xref: /openbmc/linux/tools/perf/arch/x86/util/intel-pt.c (revision 6f69e2a3)
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/evsel_config.h"
21 #include "../../util/cpumap.h"
22 #include "../../util/mmap.h"
23 #include <subcmd/parse-options.h>
24 #include "../../util/parse-events.h"
25 #include "../../util/pmu.h"
26 #include "../../util/debug.h"
27 #include "../../util/auxtrace.h"
28 #include "../../util/record.h"
29 #include "../../util/target.h"
30 #include "../../util/tsc.h"
31 #include <internal/lib.h> // page_size
32 #include "../../util/intel-pt.h"
33 
34 #define KiB(x) ((x) * 1024)
35 #define MiB(x) ((x) * 1024 * 1024)
36 #define KiB_MASK(x) (KiB(x) - 1)
37 #define MiB_MASK(x) (MiB(x) - 1)
38 
39 #define INTEL_PT_PSB_PERIOD_NEAR	256
40 
41 struct intel_pt_snapshot_ref {
42 	void *ref_buf;
43 	size_t ref_offset;
44 	bool wrapped;
45 };
46 
47 struct intel_pt_recording {
48 	struct auxtrace_record		itr;
49 	struct perf_pmu			*intel_pt_pmu;
50 	int				have_sched_switch;
51 	struct evlist		*evlist;
52 	bool				snapshot_mode;
53 	bool				snapshot_init_done;
54 	size_t				snapshot_size;
55 	size_t				snapshot_ref_buf_size;
56 	int				snapshot_ref_cnt;
57 	struct intel_pt_snapshot_ref	*snapshot_refs;
58 	size_t				priv_size;
59 };
60 
61 static int intel_pt_parse_terms_with_default(struct list_head *formats,
62 					     const char *str,
63 					     u64 *config)
64 {
65 	struct list_head *terms;
66 	struct perf_event_attr attr = { .size = 0, };
67 	int err;
68 
69 	terms = malloc(sizeof(struct list_head));
70 	if (!terms)
71 		return -ENOMEM;
72 
73 	INIT_LIST_HEAD(terms);
74 
75 	err = parse_events_terms(terms, str);
76 	if (err)
77 		goto out_free;
78 
79 	attr.config = *config;
80 	err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
81 	if (err)
82 		goto out_free;
83 
84 	*config = attr.config;
85 out_free:
86 	parse_events_terms__delete(terms);
87 	return err;
88 }
89 
90 static int intel_pt_parse_terms(struct list_head *formats, const char *str,
91 				u64 *config)
92 {
93 	*config = 0;
94 	return intel_pt_parse_terms_with_default(formats, str, config);
95 }
96 
97 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
98 {
99 	const u64 top_bit = 1ULL << 63;
100 	u64 res = 0;
101 	int i;
102 
103 	for (i = 0; i < 64; i++) {
104 		if (mask & top_bit) {
105 			res <<= 1;
106 			if (bits & top_bit)
107 				res |= 1;
108 		}
109 		mask <<= 1;
110 		bits <<= 1;
111 	}
112 
113 	return res;
114 }
115 
116 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
117 				struct evlist *evlist, u64 *res)
118 {
119 	struct evsel *evsel;
120 	u64 mask;
121 
122 	*res = 0;
123 
124 	mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
125 	if (!mask)
126 		return -EINVAL;
127 
128 	evlist__for_each_entry(evlist, evsel) {
129 		if (evsel->core.attr.type == intel_pt_pmu->type) {
130 			*res = intel_pt_masked_bits(mask, evsel->core.attr.config);
131 			return 0;
132 		}
133 	}
134 
135 	return -EINVAL;
136 }
137 
138 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
139 				  struct evlist *evlist)
140 {
141 	u64 val;
142 	int err, topa_multiple_entries;
143 	size_t psb_period;
144 
145 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
146 				"%d", &topa_multiple_entries) != 1)
147 		topa_multiple_entries = 0;
148 
149 	/*
150 	 * Use caps/topa_multiple_entries to indicate early hardware that had
151 	 * extra frequent PSBs.
152 	 */
153 	if (!topa_multiple_entries) {
154 		psb_period = 256;
155 		goto out;
156 	}
157 
158 	err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
159 	if (err)
160 		val = 0;
161 
162 	psb_period = 1 << (val + 11);
163 out:
164 	pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
165 	return psb_period;
166 }
167 
168 static int intel_pt_pick_bit(int bits, int target)
169 {
170 	int pos, pick = -1;
171 
172 	for (pos = 0; bits; bits >>= 1, pos++) {
173 		if (bits & 1) {
174 			if (pos <= target || pick < 0)
175 				pick = pos;
176 			if (pos >= target)
177 				break;
178 		}
179 	}
180 
181 	return pick;
182 }
183 
184 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
185 {
186 	char buf[256];
187 	int mtc, mtc_periods = 0, mtc_period;
188 	int psb_cyc, psb_periods, psb_period;
189 	int pos = 0;
190 	u64 config;
191 	char c;
192 
193 	pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
194 
195 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
196 				&mtc) != 1)
197 		mtc = 1;
198 
199 	if (mtc) {
200 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
201 					&mtc_periods) != 1)
202 			mtc_periods = 0;
203 		if (mtc_periods) {
204 			mtc_period = intel_pt_pick_bit(mtc_periods, 3);
205 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
206 					 ",mtc,mtc_period=%d", mtc_period);
207 		}
208 	}
209 
210 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
211 				&psb_cyc) != 1)
212 		psb_cyc = 1;
213 
214 	if (psb_cyc && mtc_periods) {
215 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
216 					&psb_periods) != 1)
217 			psb_periods = 0;
218 		if (psb_periods) {
219 			psb_period = intel_pt_pick_bit(psb_periods, 3);
220 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
221 					 ",psb_period=%d", psb_period);
222 		}
223 	}
224 
225 	if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
226 	    perf_pmu__scan_file(intel_pt_pmu, "format/branch", "%c", &c) == 1)
227 		pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
228 
229 	pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
230 
231 	intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
232 
233 	return config;
234 }
235 
236 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
237 					   struct record_opts *opts,
238 					   const char *str)
239 {
240 	struct intel_pt_recording *ptr =
241 			container_of(itr, struct intel_pt_recording, itr);
242 	unsigned long long snapshot_size = 0;
243 	char *endptr;
244 
245 	if (str) {
246 		snapshot_size = strtoull(str, &endptr, 0);
247 		if (*endptr || snapshot_size > SIZE_MAX)
248 			return -1;
249 	}
250 
251 	opts->auxtrace_snapshot_mode = true;
252 	opts->auxtrace_snapshot_size = snapshot_size;
253 
254 	ptr->snapshot_size = snapshot_size;
255 
256 	return 0;
257 }
258 
259 struct perf_event_attr *
260 intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
261 {
262 	struct perf_event_attr *attr;
263 
264 	attr = zalloc(sizeof(struct perf_event_attr));
265 	if (!attr)
266 		return NULL;
267 
268 	attr->config = intel_pt_default_config(intel_pt_pmu);
269 
270 	intel_pt_pmu->selectable = true;
271 
272 	return attr;
273 }
274 
275 static const char *intel_pt_find_filter(struct evlist *evlist,
276 					struct perf_pmu *intel_pt_pmu)
277 {
278 	struct evsel *evsel;
279 
280 	evlist__for_each_entry(evlist, evsel) {
281 		if (evsel->core.attr.type == intel_pt_pmu->type)
282 			return evsel->filter;
283 	}
284 
285 	return NULL;
286 }
287 
288 static size_t intel_pt_filter_bytes(const char *filter)
289 {
290 	size_t len = filter ? strlen(filter) : 0;
291 
292 	return len ? roundup(len + 1, 8) : 0;
293 }
294 
295 static size_t
296 intel_pt_info_priv_size(struct auxtrace_record *itr, struct evlist *evlist)
297 {
298 	struct intel_pt_recording *ptr =
299 			container_of(itr, struct intel_pt_recording, itr);
300 	const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
301 
302 	ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
303 			 intel_pt_filter_bytes(filter);
304 
305 	return ptr->priv_size;
306 }
307 
308 static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
309 {
310 	unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
311 
312 	__get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
313 	*n = ebx;
314 	*d = eax;
315 }
316 
317 static int intel_pt_info_fill(struct auxtrace_record *itr,
318 			      struct perf_session *session,
319 			      struct perf_record_auxtrace_info *auxtrace_info,
320 			      size_t priv_size)
321 {
322 	struct intel_pt_recording *ptr =
323 			container_of(itr, struct intel_pt_recording, itr);
324 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
325 	struct perf_event_mmap_page *pc;
326 	struct perf_tsc_conversion tc = { .time_mult = 0, };
327 	bool cap_user_time_zero = false, per_cpu_mmaps;
328 	u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
329 	u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
330 	unsigned long max_non_turbo_ratio;
331 	size_t filter_str_len;
332 	const char *filter;
333 	__u64 *info;
334 	int err;
335 
336 	if (priv_size != ptr->priv_size)
337 		return -EINVAL;
338 
339 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
340 	intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
341 			     &noretcomp_bit);
342 	intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
343 	mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
344 					      "mtc_period");
345 	intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
346 
347 	intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
348 
349 	if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
350 				"%lu", &max_non_turbo_ratio) != 1)
351 		max_non_turbo_ratio = 0;
352 
353 	filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
354 	filter_str_len = filter ? strlen(filter) : 0;
355 
356 	if (!session->evlist->core.nr_mmaps)
357 		return -EINVAL;
358 
359 	pc = session->evlist->mmap[0].core.base;
360 	if (pc) {
361 		err = perf_read_tsc_conversion(pc, &tc);
362 		if (err) {
363 			if (err != -EOPNOTSUPP)
364 				return err;
365 		} else {
366 			cap_user_time_zero = tc.time_mult != 0;
367 		}
368 		if (!cap_user_time_zero)
369 			ui__warning("Intel Processor Trace: TSC not available\n");
370 	}
371 
372 	per_cpu_mmaps = !perf_cpu_map__empty(session->evlist->core.cpus);
373 
374 	auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
375 	auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
376 	auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
377 	auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
378 	auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
379 	auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
380 	auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
381 	auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
382 	auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
383 	auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
384 	auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
385 	auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
386 	auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
387 	auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
388 	auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
389 	auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
390 	auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
391 	auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
392 
393 	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
394 
395 	if (filter_str_len) {
396 		size_t len = intel_pt_filter_bytes(filter);
397 
398 		strncpy((char *)info, filter, len);
399 		info += len >> 3;
400 	}
401 
402 	return 0;
403 }
404 
405 static int intel_pt_track_switches(struct evlist *evlist)
406 {
407 	const char *sched_switch = "sched:sched_switch";
408 	struct evsel *evsel;
409 	int err;
410 
411 	if (!perf_evlist__can_select_event(evlist, sched_switch))
412 		return -EPERM;
413 
414 	err = parse_events(evlist, sched_switch, NULL);
415 	if (err) {
416 		pr_debug2("%s: failed to parse %s, error %d\n",
417 			  __func__, sched_switch, err);
418 		return err;
419 	}
420 
421 	evsel = evlist__last(evlist);
422 
423 	perf_evsel__set_sample_bit(evsel, CPU);
424 	perf_evsel__set_sample_bit(evsel, TIME);
425 
426 	evsel->core.system_wide = true;
427 	evsel->no_aux_samples = true;
428 	evsel->immediate = true;
429 
430 	return 0;
431 }
432 
433 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
434 {
435 	unsigned int val, last = 0, state = 1;
436 	int p = 0;
437 
438 	str[0] = '\0';
439 
440 	for (val = 0; val <= 64; val++, valid >>= 1) {
441 		if (valid & 1) {
442 			last = val;
443 			switch (state) {
444 			case 0:
445 				p += scnprintf(str + p, len - p, ",");
446 				/* Fall through */
447 			case 1:
448 				p += scnprintf(str + p, len - p, "%u", val);
449 				state = 2;
450 				break;
451 			case 2:
452 				state = 3;
453 				break;
454 			case 3:
455 				state = 4;
456 				break;
457 			default:
458 				break;
459 			}
460 		} else {
461 			switch (state) {
462 			case 3:
463 				p += scnprintf(str + p, len - p, ",%u", last);
464 				state = 0;
465 				break;
466 			case 4:
467 				p += scnprintf(str + p, len - p, "-%u", last);
468 				state = 0;
469 				break;
470 			default:
471 				break;
472 			}
473 			if (state != 1)
474 				state = 0;
475 		}
476 	}
477 }
478 
479 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
480 				    const char *caps, const char *name,
481 				    const char *supported, u64 config)
482 {
483 	char valid_str[256];
484 	unsigned int shift;
485 	unsigned long long valid;
486 	u64 bits;
487 	int ok;
488 
489 	if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
490 		valid = 0;
491 
492 	if (supported &&
493 	    perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
494 		valid = 0;
495 
496 	valid |= 1;
497 
498 	bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
499 
500 	config &= bits;
501 
502 	for (shift = 0; bits && !(bits & 1); shift++)
503 		bits >>= 1;
504 
505 	config >>= shift;
506 
507 	if (config > 63)
508 		goto out_err;
509 
510 	if (valid & (1 << config))
511 		return 0;
512 out_err:
513 	intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
514 	pr_err("Invalid %s for %s. Valid values are: %s\n",
515 	       name, INTEL_PT_PMU_NAME, valid_str);
516 	return -EINVAL;
517 }
518 
519 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
520 				    struct evsel *evsel)
521 {
522 	int err;
523 	char c;
524 
525 	if (!evsel)
526 		return 0;
527 
528 	/*
529 	 * If supported, force pass-through config term (pt=1) even if user
530 	 * sets pt=0, which avoids senseless kernel errors.
531 	 */
532 	if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
533 	    !(evsel->core.attr.config & 1)) {
534 		pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
535 		evsel->core.attr.config |= 1;
536 	}
537 
538 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
539 				       "cyc_thresh", "caps/psb_cyc",
540 				       evsel->core.attr.config);
541 	if (err)
542 		return err;
543 
544 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
545 				       "mtc_period", "caps/mtc",
546 				       evsel->core.attr.config);
547 	if (err)
548 		return err;
549 
550 	return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
551 					"psb_period", "caps/psb_cyc",
552 					evsel->core.attr.config);
553 }
554 
555 static void intel_pt_config_sample_mode(struct perf_pmu *intel_pt_pmu,
556 					struct evsel *evsel)
557 {
558 	struct perf_evsel_config_term *term;
559 	u64 user_bits = 0, bits;
560 
561 	term = perf_evsel__get_config_term(evsel, CFG_CHG);
562 	if (term)
563 		user_bits = term->val.cfg_chg;
564 
565 	bits = perf_pmu__format_bits(&intel_pt_pmu->format, "psb_period");
566 
567 	/* Did user change psb_period */
568 	if (bits & user_bits)
569 		return;
570 
571 	/* Set psb_period to 0 */
572 	evsel->core.attr.config &= ~bits;
573 }
574 
575 static void intel_pt_min_max_sample_sz(struct evlist *evlist,
576 				       size_t *min_sz, size_t *max_sz)
577 {
578 	struct evsel *evsel;
579 
580 	evlist__for_each_entry(evlist, evsel) {
581 		size_t sz = evsel->core.attr.aux_sample_size;
582 
583 		if (!sz)
584 			continue;
585 		if (min_sz && (sz < *min_sz || !*min_sz))
586 			*min_sz = sz;
587 		if (max_sz && sz > *max_sz)
588 			*max_sz = sz;
589 	}
590 }
591 
592 /*
593  * Currently, there is not enough information to disambiguate different PEBS
594  * events, so only allow one.
595  */
596 static bool intel_pt_too_many_aux_output(struct evlist *evlist)
597 {
598 	struct evsel *evsel;
599 	int aux_output_cnt = 0;
600 
601 	evlist__for_each_entry(evlist, evsel)
602 		aux_output_cnt += !!evsel->core.attr.aux_output;
603 
604 	if (aux_output_cnt > 1) {
605 		pr_err(INTEL_PT_PMU_NAME " supports at most one event with aux-output\n");
606 		return true;
607 	}
608 
609 	return false;
610 }
611 
612 static int intel_pt_recording_options(struct auxtrace_record *itr,
613 				      struct evlist *evlist,
614 				      struct record_opts *opts)
615 {
616 	struct intel_pt_recording *ptr =
617 			container_of(itr, struct intel_pt_recording, itr);
618 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
619 	bool have_timing_info, need_immediate = false;
620 	struct evsel *evsel, *intel_pt_evsel = NULL;
621 	const struct perf_cpu_map *cpus = evlist->core.cpus;
622 	bool privileged = perf_event_paranoid_check(-1);
623 	u64 tsc_bit;
624 	int err;
625 
626 	ptr->evlist = evlist;
627 	ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
628 
629 	evlist__for_each_entry(evlist, evsel) {
630 		if (evsel->core.attr.type == intel_pt_pmu->type) {
631 			if (intel_pt_evsel) {
632 				pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
633 				return -EINVAL;
634 			}
635 			evsel->core.attr.freq = 0;
636 			evsel->core.attr.sample_period = 1;
637 			intel_pt_evsel = evsel;
638 			opts->full_auxtrace = true;
639 		}
640 	}
641 
642 	if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
643 		pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
644 		return -EINVAL;
645 	}
646 
647 	if (opts->auxtrace_snapshot_mode && opts->auxtrace_sample_mode) {
648 		pr_err("Snapshot mode (" INTEL_PT_PMU_NAME " PMU) and sample trace cannot be used together\n");
649 		return -EINVAL;
650 	}
651 
652 	if (opts->use_clockid) {
653 		pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
654 		return -EINVAL;
655 	}
656 
657 	if (intel_pt_too_many_aux_output(evlist))
658 		return -EINVAL;
659 
660 	if (!opts->full_auxtrace)
661 		return 0;
662 
663 	if (opts->auxtrace_sample_mode)
664 		intel_pt_config_sample_mode(intel_pt_pmu, intel_pt_evsel);
665 
666 	err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
667 	if (err)
668 		return err;
669 
670 	/* Set default sizes for snapshot mode */
671 	if (opts->auxtrace_snapshot_mode) {
672 		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
673 
674 		if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
675 			if (privileged) {
676 				opts->auxtrace_mmap_pages = MiB(4) / page_size;
677 			} else {
678 				opts->auxtrace_mmap_pages = KiB(128) / page_size;
679 				if (opts->mmap_pages == UINT_MAX)
680 					opts->mmap_pages = KiB(256) / page_size;
681 			}
682 		} else if (!opts->auxtrace_mmap_pages && !privileged &&
683 			   opts->mmap_pages == UINT_MAX) {
684 			opts->mmap_pages = KiB(256) / page_size;
685 		}
686 		if (!opts->auxtrace_snapshot_size)
687 			opts->auxtrace_snapshot_size =
688 				opts->auxtrace_mmap_pages * (size_t)page_size;
689 		if (!opts->auxtrace_mmap_pages) {
690 			size_t sz = opts->auxtrace_snapshot_size;
691 
692 			sz = round_up(sz, page_size) / page_size;
693 			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
694 		}
695 		if (opts->auxtrace_snapshot_size >
696 				opts->auxtrace_mmap_pages * (size_t)page_size) {
697 			pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
698 			       opts->auxtrace_snapshot_size,
699 			       opts->auxtrace_mmap_pages * (size_t)page_size);
700 			return -EINVAL;
701 		}
702 		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
703 			pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
704 			return -EINVAL;
705 		}
706 		pr_debug2("Intel PT snapshot size: %zu\n",
707 			  opts->auxtrace_snapshot_size);
708 		if (psb_period &&
709 		    opts->auxtrace_snapshot_size <= psb_period +
710 						  INTEL_PT_PSB_PERIOD_NEAR)
711 			ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
712 				    opts->auxtrace_snapshot_size, psb_period);
713 	}
714 
715 	/* Set default sizes for sample mode */
716 	if (opts->auxtrace_sample_mode) {
717 		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
718 		size_t min_sz = 0, max_sz = 0;
719 
720 		intel_pt_min_max_sample_sz(evlist, &min_sz, &max_sz);
721 		if (!opts->auxtrace_mmap_pages && !privileged &&
722 		    opts->mmap_pages == UINT_MAX)
723 			opts->mmap_pages = KiB(256) / page_size;
724 		if (!opts->auxtrace_mmap_pages) {
725 			size_t sz = round_up(max_sz, page_size) / page_size;
726 
727 			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
728 		}
729 		if (max_sz > opts->auxtrace_mmap_pages * (size_t)page_size) {
730 			pr_err("Sample size %zu must not be greater than AUX area tracing mmap size %zu\n",
731 			       max_sz,
732 			       opts->auxtrace_mmap_pages * (size_t)page_size);
733 			return -EINVAL;
734 		}
735 		pr_debug2("Intel PT min. sample size: %zu max. sample size: %zu\n",
736 			  min_sz, max_sz);
737 		if (psb_period &&
738 		    min_sz <= psb_period + INTEL_PT_PSB_PERIOD_NEAR)
739 			ui__warning("Intel PT sample size (%zu) may be too small for PSB period (%zu)\n",
740 				    min_sz, psb_period);
741 	}
742 
743 	/* Set default sizes for full trace mode */
744 	if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
745 		if (privileged) {
746 			opts->auxtrace_mmap_pages = MiB(4) / page_size;
747 		} else {
748 			opts->auxtrace_mmap_pages = KiB(128) / page_size;
749 			if (opts->mmap_pages == UINT_MAX)
750 				opts->mmap_pages = KiB(256) / page_size;
751 		}
752 	}
753 
754 	/* Validate auxtrace_mmap_pages */
755 	if (opts->auxtrace_mmap_pages) {
756 		size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
757 		size_t min_sz;
758 
759 		if (opts->auxtrace_snapshot_mode || opts->auxtrace_sample_mode)
760 			min_sz = KiB(4);
761 		else
762 			min_sz = KiB(8);
763 
764 		if (sz < min_sz || !is_power_of_2(sz)) {
765 			pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
766 			       min_sz / 1024);
767 			return -EINVAL;
768 		}
769 	}
770 
771 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
772 
773 	if (opts->full_auxtrace && (intel_pt_evsel->core.attr.config & tsc_bit))
774 		have_timing_info = true;
775 	else
776 		have_timing_info = false;
777 
778 	/*
779 	 * Per-cpu recording needs sched_switch events to distinguish different
780 	 * threads.
781 	 */
782 	if (have_timing_info && !perf_cpu_map__empty(cpus)) {
783 		if (perf_can_record_switch_events()) {
784 			bool cpu_wide = !target__none(&opts->target) &&
785 					!target__has_task(&opts->target);
786 
787 			if (!cpu_wide && perf_can_record_cpu_wide()) {
788 				struct evsel *switch_evsel;
789 
790 				err = parse_events(evlist, "dummy:u", NULL);
791 				if (err)
792 					return err;
793 
794 				switch_evsel = evlist__last(evlist);
795 
796 				switch_evsel->core.attr.freq = 0;
797 				switch_evsel->core.attr.sample_period = 1;
798 				switch_evsel->core.attr.context_switch = 1;
799 
800 				switch_evsel->core.system_wide = true;
801 				switch_evsel->no_aux_samples = true;
802 				switch_evsel->immediate = true;
803 
804 				perf_evsel__set_sample_bit(switch_evsel, TID);
805 				perf_evsel__set_sample_bit(switch_evsel, TIME);
806 				perf_evsel__set_sample_bit(switch_evsel, CPU);
807 				perf_evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
808 
809 				opts->record_switch_events = false;
810 				ptr->have_sched_switch = 3;
811 			} else {
812 				opts->record_switch_events = true;
813 				need_immediate = true;
814 				if (cpu_wide)
815 					ptr->have_sched_switch = 3;
816 				else
817 					ptr->have_sched_switch = 2;
818 			}
819 		} else {
820 			err = intel_pt_track_switches(evlist);
821 			if (err == -EPERM)
822 				pr_debug2("Unable to select sched:sched_switch\n");
823 			else if (err)
824 				return err;
825 			else
826 				ptr->have_sched_switch = 1;
827 		}
828 	}
829 
830 	if (intel_pt_evsel) {
831 		/*
832 		 * To obtain the auxtrace buffer file descriptor, the auxtrace
833 		 * event must come first.
834 		 */
835 		perf_evlist__to_front(evlist, intel_pt_evsel);
836 		/*
837 		 * In the case of per-cpu mmaps, we need the CPU on the
838 		 * AUX event.
839 		 */
840 		if (!perf_cpu_map__empty(cpus))
841 			perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
842 	}
843 
844 	/* Add dummy event to keep tracking */
845 	if (opts->full_auxtrace) {
846 		struct evsel *tracking_evsel;
847 
848 		err = parse_events(evlist, "dummy:u", NULL);
849 		if (err)
850 			return err;
851 
852 		tracking_evsel = evlist__last(evlist);
853 
854 		perf_evlist__set_tracking_event(evlist, tracking_evsel);
855 
856 		tracking_evsel->core.attr.freq = 0;
857 		tracking_evsel->core.attr.sample_period = 1;
858 
859 		tracking_evsel->no_aux_samples = true;
860 		if (need_immediate)
861 			tracking_evsel->immediate = true;
862 
863 		/* In per-cpu case, always need the time of mmap events etc */
864 		if (!perf_cpu_map__empty(cpus)) {
865 			perf_evsel__set_sample_bit(tracking_evsel, TIME);
866 			/* And the CPU for switch events */
867 			perf_evsel__set_sample_bit(tracking_evsel, CPU);
868 		}
869 		perf_evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
870 	}
871 
872 	/*
873 	 * Warn the user when we do not have enough information to decode i.e.
874 	 * per-cpu with no sched_switch (except workload-only).
875 	 */
876 	if (!ptr->have_sched_switch && !perf_cpu_map__empty(cpus) &&
877 	    !target__none(&opts->target))
878 		ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
879 
880 	return 0;
881 }
882 
883 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
884 {
885 	struct intel_pt_recording *ptr =
886 			container_of(itr, struct intel_pt_recording, itr);
887 	struct evsel *evsel;
888 
889 	evlist__for_each_entry(ptr->evlist, evsel) {
890 		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
891 			return evsel__disable(evsel);
892 	}
893 	return -EINVAL;
894 }
895 
896 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
897 {
898 	struct intel_pt_recording *ptr =
899 			container_of(itr, struct intel_pt_recording, itr);
900 	struct evsel *evsel;
901 
902 	evlist__for_each_entry(ptr->evlist, evsel) {
903 		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
904 			return evsel__enable(evsel);
905 	}
906 	return -EINVAL;
907 }
908 
909 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
910 {
911 	const size_t sz = sizeof(struct intel_pt_snapshot_ref);
912 	int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
913 	struct intel_pt_snapshot_ref *refs;
914 
915 	if (!new_cnt)
916 		new_cnt = 16;
917 
918 	while (new_cnt <= idx)
919 		new_cnt *= 2;
920 
921 	refs = calloc(new_cnt, sz);
922 	if (!refs)
923 		return -ENOMEM;
924 
925 	memcpy(refs, ptr->snapshot_refs, cnt * sz);
926 
927 	ptr->snapshot_refs = refs;
928 	ptr->snapshot_ref_cnt = new_cnt;
929 
930 	return 0;
931 }
932 
933 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
934 {
935 	int i;
936 
937 	for (i = 0; i < ptr->snapshot_ref_cnt; i++)
938 		zfree(&ptr->snapshot_refs[i].ref_buf);
939 	zfree(&ptr->snapshot_refs);
940 }
941 
942 static void intel_pt_recording_free(struct auxtrace_record *itr)
943 {
944 	struct intel_pt_recording *ptr =
945 			container_of(itr, struct intel_pt_recording, itr);
946 
947 	intel_pt_free_snapshot_refs(ptr);
948 	free(ptr);
949 }
950 
951 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
952 				       size_t snapshot_buf_size)
953 {
954 	size_t ref_buf_size = ptr->snapshot_ref_buf_size;
955 	void *ref_buf;
956 
957 	ref_buf = zalloc(ref_buf_size);
958 	if (!ref_buf)
959 		return -ENOMEM;
960 
961 	ptr->snapshot_refs[idx].ref_buf = ref_buf;
962 	ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
963 
964 	return 0;
965 }
966 
967 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
968 					     size_t snapshot_buf_size)
969 {
970 	const size_t max_size = 256 * 1024;
971 	size_t buf_size = 0, psb_period;
972 
973 	if (ptr->snapshot_size <= 64 * 1024)
974 		return 0;
975 
976 	psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
977 	if (psb_period)
978 		buf_size = psb_period * 2;
979 
980 	if (!buf_size || buf_size > max_size)
981 		buf_size = max_size;
982 
983 	if (buf_size >= snapshot_buf_size)
984 		return 0;
985 
986 	if (buf_size >= ptr->snapshot_size / 2)
987 		return 0;
988 
989 	return buf_size;
990 }
991 
992 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
993 				  size_t snapshot_buf_size)
994 {
995 	if (ptr->snapshot_init_done)
996 		return 0;
997 
998 	ptr->snapshot_init_done = true;
999 
1000 	ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
1001 							snapshot_buf_size);
1002 
1003 	return 0;
1004 }
1005 
1006 /**
1007  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
1008  * @buf1: first buffer
1009  * @compare_size: number of bytes to compare
1010  * @buf2: second buffer (a circular buffer)
1011  * @offs2: offset in second buffer
1012  * @buf2_size: size of second buffer
1013  *
1014  * The comparison allows for the possibility that the bytes to compare in the
1015  * circular buffer are not contiguous.  It is assumed that @compare_size <=
1016  * @buf2_size.  This function returns %false if the bytes are identical, %true
1017  * otherwise.
1018  */
1019 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
1020 				     void *buf2, size_t offs2, size_t buf2_size)
1021 {
1022 	size_t end2 = offs2 + compare_size, part_size;
1023 
1024 	if (end2 <= buf2_size)
1025 		return memcmp(buf1, buf2 + offs2, compare_size);
1026 
1027 	part_size = end2 - buf2_size;
1028 	if (memcmp(buf1, buf2 + offs2, part_size))
1029 		return true;
1030 
1031 	compare_size -= part_size;
1032 
1033 	return memcmp(buf1 + part_size, buf2, compare_size);
1034 }
1035 
1036 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
1037 				 size_t ref_size, size_t buf_size,
1038 				 void *data, size_t head)
1039 {
1040 	size_t ref_end = ref_offset + ref_size;
1041 
1042 	if (ref_end > buf_size) {
1043 		if (head > ref_offset || head < ref_end - buf_size)
1044 			return true;
1045 	} else if (head > ref_offset && head < ref_end) {
1046 		return true;
1047 	}
1048 
1049 	return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
1050 					buf_size);
1051 }
1052 
1053 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
1054 			      void *data, size_t head)
1055 {
1056 	if (head >= ref_size) {
1057 		memcpy(ref_buf, data + head - ref_size, ref_size);
1058 	} else {
1059 		memcpy(ref_buf, data, head);
1060 		ref_size -= head;
1061 		memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
1062 	}
1063 }
1064 
1065 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
1066 			     struct auxtrace_mmap *mm, unsigned char *data,
1067 			     u64 head)
1068 {
1069 	struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
1070 	bool wrapped;
1071 
1072 	wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
1073 				       ptr->snapshot_ref_buf_size, mm->len,
1074 				       data, head);
1075 
1076 	intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
1077 			  data, head);
1078 
1079 	return wrapped;
1080 }
1081 
1082 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
1083 {
1084 	int i, a, b;
1085 
1086 	b = buf_size >> 3;
1087 	a = b - 512;
1088 	if (a < 0)
1089 		a = 0;
1090 
1091 	for (i = a; i < b; i++) {
1092 		if (data[i])
1093 			return true;
1094 	}
1095 
1096 	return false;
1097 }
1098 
1099 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1100 				  struct auxtrace_mmap *mm, unsigned char *data,
1101 				  u64 *head, u64 *old)
1102 {
1103 	struct intel_pt_recording *ptr =
1104 			container_of(itr, struct intel_pt_recording, itr);
1105 	bool wrapped;
1106 	int err;
1107 
1108 	pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1109 		  __func__, idx, (size_t)*old, (size_t)*head);
1110 
1111 	err = intel_pt_snapshot_init(ptr, mm->len);
1112 	if (err)
1113 		goto out_err;
1114 
1115 	if (idx >= ptr->snapshot_ref_cnt) {
1116 		err = intel_pt_alloc_snapshot_refs(ptr, idx);
1117 		if (err)
1118 			goto out_err;
1119 	}
1120 
1121 	if (ptr->snapshot_ref_buf_size) {
1122 		if (!ptr->snapshot_refs[idx].ref_buf) {
1123 			err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1124 			if (err)
1125 				goto out_err;
1126 		}
1127 		wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1128 	} else {
1129 		wrapped = ptr->snapshot_refs[idx].wrapped;
1130 		if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1131 			ptr->snapshot_refs[idx].wrapped = true;
1132 			wrapped = true;
1133 		}
1134 	}
1135 
1136 	/*
1137 	 * In full trace mode 'head' continually increases.  However in snapshot
1138 	 * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1139 	 * are adjusted to match the full trace case which expects that 'old' is
1140 	 * always less than 'head'.
1141 	 */
1142 	if (wrapped) {
1143 		*old = *head;
1144 		*head += mm->len;
1145 	} else {
1146 		if (mm->mask)
1147 			*old &= mm->mask;
1148 		else
1149 			*old %= mm->len;
1150 		if (*old > *head)
1151 			*head += mm->len;
1152 	}
1153 
1154 	pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1155 		  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1156 
1157 	return 0;
1158 
1159 out_err:
1160 	pr_err("%s: failed, error %d\n", __func__, err);
1161 	return err;
1162 }
1163 
1164 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1165 {
1166 	return rdtsc();
1167 }
1168 
1169 static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1170 {
1171 	struct intel_pt_recording *ptr =
1172 			container_of(itr, struct intel_pt_recording, itr);
1173 	struct evsel *evsel;
1174 
1175 	evlist__for_each_entry(ptr->evlist, evsel) {
1176 		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
1177 			return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1178 							     idx);
1179 	}
1180 	return -EINVAL;
1181 }
1182 
1183 struct auxtrace_record *intel_pt_recording_init(int *err)
1184 {
1185 	struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1186 	struct intel_pt_recording *ptr;
1187 
1188 	if (!intel_pt_pmu)
1189 		return NULL;
1190 
1191 	if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1192 		*err = -errno;
1193 		return NULL;
1194 	}
1195 
1196 	ptr = zalloc(sizeof(struct intel_pt_recording));
1197 	if (!ptr) {
1198 		*err = -ENOMEM;
1199 		return NULL;
1200 	}
1201 
1202 	ptr->intel_pt_pmu = intel_pt_pmu;
1203 	ptr->itr.recording_options = intel_pt_recording_options;
1204 	ptr->itr.info_priv_size = intel_pt_info_priv_size;
1205 	ptr->itr.info_fill = intel_pt_info_fill;
1206 	ptr->itr.free = intel_pt_recording_free;
1207 	ptr->itr.snapshot_start = intel_pt_snapshot_start;
1208 	ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1209 	ptr->itr.find_snapshot = intel_pt_find_snapshot;
1210 	ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1211 	ptr->itr.reference = intel_pt_reference;
1212 	ptr->itr.read_finish = intel_pt_read_finish;
1213 	/*
1214 	 * Decoding starts at a PSB packet. Minimum PSB period is 2K so 4K
1215 	 * should give at least 1 PSB per sample.
1216 	 */
1217 	ptr->itr.default_aux_sample_size = 4096;
1218 	return &ptr->itr;
1219 }
1220