xref: /openbmc/linux/tools/perf/arch/x86/util/intel-pt.c (revision be709d48)
1 /*
2  * intel_pt.c: Intel Processor Trace support
3  * Copyright (c) 2013-2015, Intel Corporation.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  */
15 
16 #include <errno.h>
17 #include <stdbool.h>
18 #include <linux/kernel.h>
19 #include <linux/types.h>
20 #include <linux/bitops.h>
21 #include <linux/log2.h>
22 #include <cpuid.h>
23 
24 #include "../../perf.h"
25 #include "../../util/session.h"
26 #include "../../util/event.h"
27 #include "../../util/evlist.h"
28 #include "../../util/evsel.h"
29 #include "../../util/cpumap.h"
30 #include <subcmd/parse-options.h>
31 #include "../../util/parse-events.h"
32 #include "../../util/pmu.h"
33 #include "../../util/debug.h"
34 #include "../../util/auxtrace.h"
35 #include "../../util/tsc.h"
36 #include "../../util/intel-pt.h"
37 
38 #define KiB(x) ((x) * 1024)
39 #define MiB(x) ((x) * 1024 * 1024)
40 #define KiB_MASK(x) (KiB(x) - 1)
41 #define MiB_MASK(x) (MiB(x) - 1)
42 
43 #define INTEL_PT_PSB_PERIOD_NEAR	256
44 
45 struct intel_pt_snapshot_ref {
46 	void *ref_buf;
47 	size_t ref_offset;
48 	bool wrapped;
49 };
50 
51 struct intel_pt_recording {
52 	struct auxtrace_record		itr;
53 	struct perf_pmu			*intel_pt_pmu;
54 	int				have_sched_switch;
55 	struct perf_evlist		*evlist;
56 	bool				snapshot_mode;
57 	bool				snapshot_init_done;
58 	size_t				snapshot_size;
59 	size_t				snapshot_ref_buf_size;
60 	int				snapshot_ref_cnt;
61 	struct intel_pt_snapshot_ref	*snapshot_refs;
62 	size_t				priv_size;
63 };
64 
65 static int intel_pt_parse_terms_with_default(struct list_head *formats,
66 					     const char *str,
67 					     u64 *config)
68 {
69 	struct list_head *terms;
70 	struct perf_event_attr attr = { .size = 0, };
71 	int err;
72 
73 	terms = malloc(sizeof(struct list_head));
74 	if (!terms)
75 		return -ENOMEM;
76 
77 	INIT_LIST_HEAD(terms);
78 
79 	err = parse_events_terms(terms, str);
80 	if (err)
81 		goto out_free;
82 
83 	attr.config = *config;
84 	err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
85 	if (err)
86 		goto out_free;
87 
88 	*config = attr.config;
89 out_free:
90 	parse_events_terms__delete(terms);
91 	return err;
92 }
93 
94 static int intel_pt_parse_terms(struct list_head *formats, const char *str,
95 				u64 *config)
96 {
97 	*config = 0;
98 	return intel_pt_parse_terms_with_default(formats, str, config);
99 }
100 
101 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
102 {
103 	const u64 top_bit = 1ULL << 63;
104 	u64 res = 0;
105 	int i;
106 
107 	for (i = 0; i < 64; i++) {
108 		if (mask & top_bit) {
109 			res <<= 1;
110 			if (bits & top_bit)
111 				res |= 1;
112 		}
113 		mask <<= 1;
114 		bits <<= 1;
115 	}
116 
117 	return res;
118 }
119 
120 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
121 				struct perf_evlist *evlist, u64 *res)
122 {
123 	struct perf_evsel *evsel;
124 	u64 mask;
125 
126 	*res = 0;
127 
128 	mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
129 	if (!mask)
130 		return -EINVAL;
131 
132 	evlist__for_each_entry(evlist, evsel) {
133 		if (evsel->attr.type == intel_pt_pmu->type) {
134 			*res = intel_pt_masked_bits(mask, evsel->attr.config);
135 			return 0;
136 		}
137 	}
138 
139 	return -EINVAL;
140 }
141 
142 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
143 				  struct perf_evlist *evlist)
144 {
145 	u64 val;
146 	int err, topa_multiple_entries;
147 	size_t psb_period;
148 
149 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
150 				"%d", &topa_multiple_entries) != 1)
151 		topa_multiple_entries = 0;
152 
153 	/*
154 	 * Use caps/topa_multiple_entries to indicate early hardware that had
155 	 * extra frequent PSBs.
156 	 */
157 	if (!topa_multiple_entries) {
158 		psb_period = 256;
159 		goto out;
160 	}
161 
162 	err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
163 	if (err)
164 		val = 0;
165 
166 	psb_period = 1 << (val + 11);
167 out:
168 	pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
169 	return psb_period;
170 }
171 
172 static int intel_pt_pick_bit(int bits, int target)
173 {
174 	int pos, pick = -1;
175 
176 	for (pos = 0; bits; bits >>= 1, pos++) {
177 		if (bits & 1) {
178 			if (pos <= target || pick < 0)
179 				pick = pos;
180 			if (pos >= target)
181 				break;
182 		}
183 	}
184 
185 	return pick;
186 }
187 
188 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
189 {
190 	char buf[256];
191 	int mtc, mtc_periods = 0, mtc_period;
192 	int psb_cyc, psb_periods, psb_period;
193 	int pos = 0;
194 	u64 config;
195 	char c;
196 
197 	pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
198 
199 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
200 				&mtc) != 1)
201 		mtc = 1;
202 
203 	if (mtc) {
204 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
205 					&mtc_periods) != 1)
206 			mtc_periods = 0;
207 		if (mtc_periods) {
208 			mtc_period = intel_pt_pick_bit(mtc_periods, 3);
209 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
210 					 ",mtc,mtc_period=%d", mtc_period);
211 		}
212 	}
213 
214 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
215 				&psb_cyc) != 1)
216 		psb_cyc = 1;
217 
218 	if (psb_cyc && mtc_periods) {
219 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
220 					&psb_periods) != 1)
221 			psb_periods = 0;
222 		if (psb_periods) {
223 			psb_period = intel_pt_pick_bit(psb_periods, 3);
224 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
225 					 ",psb_period=%d", psb_period);
226 		}
227 	}
228 
229 	if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
230 	    perf_pmu__scan_file(intel_pt_pmu, "format/branch", "%c", &c) == 1)
231 		pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
232 
233 	pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
234 
235 	intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
236 
237 	return config;
238 }
239 
240 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
241 					   struct record_opts *opts,
242 					   const char *str)
243 {
244 	struct intel_pt_recording *ptr =
245 			container_of(itr, struct intel_pt_recording, itr);
246 	unsigned long long snapshot_size = 0;
247 	char *endptr;
248 
249 	if (str) {
250 		snapshot_size = strtoull(str, &endptr, 0);
251 		if (*endptr || snapshot_size > SIZE_MAX)
252 			return -1;
253 	}
254 
255 	opts->auxtrace_snapshot_mode = true;
256 	opts->auxtrace_snapshot_size = snapshot_size;
257 
258 	ptr->snapshot_size = snapshot_size;
259 
260 	return 0;
261 }
262 
263 struct perf_event_attr *
264 intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
265 {
266 	struct perf_event_attr *attr;
267 
268 	attr = zalloc(sizeof(struct perf_event_attr));
269 	if (!attr)
270 		return NULL;
271 
272 	attr->config = intel_pt_default_config(intel_pt_pmu);
273 
274 	intel_pt_pmu->selectable = true;
275 
276 	return attr;
277 }
278 
279 static const char *intel_pt_find_filter(struct perf_evlist *evlist,
280 					struct perf_pmu *intel_pt_pmu)
281 {
282 	struct perf_evsel *evsel;
283 
284 	evlist__for_each_entry(evlist, evsel) {
285 		if (evsel->attr.type == intel_pt_pmu->type)
286 			return evsel->filter;
287 	}
288 
289 	return NULL;
290 }
291 
292 static size_t intel_pt_filter_bytes(const char *filter)
293 {
294 	size_t len = filter ? strlen(filter) : 0;
295 
296 	return len ? roundup(len + 1, 8) : 0;
297 }
298 
299 static size_t
300 intel_pt_info_priv_size(struct auxtrace_record *itr, struct perf_evlist *evlist)
301 {
302 	struct intel_pt_recording *ptr =
303 			container_of(itr, struct intel_pt_recording, itr);
304 	const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
305 
306 	ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
307 			 intel_pt_filter_bytes(filter);
308 
309 	return ptr->priv_size;
310 }
311 
312 static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
313 {
314 	unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
315 
316 	__get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
317 	*n = ebx;
318 	*d = eax;
319 }
320 
321 static int intel_pt_info_fill(struct auxtrace_record *itr,
322 			      struct perf_session *session,
323 			      struct auxtrace_info_event *auxtrace_info,
324 			      size_t priv_size)
325 {
326 	struct intel_pt_recording *ptr =
327 			container_of(itr, struct intel_pt_recording, itr);
328 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
329 	struct perf_event_mmap_page *pc;
330 	struct perf_tsc_conversion tc = { .time_mult = 0, };
331 	bool cap_user_time_zero = false, per_cpu_mmaps;
332 	u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
333 	u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
334 	unsigned long max_non_turbo_ratio;
335 	size_t filter_str_len;
336 	const char *filter;
337 	u64 *info;
338 	int err;
339 
340 	if (priv_size != ptr->priv_size)
341 		return -EINVAL;
342 
343 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
344 	intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
345 			     &noretcomp_bit);
346 	intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
347 	mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
348 					      "mtc_period");
349 	intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
350 
351 	intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
352 
353 	if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
354 				"%lu", &max_non_turbo_ratio) != 1)
355 		max_non_turbo_ratio = 0;
356 
357 	filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
358 	filter_str_len = filter ? strlen(filter) : 0;
359 
360 	if (!session->evlist->nr_mmaps)
361 		return -EINVAL;
362 
363 	pc = session->evlist->mmap[0].base;
364 	if (pc) {
365 		err = perf_read_tsc_conversion(pc, &tc);
366 		if (err) {
367 			if (err != -EOPNOTSUPP)
368 				return err;
369 		} else {
370 			cap_user_time_zero = tc.time_mult != 0;
371 		}
372 		if (!cap_user_time_zero)
373 			ui__warning("Intel Processor Trace: TSC not available\n");
374 	}
375 
376 	per_cpu_mmaps = !cpu_map__empty(session->evlist->cpus);
377 
378 	auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
379 	auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
380 	auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
381 	auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
382 	auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
383 	auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
384 	auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
385 	auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
386 	auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
387 	auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
388 	auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
389 	auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
390 	auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
391 	auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
392 	auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
393 	auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
394 	auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
395 	auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
396 
397 	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
398 
399 	if (filter_str_len) {
400 		size_t len = intel_pt_filter_bytes(filter);
401 
402 		strncpy((char *)info, filter, len);
403 		info += len >> 3;
404 	}
405 
406 	return 0;
407 }
408 
409 static int intel_pt_track_switches(struct perf_evlist *evlist)
410 {
411 	const char *sched_switch = "sched:sched_switch";
412 	struct perf_evsel *evsel;
413 	int err;
414 
415 	if (!perf_evlist__can_select_event(evlist, sched_switch))
416 		return -EPERM;
417 
418 	err = parse_events(evlist, sched_switch, NULL);
419 	if (err) {
420 		pr_debug2("%s: failed to parse %s, error %d\n",
421 			  __func__, sched_switch, err);
422 		return err;
423 	}
424 
425 	evsel = perf_evlist__last(evlist);
426 
427 	perf_evsel__set_sample_bit(evsel, CPU);
428 	perf_evsel__set_sample_bit(evsel, TIME);
429 
430 	evsel->system_wide = true;
431 	evsel->no_aux_samples = true;
432 	evsel->immediate = true;
433 
434 	return 0;
435 }
436 
437 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
438 {
439 	unsigned int val, last = 0, state = 1;
440 	int p = 0;
441 
442 	str[0] = '\0';
443 
444 	for (val = 0; val <= 64; val++, valid >>= 1) {
445 		if (valid & 1) {
446 			last = val;
447 			switch (state) {
448 			case 0:
449 				p += scnprintf(str + p, len - p, ",");
450 				/* Fall through */
451 			case 1:
452 				p += scnprintf(str + p, len - p, "%u", val);
453 				state = 2;
454 				break;
455 			case 2:
456 				state = 3;
457 				break;
458 			case 3:
459 				state = 4;
460 				break;
461 			default:
462 				break;
463 			}
464 		} else {
465 			switch (state) {
466 			case 3:
467 				p += scnprintf(str + p, len - p, ",%u", last);
468 				state = 0;
469 				break;
470 			case 4:
471 				p += scnprintf(str + p, len - p, "-%u", last);
472 				state = 0;
473 				break;
474 			default:
475 				break;
476 			}
477 			if (state != 1)
478 				state = 0;
479 		}
480 	}
481 }
482 
483 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
484 				    const char *caps, const char *name,
485 				    const char *supported, u64 config)
486 {
487 	char valid_str[256];
488 	unsigned int shift;
489 	unsigned long long valid;
490 	u64 bits;
491 	int ok;
492 
493 	if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
494 		valid = 0;
495 
496 	if (supported &&
497 	    perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
498 		valid = 0;
499 
500 	valid |= 1;
501 
502 	bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
503 
504 	config &= bits;
505 
506 	for (shift = 0; bits && !(bits & 1); shift++)
507 		bits >>= 1;
508 
509 	config >>= shift;
510 
511 	if (config > 63)
512 		goto out_err;
513 
514 	if (valid & (1 << config))
515 		return 0;
516 out_err:
517 	intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
518 	pr_err("Invalid %s for %s. Valid values are: %s\n",
519 	       name, INTEL_PT_PMU_NAME, valid_str);
520 	return -EINVAL;
521 }
522 
523 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
524 				    struct perf_evsel *evsel)
525 {
526 	int err;
527 	char c;
528 
529 	if (!evsel)
530 		return 0;
531 
532 	/*
533 	 * If supported, force pass-through config term (pt=1) even if user
534 	 * sets pt=0, which avoids senseless kernel errors.
535 	 */
536 	if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
537 	    !(evsel->attr.config & 1)) {
538 		pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
539 		evsel->attr.config |= 1;
540 	}
541 
542 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
543 				       "cyc_thresh", "caps/psb_cyc",
544 				       evsel->attr.config);
545 	if (err)
546 		return err;
547 
548 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
549 				       "mtc_period", "caps/mtc",
550 				       evsel->attr.config);
551 	if (err)
552 		return err;
553 
554 	return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
555 					"psb_period", "caps/psb_cyc",
556 					evsel->attr.config);
557 }
558 
559 static int intel_pt_recording_options(struct auxtrace_record *itr,
560 				      struct perf_evlist *evlist,
561 				      struct record_opts *opts)
562 {
563 	struct intel_pt_recording *ptr =
564 			container_of(itr, struct intel_pt_recording, itr);
565 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
566 	bool have_timing_info, need_immediate = false;
567 	struct perf_evsel *evsel, *intel_pt_evsel = NULL;
568 	const struct cpu_map *cpus = evlist->cpus;
569 	bool privileged = geteuid() == 0 || perf_event_paranoid() < 0;
570 	u64 tsc_bit;
571 	int err;
572 
573 	ptr->evlist = evlist;
574 	ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
575 
576 	evlist__for_each_entry(evlist, evsel) {
577 		if (evsel->attr.type == intel_pt_pmu->type) {
578 			if (intel_pt_evsel) {
579 				pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
580 				return -EINVAL;
581 			}
582 			evsel->attr.freq = 0;
583 			evsel->attr.sample_period = 1;
584 			intel_pt_evsel = evsel;
585 			opts->full_auxtrace = true;
586 		}
587 	}
588 
589 	if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
590 		pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
591 		return -EINVAL;
592 	}
593 
594 	if (opts->use_clockid) {
595 		pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
596 		return -EINVAL;
597 	}
598 
599 	if (!opts->full_auxtrace)
600 		return 0;
601 
602 	err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
603 	if (err)
604 		return err;
605 
606 	/* Set default sizes for snapshot mode */
607 	if (opts->auxtrace_snapshot_mode) {
608 		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
609 
610 		if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
611 			if (privileged) {
612 				opts->auxtrace_mmap_pages = MiB(4) / page_size;
613 			} else {
614 				opts->auxtrace_mmap_pages = KiB(128) / page_size;
615 				if (opts->mmap_pages == UINT_MAX)
616 					opts->mmap_pages = KiB(256) / page_size;
617 			}
618 		} else if (!opts->auxtrace_mmap_pages && !privileged &&
619 			   opts->mmap_pages == UINT_MAX) {
620 			opts->mmap_pages = KiB(256) / page_size;
621 		}
622 		if (!opts->auxtrace_snapshot_size)
623 			opts->auxtrace_snapshot_size =
624 				opts->auxtrace_mmap_pages * (size_t)page_size;
625 		if (!opts->auxtrace_mmap_pages) {
626 			size_t sz = opts->auxtrace_snapshot_size;
627 
628 			sz = round_up(sz, page_size) / page_size;
629 			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
630 		}
631 		if (opts->auxtrace_snapshot_size >
632 				opts->auxtrace_mmap_pages * (size_t)page_size) {
633 			pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
634 			       opts->auxtrace_snapshot_size,
635 			       opts->auxtrace_mmap_pages * (size_t)page_size);
636 			return -EINVAL;
637 		}
638 		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
639 			pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
640 			return -EINVAL;
641 		}
642 		pr_debug2("Intel PT snapshot size: %zu\n",
643 			  opts->auxtrace_snapshot_size);
644 		if (psb_period &&
645 		    opts->auxtrace_snapshot_size <= psb_period +
646 						  INTEL_PT_PSB_PERIOD_NEAR)
647 			ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
648 				    opts->auxtrace_snapshot_size, psb_period);
649 	}
650 
651 	/* Set default sizes for full trace mode */
652 	if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
653 		if (privileged) {
654 			opts->auxtrace_mmap_pages = MiB(4) / page_size;
655 		} else {
656 			opts->auxtrace_mmap_pages = KiB(128) / page_size;
657 			if (opts->mmap_pages == UINT_MAX)
658 				opts->mmap_pages = KiB(256) / page_size;
659 		}
660 	}
661 
662 	/* Validate auxtrace_mmap_pages */
663 	if (opts->auxtrace_mmap_pages) {
664 		size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
665 		size_t min_sz;
666 
667 		if (opts->auxtrace_snapshot_mode)
668 			min_sz = KiB(4);
669 		else
670 			min_sz = KiB(8);
671 
672 		if (sz < min_sz || !is_power_of_2(sz)) {
673 			pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
674 			       min_sz / 1024);
675 			return -EINVAL;
676 		}
677 	}
678 
679 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
680 
681 	if (opts->full_auxtrace && (intel_pt_evsel->attr.config & tsc_bit))
682 		have_timing_info = true;
683 	else
684 		have_timing_info = false;
685 
686 	/*
687 	 * Per-cpu recording needs sched_switch events to distinguish different
688 	 * threads.
689 	 */
690 	if (have_timing_info && !cpu_map__empty(cpus)) {
691 		if (perf_can_record_switch_events()) {
692 			bool cpu_wide = !target__none(&opts->target) &&
693 					!target__has_task(&opts->target);
694 
695 			if (!cpu_wide && perf_can_record_cpu_wide()) {
696 				struct perf_evsel *switch_evsel;
697 
698 				err = parse_events(evlist, "dummy:u", NULL);
699 				if (err)
700 					return err;
701 
702 				switch_evsel = perf_evlist__last(evlist);
703 
704 				switch_evsel->attr.freq = 0;
705 				switch_evsel->attr.sample_period = 1;
706 				switch_evsel->attr.context_switch = 1;
707 
708 				switch_evsel->system_wide = true;
709 				switch_evsel->no_aux_samples = true;
710 				switch_evsel->immediate = true;
711 
712 				perf_evsel__set_sample_bit(switch_evsel, TID);
713 				perf_evsel__set_sample_bit(switch_evsel, TIME);
714 				perf_evsel__set_sample_bit(switch_evsel, CPU);
715 				perf_evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
716 
717 				opts->record_switch_events = false;
718 				ptr->have_sched_switch = 3;
719 			} else {
720 				opts->record_switch_events = true;
721 				need_immediate = true;
722 				if (cpu_wide)
723 					ptr->have_sched_switch = 3;
724 				else
725 					ptr->have_sched_switch = 2;
726 			}
727 		} else {
728 			err = intel_pt_track_switches(evlist);
729 			if (err == -EPERM)
730 				pr_debug2("Unable to select sched:sched_switch\n");
731 			else if (err)
732 				return err;
733 			else
734 				ptr->have_sched_switch = 1;
735 		}
736 	}
737 
738 	if (intel_pt_evsel) {
739 		/*
740 		 * To obtain the auxtrace buffer file descriptor, the auxtrace
741 		 * event must come first.
742 		 */
743 		perf_evlist__to_front(evlist, intel_pt_evsel);
744 		/*
745 		 * In the case of per-cpu mmaps, we need the CPU on the
746 		 * AUX event.
747 		 */
748 		if (!cpu_map__empty(cpus))
749 			perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
750 	}
751 
752 	/* Add dummy event to keep tracking */
753 	if (opts->full_auxtrace) {
754 		struct perf_evsel *tracking_evsel;
755 
756 		err = parse_events(evlist, "dummy:u", NULL);
757 		if (err)
758 			return err;
759 
760 		tracking_evsel = perf_evlist__last(evlist);
761 
762 		perf_evlist__set_tracking_event(evlist, tracking_evsel);
763 
764 		tracking_evsel->attr.freq = 0;
765 		tracking_evsel->attr.sample_period = 1;
766 
767 		tracking_evsel->no_aux_samples = true;
768 		if (need_immediate)
769 			tracking_evsel->immediate = true;
770 
771 		/* In per-cpu case, always need the time of mmap events etc */
772 		if (!cpu_map__empty(cpus)) {
773 			perf_evsel__set_sample_bit(tracking_evsel, TIME);
774 			/* And the CPU for switch events */
775 			perf_evsel__set_sample_bit(tracking_evsel, CPU);
776 		}
777 		perf_evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
778 	}
779 
780 	/*
781 	 * Warn the user when we do not have enough information to decode i.e.
782 	 * per-cpu with no sched_switch (except workload-only).
783 	 */
784 	if (!ptr->have_sched_switch && !cpu_map__empty(cpus) &&
785 	    !target__none(&opts->target))
786 		ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
787 
788 	return 0;
789 }
790 
791 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
792 {
793 	struct intel_pt_recording *ptr =
794 			container_of(itr, struct intel_pt_recording, itr);
795 	struct perf_evsel *evsel;
796 
797 	evlist__for_each_entry(ptr->evlist, evsel) {
798 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
799 			return perf_evsel__disable(evsel);
800 	}
801 	return -EINVAL;
802 }
803 
804 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
805 {
806 	struct intel_pt_recording *ptr =
807 			container_of(itr, struct intel_pt_recording, itr);
808 	struct perf_evsel *evsel;
809 
810 	evlist__for_each_entry(ptr->evlist, evsel) {
811 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
812 			return perf_evsel__enable(evsel);
813 	}
814 	return -EINVAL;
815 }
816 
817 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
818 {
819 	const size_t sz = sizeof(struct intel_pt_snapshot_ref);
820 	int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
821 	struct intel_pt_snapshot_ref *refs;
822 
823 	if (!new_cnt)
824 		new_cnt = 16;
825 
826 	while (new_cnt <= idx)
827 		new_cnt *= 2;
828 
829 	refs = calloc(new_cnt, sz);
830 	if (!refs)
831 		return -ENOMEM;
832 
833 	memcpy(refs, ptr->snapshot_refs, cnt * sz);
834 
835 	ptr->snapshot_refs = refs;
836 	ptr->snapshot_ref_cnt = new_cnt;
837 
838 	return 0;
839 }
840 
841 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
842 {
843 	int i;
844 
845 	for (i = 0; i < ptr->snapshot_ref_cnt; i++)
846 		zfree(&ptr->snapshot_refs[i].ref_buf);
847 	zfree(&ptr->snapshot_refs);
848 }
849 
850 static void intel_pt_recording_free(struct auxtrace_record *itr)
851 {
852 	struct intel_pt_recording *ptr =
853 			container_of(itr, struct intel_pt_recording, itr);
854 
855 	intel_pt_free_snapshot_refs(ptr);
856 	free(ptr);
857 }
858 
859 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
860 				       size_t snapshot_buf_size)
861 {
862 	size_t ref_buf_size = ptr->snapshot_ref_buf_size;
863 	void *ref_buf;
864 
865 	ref_buf = zalloc(ref_buf_size);
866 	if (!ref_buf)
867 		return -ENOMEM;
868 
869 	ptr->snapshot_refs[idx].ref_buf = ref_buf;
870 	ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
871 
872 	return 0;
873 }
874 
875 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
876 					     size_t snapshot_buf_size)
877 {
878 	const size_t max_size = 256 * 1024;
879 	size_t buf_size = 0, psb_period;
880 
881 	if (ptr->snapshot_size <= 64 * 1024)
882 		return 0;
883 
884 	psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
885 	if (psb_period)
886 		buf_size = psb_period * 2;
887 
888 	if (!buf_size || buf_size > max_size)
889 		buf_size = max_size;
890 
891 	if (buf_size >= snapshot_buf_size)
892 		return 0;
893 
894 	if (buf_size >= ptr->snapshot_size / 2)
895 		return 0;
896 
897 	return buf_size;
898 }
899 
900 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
901 				  size_t snapshot_buf_size)
902 {
903 	if (ptr->snapshot_init_done)
904 		return 0;
905 
906 	ptr->snapshot_init_done = true;
907 
908 	ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
909 							snapshot_buf_size);
910 
911 	return 0;
912 }
913 
914 /**
915  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
916  * @buf1: first buffer
917  * @compare_size: number of bytes to compare
918  * @buf2: second buffer (a circular buffer)
919  * @offs2: offset in second buffer
920  * @buf2_size: size of second buffer
921  *
922  * The comparison allows for the possibility that the bytes to compare in the
923  * circular buffer are not contiguous.  It is assumed that @compare_size <=
924  * @buf2_size.  This function returns %false if the bytes are identical, %true
925  * otherwise.
926  */
927 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
928 				     void *buf2, size_t offs2, size_t buf2_size)
929 {
930 	size_t end2 = offs2 + compare_size, part_size;
931 
932 	if (end2 <= buf2_size)
933 		return memcmp(buf1, buf2 + offs2, compare_size);
934 
935 	part_size = end2 - buf2_size;
936 	if (memcmp(buf1, buf2 + offs2, part_size))
937 		return true;
938 
939 	compare_size -= part_size;
940 
941 	return memcmp(buf1 + part_size, buf2, compare_size);
942 }
943 
944 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
945 				 size_t ref_size, size_t buf_size,
946 				 void *data, size_t head)
947 {
948 	size_t ref_end = ref_offset + ref_size;
949 
950 	if (ref_end > buf_size) {
951 		if (head > ref_offset || head < ref_end - buf_size)
952 			return true;
953 	} else if (head > ref_offset && head < ref_end) {
954 		return true;
955 	}
956 
957 	return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
958 					buf_size);
959 }
960 
961 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
962 			      void *data, size_t head)
963 {
964 	if (head >= ref_size) {
965 		memcpy(ref_buf, data + head - ref_size, ref_size);
966 	} else {
967 		memcpy(ref_buf, data, head);
968 		ref_size -= head;
969 		memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
970 	}
971 }
972 
973 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
974 			     struct auxtrace_mmap *mm, unsigned char *data,
975 			     u64 head)
976 {
977 	struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
978 	bool wrapped;
979 
980 	wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
981 				       ptr->snapshot_ref_buf_size, mm->len,
982 				       data, head);
983 
984 	intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
985 			  data, head);
986 
987 	return wrapped;
988 }
989 
990 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
991 {
992 	int i, a, b;
993 
994 	b = buf_size >> 3;
995 	a = b - 512;
996 	if (a < 0)
997 		a = 0;
998 
999 	for (i = a; i < b; i++) {
1000 		if (data[i])
1001 			return true;
1002 	}
1003 
1004 	return false;
1005 }
1006 
1007 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1008 				  struct auxtrace_mmap *mm, unsigned char *data,
1009 				  u64 *head, u64 *old)
1010 {
1011 	struct intel_pt_recording *ptr =
1012 			container_of(itr, struct intel_pt_recording, itr);
1013 	bool wrapped;
1014 	int err;
1015 
1016 	pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1017 		  __func__, idx, (size_t)*old, (size_t)*head);
1018 
1019 	err = intel_pt_snapshot_init(ptr, mm->len);
1020 	if (err)
1021 		goto out_err;
1022 
1023 	if (idx >= ptr->snapshot_ref_cnt) {
1024 		err = intel_pt_alloc_snapshot_refs(ptr, idx);
1025 		if (err)
1026 			goto out_err;
1027 	}
1028 
1029 	if (ptr->snapshot_ref_buf_size) {
1030 		if (!ptr->snapshot_refs[idx].ref_buf) {
1031 			err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1032 			if (err)
1033 				goto out_err;
1034 		}
1035 		wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1036 	} else {
1037 		wrapped = ptr->snapshot_refs[idx].wrapped;
1038 		if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1039 			ptr->snapshot_refs[idx].wrapped = true;
1040 			wrapped = true;
1041 		}
1042 	}
1043 
1044 	/*
1045 	 * In full trace mode 'head' continually increases.  However in snapshot
1046 	 * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1047 	 * are adjusted to match the full trace case which expects that 'old' is
1048 	 * always less than 'head'.
1049 	 */
1050 	if (wrapped) {
1051 		*old = *head;
1052 		*head += mm->len;
1053 	} else {
1054 		if (mm->mask)
1055 			*old &= mm->mask;
1056 		else
1057 			*old %= mm->len;
1058 		if (*old > *head)
1059 			*head += mm->len;
1060 	}
1061 
1062 	pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1063 		  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1064 
1065 	return 0;
1066 
1067 out_err:
1068 	pr_err("%s: failed, error %d\n", __func__, err);
1069 	return err;
1070 }
1071 
1072 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1073 {
1074 	return rdtsc();
1075 }
1076 
1077 static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1078 {
1079 	struct intel_pt_recording *ptr =
1080 			container_of(itr, struct intel_pt_recording, itr);
1081 	struct perf_evsel *evsel;
1082 
1083 	evlist__for_each_entry(ptr->evlist, evsel) {
1084 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
1085 			return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1086 							     idx);
1087 	}
1088 	return -EINVAL;
1089 }
1090 
1091 struct auxtrace_record *intel_pt_recording_init(int *err)
1092 {
1093 	struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1094 	struct intel_pt_recording *ptr;
1095 
1096 	if (!intel_pt_pmu)
1097 		return NULL;
1098 
1099 	if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1100 		*err = -errno;
1101 		return NULL;
1102 	}
1103 
1104 	ptr = zalloc(sizeof(struct intel_pt_recording));
1105 	if (!ptr) {
1106 		*err = -ENOMEM;
1107 		return NULL;
1108 	}
1109 
1110 	ptr->intel_pt_pmu = intel_pt_pmu;
1111 	ptr->itr.recording_options = intel_pt_recording_options;
1112 	ptr->itr.info_priv_size = intel_pt_info_priv_size;
1113 	ptr->itr.info_fill = intel_pt_info_fill;
1114 	ptr->itr.free = intel_pt_recording_free;
1115 	ptr->itr.snapshot_start = intel_pt_snapshot_start;
1116 	ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1117 	ptr->itr.find_snapshot = intel_pt_find_snapshot;
1118 	ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1119 	ptr->itr.reference = intel_pt_reference;
1120 	ptr->itr.read_finish = intel_pt_read_finish;
1121 	return &ptr->itr;
1122 }
1123