xref: /openbmc/linux/tools/perf/arch/x86/util/intel-pt.c (revision b830f94f)
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 "../../perf.h"
17 #include "../../util/session.h"
18 #include "../../util/event.h"
19 #include "../../util/evlist.h"
20 #include "../../util/evsel.h"
21 #include "../../util/cpumap.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/tsc.h"
28 #include "../../util/intel-pt.h"
29 
30 #define KiB(x) ((x) * 1024)
31 #define MiB(x) ((x) * 1024 * 1024)
32 #define KiB_MASK(x) (KiB(x) - 1)
33 #define MiB_MASK(x) (MiB(x) - 1)
34 
35 #define INTEL_PT_PSB_PERIOD_NEAR	256
36 
37 struct intel_pt_snapshot_ref {
38 	void *ref_buf;
39 	size_t ref_offset;
40 	bool wrapped;
41 };
42 
43 struct intel_pt_recording {
44 	struct auxtrace_record		itr;
45 	struct perf_pmu			*intel_pt_pmu;
46 	int				have_sched_switch;
47 	struct perf_evlist		*evlist;
48 	bool				snapshot_mode;
49 	bool				snapshot_init_done;
50 	size_t				snapshot_size;
51 	size_t				snapshot_ref_buf_size;
52 	int				snapshot_ref_cnt;
53 	struct intel_pt_snapshot_ref	*snapshot_refs;
54 	size_t				priv_size;
55 };
56 
57 static int intel_pt_parse_terms_with_default(struct list_head *formats,
58 					     const char *str,
59 					     u64 *config)
60 {
61 	struct list_head *terms;
62 	struct perf_event_attr attr = { .size = 0, };
63 	int err;
64 
65 	terms = malloc(sizeof(struct list_head));
66 	if (!terms)
67 		return -ENOMEM;
68 
69 	INIT_LIST_HEAD(terms);
70 
71 	err = parse_events_terms(terms, str);
72 	if (err)
73 		goto out_free;
74 
75 	attr.config = *config;
76 	err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
77 	if (err)
78 		goto out_free;
79 
80 	*config = attr.config;
81 out_free:
82 	parse_events_terms__delete(terms);
83 	return err;
84 }
85 
86 static int intel_pt_parse_terms(struct list_head *formats, const char *str,
87 				u64 *config)
88 {
89 	*config = 0;
90 	return intel_pt_parse_terms_with_default(formats, str, config);
91 }
92 
93 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
94 {
95 	const u64 top_bit = 1ULL << 63;
96 	u64 res = 0;
97 	int i;
98 
99 	for (i = 0; i < 64; i++) {
100 		if (mask & top_bit) {
101 			res <<= 1;
102 			if (bits & top_bit)
103 				res |= 1;
104 		}
105 		mask <<= 1;
106 		bits <<= 1;
107 	}
108 
109 	return res;
110 }
111 
112 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
113 				struct perf_evlist *evlist, u64 *res)
114 {
115 	struct perf_evsel *evsel;
116 	u64 mask;
117 
118 	*res = 0;
119 
120 	mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
121 	if (!mask)
122 		return -EINVAL;
123 
124 	evlist__for_each_entry(evlist, evsel) {
125 		if (evsel->attr.type == intel_pt_pmu->type) {
126 			*res = intel_pt_masked_bits(mask, evsel->attr.config);
127 			return 0;
128 		}
129 	}
130 
131 	return -EINVAL;
132 }
133 
134 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
135 				  struct perf_evlist *evlist)
136 {
137 	u64 val;
138 	int err, topa_multiple_entries;
139 	size_t psb_period;
140 
141 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
142 				"%d", &topa_multiple_entries) != 1)
143 		topa_multiple_entries = 0;
144 
145 	/*
146 	 * Use caps/topa_multiple_entries to indicate early hardware that had
147 	 * extra frequent PSBs.
148 	 */
149 	if (!topa_multiple_entries) {
150 		psb_period = 256;
151 		goto out;
152 	}
153 
154 	err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
155 	if (err)
156 		val = 0;
157 
158 	psb_period = 1 << (val + 11);
159 out:
160 	pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
161 	return psb_period;
162 }
163 
164 static int intel_pt_pick_bit(int bits, int target)
165 {
166 	int pos, pick = -1;
167 
168 	for (pos = 0; bits; bits >>= 1, pos++) {
169 		if (bits & 1) {
170 			if (pos <= target || pick < 0)
171 				pick = pos;
172 			if (pos >= target)
173 				break;
174 		}
175 	}
176 
177 	return pick;
178 }
179 
180 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
181 {
182 	char buf[256];
183 	int mtc, mtc_periods = 0, mtc_period;
184 	int psb_cyc, psb_periods, psb_period;
185 	int pos = 0;
186 	u64 config;
187 	char c;
188 
189 	pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
190 
191 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
192 				&mtc) != 1)
193 		mtc = 1;
194 
195 	if (mtc) {
196 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
197 					&mtc_periods) != 1)
198 			mtc_periods = 0;
199 		if (mtc_periods) {
200 			mtc_period = intel_pt_pick_bit(mtc_periods, 3);
201 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
202 					 ",mtc,mtc_period=%d", mtc_period);
203 		}
204 	}
205 
206 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
207 				&psb_cyc) != 1)
208 		psb_cyc = 1;
209 
210 	if (psb_cyc && mtc_periods) {
211 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
212 					&psb_periods) != 1)
213 			psb_periods = 0;
214 		if (psb_periods) {
215 			psb_period = intel_pt_pick_bit(psb_periods, 3);
216 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
217 					 ",psb_period=%d", psb_period);
218 		}
219 	}
220 
221 	if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
222 	    perf_pmu__scan_file(intel_pt_pmu, "format/branch", "%c", &c) == 1)
223 		pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
224 
225 	pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
226 
227 	intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
228 
229 	return config;
230 }
231 
232 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
233 					   struct record_opts *opts,
234 					   const char *str)
235 {
236 	struct intel_pt_recording *ptr =
237 			container_of(itr, struct intel_pt_recording, itr);
238 	unsigned long long snapshot_size = 0;
239 	char *endptr;
240 
241 	if (str) {
242 		snapshot_size = strtoull(str, &endptr, 0);
243 		if (*endptr || snapshot_size > SIZE_MAX)
244 			return -1;
245 	}
246 
247 	opts->auxtrace_snapshot_mode = true;
248 	opts->auxtrace_snapshot_size = snapshot_size;
249 
250 	ptr->snapshot_size = snapshot_size;
251 
252 	return 0;
253 }
254 
255 struct perf_event_attr *
256 intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
257 {
258 	struct perf_event_attr *attr;
259 
260 	attr = zalloc(sizeof(struct perf_event_attr));
261 	if (!attr)
262 		return NULL;
263 
264 	attr->config = intel_pt_default_config(intel_pt_pmu);
265 
266 	intel_pt_pmu->selectable = true;
267 
268 	return attr;
269 }
270 
271 static const char *intel_pt_find_filter(struct perf_evlist *evlist,
272 					struct perf_pmu *intel_pt_pmu)
273 {
274 	struct perf_evsel *evsel;
275 
276 	evlist__for_each_entry(evlist, evsel) {
277 		if (evsel->attr.type == intel_pt_pmu->type)
278 			return evsel->filter;
279 	}
280 
281 	return NULL;
282 }
283 
284 static size_t intel_pt_filter_bytes(const char *filter)
285 {
286 	size_t len = filter ? strlen(filter) : 0;
287 
288 	return len ? roundup(len + 1, 8) : 0;
289 }
290 
291 static size_t
292 intel_pt_info_priv_size(struct auxtrace_record *itr, struct perf_evlist *evlist)
293 {
294 	struct intel_pt_recording *ptr =
295 			container_of(itr, struct intel_pt_recording, itr);
296 	const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
297 
298 	ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
299 			 intel_pt_filter_bytes(filter);
300 
301 	return ptr->priv_size;
302 }
303 
304 static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
305 {
306 	unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
307 
308 	__get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
309 	*n = ebx;
310 	*d = eax;
311 }
312 
313 static int intel_pt_info_fill(struct auxtrace_record *itr,
314 			      struct perf_session *session,
315 			      struct auxtrace_info_event *auxtrace_info,
316 			      size_t priv_size)
317 {
318 	struct intel_pt_recording *ptr =
319 			container_of(itr, struct intel_pt_recording, itr);
320 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
321 	struct perf_event_mmap_page *pc;
322 	struct perf_tsc_conversion tc = { .time_mult = 0, };
323 	bool cap_user_time_zero = false, per_cpu_mmaps;
324 	u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
325 	u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
326 	unsigned long max_non_turbo_ratio;
327 	size_t filter_str_len;
328 	const char *filter;
329 	u64 *info;
330 	int err;
331 
332 	if (priv_size != ptr->priv_size)
333 		return -EINVAL;
334 
335 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
336 	intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
337 			     &noretcomp_bit);
338 	intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
339 	mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
340 					      "mtc_period");
341 	intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
342 
343 	intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
344 
345 	if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
346 				"%lu", &max_non_turbo_ratio) != 1)
347 		max_non_turbo_ratio = 0;
348 
349 	filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
350 	filter_str_len = filter ? strlen(filter) : 0;
351 
352 	if (!session->evlist->nr_mmaps)
353 		return -EINVAL;
354 
355 	pc = session->evlist->mmap[0].base;
356 	if (pc) {
357 		err = perf_read_tsc_conversion(pc, &tc);
358 		if (err) {
359 			if (err != -EOPNOTSUPP)
360 				return err;
361 		} else {
362 			cap_user_time_zero = tc.time_mult != 0;
363 		}
364 		if (!cap_user_time_zero)
365 			ui__warning("Intel Processor Trace: TSC not available\n");
366 	}
367 
368 	per_cpu_mmaps = !cpu_map__empty(session->evlist->cpus);
369 
370 	auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
371 	auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
372 	auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
373 	auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
374 	auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
375 	auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
376 	auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
377 	auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
378 	auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
379 	auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
380 	auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
381 	auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
382 	auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
383 	auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
384 	auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
385 	auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
386 	auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
387 	auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
388 
389 	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
390 
391 	if (filter_str_len) {
392 		size_t len = intel_pt_filter_bytes(filter);
393 
394 		strncpy((char *)info, filter, len);
395 		info += len >> 3;
396 	}
397 
398 	return 0;
399 }
400 
401 static int intel_pt_track_switches(struct perf_evlist *evlist)
402 {
403 	const char *sched_switch = "sched:sched_switch";
404 	struct perf_evsel *evsel;
405 	int err;
406 
407 	if (!perf_evlist__can_select_event(evlist, sched_switch))
408 		return -EPERM;
409 
410 	err = parse_events(evlist, sched_switch, NULL);
411 	if (err) {
412 		pr_debug2("%s: failed to parse %s, error %d\n",
413 			  __func__, sched_switch, err);
414 		return err;
415 	}
416 
417 	evsel = perf_evlist__last(evlist);
418 
419 	perf_evsel__set_sample_bit(evsel, CPU);
420 	perf_evsel__set_sample_bit(evsel, TIME);
421 
422 	evsel->system_wide = true;
423 	evsel->no_aux_samples = true;
424 	evsel->immediate = true;
425 
426 	return 0;
427 }
428 
429 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
430 {
431 	unsigned int val, last = 0, state = 1;
432 	int p = 0;
433 
434 	str[0] = '\0';
435 
436 	for (val = 0; val <= 64; val++, valid >>= 1) {
437 		if (valid & 1) {
438 			last = val;
439 			switch (state) {
440 			case 0:
441 				p += scnprintf(str + p, len - p, ",");
442 				/* Fall through */
443 			case 1:
444 				p += scnprintf(str + p, len - p, "%u", val);
445 				state = 2;
446 				break;
447 			case 2:
448 				state = 3;
449 				break;
450 			case 3:
451 				state = 4;
452 				break;
453 			default:
454 				break;
455 			}
456 		} else {
457 			switch (state) {
458 			case 3:
459 				p += scnprintf(str + p, len - p, ",%u", last);
460 				state = 0;
461 				break;
462 			case 4:
463 				p += scnprintf(str + p, len - p, "-%u", last);
464 				state = 0;
465 				break;
466 			default:
467 				break;
468 			}
469 			if (state != 1)
470 				state = 0;
471 		}
472 	}
473 }
474 
475 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
476 				    const char *caps, const char *name,
477 				    const char *supported, u64 config)
478 {
479 	char valid_str[256];
480 	unsigned int shift;
481 	unsigned long long valid;
482 	u64 bits;
483 	int ok;
484 
485 	if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
486 		valid = 0;
487 
488 	if (supported &&
489 	    perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
490 		valid = 0;
491 
492 	valid |= 1;
493 
494 	bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
495 
496 	config &= bits;
497 
498 	for (shift = 0; bits && !(bits & 1); shift++)
499 		bits >>= 1;
500 
501 	config >>= shift;
502 
503 	if (config > 63)
504 		goto out_err;
505 
506 	if (valid & (1 << config))
507 		return 0;
508 out_err:
509 	intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
510 	pr_err("Invalid %s for %s. Valid values are: %s\n",
511 	       name, INTEL_PT_PMU_NAME, valid_str);
512 	return -EINVAL;
513 }
514 
515 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
516 				    struct perf_evsel *evsel)
517 {
518 	int err;
519 	char c;
520 
521 	if (!evsel)
522 		return 0;
523 
524 	/*
525 	 * If supported, force pass-through config term (pt=1) even if user
526 	 * sets pt=0, which avoids senseless kernel errors.
527 	 */
528 	if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
529 	    !(evsel->attr.config & 1)) {
530 		pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
531 		evsel->attr.config |= 1;
532 	}
533 
534 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
535 				       "cyc_thresh", "caps/psb_cyc",
536 				       evsel->attr.config);
537 	if (err)
538 		return err;
539 
540 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
541 				       "mtc_period", "caps/mtc",
542 				       evsel->attr.config);
543 	if (err)
544 		return err;
545 
546 	return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
547 					"psb_period", "caps/psb_cyc",
548 					evsel->attr.config);
549 }
550 
551 static int intel_pt_recording_options(struct auxtrace_record *itr,
552 				      struct perf_evlist *evlist,
553 				      struct record_opts *opts)
554 {
555 	struct intel_pt_recording *ptr =
556 			container_of(itr, struct intel_pt_recording, itr);
557 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
558 	bool have_timing_info, need_immediate = false;
559 	struct perf_evsel *evsel, *intel_pt_evsel = NULL;
560 	const struct cpu_map *cpus = evlist->cpus;
561 	bool privileged = geteuid() == 0 || perf_event_paranoid() < 0;
562 	u64 tsc_bit;
563 	int err;
564 
565 	ptr->evlist = evlist;
566 	ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
567 
568 	evlist__for_each_entry(evlist, evsel) {
569 		if (evsel->attr.type == intel_pt_pmu->type) {
570 			if (intel_pt_evsel) {
571 				pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
572 				return -EINVAL;
573 			}
574 			evsel->attr.freq = 0;
575 			evsel->attr.sample_period = 1;
576 			intel_pt_evsel = evsel;
577 			opts->full_auxtrace = true;
578 		}
579 	}
580 
581 	if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
582 		pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
583 		return -EINVAL;
584 	}
585 
586 	if (opts->use_clockid) {
587 		pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
588 		return -EINVAL;
589 	}
590 
591 	if (!opts->full_auxtrace)
592 		return 0;
593 
594 	err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
595 	if (err)
596 		return err;
597 
598 	/* Set default sizes for snapshot mode */
599 	if (opts->auxtrace_snapshot_mode) {
600 		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
601 
602 		if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
603 			if (privileged) {
604 				opts->auxtrace_mmap_pages = MiB(4) / page_size;
605 			} else {
606 				opts->auxtrace_mmap_pages = KiB(128) / page_size;
607 				if (opts->mmap_pages == UINT_MAX)
608 					opts->mmap_pages = KiB(256) / page_size;
609 			}
610 		} else if (!opts->auxtrace_mmap_pages && !privileged &&
611 			   opts->mmap_pages == UINT_MAX) {
612 			opts->mmap_pages = KiB(256) / page_size;
613 		}
614 		if (!opts->auxtrace_snapshot_size)
615 			opts->auxtrace_snapshot_size =
616 				opts->auxtrace_mmap_pages * (size_t)page_size;
617 		if (!opts->auxtrace_mmap_pages) {
618 			size_t sz = opts->auxtrace_snapshot_size;
619 
620 			sz = round_up(sz, page_size) / page_size;
621 			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
622 		}
623 		if (opts->auxtrace_snapshot_size >
624 				opts->auxtrace_mmap_pages * (size_t)page_size) {
625 			pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
626 			       opts->auxtrace_snapshot_size,
627 			       opts->auxtrace_mmap_pages * (size_t)page_size);
628 			return -EINVAL;
629 		}
630 		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
631 			pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
632 			return -EINVAL;
633 		}
634 		pr_debug2("Intel PT snapshot size: %zu\n",
635 			  opts->auxtrace_snapshot_size);
636 		if (psb_period &&
637 		    opts->auxtrace_snapshot_size <= psb_period +
638 						  INTEL_PT_PSB_PERIOD_NEAR)
639 			ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
640 				    opts->auxtrace_snapshot_size, psb_period);
641 	}
642 
643 	/* Set default sizes for full trace mode */
644 	if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
645 		if (privileged) {
646 			opts->auxtrace_mmap_pages = MiB(4) / page_size;
647 		} else {
648 			opts->auxtrace_mmap_pages = KiB(128) / page_size;
649 			if (opts->mmap_pages == UINT_MAX)
650 				opts->mmap_pages = KiB(256) / page_size;
651 		}
652 	}
653 
654 	/* Validate auxtrace_mmap_pages */
655 	if (opts->auxtrace_mmap_pages) {
656 		size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
657 		size_t min_sz;
658 
659 		if (opts->auxtrace_snapshot_mode)
660 			min_sz = KiB(4);
661 		else
662 			min_sz = KiB(8);
663 
664 		if (sz < min_sz || !is_power_of_2(sz)) {
665 			pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
666 			       min_sz / 1024);
667 			return -EINVAL;
668 		}
669 	}
670 
671 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
672 
673 	if (opts->full_auxtrace && (intel_pt_evsel->attr.config & tsc_bit))
674 		have_timing_info = true;
675 	else
676 		have_timing_info = false;
677 
678 	/*
679 	 * Per-cpu recording needs sched_switch events to distinguish different
680 	 * threads.
681 	 */
682 	if (have_timing_info && !cpu_map__empty(cpus)) {
683 		if (perf_can_record_switch_events()) {
684 			bool cpu_wide = !target__none(&opts->target) &&
685 					!target__has_task(&opts->target);
686 
687 			if (!cpu_wide && perf_can_record_cpu_wide()) {
688 				struct perf_evsel *switch_evsel;
689 
690 				err = parse_events(evlist, "dummy:u", NULL);
691 				if (err)
692 					return err;
693 
694 				switch_evsel = perf_evlist__last(evlist);
695 
696 				switch_evsel->attr.freq = 0;
697 				switch_evsel->attr.sample_period = 1;
698 				switch_evsel->attr.context_switch = 1;
699 
700 				switch_evsel->system_wide = true;
701 				switch_evsel->no_aux_samples = true;
702 				switch_evsel->immediate = true;
703 
704 				perf_evsel__set_sample_bit(switch_evsel, TID);
705 				perf_evsel__set_sample_bit(switch_evsel, TIME);
706 				perf_evsel__set_sample_bit(switch_evsel, CPU);
707 				perf_evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
708 
709 				opts->record_switch_events = false;
710 				ptr->have_sched_switch = 3;
711 			} else {
712 				opts->record_switch_events = true;
713 				need_immediate = true;
714 				if (cpu_wide)
715 					ptr->have_sched_switch = 3;
716 				else
717 					ptr->have_sched_switch = 2;
718 			}
719 		} else {
720 			err = intel_pt_track_switches(evlist);
721 			if (err == -EPERM)
722 				pr_debug2("Unable to select sched:sched_switch\n");
723 			else if (err)
724 				return err;
725 			else
726 				ptr->have_sched_switch = 1;
727 		}
728 	}
729 
730 	if (intel_pt_evsel) {
731 		/*
732 		 * To obtain the auxtrace buffer file descriptor, the auxtrace
733 		 * event must come first.
734 		 */
735 		perf_evlist__to_front(evlist, intel_pt_evsel);
736 		/*
737 		 * In the case of per-cpu mmaps, we need the CPU on the
738 		 * AUX event.
739 		 */
740 		if (!cpu_map__empty(cpus))
741 			perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
742 	}
743 
744 	/* Add dummy event to keep tracking */
745 	if (opts->full_auxtrace) {
746 		struct perf_evsel *tracking_evsel;
747 
748 		err = parse_events(evlist, "dummy:u", NULL);
749 		if (err)
750 			return err;
751 
752 		tracking_evsel = perf_evlist__last(evlist);
753 
754 		perf_evlist__set_tracking_event(evlist, tracking_evsel);
755 
756 		tracking_evsel->attr.freq = 0;
757 		tracking_evsel->attr.sample_period = 1;
758 
759 		tracking_evsel->no_aux_samples = true;
760 		if (need_immediate)
761 			tracking_evsel->immediate = true;
762 
763 		/* In per-cpu case, always need the time of mmap events etc */
764 		if (!cpu_map__empty(cpus)) {
765 			perf_evsel__set_sample_bit(tracking_evsel, TIME);
766 			/* And the CPU for switch events */
767 			perf_evsel__set_sample_bit(tracking_evsel, CPU);
768 		}
769 		perf_evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
770 	}
771 
772 	/*
773 	 * Warn the user when we do not have enough information to decode i.e.
774 	 * per-cpu with no sched_switch (except workload-only).
775 	 */
776 	if (!ptr->have_sched_switch && !cpu_map__empty(cpus) &&
777 	    !target__none(&opts->target))
778 		ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
779 
780 	return 0;
781 }
782 
783 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
784 {
785 	struct intel_pt_recording *ptr =
786 			container_of(itr, struct intel_pt_recording, itr);
787 	struct perf_evsel *evsel;
788 
789 	evlist__for_each_entry(ptr->evlist, evsel) {
790 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
791 			return perf_evsel__disable(evsel);
792 	}
793 	return -EINVAL;
794 }
795 
796 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
797 {
798 	struct intel_pt_recording *ptr =
799 			container_of(itr, struct intel_pt_recording, itr);
800 	struct perf_evsel *evsel;
801 
802 	evlist__for_each_entry(ptr->evlist, evsel) {
803 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
804 			return perf_evsel__enable(evsel);
805 	}
806 	return -EINVAL;
807 }
808 
809 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
810 {
811 	const size_t sz = sizeof(struct intel_pt_snapshot_ref);
812 	int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
813 	struct intel_pt_snapshot_ref *refs;
814 
815 	if (!new_cnt)
816 		new_cnt = 16;
817 
818 	while (new_cnt <= idx)
819 		new_cnt *= 2;
820 
821 	refs = calloc(new_cnt, sz);
822 	if (!refs)
823 		return -ENOMEM;
824 
825 	memcpy(refs, ptr->snapshot_refs, cnt * sz);
826 
827 	ptr->snapshot_refs = refs;
828 	ptr->snapshot_ref_cnt = new_cnt;
829 
830 	return 0;
831 }
832 
833 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
834 {
835 	int i;
836 
837 	for (i = 0; i < ptr->snapshot_ref_cnt; i++)
838 		zfree(&ptr->snapshot_refs[i].ref_buf);
839 	zfree(&ptr->snapshot_refs);
840 }
841 
842 static void intel_pt_recording_free(struct auxtrace_record *itr)
843 {
844 	struct intel_pt_recording *ptr =
845 			container_of(itr, struct intel_pt_recording, itr);
846 
847 	intel_pt_free_snapshot_refs(ptr);
848 	free(ptr);
849 }
850 
851 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
852 				       size_t snapshot_buf_size)
853 {
854 	size_t ref_buf_size = ptr->snapshot_ref_buf_size;
855 	void *ref_buf;
856 
857 	ref_buf = zalloc(ref_buf_size);
858 	if (!ref_buf)
859 		return -ENOMEM;
860 
861 	ptr->snapshot_refs[idx].ref_buf = ref_buf;
862 	ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
863 
864 	return 0;
865 }
866 
867 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
868 					     size_t snapshot_buf_size)
869 {
870 	const size_t max_size = 256 * 1024;
871 	size_t buf_size = 0, psb_period;
872 
873 	if (ptr->snapshot_size <= 64 * 1024)
874 		return 0;
875 
876 	psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
877 	if (psb_period)
878 		buf_size = psb_period * 2;
879 
880 	if (!buf_size || buf_size > max_size)
881 		buf_size = max_size;
882 
883 	if (buf_size >= snapshot_buf_size)
884 		return 0;
885 
886 	if (buf_size >= ptr->snapshot_size / 2)
887 		return 0;
888 
889 	return buf_size;
890 }
891 
892 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
893 				  size_t snapshot_buf_size)
894 {
895 	if (ptr->snapshot_init_done)
896 		return 0;
897 
898 	ptr->snapshot_init_done = true;
899 
900 	ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
901 							snapshot_buf_size);
902 
903 	return 0;
904 }
905 
906 /**
907  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
908  * @buf1: first buffer
909  * @compare_size: number of bytes to compare
910  * @buf2: second buffer (a circular buffer)
911  * @offs2: offset in second buffer
912  * @buf2_size: size of second buffer
913  *
914  * The comparison allows for the possibility that the bytes to compare in the
915  * circular buffer are not contiguous.  It is assumed that @compare_size <=
916  * @buf2_size.  This function returns %false if the bytes are identical, %true
917  * otherwise.
918  */
919 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
920 				     void *buf2, size_t offs2, size_t buf2_size)
921 {
922 	size_t end2 = offs2 + compare_size, part_size;
923 
924 	if (end2 <= buf2_size)
925 		return memcmp(buf1, buf2 + offs2, compare_size);
926 
927 	part_size = end2 - buf2_size;
928 	if (memcmp(buf1, buf2 + offs2, part_size))
929 		return true;
930 
931 	compare_size -= part_size;
932 
933 	return memcmp(buf1 + part_size, buf2, compare_size);
934 }
935 
936 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
937 				 size_t ref_size, size_t buf_size,
938 				 void *data, size_t head)
939 {
940 	size_t ref_end = ref_offset + ref_size;
941 
942 	if (ref_end > buf_size) {
943 		if (head > ref_offset || head < ref_end - buf_size)
944 			return true;
945 	} else if (head > ref_offset && head < ref_end) {
946 		return true;
947 	}
948 
949 	return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
950 					buf_size);
951 }
952 
953 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
954 			      void *data, size_t head)
955 {
956 	if (head >= ref_size) {
957 		memcpy(ref_buf, data + head - ref_size, ref_size);
958 	} else {
959 		memcpy(ref_buf, data, head);
960 		ref_size -= head;
961 		memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
962 	}
963 }
964 
965 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
966 			     struct auxtrace_mmap *mm, unsigned char *data,
967 			     u64 head)
968 {
969 	struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
970 	bool wrapped;
971 
972 	wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
973 				       ptr->snapshot_ref_buf_size, mm->len,
974 				       data, head);
975 
976 	intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
977 			  data, head);
978 
979 	return wrapped;
980 }
981 
982 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
983 {
984 	int i, a, b;
985 
986 	b = buf_size >> 3;
987 	a = b - 512;
988 	if (a < 0)
989 		a = 0;
990 
991 	for (i = a; i < b; i++) {
992 		if (data[i])
993 			return true;
994 	}
995 
996 	return false;
997 }
998 
999 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1000 				  struct auxtrace_mmap *mm, unsigned char *data,
1001 				  u64 *head, u64 *old)
1002 {
1003 	struct intel_pt_recording *ptr =
1004 			container_of(itr, struct intel_pt_recording, itr);
1005 	bool wrapped;
1006 	int err;
1007 
1008 	pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1009 		  __func__, idx, (size_t)*old, (size_t)*head);
1010 
1011 	err = intel_pt_snapshot_init(ptr, mm->len);
1012 	if (err)
1013 		goto out_err;
1014 
1015 	if (idx >= ptr->snapshot_ref_cnt) {
1016 		err = intel_pt_alloc_snapshot_refs(ptr, idx);
1017 		if (err)
1018 			goto out_err;
1019 	}
1020 
1021 	if (ptr->snapshot_ref_buf_size) {
1022 		if (!ptr->snapshot_refs[idx].ref_buf) {
1023 			err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1024 			if (err)
1025 				goto out_err;
1026 		}
1027 		wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1028 	} else {
1029 		wrapped = ptr->snapshot_refs[idx].wrapped;
1030 		if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1031 			ptr->snapshot_refs[idx].wrapped = true;
1032 			wrapped = true;
1033 		}
1034 	}
1035 
1036 	/*
1037 	 * In full trace mode 'head' continually increases.  However in snapshot
1038 	 * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1039 	 * are adjusted to match the full trace case which expects that 'old' is
1040 	 * always less than 'head'.
1041 	 */
1042 	if (wrapped) {
1043 		*old = *head;
1044 		*head += mm->len;
1045 	} else {
1046 		if (mm->mask)
1047 			*old &= mm->mask;
1048 		else
1049 			*old %= mm->len;
1050 		if (*old > *head)
1051 			*head += mm->len;
1052 	}
1053 
1054 	pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1055 		  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1056 
1057 	return 0;
1058 
1059 out_err:
1060 	pr_err("%s: failed, error %d\n", __func__, err);
1061 	return err;
1062 }
1063 
1064 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1065 {
1066 	return rdtsc();
1067 }
1068 
1069 static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1070 {
1071 	struct intel_pt_recording *ptr =
1072 			container_of(itr, struct intel_pt_recording, itr);
1073 	struct perf_evsel *evsel;
1074 
1075 	evlist__for_each_entry(ptr->evlist, evsel) {
1076 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
1077 			return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1078 							     idx);
1079 	}
1080 	return -EINVAL;
1081 }
1082 
1083 struct auxtrace_record *intel_pt_recording_init(int *err)
1084 {
1085 	struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1086 	struct intel_pt_recording *ptr;
1087 
1088 	if (!intel_pt_pmu)
1089 		return NULL;
1090 
1091 	if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1092 		*err = -errno;
1093 		return NULL;
1094 	}
1095 
1096 	ptr = zalloc(sizeof(struct intel_pt_recording));
1097 	if (!ptr) {
1098 		*err = -ENOMEM;
1099 		return NULL;
1100 	}
1101 
1102 	ptr->intel_pt_pmu = intel_pt_pmu;
1103 	ptr->itr.recording_options = intel_pt_recording_options;
1104 	ptr->itr.info_priv_size = intel_pt_info_priv_size;
1105 	ptr->itr.info_fill = intel_pt_info_fill;
1106 	ptr->itr.free = intel_pt_recording_free;
1107 	ptr->itr.snapshot_start = intel_pt_snapshot_start;
1108 	ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1109 	ptr->itr.find_snapshot = intel_pt_find_snapshot;
1110 	ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1111 	ptr->itr.reference = intel_pt_reference;
1112 	ptr->itr.read_finish = intel_pt_read_finish;
1113 	return &ptr->itr;
1114 }
1115