xref: /openbmc/linux/tools/perf/arch/x86/util/intel-pt.c (revision 0883c2c0)
1 /*
2  * intel_pt.c: Intel Processor Trace support
3  * Copyright (c) 2013-2015, Intel Corporation.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  */
15 
16 #include <stdbool.h>
17 #include <linux/kernel.h>
18 #include <linux/types.h>
19 #include <linux/bitops.h>
20 #include <linux/log2.h>
21 #include <cpuid.h>
22 
23 #include "../../perf.h"
24 #include "../../util/session.h"
25 #include "../../util/event.h"
26 #include "../../util/evlist.h"
27 #include "../../util/evsel.h"
28 #include "../../util/cpumap.h"
29 #include <subcmd/parse-options.h>
30 #include "../../util/parse-events.h"
31 #include "../../util/pmu.h"
32 #include "../../util/debug.h"
33 #include "../../util/auxtrace.h"
34 #include "../../util/tsc.h"
35 #include "../../util/intel-pt.h"
36 
37 #define KiB(x) ((x) * 1024)
38 #define MiB(x) ((x) * 1024 * 1024)
39 #define KiB_MASK(x) (KiB(x) - 1)
40 #define MiB_MASK(x) (MiB(x) - 1)
41 
42 #define INTEL_PT_DEFAULT_SAMPLE_SIZE	KiB(4)
43 
44 #define INTEL_PT_MAX_SAMPLE_SIZE	KiB(60)
45 
46 #define INTEL_PT_PSB_PERIOD_NEAR	256
47 
48 struct intel_pt_snapshot_ref {
49 	void *ref_buf;
50 	size_t ref_offset;
51 	bool wrapped;
52 };
53 
54 struct intel_pt_recording {
55 	struct auxtrace_record		itr;
56 	struct perf_pmu			*intel_pt_pmu;
57 	int				have_sched_switch;
58 	struct perf_evlist		*evlist;
59 	bool				snapshot_mode;
60 	bool				snapshot_init_done;
61 	size_t				snapshot_size;
62 	size_t				snapshot_ref_buf_size;
63 	int				snapshot_ref_cnt;
64 	struct intel_pt_snapshot_ref	*snapshot_refs;
65 };
66 
67 static int intel_pt_parse_terms_with_default(struct list_head *formats,
68 					     const char *str,
69 					     u64 *config)
70 {
71 	struct list_head *terms;
72 	struct perf_event_attr attr = { .size = 0, };
73 	int err;
74 
75 	terms = malloc(sizeof(struct list_head));
76 	if (!terms)
77 		return -ENOMEM;
78 
79 	INIT_LIST_HEAD(terms);
80 
81 	err = parse_events_terms(terms, str);
82 	if (err)
83 		goto out_free;
84 
85 	attr.config = *config;
86 	err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
87 	if (err)
88 		goto out_free;
89 
90 	*config = attr.config;
91 out_free:
92 	parse_events_terms__delete(terms);
93 	return err;
94 }
95 
96 static int intel_pt_parse_terms(struct list_head *formats, const char *str,
97 				u64 *config)
98 {
99 	*config = 0;
100 	return intel_pt_parse_terms_with_default(formats, str, config);
101 }
102 
103 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
104 {
105 	const u64 top_bit = 1ULL << 63;
106 	u64 res = 0;
107 	int i;
108 
109 	for (i = 0; i < 64; i++) {
110 		if (mask & top_bit) {
111 			res <<= 1;
112 			if (bits & top_bit)
113 				res |= 1;
114 		}
115 		mask <<= 1;
116 		bits <<= 1;
117 	}
118 
119 	return res;
120 }
121 
122 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
123 				struct perf_evlist *evlist, u64 *res)
124 {
125 	struct perf_evsel *evsel;
126 	u64 mask;
127 
128 	*res = 0;
129 
130 	mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
131 	if (!mask)
132 		return -EINVAL;
133 
134 	evlist__for_each(evlist, evsel) {
135 		if (evsel->attr.type == intel_pt_pmu->type) {
136 			*res = intel_pt_masked_bits(mask, evsel->attr.config);
137 			return 0;
138 		}
139 	}
140 
141 	return -EINVAL;
142 }
143 
144 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
145 				  struct perf_evlist *evlist)
146 {
147 	u64 val;
148 	int err, topa_multiple_entries;
149 	size_t psb_period;
150 
151 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
152 				"%d", &topa_multiple_entries) != 1)
153 		topa_multiple_entries = 0;
154 
155 	/*
156 	 * Use caps/topa_multiple_entries to indicate early hardware that had
157 	 * extra frequent PSBs.
158 	 */
159 	if (!topa_multiple_entries) {
160 		psb_period = 256;
161 		goto out;
162 	}
163 
164 	err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
165 	if (err)
166 		val = 0;
167 
168 	psb_period = 1 << (val + 11);
169 out:
170 	pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
171 	return psb_period;
172 }
173 
174 static int intel_pt_pick_bit(int bits, int target)
175 {
176 	int pos, pick = -1;
177 
178 	for (pos = 0; bits; bits >>= 1, pos++) {
179 		if (bits & 1) {
180 			if (pos <= target || pick < 0)
181 				pick = pos;
182 			if (pos >= target)
183 				break;
184 		}
185 	}
186 
187 	return pick;
188 }
189 
190 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
191 {
192 	char buf[256];
193 	int mtc, mtc_periods = 0, mtc_period;
194 	int psb_cyc, psb_periods, psb_period;
195 	int pos = 0;
196 	u64 config;
197 
198 	pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
199 
200 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
201 				&mtc) != 1)
202 		mtc = 1;
203 
204 	if (mtc) {
205 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
206 					&mtc_periods) != 1)
207 			mtc_periods = 0;
208 		if (mtc_periods) {
209 			mtc_period = intel_pt_pick_bit(mtc_periods, 3);
210 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
211 					 ",mtc,mtc_period=%d", mtc_period);
212 		}
213 	}
214 
215 	if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
216 				&psb_cyc) != 1)
217 		psb_cyc = 1;
218 
219 	if (psb_cyc && mtc_periods) {
220 		if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
221 					&psb_periods) != 1)
222 			psb_periods = 0;
223 		if (psb_periods) {
224 			psb_period = intel_pt_pick_bit(psb_periods, 3);
225 			pos += scnprintf(buf + pos, sizeof(buf) - pos,
226 					 ",psb_period=%d", psb_period);
227 		}
228 	}
229 
230 	pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
231 
232 	intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
233 
234 	return config;
235 }
236 
237 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
238 					   struct record_opts *opts,
239 					   const char *str)
240 {
241 	struct intel_pt_recording *ptr =
242 			container_of(itr, struct intel_pt_recording, itr);
243 	unsigned long long snapshot_size = 0;
244 	char *endptr;
245 
246 	if (str) {
247 		snapshot_size = strtoull(str, &endptr, 0);
248 		if (*endptr || snapshot_size > SIZE_MAX)
249 			return -1;
250 	}
251 
252 	opts->auxtrace_snapshot_mode = true;
253 	opts->auxtrace_snapshot_size = snapshot_size;
254 
255 	ptr->snapshot_size = snapshot_size;
256 
257 	return 0;
258 }
259 
260 struct perf_event_attr *
261 intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
262 {
263 	struct perf_event_attr *attr;
264 
265 	attr = zalloc(sizeof(struct perf_event_attr));
266 	if (!attr)
267 		return NULL;
268 
269 	attr->config = intel_pt_default_config(intel_pt_pmu);
270 
271 	intel_pt_pmu->selectable = true;
272 
273 	return attr;
274 }
275 
276 static size_t
277 intel_pt_info_priv_size(struct auxtrace_record *itr __maybe_unused,
278 			struct perf_evlist *evlist __maybe_unused)
279 {
280 	return INTEL_PT_AUXTRACE_PRIV_SIZE;
281 }
282 
283 static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
284 {
285 	unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
286 
287 	__get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
288 	*n = ebx;
289 	*d = eax;
290 }
291 
292 static int intel_pt_info_fill(struct auxtrace_record *itr,
293 			      struct perf_session *session,
294 			      struct auxtrace_info_event *auxtrace_info,
295 			      size_t priv_size)
296 {
297 	struct intel_pt_recording *ptr =
298 			container_of(itr, struct intel_pt_recording, itr);
299 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
300 	struct perf_event_mmap_page *pc;
301 	struct perf_tsc_conversion tc = { .time_mult = 0, };
302 	bool cap_user_time_zero = false, per_cpu_mmaps;
303 	u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
304 	u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
305 	int err;
306 
307 	if (priv_size != INTEL_PT_AUXTRACE_PRIV_SIZE)
308 		return -EINVAL;
309 
310 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
311 	intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
312 			     &noretcomp_bit);
313 	intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
314 	mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
315 					      "mtc_period");
316 	intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
317 
318 	intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
319 
320 	if (!session->evlist->nr_mmaps)
321 		return -EINVAL;
322 
323 	pc = session->evlist->mmap[0].base;
324 	if (pc) {
325 		err = perf_read_tsc_conversion(pc, &tc);
326 		if (err) {
327 			if (err != -EOPNOTSUPP)
328 				return err;
329 		} else {
330 			cap_user_time_zero = tc.time_mult != 0;
331 		}
332 		if (!cap_user_time_zero)
333 			ui__warning("Intel Processor Trace: TSC not available\n");
334 	}
335 
336 	per_cpu_mmaps = !cpu_map__empty(session->evlist->cpus);
337 
338 	auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
339 	auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
340 	auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
341 	auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
342 	auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
343 	auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
344 	auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
345 	auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
346 	auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
347 	auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
348 	auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
349 	auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
350 	auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
351 	auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
352 	auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
353 	auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
354 
355 	return 0;
356 }
357 
358 static int intel_pt_track_switches(struct perf_evlist *evlist)
359 {
360 	const char *sched_switch = "sched:sched_switch";
361 	struct perf_evsel *evsel;
362 	int err;
363 
364 	if (!perf_evlist__can_select_event(evlist, sched_switch))
365 		return -EPERM;
366 
367 	err = parse_events(evlist, sched_switch, NULL);
368 	if (err) {
369 		pr_debug2("%s: failed to parse %s, error %d\n",
370 			  __func__, sched_switch, err);
371 		return err;
372 	}
373 
374 	evsel = perf_evlist__last(evlist);
375 
376 	perf_evsel__set_sample_bit(evsel, CPU);
377 	perf_evsel__set_sample_bit(evsel, TIME);
378 
379 	evsel->system_wide = true;
380 	evsel->no_aux_samples = true;
381 	evsel->immediate = true;
382 
383 	return 0;
384 }
385 
386 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
387 {
388 	unsigned int val, last = 0, state = 1;
389 	int p = 0;
390 
391 	str[0] = '\0';
392 
393 	for (val = 0; val <= 64; val++, valid >>= 1) {
394 		if (valid & 1) {
395 			last = val;
396 			switch (state) {
397 			case 0:
398 				p += scnprintf(str + p, len - p, ",");
399 				/* Fall through */
400 			case 1:
401 				p += scnprintf(str + p, len - p, "%u", val);
402 				state = 2;
403 				break;
404 			case 2:
405 				state = 3;
406 				break;
407 			case 3:
408 				state = 4;
409 				break;
410 			default:
411 				break;
412 			}
413 		} else {
414 			switch (state) {
415 			case 3:
416 				p += scnprintf(str + p, len - p, ",%u", last);
417 				state = 0;
418 				break;
419 			case 4:
420 				p += scnprintf(str + p, len - p, "-%u", last);
421 				state = 0;
422 				break;
423 			default:
424 				break;
425 			}
426 			if (state != 1)
427 				state = 0;
428 		}
429 	}
430 }
431 
432 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
433 				    const char *caps, const char *name,
434 				    const char *supported, u64 config)
435 {
436 	char valid_str[256];
437 	unsigned int shift;
438 	unsigned long long valid;
439 	u64 bits;
440 	int ok;
441 
442 	if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
443 		valid = 0;
444 
445 	if (supported &&
446 	    perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
447 		valid = 0;
448 
449 	valid |= 1;
450 
451 	bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
452 
453 	config &= bits;
454 
455 	for (shift = 0; bits && !(bits & 1); shift++)
456 		bits >>= 1;
457 
458 	config >>= shift;
459 
460 	if (config > 63)
461 		goto out_err;
462 
463 	if (valid & (1 << config))
464 		return 0;
465 out_err:
466 	intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
467 	pr_err("Invalid %s for %s. Valid values are: %s\n",
468 	       name, INTEL_PT_PMU_NAME, valid_str);
469 	return -EINVAL;
470 }
471 
472 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
473 				    struct perf_evsel *evsel)
474 {
475 	int err;
476 
477 	if (!evsel)
478 		return 0;
479 
480 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
481 				       "cyc_thresh", "caps/psb_cyc",
482 				       evsel->attr.config);
483 	if (err)
484 		return err;
485 
486 	err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
487 				       "mtc_period", "caps/mtc",
488 				       evsel->attr.config);
489 	if (err)
490 		return err;
491 
492 	return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
493 					"psb_period", "caps/psb_cyc",
494 					evsel->attr.config);
495 }
496 
497 static int intel_pt_recording_options(struct auxtrace_record *itr,
498 				      struct perf_evlist *evlist,
499 				      struct record_opts *opts)
500 {
501 	struct intel_pt_recording *ptr =
502 			container_of(itr, struct intel_pt_recording, itr);
503 	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
504 	bool have_timing_info;
505 	struct perf_evsel *evsel, *intel_pt_evsel = NULL;
506 	const struct cpu_map *cpus = evlist->cpus;
507 	bool privileged = geteuid() == 0 || perf_event_paranoid() < 0;
508 	u64 tsc_bit;
509 	int err;
510 
511 	ptr->evlist = evlist;
512 	ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
513 
514 	evlist__for_each(evlist, evsel) {
515 		if (evsel->attr.type == intel_pt_pmu->type) {
516 			if (intel_pt_evsel) {
517 				pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
518 				return -EINVAL;
519 			}
520 			evsel->attr.freq = 0;
521 			evsel->attr.sample_period = 1;
522 			intel_pt_evsel = evsel;
523 			opts->full_auxtrace = true;
524 		}
525 	}
526 
527 	if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
528 		pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
529 		return -EINVAL;
530 	}
531 
532 	if (opts->use_clockid) {
533 		pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
534 		return -EINVAL;
535 	}
536 
537 	if (!opts->full_auxtrace)
538 		return 0;
539 
540 	err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
541 	if (err)
542 		return err;
543 
544 	/* Set default sizes for snapshot mode */
545 	if (opts->auxtrace_snapshot_mode) {
546 		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
547 
548 		if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
549 			if (privileged) {
550 				opts->auxtrace_mmap_pages = MiB(4) / page_size;
551 			} else {
552 				opts->auxtrace_mmap_pages = KiB(128) / page_size;
553 				if (opts->mmap_pages == UINT_MAX)
554 					opts->mmap_pages = KiB(256) / page_size;
555 			}
556 		} else if (!opts->auxtrace_mmap_pages && !privileged &&
557 			   opts->mmap_pages == UINT_MAX) {
558 			opts->mmap_pages = KiB(256) / page_size;
559 		}
560 		if (!opts->auxtrace_snapshot_size)
561 			opts->auxtrace_snapshot_size =
562 				opts->auxtrace_mmap_pages * (size_t)page_size;
563 		if (!opts->auxtrace_mmap_pages) {
564 			size_t sz = opts->auxtrace_snapshot_size;
565 
566 			sz = round_up(sz, page_size) / page_size;
567 			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
568 		}
569 		if (opts->auxtrace_snapshot_size >
570 				opts->auxtrace_mmap_pages * (size_t)page_size) {
571 			pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
572 			       opts->auxtrace_snapshot_size,
573 			       opts->auxtrace_mmap_pages * (size_t)page_size);
574 			return -EINVAL;
575 		}
576 		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
577 			pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
578 			return -EINVAL;
579 		}
580 		pr_debug2("Intel PT snapshot size: %zu\n",
581 			  opts->auxtrace_snapshot_size);
582 		if (psb_period &&
583 		    opts->auxtrace_snapshot_size <= psb_period +
584 						  INTEL_PT_PSB_PERIOD_NEAR)
585 			ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
586 				    opts->auxtrace_snapshot_size, psb_period);
587 	}
588 
589 	/* Set default sizes for full trace mode */
590 	if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
591 		if (privileged) {
592 			opts->auxtrace_mmap_pages = MiB(4) / page_size;
593 		} else {
594 			opts->auxtrace_mmap_pages = KiB(128) / page_size;
595 			if (opts->mmap_pages == UINT_MAX)
596 				opts->mmap_pages = KiB(256) / page_size;
597 		}
598 	}
599 
600 	/* Validate auxtrace_mmap_pages */
601 	if (opts->auxtrace_mmap_pages) {
602 		size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
603 		size_t min_sz;
604 
605 		if (opts->auxtrace_snapshot_mode)
606 			min_sz = KiB(4);
607 		else
608 			min_sz = KiB(8);
609 
610 		if (sz < min_sz || !is_power_of_2(sz)) {
611 			pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
612 			       min_sz / 1024);
613 			return -EINVAL;
614 		}
615 	}
616 
617 	intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
618 
619 	if (opts->full_auxtrace && (intel_pt_evsel->attr.config & tsc_bit))
620 		have_timing_info = true;
621 	else
622 		have_timing_info = false;
623 
624 	/*
625 	 * Per-cpu recording needs sched_switch events to distinguish different
626 	 * threads.
627 	 */
628 	if (have_timing_info && !cpu_map__empty(cpus)) {
629 		if (perf_can_record_switch_events()) {
630 			bool cpu_wide = !target__none(&opts->target) &&
631 					!target__has_task(&opts->target);
632 
633 			if (!cpu_wide && perf_can_record_cpu_wide()) {
634 				struct perf_evsel *switch_evsel;
635 
636 				err = parse_events(evlist, "dummy:u", NULL);
637 				if (err)
638 					return err;
639 
640 				switch_evsel = perf_evlist__last(evlist);
641 
642 				switch_evsel->attr.freq = 0;
643 				switch_evsel->attr.sample_period = 1;
644 				switch_evsel->attr.context_switch = 1;
645 
646 				switch_evsel->system_wide = true;
647 				switch_evsel->no_aux_samples = true;
648 				switch_evsel->immediate = true;
649 
650 				perf_evsel__set_sample_bit(switch_evsel, TID);
651 				perf_evsel__set_sample_bit(switch_evsel, TIME);
652 				perf_evsel__set_sample_bit(switch_evsel, CPU);
653 
654 				opts->record_switch_events = false;
655 				ptr->have_sched_switch = 3;
656 			} else {
657 				opts->record_switch_events = true;
658 				if (cpu_wide)
659 					ptr->have_sched_switch = 3;
660 				else
661 					ptr->have_sched_switch = 2;
662 			}
663 		} else {
664 			err = intel_pt_track_switches(evlist);
665 			if (err == -EPERM)
666 				pr_debug2("Unable to select sched:sched_switch\n");
667 			else if (err)
668 				return err;
669 			else
670 				ptr->have_sched_switch = 1;
671 		}
672 	}
673 
674 	if (intel_pt_evsel) {
675 		/*
676 		 * To obtain the auxtrace buffer file descriptor, the auxtrace
677 		 * event must come first.
678 		 */
679 		perf_evlist__to_front(evlist, intel_pt_evsel);
680 		/*
681 		 * In the case of per-cpu mmaps, we need the CPU on the
682 		 * AUX event.
683 		 */
684 		if (!cpu_map__empty(cpus))
685 			perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
686 	}
687 
688 	/* Add dummy event to keep tracking */
689 	if (opts->full_auxtrace) {
690 		struct perf_evsel *tracking_evsel;
691 
692 		err = parse_events(evlist, "dummy:u", NULL);
693 		if (err)
694 			return err;
695 
696 		tracking_evsel = perf_evlist__last(evlist);
697 
698 		perf_evlist__set_tracking_event(evlist, tracking_evsel);
699 
700 		tracking_evsel->attr.freq = 0;
701 		tracking_evsel->attr.sample_period = 1;
702 
703 		/* In per-cpu case, always need the time of mmap events etc */
704 		if (!cpu_map__empty(cpus)) {
705 			perf_evsel__set_sample_bit(tracking_evsel, TIME);
706 			/* And the CPU for switch events */
707 			perf_evsel__set_sample_bit(tracking_evsel, CPU);
708 		}
709 	}
710 
711 	/*
712 	 * Warn the user when we do not have enough information to decode i.e.
713 	 * per-cpu with no sched_switch (except workload-only).
714 	 */
715 	if (!ptr->have_sched_switch && !cpu_map__empty(cpus) &&
716 	    !target__none(&opts->target))
717 		ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
718 
719 	return 0;
720 }
721 
722 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
723 {
724 	struct intel_pt_recording *ptr =
725 			container_of(itr, struct intel_pt_recording, itr);
726 	struct perf_evsel *evsel;
727 
728 	evlist__for_each(ptr->evlist, evsel) {
729 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
730 			return perf_evsel__disable(evsel);
731 	}
732 	return -EINVAL;
733 }
734 
735 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
736 {
737 	struct intel_pt_recording *ptr =
738 			container_of(itr, struct intel_pt_recording, itr);
739 	struct perf_evsel *evsel;
740 
741 	evlist__for_each(ptr->evlist, evsel) {
742 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
743 			return perf_evsel__enable(evsel);
744 	}
745 	return -EINVAL;
746 }
747 
748 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
749 {
750 	const size_t sz = sizeof(struct intel_pt_snapshot_ref);
751 	int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
752 	struct intel_pt_snapshot_ref *refs;
753 
754 	if (!new_cnt)
755 		new_cnt = 16;
756 
757 	while (new_cnt <= idx)
758 		new_cnt *= 2;
759 
760 	refs = calloc(new_cnt, sz);
761 	if (!refs)
762 		return -ENOMEM;
763 
764 	memcpy(refs, ptr->snapshot_refs, cnt * sz);
765 
766 	ptr->snapshot_refs = refs;
767 	ptr->snapshot_ref_cnt = new_cnt;
768 
769 	return 0;
770 }
771 
772 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
773 {
774 	int i;
775 
776 	for (i = 0; i < ptr->snapshot_ref_cnt; i++)
777 		zfree(&ptr->snapshot_refs[i].ref_buf);
778 	zfree(&ptr->snapshot_refs);
779 }
780 
781 static void intel_pt_recording_free(struct auxtrace_record *itr)
782 {
783 	struct intel_pt_recording *ptr =
784 			container_of(itr, struct intel_pt_recording, itr);
785 
786 	intel_pt_free_snapshot_refs(ptr);
787 	free(ptr);
788 }
789 
790 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
791 				       size_t snapshot_buf_size)
792 {
793 	size_t ref_buf_size = ptr->snapshot_ref_buf_size;
794 	void *ref_buf;
795 
796 	ref_buf = zalloc(ref_buf_size);
797 	if (!ref_buf)
798 		return -ENOMEM;
799 
800 	ptr->snapshot_refs[idx].ref_buf = ref_buf;
801 	ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
802 
803 	return 0;
804 }
805 
806 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
807 					     size_t snapshot_buf_size)
808 {
809 	const size_t max_size = 256 * 1024;
810 	size_t buf_size = 0, psb_period;
811 
812 	if (ptr->snapshot_size <= 64 * 1024)
813 		return 0;
814 
815 	psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
816 	if (psb_period)
817 		buf_size = psb_period * 2;
818 
819 	if (!buf_size || buf_size > max_size)
820 		buf_size = max_size;
821 
822 	if (buf_size >= snapshot_buf_size)
823 		return 0;
824 
825 	if (buf_size >= ptr->snapshot_size / 2)
826 		return 0;
827 
828 	return buf_size;
829 }
830 
831 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
832 				  size_t snapshot_buf_size)
833 {
834 	if (ptr->snapshot_init_done)
835 		return 0;
836 
837 	ptr->snapshot_init_done = true;
838 
839 	ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
840 							snapshot_buf_size);
841 
842 	return 0;
843 }
844 
845 /**
846  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
847  * @buf1: first buffer
848  * @compare_size: number of bytes to compare
849  * @buf2: second buffer (a circular buffer)
850  * @offs2: offset in second buffer
851  * @buf2_size: size of second buffer
852  *
853  * The comparison allows for the possibility that the bytes to compare in the
854  * circular buffer are not contiguous.  It is assumed that @compare_size <=
855  * @buf2_size.  This function returns %false if the bytes are identical, %true
856  * otherwise.
857  */
858 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
859 				     void *buf2, size_t offs2, size_t buf2_size)
860 {
861 	size_t end2 = offs2 + compare_size, part_size;
862 
863 	if (end2 <= buf2_size)
864 		return memcmp(buf1, buf2 + offs2, compare_size);
865 
866 	part_size = end2 - buf2_size;
867 	if (memcmp(buf1, buf2 + offs2, part_size))
868 		return true;
869 
870 	compare_size -= part_size;
871 
872 	return memcmp(buf1 + part_size, buf2, compare_size);
873 }
874 
875 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
876 				 size_t ref_size, size_t buf_size,
877 				 void *data, size_t head)
878 {
879 	size_t ref_end = ref_offset + ref_size;
880 
881 	if (ref_end > buf_size) {
882 		if (head > ref_offset || head < ref_end - buf_size)
883 			return true;
884 	} else if (head > ref_offset && head < ref_end) {
885 		return true;
886 	}
887 
888 	return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
889 					buf_size);
890 }
891 
892 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
893 			      void *data, size_t head)
894 {
895 	if (head >= ref_size) {
896 		memcpy(ref_buf, data + head - ref_size, ref_size);
897 	} else {
898 		memcpy(ref_buf, data, head);
899 		ref_size -= head;
900 		memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
901 	}
902 }
903 
904 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
905 			     struct auxtrace_mmap *mm, unsigned char *data,
906 			     u64 head)
907 {
908 	struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
909 	bool wrapped;
910 
911 	wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
912 				       ptr->snapshot_ref_buf_size, mm->len,
913 				       data, head);
914 
915 	intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
916 			  data, head);
917 
918 	return wrapped;
919 }
920 
921 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
922 {
923 	int i, a, b;
924 
925 	b = buf_size >> 3;
926 	a = b - 512;
927 	if (a < 0)
928 		a = 0;
929 
930 	for (i = a; i < b; i++) {
931 		if (data[i])
932 			return true;
933 	}
934 
935 	return false;
936 }
937 
938 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
939 				  struct auxtrace_mmap *mm, unsigned char *data,
940 				  u64 *head, u64 *old)
941 {
942 	struct intel_pt_recording *ptr =
943 			container_of(itr, struct intel_pt_recording, itr);
944 	bool wrapped;
945 	int err;
946 
947 	pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
948 		  __func__, idx, (size_t)*old, (size_t)*head);
949 
950 	err = intel_pt_snapshot_init(ptr, mm->len);
951 	if (err)
952 		goto out_err;
953 
954 	if (idx >= ptr->snapshot_ref_cnt) {
955 		err = intel_pt_alloc_snapshot_refs(ptr, idx);
956 		if (err)
957 			goto out_err;
958 	}
959 
960 	if (ptr->snapshot_ref_buf_size) {
961 		if (!ptr->snapshot_refs[idx].ref_buf) {
962 			err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
963 			if (err)
964 				goto out_err;
965 		}
966 		wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
967 	} else {
968 		wrapped = ptr->snapshot_refs[idx].wrapped;
969 		if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
970 			ptr->snapshot_refs[idx].wrapped = true;
971 			wrapped = true;
972 		}
973 	}
974 
975 	/*
976 	 * In full trace mode 'head' continually increases.  However in snapshot
977 	 * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
978 	 * are adjusted to match the full trace case which expects that 'old' is
979 	 * always less than 'head'.
980 	 */
981 	if (wrapped) {
982 		*old = *head;
983 		*head += mm->len;
984 	} else {
985 		if (mm->mask)
986 			*old &= mm->mask;
987 		else
988 			*old %= mm->len;
989 		if (*old > *head)
990 			*head += mm->len;
991 	}
992 
993 	pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
994 		  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
995 
996 	return 0;
997 
998 out_err:
999 	pr_err("%s: failed, error %d\n", __func__, err);
1000 	return err;
1001 }
1002 
1003 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1004 {
1005 	return rdtsc();
1006 }
1007 
1008 static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1009 {
1010 	struct intel_pt_recording *ptr =
1011 			container_of(itr, struct intel_pt_recording, itr);
1012 	struct perf_evsel *evsel;
1013 
1014 	evlist__for_each(ptr->evlist, evsel) {
1015 		if (evsel->attr.type == ptr->intel_pt_pmu->type)
1016 			return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1017 							     idx);
1018 	}
1019 	return -EINVAL;
1020 }
1021 
1022 struct auxtrace_record *intel_pt_recording_init(int *err)
1023 {
1024 	struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1025 	struct intel_pt_recording *ptr;
1026 
1027 	if (!intel_pt_pmu)
1028 		return NULL;
1029 
1030 	if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1031 		*err = -errno;
1032 		return NULL;
1033 	}
1034 
1035 	ptr = zalloc(sizeof(struct intel_pt_recording));
1036 	if (!ptr) {
1037 		*err = -ENOMEM;
1038 		return NULL;
1039 	}
1040 
1041 	ptr->intel_pt_pmu = intel_pt_pmu;
1042 	ptr->itr.recording_options = intel_pt_recording_options;
1043 	ptr->itr.info_priv_size = intel_pt_info_priv_size;
1044 	ptr->itr.info_fill = intel_pt_info_fill;
1045 	ptr->itr.free = intel_pt_recording_free;
1046 	ptr->itr.snapshot_start = intel_pt_snapshot_start;
1047 	ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1048 	ptr->itr.find_snapshot = intel_pt_find_snapshot;
1049 	ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1050 	ptr->itr.reference = intel_pt_reference;
1051 	ptr->itr.read_finish = intel_pt_read_finish;
1052 	return &ptr->itr;
1053 }
1054