113 {
2114 	struct intel_pt *pt = ptq->pt;
2115 
2116 	if (ptq->sel_start) {
2117 		/* Next time is an end time */
2118 		ptq->sel_start = false;
2119 		ptq->sel_timestamp = pt->time_ranges[ptq->sel_idx].end;
2120 		return true;
2121 	} else if (ptq->sel_idx + 1 < pt->range_cnt) {
2122 		/* Next time is a start time */
2123 		ptq->sel_start = true;
2124 		ptq->sel_idx += 1;
2125 		ptq->sel_timestamp = pt->time_ranges[ptq->sel_idx].start;
2126 		return true;
2127 	}
2128 
2129 	/* No next time */
2130 	return false;
2131 }
2132 
2133 static int intel_pt_time_filter(struct intel_pt_queue *ptq, u64 *ff_timestamp)
2134 {
2135 	int err;
2136 
2137 	while (1) {
2138 		if (ptq->sel_start) {
2139 			if (ptq->timestamp >= ptq->sel_timestamp) {
2140 				/* After start time, so consider next time */
2141 				intel_pt_next_time(ptq);
2142 				if (!ptq->sel_timestamp) {
2143 					/* No end time */
2144 					return 0;
2145 				}
2146 				/* Check against end time */
2147 				continue;
2148 			}
2149 			/* Before start time, so fast forward */
2150 			ptq->have_sample = false;
2151 			if (ptq->sel_timestamp > *ff_timestamp) {
2152 				if (ptq->sync_switch) {
2153 					intel_pt_next_tid(ptq->pt, ptq);
2154 					ptq->switch_state = INTEL_PT_SS_UNKNOWN;
2155 				}
2156 				*ff_timestamp = ptq->sel_timestamp;
2157 				err = intel_pt_fast_forward(ptq->decoder,
2158 							    ptq->sel_timestamp);
2159 				if (err)
2160 					return err;
2161 			}
2162 			return 0;
2163 		} else if (ptq->timestamp > ptq->sel_timestamp) {
2164 			/* After end time, so consider next time */
2165 			if (!intel_pt_next_time(ptq)) {
2166 				/* No next time range, so stop decoding */
2167 				ptq->have_sample = false;
2168 				ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
2169 				return 1;
2170 			}
2171 			/* Check against next start time */
2172 			continue;
2173 		} else {
2174 			/* Before end time */
2175 			return 0;
2176 		}
2177 	}
2178 }
2179 
2180 static int intel_pt_run_decoder(struct intel_pt_queue *ptq, u64 *timestamp)
2181 {
2182 	const struct intel_pt_state *state = ptq->state;
2183 	struct intel_pt *pt = ptq->pt;
2184 	u64 ff_timestamp = 0;
2185 	int err;
2186 
2187 	if (!pt->kernel_start) {
2188 		pt->kernel_start = machine__kernel_start(pt->machine);
2189 		if (pt->per_cpu_mmaps &&
2190 		    (pt->have_sched_switch == 1 || pt->have_sched_switch == 3) &&
2191 		    !pt->timeless_decoding && intel_pt_tracing_kernel(pt) &&
2192 		    !pt->sampling_mode) {
2193 			pt->switch_ip = intel_pt_switch_ip(pt, &pt->ptss_ip);
2194 			if (pt->switch_ip) {
2195 				intel_pt_log("switch_ip: %"PRIx64" ptss_ip: %"PRIx64"\n",
2196 					     pt->switch_ip, pt->ptss_ip);
2197 				intel_pt_enable_sync_switch(pt);
2198 			}
2199 		}
2200 	}
2201 
2202 	intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
2203 		     ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
2204 	while (1) {
2205 		err = intel_pt_sample(ptq);
2206 		if (err)
2207 			return err;
2208 
2209 		state = intel_pt_decode(ptq->decoder);
2210 		if (state->err) {
2211 			if (state->err == INTEL_PT_ERR_NODATA)
2212 				return 1;
2213 			if (ptq->sync_switch &&
2214 			    state->from_ip >= pt->kernel_start) {
2215 				ptq->sync_switch = false;
2216 				intel_pt_next_tid(pt, ptq);
2217 			}
2218 			if (pt->synth_opts.errors) {
2219 				err = intel_ptq_synth_error(ptq, state);
2220 				if (err)
2221 					return err;
2222 			}
2223 			continue;
2224 		}
2225 
2226 		ptq->state = state;
2227 		ptq->have_sample = true;
2228 		intel_pt_sample_flags(ptq);
2229 
2230 		/* Use estimated TSC upon return to user space */
2231 		if (pt->est_tsc &&
2232 		    (state->from_ip >= pt->kernel_start || !state->from_ip) &&
2233 		    state->to_ip && state->to_ip < pt->kernel_start) {
2234 			intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
2235 				     state->timestamp, state->est_timestamp);
2236 			ptq->timestamp = state->est_timestamp;
2237 		/* Use estimated TSC in unknown switch state */
2238 		} else if (ptq->sync_switch &&
2239 			   ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
2240 			   intel_pt_is_switch_ip(ptq, state->to_ip) &&
2241 			   ptq->next_tid == -1) {
2242 			intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
2243 				     state->timestamp, state->est_timestamp);
2244 			ptq->timestamp = state->est_timestamp;
2245 		} else if (state->timestamp > ptq->timestamp) {
2246 			ptq->timestamp = state->timestamp;
2247 		}
2248 
2249 		if (ptq->sel_timestamp) {
2250 			err = intel_pt_time_filter(ptq, &ff_timestamp);
2251 			if (err)
2252 				return err;
2253 		}
2254 
2255 		if (!pt->timeless_decoding && ptq->timestamp >= *timestamp) {
2256 			*timestamp = ptq->timestamp;
2257 			return 0;
2258 		}
2259 	}
2260 	return 0;
2261 }
2262 
2263 static inline int intel_pt_update_queues(struct intel_pt *pt)
2264 {
2265 	if (pt->queues.new_data) {
2266 		pt->queues.new_data = false;
2267 		return intel_pt_setup_queues(pt);
2268 	}
2269 	return 0;
2270 }
2271 
2272 static int intel_pt_process_queues(struct intel_pt *pt, u64 timestamp)
2273 {
2274 	unsigned int queue_nr;
2275 	u64 ts;
2276 	int ret;
2277 
2278 	while (1) {
2279 		struct auxtrace_queue *queue;
2280 		struct intel_pt_queue *ptq;
2281 
2282 		if (!pt->heap.heap_cnt)
2283 			return 0;
2284 
2285 		if (pt->heap.heap_array[0].ordinal >= timestamp)
2286 			return 0;
2287 
2288 		queue_nr = pt->heap.heap_array[0].queue_nr;
2289 		queue = &pt->queues.queue_array[queue_nr];
2290 		ptq = queue->priv;
2291 
2292 		intel_pt_log("queue %u processing 0x%" PRIx64 " to 0x%" PRIx64 "\n",
2293 			     queue_nr, pt->heap.heap_array[0].ordinal,
2294 			     timestamp);
2295 
2296 		auxtrace_heap__pop(&pt->heap);
2297 
2298 		if (pt->heap.heap_cnt) {
2299 			ts = pt->heap.heap_array[0].ordinal + 1;
2300 			if (ts > timestamp)
2301 				ts = timestamp;
2302 		} else {
2303 			ts = timestamp;
2304 		}
2305 
2306 		intel_pt_set_pid_tid_cpu(pt, queue);
2307 
2308 		ret = intel_pt_run_decoder(ptq, &ts);
2309 
2310 		if (ret < 0) {
2311 			auxtrace_heap__add(&pt->heap, queue_nr, ts);
2312 			return ret;
2313 		}
2314 
2315 		if (!ret) {
2316 			ret = auxtrace_heap__add(&pt->heap, queue_nr, ts);
2317 			if (ret < 0)
2318 				return ret;
2319 		} else {
2320 			ptq->on_heap = false;
2321 		}
2322 	}
2323 
2324 	return 0;
2325 }
2326 
2327 static int intel_pt_process_timeless_queues(struct intel_pt *pt, pid_t tid,
2328 					    u64 time_)
2329 {
2330 	struct auxtrace_queues *queues = &pt->queues;
2331 	unsigned int i;
2332 	u64 ts = 0;
2333 
2334 	for (i = 0; i < queues->nr_queues; i++) {
2335 		struct auxtrace_queue *queue = &pt->queues.queue_array[i];
2336 		struct intel_pt_queue *ptq = queue->priv;
2337 
2338 		if (ptq && (tid == -1 || ptq->tid == tid)) {
2339 			ptq->time = time_;
2340 			intel_pt_set_pid_tid_cpu(pt, queue);
2341 			intel_pt_run_decoder(ptq, &ts);
2342 		}
2343 	}
2344 	return 0;
2345 }
2346 
2347 static void intel_pt_sample_set_pid_tid_cpu(struct intel_pt_queue *ptq,
2348 					    struct auxtrace_queue *queue,
2349 					    struct perf_sample *sample)
2350 {
2351 	struct machine *m = ptq->pt->machine;
2352 
2353 	ptq->pid = sample->pid;
2354 	ptq->tid = sample->tid;
2355 	ptq->cpu = queue->cpu;
2356 
2357 	intel_pt_log("queue %u cpu %d pid %d tid %d\n",
2358 		     ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
2359 
2360 	thread__zput(ptq->thread);
2361 
2362 	if (ptq->tid == -1)
2363 		return;
2364 
2365 	if (ptq->pid == -1) {
2366 		ptq->thread = machine__find_thread(m, -1, ptq->tid);
2367 		if (ptq->thread)
2368 			ptq->pid = ptq->thread->pid_;
2369 		return;
2370 	}
2371 
2372 	ptq->thread = machine__findnew_thread(m, ptq->pid, ptq->tid);
2373 }
2374 
2375 static int intel_pt_process_timeless_sample(struct intel_pt *pt,
2376 					    struct perf_sample *sample)
2377 {
2378 	struct auxtrace_queue *queue;
2379 	struct intel_pt_queue *ptq;
2380 	u64 ts = 0;
2381 
2382 	queue = auxtrace_queues__sample_queue(&pt->queues, sample, pt->session);
2383 	if (!queue)
2384 		return -EINVAL;
2385 
2386 	ptq = queue->priv;
2387 	if (!ptq)
2388 		return 0;
2389 
2390 	ptq->stop = false;
2391 	ptq->time = sample->time;
2392 	intel_pt_sample_set_pid_tid_cpu(ptq, queue, sample);
2393 	intel_pt_run_decoder(ptq, &ts);
2394 	return 0;
2395 }
2396 
2397 static int intel_pt_lost(struct intel_pt *pt, struct perf_sample *sample)
2398 {
2399 	return intel_pt_synth_error(pt, INTEL_PT_ERR_LOST, sample->cpu,
2400 				    sample->pid, sample->tid, 0, sample->time);
2401 }
2402 
2403 static struct intel_pt_queue *intel_pt_cpu_to_ptq(struct intel_pt *pt, int cpu)
2404 {
2405 	unsigned i, j;
2406 
2407 	if (cpu < 0 || !pt->queues.nr_queues)
2408 		return NULL;
2409 
2410 	if ((unsigned)cpu >= pt->queues.nr_queues)
2411 		i = pt->queues.nr_queues - 1;
2412 	else
2413 		i = cpu;
2414 
2415 	if (pt->queues.queue_array[i].cpu == cpu)
2416 		return pt->queues.queue_array[i].priv;
2417 
2418 	for (j = 0; i > 0; j++) {
2419 		if (pt->queues.queue_array[--i].cpu == cpu)
2420 			return pt->queues.queue_array[i].priv;
2421 	}
2422 
2423 	for (; j < pt->queues.nr_queues; j++) {
2424 		if (pt->queues.queue_array[j].cpu == cpu)
2425 			return pt->queues.queue_array[j].priv;
2426 	}
2427 
2428 	return NULL;
2429 }
2430 
2431 static int intel_pt_sync_switch(struct intel_pt *pt, int cpu, pid_t tid,
2432 				u64 timestamp)
2433 {
2434 	struct intel_pt_queue *ptq;
2435 	int err;
2436 
2437 	if (!pt->sync_switch)
2438 		return 1;
2439 
2440 	ptq = intel_pt_cpu_to_ptq(pt, cpu);
2441 	if (!ptq || !ptq->sync_switch)
2442 		return 1;
2443 
2444 	switch (ptq->switch_state) {
2445 	case INTEL_PT_SS_NOT_TRACING:
2446 		break;
2447 	case INTEL_PT_SS_UNKNOWN:
2448 	case INTEL_PT_SS_TRACING:
2449 		ptq->next_tid = tid;
2450 		ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_IP;
2451 		return 0;
2452 	case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
2453 		if (!ptq->on_heap) {
2454 			ptq->timestamp = perf_time_to_tsc(timestamp,
2455 							  &pt->tc);
2456 			err = auxtrace_heap__add(&pt->heap, ptq->queue_nr,
2457 						 ptq->timestamp);
2458 			if (err)
2459 				return err;
2460 			ptq->on_heap = true;
2461 		}
2462 		ptq->switch_state = INTEL_PT_SS_TRACING;
2463 		break;
2464 	case INTEL_PT_SS_EXPECTING_SWITCH_IP:
2465 		intel_pt_log("ERROR: cpu %d expecting switch ip\n", cpu);
2466 		break;
2467 	default:
2468 		break;
2469 	}
2470 
2471 	ptq->next_tid = -1;
2472 
2473 	return 1;
2474 }
2475 
2476 static int intel_pt_process_switch(struct intel_pt *pt,
2477 				   struct perf_sample *sample)
2478 {
2479 	struct evsel *evsel;
2480 	pid_t tid;
2481 	int cpu, ret;
2482 
2483 	evsel = perf_evlist__id2evsel(pt->session->evlist, sample->id);
2484 	if (evsel != pt->switch_evsel)
2485 		return 0;
2486 
2487 	tid = evsel__intval(evsel, sample, "next_pid");
2488 	cpu = sample->cpu;
2489 
2490 	intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
2491 		     cpu, tid, sample->time, perf_time_to_tsc(sample->time,
2492 		     &pt->tc));
2493 
2494 	ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
2495 	if (ret <= 0)
2496 		return ret;
2497 
2498 	return machine__set_current_tid(pt->machine, cpu, -1, tid);
2499 }
2500 
2501 static int intel_pt_context_switch_in(struct intel_pt *pt,
2502 				      struct perf_sample *sample)
2503 {
2504 	pid_t pid = sample->pid;
2505 	pid_t tid = sample->tid;
2506 	int cpu = sample->cpu;
2507 
2508 	if (pt->sync_switch) {
2509 		struct intel_pt_queue *ptq;
2510 
2511 		ptq = intel_pt_cpu_to_ptq(pt, cpu);
2512 		if (ptq && ptq->sync_switch) {
2513 			ptq->next_tid = -1;
2514 			switch (ptq->switch_state) {
2515 			case INTEL_PT_SS_NOT_TRACING:
2516 			case INTEL_PT_SS_UNKNOWN:
2517 			case INTEL_PT_SS_TRACING:
2518 				break;
2519 			case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
2520 			case INTEL_PT_SS_EXPECTING_SWITCH_IP:
2521 				ptq->switch_state = INTEL_PT_SS_TRACING;
2522 				break;
2523 			default:
2524 				break;
2525 			}
2526 		}
2527 	}
2528 
2529 	/*
2530 	 * If the current tid has not been updated yet, ensure it is now that
2531 	 * a "switch in" event has occurred.
2532 	 */
2533 	if (machine__get_current_tid(pt->machine, cpu) == tid)
2534 		return 0;
2535 
2536 	return machine__set_current_tid(pt->machine, cpu, pid, tid);
2537 }
2538 
2539 static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,
2540 				   struct perf_sample *sample)
2541 {
2542 	bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
2543 	pid_t pid, tid;
2544 	int cpu, ret;
2545 
2546 	cpu = sample->cpu;
2547 
2548 	if (pt->have_sched_switch == 3) {
2549 		if (!out)
2550 			return intel_pt_context_switch_in(pt, sample);
2551 		if (event->header.type != PERF_RECORD_SWITCH_CPU_WIDE) {
2552 			pr_err("Expecting CPU-wide context switch event\n");
2553 			return -EINVAL;
2554 		}
2555 		pid = event->context_switch.next_prev_pid;
2556 		tid = event->context_switch.next_prev_tid;
2557 	} else {
2558 		if (out)
2559 			return 0;
2560 		pid = sample->pid;
2561 		tid = sample->tid;
2562 	}
2563 
2564 	if (tid == -1) {
2565 		pr_err("context_switch event has no tid\n");
2566 		return -EINVAL;
2567 	}
2568 
2569 	intel_pt_log("context_switch: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
2570 		     cpu, pid, tid, sample->time, perf_time_to_tsc(sample->time,
2571 		     &pt->tc));
2572 
2573 	ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
2574 	if (ret <= 0)
2575 		return ret;
2576 
2577 	return machine__set_current_tid(pt->machine, cpu, pid, tid);
2578 }
2579 
2580 static int intel_pt_process_itrace_start(struct intel_pt *pt,
2581 					 union perf_event *event,
2582 					 struct perf_sample *sample)
2583 {
2584 	if (!pt->per_cpu_mmaps)
2585 		return 0;
2586 
2587 	intel_pt_log("itrace_start: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
2588 		     sample->cpu, event->itrace_start.pid,
2589 		     event->itrace_start.tid, sample->time,
2590 		     perf_time_to_tsc(sample->time, &pt->tc));
2591 
2592 	return machine__set_current_tid(pt->machine, sample->cpu,
2593 					event->itrace_start.pid,
2594 					event->itrace_start.tid);
2595 }
2596 
2597 static int intel_pt_process_event(struct perf_session *session,
2598 				  union perf_event *event,
2599 				  struct perf_sample *sample,
2600 				  struct perf_tool *tool)
2601 {
2602 	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2603 					   auxtrace);
2604 	u64 timestamp;
2605 	int err = 0;
2606 
2607 	if (dump_trace)
2608 		return 0;
2609 
2610 	if (!tool->ordered_events) {
2611 		pr_err("Intel Processor Trace requires ordered events\n");
2612 		return -EINVAL;
2613 	}
2614 
2615 	if (sample->time && sample->time != (u64)-1)
2616 		timestamp = perf_time_to_tsc(sample->time, &pt->tc);
2617 	else
2618 		timestamp = 0;
2619 
2620 	if (timestamp || pt->timeless_decoding) {
2621 		err = intel_pt_update_queues(pt);
2622 		if (err)
2623 			return err;
2624 	}
2625 
2626 	if (pt->timeless_decoding) {
2627 		if (pt->sampling_mode) {
2628 			if (sample->aux_sample.size)
2629 				err = intel_pt_process_timeless_sample(pt,
2630 								       sample);
2631 		} else if (event->header.type == PERF_RECORD_EXIT) {
2632 			err = intel_pt_process_timeless_queues(pt,
2633 							       event->fork.tid,
2634 							       sample->time);
2635 		}
2636 	} else if (timestamp) {
2637 		err = intel_pt_process_queues(pt, timestamp);
2638 	}
2639 	if (err)
2640 		return err;
2641 
2642 	if (event->header.type == PERF_RECORD_SAMPLE) {
2643 		if (pt->synth_opts.add_callchain && !sample->callchain)
2644 			intel_pt_add_callchain(pt, sample);
2645 		if (pt->synth_opts.add_last_branch && !sample->branch_stack)
2646 			intel_pt_add_br_stack(pt, sample);
2647 	}
2648 
2649 	if (event->header.type == PERF_RECORD_AUX &&
2650 	    (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
2651 	    pt->synth_opts.errors) {
2652 		err = intel_pt_lost(pt, sample);
2653 		if (err)
2654 			return err;
2655 	}
2656 
2657 	if (pt->switch_evsel && event->header.type == PERF_RECORD_SAMPLE)
2658 		err = intel_pt_process_switch(pt, sample);
2659 	else if (event->header.type == PERF_RECORD_ITRACE_START)
2660 		err = intel_pt_process_itrace_start(pt, event, sample);
2661 	else if (event->header.type == PERF_RECORD_SWITCH ||
2662 		 event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
2663 		err = intel_pt_context_switch(pt, event, sample);
2664 
2665 	intel_pt_log("event %u: cpu %d time %"PRIu64" tsc %#"PRIx64" ",
2666 		     event->header.type, sample->cpu, sample->time, timestamp);
2667 	intel_pt_log_event(event);
2668 
2669 	return err;
2670 }
2671 
2672 static int intel_pt_flush(struct perf_session *session, struct perf_tool *tool)
2673 {
2674 	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2675 					   auxtrace);
2676 	int ret;
2677 
2678 	if (dump_trace)
2679 		return 0;
2680 
2681 	if (!tool->ordered_events)
2682 		return -EINVAL;
2683 
2684 	ret = intel_pt_update_queues(pt);
2685 	if (ret < 0)
2686 		return ret;
2687 
2688 	if (pt->timeless_decoding)
2689 		return intel_pt_process_timeless_queues(pt, -1,
2690 							MAX_TIMESTAMP - 1);
2691 
2692 	return intel_pt_process_queues(pt, MAX_TIMESTAMP);
2693 }
2694 
2695 static void intel_pt_free_events(struct perf_session *session)
2696 {
2697 	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2698 					   auxtrace);
2699 	struct auxtrace_queues *queues = &pt->queues;
2700 	unsigned int i;
2701 
2702 	for (i = 0; i < queues->nr_queues; i++) {
2703 		intel_pt_free_queue(queues->queue_array[i].priv);
2704 		queues->queue_array[i].priv = NULL;
2705 	}
2706 	intel_pt_log_disable();
2707 	auxtrace_queues__free(queues);
2708 }
2709 
2710 static void intel_pt_free(struct perf_session *session)
2711 {
2712 	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2713 					   auxtrace);
2714 
2715 	auxtrace_heap__free(&pt->heap);
2716 	intel_pt_free_events(session);
2717 	session->auxtrace = NULL;
2718 	thread__put(pt->unknown_thread);
2719 	addr_filters__exit(&pt->filts);
2720 	zfree(&pt->chain);
2721 	zfree(&pt->filter);
2722 	zfree(&pt->time_ranges);
2723 	free(pt);
2724 }
2725 
2726 static bool intel_pt_evsel_is_auxtrace(struct perf_session *session,
2727 				       struct evsel *evsel)
2728 {
2729 	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2730 					   auxtrace);
2731 
2732 	return evsel->core.attr.type == pt->pmu_type;
2733 }
2734 
2735 static int intel_pt_process_auxtrace_event(struct perf_session *session,
2736 					   union perf_event *event,
2737 					   struct perf_tool *tool __maybe_unused)
2738 {
2739 	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2740 					   auxtrace);
2741 
2742 	if (!pt->data_queued) {
2743 		struct auxtrace_buffer *buffer;
2744 		off_t data_offset;
2745 		int fd = perf_data__fd(session->data);
2746 		int err;
2747 
2748 		if (perf_data__is_pipe(session->data)) {
2749 			data_offset = 0;
2750 		} else {
2751 			data_offset = lseek(fd, 0, SEEK_CUR);
2752 			if (data_offset == -1)
2753 				return -errno;
2754 		}
2755 
2756 		err = auxtrace_queues__add_event(&pt->queues, session, event,
2757 						 data_offset, &buffer);
2758 		if (err)
2759 			return err;
2760 
2761 		/* Dump here now we have copied a piped trace out of the pipe */
2762 		if (dump_trace) {
2763 			if (auxtrace_buffer__get_data(buffer, fd)) {
2764 				intel_pt_dump_event(pt, buffer->data,
2765 						    buffer->size);
2766 				auxtrace_buffer__put_data(buffer);
2767 			}
2768 		}
2769 	}
2770 
2771 	return 0;
2772 }
2773 
2774 static int intel_pt_queue_data(struct perf_session *session,
2775 			       struct perf_sample *sample,
2776 			       union perf_event *event, u64 data_offset)
2777 {
2778 	struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2779 					   auxtrace);
2780 	u64 timestamp;
2781 
2782 	if (event) {
2783 		return auxtrace_queues__add_event(&pt->queues, session, event,
2784 						  data_offset, NULL);
2785 	}
2786 
2787 	if (sample->time && sample->time != (u64)-1)
2788 		timestamp = perf_time_to_tsc(sample->time, &pt->tc);
2789 	else
2790 		timestamp = 0;
2791 
2792 	return auxtrace_queues__add_sample(&pt->queues, session, sample,
2793 					   data_offset, timestamp);
2794 }
2795 
2796 struct intel_pt_synth {
2797 	struct perf_tool dummy_tool;
2798 	struct perf_session *session;
2799 };
2800 
2801 static int intel_pt_event_synth(struct perf_tool *tool,
2802 				union perf_event *event,
2803 				struct perf_sample *sample __maybe_unused,
2804 				struct machine *machine __maybe_unused)
2805 {
2806 	struct intel_pt_synth *intel_pt_synth =
2807 			container_of(tool, struct intel_pt_synth, dummy_tool);
2808 
2809 	return perf_session__deliver_synth_event(intel_pt_synth->session, event,
2810 						 NULL);
2811 }
2812 
2813 static int intel_pt_synth_event(struct perf_session *session, const char *name,
2814 				struct perf_event_attr *attr, u64 id)
2815 {
2816 	struct intel_pt_synth intel_pt_synth;
2817 	int err;
2818 
2819 	pr_debug("Synthesizing '%s' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
2820 		 name, id, (u64)attr->sample_type);
2821 
2822 	memset(&intel_pt_synth, 0, sizeof(struct intel_pt_synth));
2823 	intel_pt_synth.session = session;
2824 
2825 	err = perf_event__synthesize_attr(&intel_pt_synth.dummy_tool, attr, 1,
2826 					  &id, intel_pt_event_synth);
2827 	if (err)
2828 		pr_err("%s: failed to synthesize '%s' event type\n",
2829 		       __func__, name);
2830 
2831 	return err;
2832 }
2833 
2834 static void intel_pt_set_event_name(struct evlist *evlist, u64 id,
2835 				    const char *name)
2836 {
2837 	struct evsel *evsel;
2838 
2839 	evlist__for_each_entry(evlist, evsel) {
2840 		if (evsel->core.id && evsel->core.id[0] == id) {
2841 			if (evsel->name)
2842 				zfree(&evsel->name);
2843 			evsel->name = strdup(name);
2844 			break;
2845 		}
2846 	}
2847 }
2848 
2849 static struct evsel *intel_pt_evsel(struct intel_pt *pt,
2850 					 struct evlist *evlist)
2851 {
2852 	struct evsel *evsel;
2853 
2854 	evlist__for_each_entry(evlist, evsel) {
2855 		if (evsel->core.attr.type == pt->pmu_type && evsel->core.ids)
2856 			return evsel;
2857 	}
2858 
2859 	return NULL;
2860 }
2861 
2862 static int intel_pt_synth_events(struct intel_pt *pt,
2863 				 struct perf_session *session)
2864 {
2865 	struct evlist *evlist = session->evlist;
2866 	struct evsel *evsel = intel_pt_evsel(pt, evlist);
2867 	struct perf_event_attr attr;
2868 	u64 id;
2869 	int err;
2870 
2871 	if (!evsel) {
2872 		pr_debug("There are no selected events with Intel Processor Trace data\n");
2873 		return 0;
2874 	}
2875 
2876 	memset(&attr, 0, sizeof(struct perf_event_attr));
2877 	attr.size = sizeof(struct perf_event_attr);
2878 	attr.type = PERF_TYPE_HARDWARE;
2879 	attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK;
2880 	attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
2881 			    PERF_SAMPLE_PERIOD;
2882 	if (pt->timeless_decoding)
2883 		attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
2884 	else
2885 		attr.sample_type |= PERF_SAMPLE_TIME;
2886 	if (!pt->per_cpu_mmaps)
2887 		attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
2888 	attr.exclude_user = evsel->core.attr.exclude_user;
2889 	attr.exclude_kernel = evsel->core.attr.exclude_kernel;
2890 	attr.exclude_hv = evsel->core.attr.exclude_hv;
2891 	attr.exclude_host = evsel->core.attr.exclude_host;
2892 	attr.exclude_guest = evsel->core.attr.exclude_guest;
2893 	attr.sample_id_all = evsel->core.attr.sample_id_all;
2894 	attr.read_format = evsel->core.attr.read_format;
2895 
2896 	id = evsel->core.id[0] + 1000000000;
2897 	if (!id)
2898 		id = 1;
2899 
2900 	if (pt->synth_opts.branches) {
2901 		attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
2902 		attr.sample_period = 1;
2903 		attr.sample_type |= PERF_SAMPLE_ADDR;
2904 		err = intel_pt_synth_event(session, "branches", &attr, id);
2905 		if (err)
2906 			return err;
2907 		pt->sample_branches = true;
2908 		pt->branches_sample_type = attr.sample_type;
2909 		pt->branches_id = id;
2910 		id += 1;
2911 		attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR;
2912 	}
2913 
2914 	if (pt->synth_opts.callchain)
2915 		attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
2916 	if (pt->synth_opts.last_branch)
2917 		attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
2918 
2919 	if (pt->synth_opts.instructions) {
2920 		attr.config = PERF_COUNT_HW_INSTRUCTIONS;
2921 		if (pt->synth_opts.period_type == PERF_ITRACE_PERIOD_NANOSECS)
2922 			attr.sample_period =
2923 				intel_pt_ns_to_ticks(pt, pt->synth_opts.period);
2924 		else
2925 			attr.sample_period = pt->synth_opts.period;
2926 		err = intel_pt_synth_event(session, "instructions", &attr, id);
2927 		if (err)
2928 			return err;
2929 		pt->sample_instructions = true;
2930 		pt->instructions_sample_type = attr.sample_type;
2931 		pt->instructions_id = id;
2932 		id += 1;
2933 	}
2934 
2935 	attr.sample_type &= ~(u64)PERF_SAMPLE_PERIOD;
2936 	attr.sample_period = 1;
2937 
2938 	if (pt->synth_opts.transactions) {
2939 		attr.config = PERF_COUNT_HW_INSTRUCTIONS;
2940 		err = intel_pt_synth_event(session, "transactions", &attr, id);
2941 		if (err)
2942 			return err;
2943 		pt->sample_transactions = true;
2944 		pt->transactions_sample_type = attr.sample_type;
2945 		pt->transactions_id = id;
2946 		intel_pt_set_event_name(evlist, id, "transactions");
2947 		id += 1;
2948 	}
2949 
2950 	attr.type = PERF_TYPE_SYNTH;
2951 	attr.sample_type |= PERF_SAMPLE_RAW;
2952 
2953 	if (pt->synth_opts.ptwrites) {
2954 		attr.config = PERF_SYNTH_INTEL_PTWRITE;
2955 		err = intel_pt_synth_event(session, "ptwrite", &attr, id);
2956 		if (err)
2957 			return err;
2958 		pt->sample_ptwrites = true;
2959 		pt->ptwrites_sample_type = attr.sample_type;
2960 		pt->ptwrites_id = id;
2961 		intel_pt_set_event_name(evlist, id, "ptwrite");
2962 		id += 1;
2963 	}
2964 
2965 	if (pt->synth_opts.pwr_events) {
2966 		pt->sample_pwr_events = true;
2967 		pt->pwr_events_sample_type = attr.sample_type;
2968 
2969 		attr.config = PERF_SYNTH_INTEL_CBR;
2970 		err = intel_pt_synth_event(session, "cbr", &attr, id);
2971 		if (err)
2972 			return err;
2973 		pt->cbr_id = id;
2974 		intel_pt_set_event_name(evlist, id, "cbr");
2975 		id += 1;
2976 	}
2977 
2978 	if (pt->synth_opts.pwr_events && (evsel->core.attr.config & 0x10)) {
2979 		attr.config = PERF_SYNTH_INTEL_MWAIT;
2980 		err = intel_pt_synth_event(session, "mwait", &attr, id);
2981 		if (err)
2982 			return err;
2983 		pt->mwait_id = id;
2984 		intel_pt_set_event_name(evlist, id, "mwait");
2985 		id += 1;
2986 
2987 		attr.config = PERF_SYNTH_INTEL_PWRE;
2988 		err = intel_pt_synth_event(session, "pwre", &attr, id);
2989 		if (err)
2990 			return err;
2991 		pt->pwre_id = id;
2992 		intel_pt_set_event_name(evlist, id, "pwre");
2993 		id += 1;
2994 
2995 		attr.config = PERF_SYNTH_INTEL_EXSTOP;
2996 		err = intel_pt_synth_event(session, "exstop", &attr, id);
2997 		if (err)
2998 			return err;
2999 		pt->exstop_id = id;
3000 		intel_pt_set_event_name(evlist, id, "exstop");
3001 		id += 1;
3002 
3003 		attr.config = PERF_SYNTH_INTEL_PWRX;
3004 		err = intel_pt_synth_event(session, "pwrx", &attr, id);
3005 		if (err)
3006 			return err;
3007 		pt->pwrx_id = id;
3008 		intel_pt_set_event_name(evlist, id, "pwrx");
3009 		id += 1;
3010 	}
3011 
3012 	return 0;
3013 }
3014 
3015 static void intel_pt_setup_pebs_events(struct intel_pt *pt)
3016 {
3017 	struct evsel *evsel;
3018 
3019 	if (!pt->synth_opts.other_events)
3020 		return;
3021 
3022 	evlist__for_each_entry(pt->session->evlist, evsel) {
3023 		if (evsel->core.attr.aux_output && evsel->core.id) {
3024 			pt->sample_pebs = true;
3025 			pt->pebs_evsel = evsel;
3026 			return;
3027 		}
3028 	}
3029 }
3030 
3031 static struct evsel *intel_pt_find_sched_switch(struct evlist *evlist)
3032 {
3033 	struct evsel *evsel;
3034 
3035 	evlist__for_each_entry_reverse(evlist, evsel) {
3036 		const char *name = evsel__name(evsel);
3037 
3038 		if (!strcmp(name, "sched:sched_switch"))
3039 			return evsel;
3040 	}
3041 
3042 	return NULL;
3043 }
3044 
3045 static bool intel_pt_find_switch(struct evlist *evlist)
3046 {
3047 	struct evsel *evsel;
3048 
3049 	evlist__for_each_entry(evlist, evsel) {
3050 		if (evsel->core.attr.context_switch)
3051 			return true;
3052 	}
3053 
3054 	return false;
3055 }
3056 
3057 static int intel_pt_perf_config(const char *var, const char *value, void *data)
3058 {
3059 	struct intel_pt *pt = data;
3060 
3061 	if (!strcmp(var, "intel-pt.mispred-all"))
3062 		pt->mispred_all = perf_config_bool(var, value);
3063 
3064 	return 0;
3065 }
3066 
3067 /* Find least TSC which converts to ns or later */
3068 static u64 intel_pt_tsc_start(u64 ns, struct intel_pt *pt)
3069 {
3070 	u64 tsc, tm;
3071 
3072 	tsc = perf_time_to_tsc(ns, &pt->tc);
3073 
3074 	while (1) {
3075 		tm = tsc_to_perf_time(tsc, &pt->tc);
3076 		if (tm < ns)
3077 			break;
3078 		tsc -= 1;
3079 	}
3080 
3081 	while (tm < ns)
3082 		tm = tsc_to_perf_time(++tsc, &pt->tc);
3083 
3084 	return tsc;
3085 }
3086 
3087 /* Find greatest TSC which converts to ns or earlier */
3088 static u64 intel_pt_tsc_end(u64 ns, struct intel_pt *pt)
3089 {
3090 	u64 tsc, tm;
3091 
3092 	tsc = perf_time_to_tsc(ns, &pt->tc);
3093 
3094 	while (1) {
3095 		tm = tsc_to_perf_time(tsc, &pt->tc);
3096 		if (tm > ns)
3097 			break;
3098 		tsc += 1;
3099 	}
3100 
3101 	while (tm > ns)
3102 		tm = tsc_to_perf_time(--tsc, &pt->tc);
3103 
3104 	return tsc;
3105 }
3106 
3107 static int intel_pt_setup_time_ranges(struct intel_pt *pt,
3108 				      struct itrace_synth_opts *opts)
3109 {
3110 	struct perf_time_interval *p = opts->ptime_range;
3111 	int n = opts->range_num;
3112 	int i;
3113 
3114 	if (!n || !p || pt->timeless_decoding)
3115 		return 0;
3116 
3117 	pt->time_ranges = calloc(n, sizeof(struct range));
3118 	if (!pt->time_ranges)
3119 		return -ENOMEM;
3120 
3121 	pt->range_cnt = n;
3122 
3123 	intel_pt_log("%s: %u range(s)\n", __func__, n);
3124 
3125 	for (i = 0; i < n; i++) {
3126 		struct range *r = &pt->time_ranges[i];
3127 		u64 ts = p[i].start;
3128 		u64 te = p[i].end;
3129 
3130 		/*
3131 		 * Take care to ensure the TSC range matches the perf-time range
3132 		 * when converted back to perf-time.
3133 		 */
3134 		r->start = ts ? intel_pt_tsc_start(ts, pt) : 0;
3135 		r->end   = te ? intel_pt_tsc_end(te, pt) : 0;
3136 
3137 		intel_pt_log("range %d: perf time interval: %"PRIu64" to %"PRIu64"\n",
3138 			     i, ts, te);
3139 		intel_pt_log("range %d: TSC time interval: %#"PRIx64" to %#"PRIx64"\n",
3140 			     i, r->start, r->end);
3141 	}
3142 
3143 	return 0;
3144 }
3145 
3146 static const char * const intel_pt_info_fmts[] = {
3147 	[INTEL_PT_PMU_TYPE]		= "  PMU Type            %"PRId64"\n",
3148 	[INTEL_PT_TIME_SHIFT]		= "  Time Shift          %"PRIu64"\n",
3149 	[INTEL_PT_TIME_MULT]		= "  Time Muliplier      %"PRIu64"\n",
3150 	[INTEL_PT_TIME_ZERO]		= "  Time Zero           %"PRIu64"\n",
3151 	[INTEL_PT_CAP_USER_TIME_ZERO]	= "  Cap Time Zero       %"PRId64"\n",
3152 	[INTEL_PT_TSC_BIT]		= "  TSC bit             %#"PRIx64"\n",
3153 	[INTEL_PT_NORETCOMP_BIT]	= "  NoRETComp bit       %#"PRIx64"\n",
3154 	[INTEL_PT_HAVE_SCHED_SWITCH]	= "  Have sched_switch   %"PRId64"\n",
3155 	[INTEL_PT_SNAPSHOT_MODE]	= "  Snapshot mode       %"PRId64"\n",
3156 	[INTEL_PT_PER_CPU_MMAPS]	= "  Per-cpu maps        %"PRId64"\n",
3157 	[INTEL_PT_MTC_BIT]		= "  MTC bit             %#"PRIx64"\n",
3158 	[INTEL_PT_TSC_CTC_N]		= "  TSC:CTC numerator   %"PRIu64"\n",
3159 	[INTEL_PT_TSC_CTC_D]		= "  TSC:CTC denominator %"PRIu64"\n",
3160 	[INTEL_PT_CYC_BIT]		= "  CYC bit             %#"PRIx64"\n",
3161 	[INTEL_PT_MAX_NONTURBO_RATIO]	= "  Max non-turbo ratio %"PRIu64"\n",
3162 	[INTEL_PT_FILTER_STR_LEN]	= "  Filter string len.  %"PRIu64"\n",
3163 };
3164 
3165 static void intel_pt_print_info(__u64 *arr, int start, int finish)
3166 {
3167 	int i;
3168 
3169 	if (!dump_trace)
3170 		return;
3171 
3172 	for (i = start; i <= finish; i++)
3173 		fprintf(stdout, intel_pt_info_fmts[i], arr[i]);
3174 }
3175 
3176 static void intel_pt_print_info_str(const char *name, const char *str)
3177 {
3178 	if (!dump_trace)
3179 		return;
3180 
3181 	fprintf(stdout, "  %-20s%s\n", name, str ? str : "");
3182 }
3183 
3184 static bool intel_pt_has(struct perf_record_auxtrace_info *auxtrace_info, int pos)
3185 {
3186 	return auxtrace_info->header.size >=
3187 		sizeof(struct perf_record_auxtrace_info) + (sizeof(u64) * (pos + 1));
3188 }
3189 
3190 int intel_pt_process_auxtrace_info(union perf_event *event,
3191 				   struct perf_session *session)
3192 {
3193 	struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info;
3194 	size_t min_sz = sizeof(u64) * INTEL_PT_PER_CPU_MMAPS;
3195 	struct intel_pt *pt;
3196 	void *info_end;
3197 	__u64 *info;
3198 	int err;
3199 
3200 	if (auxtrace_info->header.size < sizeof(struct perf_record_auxtrace_info) +
3201 					min_sz)
3202 		return -EINVAL;
3203 
3204 	pt = zalloc(sizeof(struct intel_pt));
3205 	if (!pt)
3206 		return -ENOMEM;
3207 
3208 	addr_filters__init(&pt->filts);
3209 
3210 	err = perf_config(intel_pt_perf_config, pt);
3211 	if (err)
3212 		goto err_free;
3213 
3214 	err = auxtrace_queues__init(&pt->queues);
3215 	if (err)
3216 		goto err_free;
3217 
3218 	intel_pt_log_set_name(INTEL_PT_PMU_NAME);
3219 
3220 	pt->session = session;
3221 	pt->machine = &session->machines.host; /* No kvm support */
3222 	pt->auxtrace_type = auxtrace_info->type;
3223 	pt->pmu_type = auxtrace_info->priv[INTEL_PT_PMU_TYPE];
3224 	pt->tc.time_shift = auxtrace_info->priv[INTEL_PT_TIME_SHIFT];
3225 	pt->tc.time_mult = auxtrace_info->priv[INTEL_PT_TIME_MULT];
3226 	pt->tc.time_zero = auxtrace_info->priv[INTEL_PT_TIME_ZERO];
3227 	pt->cap_user_time_zero = auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO];
3228 	pt->tsc_bit = auxtrace_info->priv[INTEL_PT_TSC_BIT];
3229 	pt->noretcomp_bit = auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT];
3230 	pt->have_sched_switch = auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH];
3231 	pt->snapshot_mode = auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE];
3232 	pt->per_cpu_mmaps = auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS];
3233 	intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_PMU_TYPE,
3234 			    INTEL_PT_PER_CPU_MMAPS);
3235 
3236 	if (intel_pt_has(auxtrace_info, INTEL_PT_CYC_BIT)) {
3237 		pt->mtc_bit = auxtrace_info->priv[INTEL_PT_MTC_BIT];
3238 		pt->mtc_freq_bits = auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS];
3239 		pt->tsc_ctc_ratio_n = auxtrace_info->priv[INTEL_PT_TSC_CTC_N];
3240 		pt->tsc_ctc_ratio_d = auxtrace_info->priv[INTEL_PT_TSC_CTC_D];
3241 		pt->cyc_bit = auxtrace_info->priv[INTEL_PT_CYC_BIT];
3242 		intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_MTC_BIT,
3243 				    INTEL_PT_CYC_BIT);
3244 	}
3245 
3246 	if (intel_pt_has(auxtrace_info, INTEL_PT_MAX_NONTURBO_RATIO)) {
3247 		pt->max_non_turbo_ratio =
3248 			auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO];
3249 		intel_pt_print_info(&auxtrace_info->priv[0],
3250 				    INTEL_PT_MAX_NONTURBO_RATIO,
3251 				    INTEL_PT_MAX_NONTURBO_RATIO);
3252 	}
3253 
3254 	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
3255 	info_end = (void *)info + auxtrace_info->header.size;
3256 
3257 	if (intel_pt_has(auxtrace_info, INTEL_PT_FILTER_STR_LEN)) {
3258 		size_t len;
3259 
3260 		len = auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN];
3261 		intel_pt_print_info(&auxtrace_info->priv[0],
3262 				    INTEL_PT_FILTER_STR_LEN,
3263 				    INTEL_PT_FILTER_STR_LEN);
3264 		if (len) {
3265 			const char *filter = (const char *)info;
3266 
3267 			len = roundup(len + 1, 8);
3268 			info += len >> 3;
3269 			if ((void *)info > info_end) {
3270 				pr_err("%s: bad filter string length\n", __func__);
3271 				err = -EINVAL;
3272 				goto err_free_queues;
3273 			}
3274 			pt->filter = memdup(filter, len);
3275 			if (!pt->filter) {
3276 				err = -ENOMEM;
3277 				goto err_free_queues;
3278 			}
3279 			if (session->header.needs_swap)
3280 				mem_bswap_64(pt->filter, len);
3281 			if (pt->filter[len - 1]) {
3282 				pr_err("%s: filter string not null terminated\n", __func__);
3283 				err = -EINVAL;
3284 				goto err_free_queues;
3285 			}
3286 			err = addr_filters__parse_bare_filter(&pt->filts,
3287 							      filter);
3288 			if (err)
3289 				goto err_free_queues;
3290 		}
3291 		intel_pt_print_info_str("Filter string", pt->filter);
3292 	}
3293 
3294 	pt->timeless_decoding = intel_pt_timeless_decoding(pt);
3295 	if (pt->timeless_decoding && !pt->tc.time_mult)
3296 		pt->tc.time_mult = 1;
3297 	pt->have_tsc = intel_pt_have_tsc(pt);
3298 	pt->sampling_mode = intel_pt_sampling_mode(pt);
3299 	pt->est_tsc = !pt->timeless_decoding;
3300 
3301 	pt->unknown_thread = thread__new(999999999, 999999999);
3302 	if (!pt->unknown_thread) {
3303 		err = -ENOMEM;
3304 		goto err_free_queues;
3305 	}
3306 
3307 	/*
3308 	 * Since this thread will not be kept in any rbtree not in a
3309 	 * list, initialize its list node so that at thread__put() the
3310 	 * current thread lifetime assuption is kept and we don't segfault
3311 	 * at list_del_init().
3312 	 */
3313 	INIT_LIST_HEAD(&pt->unknown_thread->node);
3314 
3315 	err = thread__set_comm(pt->unknown_thread, "unknown", 0);
3316 	if (err)
3317 		goto err_delete_thread;
3318 	if (thread__init_maps(pt->unknown_thread, pt->machine)) {
3319 		err = -ENOMEM;
3320 		goto err_delete_thread;
3321 	}
3322 
3323 	pt->auxtrace.process_event = intel_pt_process_event;
3324 	pt->auxtrace.process_auxtrace_event = intel_pt_process_auxtrace_event;
3325 	pt->auxtrace.queue_data = intel_pt_queue_data;
3326 	pt->auxtrace.dump_auxtrace_sample = intel_pt_dump_sample;
3327 	pt->auxtrace.flush_events = intel_pt_flush;
3328 	pt->auxtrace.free_events = intel_pt_free_events;
3329 	pt->auxtrace.free = intel_pt_free;
3330 	pt->auxtrace.evsel_is_auxtrace = intel_pt_evsel_is_auxtrace;
3331 	session->auxtrace = &pt->auxtrace;
3332 
3333 	if (dump_trace)
3334 		return 0;
3335 
3336 	if (pt->have_sched_switch == 1) {
3337 		pt->switch_evsel = intel_pt_find_sched_switch(session->evlist);
3338 		if (!pt->switch_evsel) {
3339 			pr_err("%s: missing sched_switch event\n", __func__);
3340 			err = -EINVAL;
3341 			goto err_delete_thread;
3342 		}
3343 	} else if (pt->have_sched_switch == 2 &&
3344 		   !intel_pt_find_switch(session->evlist)) {
3345 		pr_err("%s: missing context_switch attribute flag\n", __func__);
3346 		err = -EINVAL;
3347 		goto err_delete_thread;
3348 	}
3349 
3350 	if (session->itrace_synth_opts->set) {
3351 		pt->synth_opts = *session->itrace_synth_opts;
3352 	} else {
3353 		itrace_synth_opts__set_default(&pt->synth_opts,
3354 				session->itrace_synth_opts->default_no_sample);
3355 		if (!session->itrace_synth_opts->default_no_sample &&
3356 		    !session->itrace_synth_opts->inject) {
3357 			pt->synth_opts.branches = false;
3358 			pt->synth_opts.callchain = true;
3359 			pt->synth_opts.add_callchain = true;
3360 		}
3361 		pt->synth_opts.thread_stack =
3362 				session->itrace_synth_opts->thread_stack;
3363 	}
3364 
3365 	if (pt->synth_opts.log)
3366 		intel_pt_log_enable();
3367 
3368 	/* Maximum non-turbo ratio is TSC freq / 100 MHz */
3369 	if (pt->tc.time_mult) {
3370 		u64 tsc_freq = intel_pt_ns_to_ticks(pt, 1000000000);
3371 
3372 		if (!pt->max_non_turbo_ratio)
3373 			pt->max_non_turbo_ratio =
3374 					(tsc_freq + 50000000) / 100000000;
3375 		intel_pt_log("TSC frequency %"PRIu64"\n", tsc_freq);
3376 		intel_pt_log("Maximum non-turbo ratio %u\n",
3377 			     pt->max_non_turbo_ratio);
3378 		pt->cbr2khz = tsc_freq / pt->max_non_turbo_ratio / 1000;
3379 	}
3380 
3381 	err = intel_pt_setup_time_ranges(pt, session->itrace_synth_opts);
3382 	if (err)
3383 		goto err_delete_thread;
3384 
3385 	if (pt->synth_opts.calls)
3386 		pt->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
3387 				       PERF_IP_FLAG_TRACE_END;
3388 	if (pt->synth_opts.returns)
3389 		pt->branches_filter |= PERF_IP_FLAG_RETURN |
3390 				       PERF_IP_FLAG_TRACE_BEGIN;
3391 
3392 	if ((pt->synth_opts.callchain || pt->synth_opts.add_callchain) &&
3393 	    !symbol_conf.use_callchain) {
3394 		symbol_conf.use_callchain = true;
3395 		if (callchain_register_param(&callchain_param) < 0) {
3396 			symbol_conf.use_callchain = false;
3397 			pt->synth_opts.callchain = false;
3398 			pt->synth_opts.add_callchain = false;
3399 		}
3400 	}
3401 
3402 	if (pt->synth_opts.add_callchain) {
3403 		err = intel_pt_callchain_init(pt);
3404 		if (err)
3405 			goto err_delete_thread;
3406 	}
3407 
3408 	if (pt->synth_opts.last_branch || pt->synth_opts.add_last_branch) {
3409 		pt->br_stack_sz = pt->synth_opts.last_branch_sz;
3410 		pt->br_stack_sz_plus = pt->br_stack_sz;
3411 	}
3412 
3413 	if (pt->synth_opts.add_last_branch) {
3414 		err = intel_pt_br_stack_init(pt);
3415 		if (err)
3416 			goto err_delete_thread;
3417 		/*
3418 		 * Additional branch stack size to cater for tracing from the
3419 		 * actual sample ip to where the sample time is recorded.
3420 		 * Measured at about 200 branches, but generously set to 1024.
3421 		 * If kernel space is not being traced, then add just 1 for the
3422 		 * branch to kernel space.
3423 		 */
3424 		if (intel_pt_tracing_kernel(pt))
3425 			pt->br_stack_sz_plus += 1024;
3426 		else
3427 			pt->br_stack_sz_plus += 1;
3428 	}
3429 
3430 	pt->use_thread_stack = pt->synth_opts.callchain ||
3431 			       pt->synth_opts.add_callchain ||
3432 			       pt->synth_opts.thread_stack ||
3433 			       pt->synth_opts.last_branch ||
3434 			       pt->synth_opts.add_last_branch;
3435 
3436 	pt->callstack = pt->synth_opts.callchain ||
3437 			pt->synth_opts.add_callchain ||
3438 			pt->synth_opts.thread_stack;
3439 
3440 	err = intel_pt_synth_events(pt, session);
3441 	if (err)
3442 		goto err_delete_thread;
3443 
3444 	intel_pt_setup_pebs_events(pt);
3445 
3446 	if (pt->sampling_mode || list_empty(&session->auxtrace_index))
3447 		err = auxtrace_queue_data(session, true, true);
3448 	else
3449 		err = auxtrace_queues__process_index(&pt->queues, session);
3450 	if (err)
3451 		goto err_delete_thread;
3452 
3453 	if (pt->queues.populated)
3454 		pt->data_queued = true;
3455 
3456 	if (pt->timeless_decoding)
3457 		pr_debug2("Intel PT decoding without timestamps\n");
3458 
3459 	return 0;
3460 
3461 err_delete_thread:
3462 	zfree(&pt->chain);
3463 	thread__zput(pt->unknown_thread);
3464 err_free_queues:
3465 	intel_pt_log_disable();
3466 	auxtrace_queues__free(&pt->queues);
3467 	session->auxtrace = NULL;
3468 err_free:
3469 	addr_filters__exit(&pt->filts);
3470 	zfree(&pt->filter);
3471 	zfree(&pt->time_ranges);
3472 	free(pt);
3473 	return err;
3474 }
3475