xref: /openbmc/linux/tools/perf/arch/x86/util/intel-pt.c (revision 50df3be7)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * intel_pt.c: Intel Processor Trace support
4  * Copyright (c) 2013-2015, Intel Corporation.
5  */
6 
7 #include <errno.h>
8 #include <stdbool.h>
9 #include <linux/kernel.h>
10 #include <linux/types.h>
11 #include <linux/bitops.h>
12 #include <linux/log2.h>
13 #include <linux/zalloc.h>
14 #include <cpuid.h>
15 
16 #include "../../util/session.h"
17 #include "../../util/event.h"
18 #include "../../util/evlist.h"
19 #include "../../util/evsel.h"
20 #include "../../util/cpumap.h"
21 #include "../../util/mmap.h"
22 #include <subcmd/parse-options.h>
23 #include "../../util/parse-events.h"
24 #include "../../util/pmu.h"
25 #include "../../util/debug.h"
26 #include "../../util/auxtrace.h"
27 #include "../../util/record.h"
28 #include "../../util/target.h"
29 #include "../../util/tsc.h"
30 #include <internal/lib.h> // page_size
31 #include "../../util/intel-pt.h"
32 
33 #define KiB(x) ((x) * 1024)
34 #define MiB(x) ((x) * 1024 * 1024)
35 #define KiB_MASK(x) (KiB(x) - 1)
36 #define MiB_MASK(x) (MiB(x) - 1)
37 
38 #define INTEL_PT_PSB_PERIOD_NEAR	256
39 
40 struct intel_pt_snapshot_ref {
41 	void *ref_buf;
42 	size_t ref_offset;
43 	bool wrapped;
44 };
45 
46 struct intel_pt_recording {
47 	struct auxtrace_record		itr;
48 	struct perf_pmu			*intel_pt_pmu;
49 	int				have_sched_switch;
50 	struct evlist		*evlist;
51 	bool				snapshot_mode;
52 	bool				snapshot_init_done;
53 	size_t				snapshot_size;
54 	size_t				snapshot_ref_buf_size;
55 	int				snapshot_ref_cnt;
56 	struct intel_pt_snapshot_ref	*snapshot_refs;
57 	size_t				priv_size;
58 };
59 
60 static int intel_pt_parse_terms_with_default(struct list_head *formats,
61 					     const char *str,
62 					     u64 *config)
63 {
64 	struct list_head *terms;
65 	struct perf_event_attr attr = { .size = 0, };
66 	int err;
67 
68 	terms = malloc(sizeof(struct list_head));
69 	if (!terms)
70 		return -ENOMEM;
71 
72 	INIT_LIST_HEAD(terms);
73 
74 	err = parse_events_terms(terms, str);
75 	if (err)
76 		goto out_free;
77 
78 	attr.config = *config;
79 	err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
80 	if (err)
81 		goto out_free;
82 
83 	*config = attr.config;
84 out_free:
85 	parse_events_terms__delete(terms);
86 	return err;
87 }
88 
89 static int intel_pt_parse_terms(struct list_head *formats, const char *str,
90 				u64 *config)
91 {
92 	*config = 0;
93 	return intel_pt_parse_terms_with_default(formats, str, config);
94 }
95 
96 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
97 {
98 	const u64 top_bit = 1ULL << 63;
99 	u64 res = 0;
100 	int i;
101 
102 	for (i = 0; i < 64; i++) {
103 		if (mask & top_bit) {
104 			res <<= 1;
105 			if (bits & top_bit)
106 				res |= 1;
107 		}
108 		mask <<= 1;
109 		bits <<= 1;
110 	}
111 
112 	return res;
113 }
114 
115 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
116 				struct evlist *evlist, u64 *res)
117 {
118 	struct evsel *evsel;
119 	u64 mask;
120 
121 	*res = 0;
122 
123 	mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
124 	if (!mask)
125 		return -EINVAL;
126 
127 	evlist__for_each_entry(evlist, evsel) {
128 		if (evsel->core.attr.type == intel_pt_pmu->type) {
129 			*res = intel_pt_masked_bits(mask, evsel->core.attr.config);
130 			return 0;
131 		}
132 	}
133 
134 	return -EINVAL;
135 }
136 
137 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
138 				  struct evlist *evlist)
139 {
140 	u64 val;
141 	int err, topa_multiple_entries;
142 	size_t psb_period;
143 
144 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
145 				"%d", &topa_multiple_entries) != 1)
146 		topa_multiple_entries = 0;
147 
148 	/*
149 	 * Use caps/topa_multiple_entries to indicate early hardware that had
150 	 * extra frequent PSBs.
151 	 */
152 	if (!topa_multiple_entries) {
153 		psb_period = 256;
154 		goto out;
155 	}
156 
157 	err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
158 	if (err)
159 		val = 0;
160 
161 	psb_period = 1 << (val + 11);
162 out:
163 	pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
164 	return psb_period;
165 }
166 
167 static int intel_pt_pick_bit(int bits, int target)
168 {
169 	int pos, pick = -1;
170 
171 	for (pos = 0; bits; bits >>= 1, pos++) {
172 		if (bits & 1) {
173 			if (pos <= target || pick < 0)
174 				pick = pos;
175 			if (pos >= target)
176 				break;
177 		}
178 	}
179 
180 	return pick;
181 }
182 
183 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
184 {
185 	char buf[256];
186 	int mtc, mtc_periods = 0, mtc_period;
187 	int psb_cyc, psb_periods, psb_period;
188 	int pos = 0;
189 	u64 config;
190 	char c;
191 
192 	pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
193 
194 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
195 				&mtc) != 1)
196 		mtc = 1;
197 
198 	if (mtc) {
199 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
200 					&mtc_periods) != 1)
201 			mtc_periods = 0;
202 		if (mtc_periods) {
203 			mtc_period = intel_pt_pick_bit(mtc_periods, 3);
204 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
205 					 ",mtc,mtc_period=%d", mtc_period);
206 		}
207 	}
208 
209 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
210 				&psb_cyc) != 1)
211 		psb_cyc = 1;
212 
213 	if (psb_cyc && mtc_periods) {
214 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
215 					&psb_periods) != 1)
216 			psb_periods = 0;
217 		if (psb_periods) {
218 			psb_period = intel_pt_pick_bit(psb_periods, 3);
219 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
220 					 ",psb_period=%d", psb_period);
221 		}
222 	}
223 
224 	if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
225 	    perf_pmu__scan_file(intel_pt_pmu, "format/branch", "%c", &c) == 1)
226 		pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
227 
228 	pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
229 
230 	intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
231 
232 	return config;
233 }
234 
235 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
236 					   struct record_opts *opts,
237 					   const char *str)
238 {
239 	struct intel_pt_recording *ptr =
240 			container_of(itr, struct intel_pt_recording, itr);
241 	unsigned long long snapshot_size = 0;
242 	char *endptr;
243 
244 	if (str) {
245 		snapshot_size = strtoull(str, &endptr, 0);
246 		if (*endptr || snapshot_size > SIZE_MAX)
247 			return -1;
248 	}
249 
250 	opts->auxtrace_snapshot_mode = true;
251 	opts->auxtrace_snapshot_size = snapshot_size;
252 
253 	ptr->snapshot_size = snapshot_size;
254 
255 	return 0;
256 }
257 
258 struct perf_event_attr *
259 intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
260 {
261 	struct perf_event_attr *attr;
262 
263 	attr = zalloc(sizeof(struct perf_event_attr));
264 	if (!attr)
265 		return NULL;
266 
267 	attr->config = intel_pt_default_config(intel_pt_pmu);
268 
269 	intel_pt_pmu->selectable = true;
270 
271 	return attr;
272 }
273 
274 static const char *intel_pt_find_filter(struct evlist *evlist,
275 					struct perf_pmu *intel_pt_pmu)
276 {
277 	struct evsel *evsel;
278 
279 	evlist__for_each_entry(evlist, evsel) {
280 		if (evsel->core.attr.type == intel_pt_pmu->type)
281 			return evsel->filter;
282 	}
283 
284 	return NULL;
285 }
286 
287 static size_t intel_pt_filter_bytes(const char *filter)
288 {
289 	size_t len = filter ? strlen(filter) : 0;
290 
291 	return len ? roundup(len + 1, 8) : 0;
292 }
293 
294 static size_t
295 intel_pt_info_priv_size(struct auxtrace_record *itr, struct evlist *evlist)
296 {
297 	struct intel_pt_recording *ptr =
298 			container_of(itr, struct intel_pt_recording, itr);
299 	const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
300 
301 	ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
302 			 intel_pt_filter_bytes(filter);
303 
304 	return ptr->priv_size;
305 }
306 
307 static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
308 {
309 	unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
310 
311 	__get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
312 	*n = ebx;
313 	*d = eax;
314 }
315 
316 static int intel_pt_info_fill(struct auxtrace_record *itr,
317 			      struct perf_session *session,
318 			      struct perf_record_auxtrace_info *auxtrace_info,
319 			      size_t priv_size)
320 {
321 	struct intel_pt_recording *ptr =
322 			container_of(itr, struct intel_pt_recording, itr);
323 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
324 	struct perf_event_mmap_page *pc;
325 	struct perf_tsc_conversion tc = { .time_mult = 0, };
326 	bool cap_user_time_zero = false, per_cpu_mmaps;
327 	u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
328 	u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
329 	unsigned long max_non_turbo_ratio;
330 	size_t filter_str_len;
331 	const char *filter;
332 	__u64 *info;
333 	int err;
334 
335 	if (priv_size != ptr->priv_size)
336 		return -EINVAL;
337 
338 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
339 	intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
340 			     &noretcomp_bit);
341 	intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
342 	mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
343 					      "mtc_period");
344 	intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
345 
346 	intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
347 
348 	if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
349 				"%lu", &max_non_turbo_ratio) != 1)
350 		max_non_turbo_ratio = 0;
351 
352 	filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
353 	filter_str_len = filter ? strlen(filter) : 0;
354 
355 	if (!session->evlist->core.nr_mmaps)
356 		return -EINVAL;
357 
358 	pc = session->evlist->mmap[0].core.base;
359 	if (pc) {
360 		err = perf_read_tsc_conversion(pc, &tc);
361 		if (err) {
362 			if (err != -EOPNOTSUPP)
363 				return err;
364 		} else {
365 			cap_user_time_zero = tc.time_mult != 0;
366 		}
367 		if (!cap_user_time_zero)
368 			ui__warning("Intel Processor Trace: TSC not available\n");
369 	}
370 
371 	per_cpu_mmaps = !perf_cpu_map__empty(session->evlist->core.cpus);
372 
373 	auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
374 	auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
375 	auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
376 	auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
377 	auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
378 	auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
379 	auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
380 	auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
381 	auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
382 	auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
383 	auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
384 	auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
385 	auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
386 	auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
387 	auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
388 	auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
389 	auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
390 	auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
391 
392 	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
393 
394 	if (filter_str_len) {
395 		size_t len = intel_pt_filter_bytes(filter);
396 
397 		strncpy((char *)info, filter, len);
398 		info += len >> 3;
399 	}
400 
401 	return 0;
402 }
403 
404 static int intel_pt_track_switches(struct evlist *evlist)
405 {
406 	const char *sched_switch = "sched:sched_switch";
407 	struct evsel *evsel;
408 	int err;
409 
410 	if (!perf_evlist__can_select_event(evlist, sched_switch))
411 		return -EPERM;
412 
413 	err = parse_events(evlist, sched_switch, NULL);
414 	if (err) {
415 		pr_debug2("%s: failed to parse %s, error %d\n",
416 			  __func__, sched_switch, err);
417 		return err;
418 	}
419 
420 	evsel = evlist__last(evlist);
421 
422 	perf_evsel__set_sample_bit(evsel, CPU);
423 	perf_evsel__set_sample_bit(evsel, TIME);
424 
425 	evsel->core.system_wide = true;
426 	evsel->no_aux_samples = true;
427 	evsel->immediate = true;
428 
429 	return 0;
430 }
431 
432 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
433 {
434 	unsigned int val, last = 0, state = 1;
435 	int p = 0;
436 
437 	str[0] = '\0';
438 
439 	for (val = 0; val <= 64; val++, valid >>= 1) {
440 		if (valid & 1) {
441 			last = val;
442 			switch (state) {
443 			case 0:
444 				p += scnprintf(str + p, len - p, ",");
445 				/* Fall through */
446 			case 1:
447 				p += scnprintf(str + p, len - p, "%u", val);
448 				state = 2;
449 				break;
450 			case 2:
451 				state = 3;
452 				break;
453 			case 3:
454 				state = 4;
455 				break;
456 			default:
457 				break;
458 			}
459 		} else {
460 			switch (state) {
461 			case 3:
462 				p += scnprintf(str + p, len - p, ",%u", last);
463 				state = 0;
464 				break;
465 			case 4:
466 				p += scnprintf(str + p, len - p, "-%u", last);
467 				state = 0;
468 				break;
469 			default:
470 				break;
471 			}
472 			if (state != 1)
473 				state = 0;
474 		}
475 	}
476 }
477 
478 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
479 				    const char *caps, const char *name,
480 				    const char *supported, u64 config)
481 {
482 	char valid_str[256];
483 	unsigned int shift;
484 	unsigned long long valid;
485 	u64 bits;
486 	int ok;
487 
488 	if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
489 		valid = 0;
490 
491 	if (supported &&
492 	    perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
493 		valid = 0;
494 
495 	valid |= 1;
496 
497 	bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
498 
499 	config &= bits;
500 
501 	for (shift = 0; bits && !(bits & 1); shift++)
502 		bits >>= 1;
503 
504 	config >>= shift;
505 
506 	if (config > 63)
507 		goto out_err;
508 
509 	if (valid & (1 << config))
510 		return 0;
511 out_err:
512 	intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
513 	pr_err("Invalid %s for %s. Valid values are: %s\n",
514 	       name, INTEL_PT_PMU_NAME, valid_str);
515 	return -EINVAL;
516 }
517 
518 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
519 				    struct evsel *evsel)
520 {
521 	int err;
522 	char c;
523 
524 	if (!evsel)
525 		return 0;
526 
527 	/*
528 	 * If supported, force pass-through config term (pt=1) even if user
529 	 * sets pt=0, which avoids senseless kernel errors.
530 	 */
531 	if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
532 	    !(evsel->core.attr.config & 1)) {
533 		pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
534 		evsel->core.attr.config |= 1;
535 	}
536 
537 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
538 				       "cyc_thresh", "caps/psb_cyc",
539 				       evsel->core.attr.config);
540 	if (err)
541 		return err;
542 
543 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
544 				       "mtc_period", "caps/mtc",
545 				       evsel->core.attr.config);
546 	if (err)
547 		return err;
548 
549 	return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
550 					"psb_period", "caps/psb_cyc",
551 					evsel->core.attr.config);
552 }
553 
554 /*
555  * Currently, there is not enough information to disambiguate different PEBS
556  * events, so only allow one.
557  */
558 static bool intel_pt_too_many_aux_output(struct evlist *evlist)
559 {
560 	struct evsel *evsel;
561 	int aux_output_cnt = 0;
562 
563 	evlist__for_each_entry(evlist, evsel)
564 		aux_output_cnt += !!evsel->core.attr.aux_output;
565 
566 	if (aux_output_cnt > 1) {
567 		pr_err(INTEL_PT_PMU_NAME " supports at most one event with aux-output\n");
568 		return true;
569 	}
570 
571 	return false;
572 }
573 
574 static int intel_pt_recording_options(struct auxtrace_record *itr,
575 				      struct evlist *evlist,
576 				      struct record_opts *opts)
577 {
578 	struct intel_pt_recording *ptr =
579 			container_of(itr, struct intel_pt_recording, itr);
580 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
581 	bool have_timing_info, need_immediate = false;
582 	struct evsel *evsel, *intel_pt_evsel = NULL;
583 	const struct perf_cpu_map *cpus = evlist->core.cpus;
584 	bool privileged = perf_event_paranoid_check(-1);
585 	u64 tsc_bit;
586 	int err;
587 
588 	ptr->evlist = evlist;
589 	ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
590 
591 	evlist__for_each_entry(evlist, evsel) {
592 		if (evsel->core.attr.type == intel_pt_pmu->type) {
593 			if (intel_pt_evsel) {
594 				pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
595 				return -EINVAL;
596 			}
597 			evsel->core.attr.freq = 0;
598 			evsel->core.attr.sample_period = 1;
599 			intel_pt_evsel = evsel;
600 			opts->full_auxtrace = true;
601 		}
602 	}
603 
604 	if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
605 		pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
606 		return -EINVAL;
607 	}
608 
609 	if (opts->use_clockid) {
610 		pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
611 		return -EINVAL;
612 	}
613 
614 	if (intel_pt_too_many_aux_output(evlist))
615 		return -EINVAL;
616 
617 	if (!opts->full_auxtrace)
618 		return 0;
619 
620 	err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
621 	if (err)
622 		return err;
623 
624 	/* Set default sizes for snapshot mode */
625 	if (opts->auxtrace_snapshot_mode) {
626 		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
627 
628 		if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
629 			if (privileged) {
630 				opts->auxtrace_mmap_pages = MiB(4) / page_size;
631 			} else {
632 				opts->auxtrace_mmap_pages = KiB(128) / page_size;
633 				if (opts->mmap_pages == UINT_MAX)
634 					opts->mmap_pages = KiB(256) / page_size;
635 			}
636 		} else if (!opts->auxtrace_mmap_pages && !privileged &&
637 			   opts->mmap_pages == UINT_MAX) {
638 			opts->mmap_pages = KiB(256) / page_size;
639 		}
640 		if (!opts->auxtrace_snapshot_size)
641 			opts->auxtrace_snapshot_size =
642 				opts->auxtrace_mmap_pages * (size_t)page_size;
643 		if (!opts->auxtrace_mmap_pages) {
644 			size_t sz = opts->auxtrace_snapshot_size;
645 
646 			sz = round_up(sz, page_size) / page_size;
647 			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
648 		}
649 		if (opts->auxtrace_snapshot_size >
650 				opts->auxtrace_mmap_pages * (size_t)page_size) {
651 			pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
652 			       opts->auxtrace_snapshot_size,
653 			       opts->auxtrace_mmap_pages * (size_t)page_size);
654 			return -EINVAL;
655 		}
656 		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
657 			pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
658 			return -EINVAL;
659 		}
660 		pr_debug2("Intel PT snapshot size: %zu\n",
661 			  opts->auxtrace_snapshot_size);
662 		if (psb_period &&
663 		    opts->auxtrace_snapshot_size <= psb_period +
664 						  INTEL_PT_PSB_PERIOD_NEAR)
665 			ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
666 				    opts->auxtrace_snapshot_size, psb_period);
667 	}
668 
669 	/* Set default sizes for full trace mode */
670 	if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
671 		if (privileged) {
672 			opts->auxtrace_mmap_pages = MiB(4) / page_size;
673 		} else {
674 			opts->auxtrace_mmap_pages = KiB(128) / page_size;
675 			if (opts->mmap_pages == UINT_MAX)
676 				opts->mmap_pages = KiB(256) / page_size;
677 		}
678 	}
679 
680 	/* Validate auxtrace_mmap_pages */
681 	if (opts->auxtrace_mmap_pages) {
682 		size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
683 		size_t min_sz;
684 
685 		if (opts->auxtrace_snapshot_mode)
686 			min_sz = KiB(4);
687 		else
688 			min_sz = KiB(8);
689 
690 		if (sz < min_sz || !is_power_of_2(sz)) {
691 			pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
692 			       min_sz / 1024);
693 			return -EINVAL;
694 		}
695 	}
696 
697 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
698 
699 	if (opts->full_auxtrace && (intel_pt_evsel->core.attr.config & tsc_bit))
700 		have_timing_info = true;
701 	else
702 		have_timing_info = false;
703 
704 	/*
705 	 * Per-cpu recording needs sched_switch events to distinguish different
706 	 * threads.
707 	 */
708 	if (have_timing_info && !perf_cpu_map__empty(cpus)) {
709 		if (perf_can_record_switch_events()) {
710 			bool cpu_wide = !target__none(&opts->target) &&
711 					!target__has_task(&opts->target);
712 
713 			if (!cpu_wide && perf_can_record_cpu_wide()) {
714 				struct evsel *switch_evsel;
715 
716 				err = parse_events(evlist, "dummy:u", NULL);
717 				if (err)
718 					return err;
719 
720 				switch_evsel = evlist__last(evlist);
721 
722 				switch_evsel->core.attr.freq = 0;
723 				switch_evsel->core.attr.sample_period = 1;
724 				switch_evsel->core.attr.context_switch = 1;
725 
726 				switch_evsel->core.system_wide = true;
727 				switch_evsel->no_aux_samples = true;
728 				switch_evsel->immediate = true;
729 
730 				perf_evsel__set_sample_bit(switch_evsel, TID);
731 				perf_evsel__set_sample_bit(switch_evsel, TIME);
732 				perf_evsel__set_sample_bit(switch_evsel, CPU);
733 				perf_evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
734 
735 				opts->record_switch_events = false;
736 				ptr->have_sched_switch = 3;
737 			} else {
738 				opts->record_switch_events = true;
739 				need_immediate = true;
740 				if (cpu_wide)
741 					ptr->have_sched_switch = 3;
742 				else
743 					ptr->have_sched_switch = 2;
744 			}
745 		} else {
746 			err = intel_pt_track_switches(evlist);
747 			if (err == -EPERM)
748 				pr_debug2("Unable to select sched:sched_switch\n");
749 			else if (err)
750 				return err;
751 			else
752 				ptr->have_sched_switch = 1;
753 		}
754 	}
755 
756 	if (intel_pt_evsel) {
757 		/*
758 		 * To obtain the auxtrace buffer file descriptor, the auxtrace
759 		 * event must come first.
760 		 */
761 		perf_evlist__to_front(evlist, intel_pt_evsel);
762 		/*
763 		 * In the case of per-cpu mmaps, we need the CPU on the
764 		 * AUX event.
765 		 */
766 		if (!perf_cpu_map__empty(cpus))
767 			perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
768 	}
769 
770 	/* Add dummy event to keep tracking */
771 	if (opts->full_auxtrace) {
772 		struct evsel *tracking_evsel;
773 
774 		err = parse_events(evlist, "dummy:u", NULL);
775 		if (err)
776 			return err;
777 
778 		tracking_evsel = evlist__last(evlist);
779 
780 		perf_evlist__set_tracking_event(evlist, tracking_evsel);
781 
782 		tracking_evsel->core.attr.freq = 0;
783 		tracking_evsel->core.attr.sample_period = 1;
784 
785 		tracking_evsel->no_aux_samples = true;
786 		if (need_immediate)
787 			tracking_evsel->immediate = true;
788 
789 		/* In per-cpu case, always need the time of mmap events etc */
790 		if (!perf_cpu_map__empty(cpus)) {
791 			perf_evsel__set_sample_bit(tracking_evsel, TIME);
792 			/* And the CPU for switch events */
793 			perf_evsel__set_sample_bit(tracking_evsel, CPU);
794 		}
795 		perf_evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
796 	}
797 
798 	/*
799 	 * Warn the user when we do not have enough information to decode i.e.
800 	 * per-cpu with no sched_switch (except workload-only).
801 	 */
802 	if (!ptr->have_sched_switch && !perf_cpu_map__empty(cpus) &&
803 	    !target__none(&opts->target))
804 		ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
805 
806 	return 0;
807 }
808 
809 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
810 {
811 	struct intel_pt_recording *ptr =
812 			container_of(itr, struct intel_pt_recording, itr);
813 	struct evsel *evsel;
814 
815 	evlist__for_each_entry(ptr->evlist, evsel) {
816 		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
817 			return evsel__disable(evsel);
818 	}
819 	return -EINVAL;
820 }
821 
822 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
823 {
824 	struct intel_pt_recording *ptr =
825 			container_of(itr, struct intel_pt_recording, itr);
826 	struct evsel *evsel;
827 
828 	evlist__for_each_entry(ptr->evlist, evsel) {
829 		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
830 			return evsel__enable(evsel);
831 	}
832 	return -EINVAL;
833 }
834 
835 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
836 {
837 	const size_t sz = sizeof(struct intel_pt_snapshot_ref);
838 	int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
839 	struct intel_pt_snapshot_ref *refs;
840 
841 	if (!new_cnt)
842 		new_cnt = 16;
843 
844 	while (new_cnt <= idx)
845 		new_cnt *= 2;
846 
847 	refs = calloc(new_cnt, sz);
848 	if (!refs)
849 		return -ENOMEM;
850 
851 	memcpy(refs, ptr->snapshot_refs, cnt * sz);
852 
853 	ptr->snapshot_refs = refs;
854 	ptr->snapshot_ref_cnt = new_cnt;
855 
856 	return 0;
857 }
858 
859 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
860 {
861 	int i;
862 
863 	for (i = 0; i < ptr->snapshot_ref_cnt; i++)
864 		zfree(&ptr->snapshot_refs[i].ref_buf);
865 	zfree(&ptr->snapshot_refs);
866 }
867 
868 static void intel_pt_recording_free(struct auxtrace_record *itr)
869 {
870 	struct intel_pt_recording *ptr =
871 			container_of(itr, struct intel_pt_recording, itr);
872 
873 	intel_pt_free_snapshot_refs(ptr);
874 	free(ptr);
875 }
876 
877 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
878 				       size_t snapshot_buf_size)
879 {
880 	size_t ref_buf_size = ptr->snapshot_ref_buf_size;
881 	void *ref_buf;
882 
883 	ref_buf = zalloc(ref_buf_size);
884 	if (!ref_buf)
885 		return -ENOMEM;
886 
887 	ptr->snapshot_refs[idx].ref_buf = ref_buf;
888 	ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
889 
890 	return 0;
891 }
892 
893 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
894 					     size_t snapshot_buf_size)
895 {
896 	const size_t max_size = 256 * 1024;
897 	size_t buf_size = 0, psb_period;
898 
899 	if (ptr->snapshot_size <= 64 * 1024)
900 		return 0;
901 
902 	psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
903 	if (psb_period)
904 		buf_size = psb_period * 2;
905 
906 	if (!buf_size || buf_size > max_size)
907 		buf_size = max_size;
908 
909 	if (buf_size >= snapshot_buf_size)
910 		return 0;
911 
912 	if (buf_size >= ptr->snapshot_size / 2)
913 		return 0;
914 
915 	return buf_size;
916 }
917 
918 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
919 				  size_t snapshot_buf_size)
920 {
921 	if (ptr->snapshot_init_done)
922 		return 0;
923 
924 	ptr->snapshot_init_done = true;
925 
926 	ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
927 							snapshot_buf_size);
928 
929 	return 0;
930 }
931 
932 /**
933  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
934  * @buf1: first buffer
935  * @compare_size: number of bytes to compare
936  * @buf2: second buffer (a circular buffer)
937  * @offs2: offset in second buffer
938  * @buf2_size: size of second buffer
939  *
940  * The comparison allows for the possibility that the bytes to compare in the
941  * circular buffer are not contiguous.  It is assumed that @compare_size <=
942  * @buf2_size.  This function returns %false if the bytes are identical, %true
943  * otherwise.
944  */
945 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
946 				     void *buf2, size_t offs2, size_t buf2_size)
947 {
948 	size_t end2 = offs2 + compare_size, part_size;
949 
950 	if (end2 <= buf2_size)
951 		return memcmp(buf1, buf2 + offs2, compare_size);
952 
953 	part_size = end2 - buf2_size;
954 	if (memcmp(buf1, buf2 + offs2, part_size))
955 		return true;
956 
957 	compare_size -= part_size;
958 
959 	return memcmp(buf1 + part_size, buf2, compare_size);
960 }
961 
962 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
963 				 size_t ref_size, size_t buf_size,
964 				 void *data, size_t head)
965 {
966 	size_t ref_end = ref_offset + ref_size;
967 
968 	if (ref_end > buf_size) {
969 		if (head > ref_offset || head < ref_end - buf_size)
970 			return true;
971 	} else if (head > ref_offset && head < ref_end) {
972 		return true;
973 	}
974 
975 	return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
976 					buf_size);
977 }
978 
979 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
980 			      void *data, size_t head)
981 {
982 	if (head >= ref_size) {
983 		memcpy(ref_buf, data + head - ref_size, ref_size);
984 	} else {
985 		memcpy(ref_buf, data, head);
986 		ref_size -= head;
987 		memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
988 	}
989 }
990 
991 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
992 			     struct auxtrace_mmap *mm, unsigned char *data,
993 			     u64 head)
994 {
995 	struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
996 	bool wrapped;
997 
998 	wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
999 				       ptr->snapshot_ref_buf_size, mm->len,
1000 				       data, head);
1001 
1002 	intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
1003 			  data, head);
1004 
1005 	return wrapped;
1006 }
1007 
1008 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
1009 {
1010 	int i, a, b;
1011 
1012 	b = buf_size >> 3;
1013 	a = b - 512;
1014 	if (a < 0)
1015 		a = 0;
1016 
1017 	for (i = a; i < b; i++) {
1018 		if (data[i])
1019 			return true;
1020 	}
1021 
1022 	return false;
1023 }
1024 
1025 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1026 				  struct auxtrace_mmap *mm, unsigned char *data,
1027 				  u64 *head, u64 *old)
1028 {
1029 	struct intel_pt_recording *ptr =
1030 			container_of(itr, struct intel_pt_recording, itr);
1031 	bool wrapped;
1032 	int err;
1033 
1034 	pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1035 		  __func__, idx, (size_t)*old, (size_t)*head);
1036 
1037 	err = intel_pt_snapshot_init(ptr, mm->len);
1038 	if (err)
1039 		goto out_err;
1040 
1041 	if (idx >= ptr->snapshot_ref_cnt) {
1042 		err = intel_pt_alloc_snapshot_refs(ptr, idx);
1043 		if (err)
1044 			goto out_err;
1045 	}
1046 
1047 	if (ptr->snapshot_ref_buf_size) {
1048 		if (!ptr->snapshot_refs[idx].ref_buf) {
1049 			err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1050 			if (err)
1051 				goto out_err;
1052 		}
1053 		wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1054 	} else {
1055 		wrapped = ptr->snapshot_refs[idx].wrapped;
1056 		if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1057 			ptr->snapshot_refs[idx].wrapped = true;
1058 			wrapped = true;
1059 		}
1060 	}
1061 
1062 	/*
1063 	 * In full trace mode 'head' continually increases.  However in snapshot
1064 	 * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1065 	 * are adjusted to match the full trace case which expects that 'old' is
1066 	 * always less than 'head'.
1067 	 */
1068 	if (wrapped) {
1069 		*old = *head;
1070 		*head += mm->len;
1071 	} else {
1072 		if (mm->mask)
1073 			*old &= mm->mask;
1074 		else
1075 			*old %= mm->len;
1076 		if (*old > *head)
1077 			*head += mm->len;
1078 	}
1079 
1080 	pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1081 		  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1082 
1083 	return 0;
1084 
1085 out_err:
1086 	pr_err("%s: failed, error %d\n", __func__, err);
1087 	return err;
1088 }
1089 
1090 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1091 {
1092 	return rdtsc();
1093 }
1094 
1095 static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1096 {
1097 	struct intel_pt_recording *ptr =
1098 			container_of(itr, struct intel_pt_recording, itr);
1099 	struct evsel *evsel;
1100 
1101 	evlist__for_each_entry(ptr->evlist, evsel) {
1102 		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
1103 			return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1104 							     idx);
1105 	}
1106 	return -EINVAL;
1107 }
1108 
1109 struct auxtrace_record *intel_pt_recording_init(int *err)
1110 {
1111 	struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1112 	struct intel_pt_recording *ptr;
1113 
1114 	if (!intel_pt_pmu)
1115 		return NULL;
1116 
1117 	if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1118 		*err = -errno;
1119 		return NULL;
1120 	}
1121 
1122 	ptr = zalloc(sizeof(struct intel_pt_recording));
1123 	if (!ptr) {
1124 		*err = -ENOMEM;
1125 		return NULL;
1126 	}
1127 
1128 	ptr->intel_pt_pmu = intel_pt_pmu;
1129 	ptr->itr.recording_options = intel_pt_recording_options;
1130 	ptr->itr.info_priv_size = intel_pt_info_priv_size;
1131 	ptr->itr.info_fill = intel_pt_info_fill;
1132 	ptr->itr.free = intel_pt_recording_free;
1133 	ptr->itr.snapshot_start = intel_pt_snapshot_start;
1134 	ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1135 	ptr->itr.find_snapshot = intel_pt_find_snapshot;
1136 	ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1137 	ptr->itr.reference = intel_pt_reference;
1138 	ptr->itr.read_finish = intel_pt_read_finish;
1139 	return &ptr->itr;
1140 }
1141