?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'trace-printf-v6.13' of git://git.kernel.org/pub/scm/linux/kernel/git/trace...
[drm/drm-misc.git] / tools / perf / arch / x86 / util / intel-pt.c
blob8f235d8b67b628177125b92880a97486e2d9de83
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 <linux/err.h>
15 #include <cpuid.h>
16
17 #include "../../../util/session.h"
18 #include "../../../util/event.h"
19 #include "../../../util/evlist.h"
20 #include "../../../util/evsel.h"
21 #include "../../../util/evsel_config.h"
22 #include "../../../util/cpumap.h"
23 #include "../../../util/mmap.h"
24 #include <subcmd/parse-options.h>
25 #include "../../../util/parse-events.h"
26 #include "../../../util/pmus.h"
27 #include "../../../util/debug.h"
28 #include "../../../util/auxtrace.h"
29 #include "../../../util/perf_api_probe.h"
30 #include "../../../util/record.h"
31 #include "../../../util/target.h"
32 #include "../../../util/tsc.h"
33 #include <internal/lib.h> // page_size
34 #include "../../../util/intel-pt.h"
35 #include <api/fs/fs.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_PSB_PERIOD_NEAR 256
43
44 struct intel_pt_snapshot_ref {
45 void *ref_buf;
46 size_t ref_offset;
47 bool wrapped;
48 };
49
50 struct intel_pt_recording {
51 struct auxtrace_record itr;
52 struct perf_pmu *intel_pt_pmu;
53 int have_sched_switch;
54 struct evlist *evlist;
55 bool snapshot_mode;
56 bool snapshot_init_done;
57 size_t snapshot_size;
58 size_t snapshot_ref_buf_size;
59 int snapshot_ref_cnt;
60 struct intel_pt_snapshot_ref *snapshot_refs;
61 size_t priv_size;
62 };
63
64 static int intel_pt_parse_terms_with_default(const struct perf_pmu *pmu,
65 const char *str,
66 u64 *config)
67 {
68 struct parse_events_terms terms;
69 struct perf_event_attr attr = { .size = 0, };
70 int err;
71
72 parse_events_terms__init(&terms);
73 err = parse_events_terms(&terms, str, /*input=*/ NULL);
74 if (err)
75 goto out_free;
76
77 attr.config = *config;
78 err = perf_pmu__config_terms(pmu, &attr, &terms, /*zero=*/true, /*apply_hardcoded=*/false,
79 /*err=*/NULL);
80 if (err)
81 goto out_free;
82
83 *config = attr.config;
84 out_free:
85 parse_events_terms__exit(&terms);
86 return err;
87 }
88
89 static int intel_pt_parse_terms(const struct perf_pmu *pmu, const char *str, u64 *config)
90 {
91 *config = 0;
92 return intel_pt_parse_terms_with_default(pmu, str, config);
93 }
94
95 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
96 {
97 const u64 top_bit = 1ULL << 63;
98 u64 res = 0;
99 int i;
100
101 for (i = 0; i < 64; i++) {
102 if (mask & top_bit) {
103 res <<= 1;
104 if (bits & top_bit)
105 res |= 1;
106 }
107 mask <<= 1;
108 bits <<= 1;
109 }
110
111 return res;
112 }
113
114 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
115 struct evlist *evlist, u64 *res)
116 {
117 struct evsel *evsel;
118 u64 mask;
119
120 *res = 0;
121
122 mask = perf_pmu__format_bits(intel_pt_pmu, str);
123 if (!mask)
124 return -EINVAL;
125
126 evlist__for_each_entry(evlist, evsel) {
127 if (evsel->core.attr.type == intel_pt_pmu->type) {
128 *res = intel_pt_masked_bits(mask, evsel->core.attr.config);
129 return 0;
130 }
131 }
132
133 return -EINVAL;
134 }
135
136 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
137 struct evlist *evlist)
138 {
139 u64 val;
140 int err, topa_multiple_entries;
141 size_t psb_period;
142
143 if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
144 "%d", &topa_multiple_entries) != 1)
145 topa_multiple_entries = 0;
146
147 /*
148 * Use caps/topa_multiple_entries to indicate early hardware that had
149 * extra frequent PSBs.
150 */
151 if (!topa_multiple_entries) {
152 psb_period = 256;
153 goto out;
154 }
155
156 err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
157 if (err)
158 val = 0;
159
160 psb_period = 1 << (val + 11);
161 out:
162 pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
163 return psb_period;
164 }
165
166 static int intel_pt_pick_bit(int bits, int target)
167 {
168 int pos, pick = -1;
169
170 for (pos = 0; bits; bits >>= 1, pos++) {
171 if (bits & 1) {
172 if (pos <= target || pick < 0)
173 pick = pos;
174 if (pos >= target)
175 break;
176 }
177 }
178
179 return pick;
180 }
181
182 static u64 intel_pt_default_config(const struct perf_pmu *intel_pt_pmu)
183 {
184 char buf[256];
185 int mtc, mtc_periods = 0, mtc_period;
186 int psb_cyc, psb_periods, psb_period;
187 int pos = 0;
188 u64 config;
189 char c;
190 int dirfd;
191
192 dirfd = perf_pmu__event_source_devices_fd();
193
194 pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
195
196 if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/mtc", "%d",
197 &mtc) != 1)
198 mtc = 1;
199
200 if (mtc) {
201 if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/mtc_periods", "%x",
202 &mtc_periods) != 1)
203 mtc_periods = 0;
204 if (mtc_periods) {
205 mtc_period = intel_pt_pick_bit(mtc_periods, 3);
206 pos += scnprintf(buf + pos, sizeof(buf) - pos,
207 ",mtc,mtc_period=%d", mtc_period);
208 }
209 }
210
211 if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/psb_cyc", "%d",
212 &psb_cyc) != 1)
213 psb_cyc = 1;
214
215 if (psb_cyc && mtc_periods) {
216 if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/psb_periods", "%x",
217 &psb_periods) != 1)
218 psb_periods = 0;
219 if (psb_periods) {
220 psb_period = intel_pt_pick_bit(psb_periods, 3);
221 pos += scnprintf(buf + pos, sizeof(buf) - pos,
222 ",psb_period=%d", psb_period);
223 }
224 }
225
226 if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/pt", "%c", &c) == 1 &&
227 perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/branch", "%c", &c) == 1)
228 pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
229
230 pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
231
232 intel_pt_parse_terms(intel_pt_pmu, buf, &config);
233
234 close(dirfd);
235 return config;
236 }
237
238 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
239 struct record_opts *opts,
240 const char *str)
241 {
242 struct intel_pt_recording *ptr =
243 container_of(itr, struct intel_pt_recording, itr);
244 unsigned long long snapshot_size = 0;
245 char *endptr;
246
247 if (str) {
248 snapshot_size = strtoull(str, &endptr, 0);
249 if (*endptr || snapshot_size > SIZE_MAX)
250 return -1;
251 }
252
253 opts->auxtrace_snapshot_mode = true;
254 opts->auxtrace_snapshot_size = snapshot_size;
255
256 ptr->snapshot_size = snapshot_size;
257
258 return 0;
259 }
260
261 void intel_pt_pmu_default_config(const struct perf_pmu *intel_pt_pmu,
262 struct perf_event_attr *attr)
263 {
264 static u64 config;
265 static bool initialized;
266
267 if (!initialized) {
268 config = intel_pt_default_config(intel_pt_pmu);
269 initialized = true;
270 }
271 attr->config = config;
272 }
273
274 static const char *intel_pt_find_filter(struct evlist *evlist,
275 struct perf_pmu *intel_pt_pmu)
276 {
277 struct evsel *evsel;
278
279 evlist__for_each_entry(evlist, evsel) {
280 if (evsel->core.attr.type == intel_pt_pmu->type)
281 return evsel->filter;
282 }
283
284 return NULL;
285 }
286
287 static size_t intel_pt_filter_bytes(const char *filter)
288 {
289 size_t len = filter ? strlen(filter) : 0;
290
291 return len ? roundup(len + 1, 8) : 0;
292 }
293
294 static size_t
295 intel_pt_info_priv_size(struct auxtrace_record *itr, struct evlist *evlist)
296 {
297 struct intel_pt_recording *ptr =
298 container_of(itr, struct intel_pt_recording, itr);
299 const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
300
301 ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
302 intel_pt_filter_bytes(filter);
303 ptr->priv_size += sizeof(u64); /* Cap Event Trace */
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 int event_trace;
334 __u64 *info;
335 int err;
336
337 if (priv_size != ptr->priv_size)
338 return -EINVAL;
339
340 intel_pt_parse_terms(intel_pt_pmu, "tsc", &tsc_bit);
341 intel_pt_parse_terms(intel_pt_pmu, "noretcomp", &noretcomp_bit);
342 intel_pt_parse_terms(intel_pt_pmu, "mtc", &mtc_bit);
343 mtc_freq_bits = perf_pmu__format_bits(intel_pt_pmu, "mtc_period");
344 intel_pt_parse_terms(intel_pt_pmu, "cyc", &cyc_bit);
345
346 intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
347
348 if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
349 "%lu", &max_non_turbo_ratio) != 1)
350 max_non_turbo_ratio = 0;
351 if (perf_pmu__scan_file(intel_pt_pmu, "caps/event_trace",
352 "%d", &event_trace) != 1)
353 event_trace = 0;
354
355 filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
356 filter_str_len = filter ? strlen(filter) : 0;
357
358 if (!session->evlist->core.nr_mmaps)
359 return -EINVAL;
360
361 pc = session->evlist->mmap[0].core.base;
362 if (pc) {
363 err = perf_read_tsc_conversion(pc, &tc);
364 if (err) {
365 if (err != -EOPNOTSUPP)
366 return err;
367 } else {
368 cap_user_time_zero = tc.time_mult != 0;
369 }
370 if (!cap_user_time_zero)
371 ui__warning("Intel Processor Trace: TSC not available\n");
372 }
373
374 per_cpu_mmaps = !perf_cpu_map__is_any_cpu_or_is_empty(session->evlist->core.user_requested_cpus);
375
376 auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
377 auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
378 auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
379 auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
380 auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
381 auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
382 auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
383 auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
384 auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
385 auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
386 auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
387 auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
388 auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
389 auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
390 auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
391 auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
392 auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
393 auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
394
395 info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
396
397 if (filter_str_len) {
398 size_t len = intel_pt_filter_bytes(filter);
399
400 strncpy((char *)info, filter, len);
401 info += len >> 3;
402 }
403
404 *info++ = event_trace;
405
406 return 0;
407 }
408
409 #ifdef HAVE_LIBTRACEEVENT
410 static int intel_pt_track_switches(struct evlist *evlist)
411 {
412 const char *sched_switch = "sched:sched_switch";
413 struct evsel *evsel;
414 int err;
415
416 if (!evlist__can_select_event(evlist, sched_switch))
417 return -EPERM;
418
419 evsel = evlist__add_sched_switch(evlist, true);
420 if (IS_ERR(evsel)) {
421 err = PTR_ERR(evsel);
422 pr_debug2("%s: failed to create %s, error = %d\n",
423 __func__, sched_switch, err);
424 return err;
425 }
426
427 evsel->immediate = true;
428
429 return 0;
430 }
431 #endif
432
433 static bool intel_pt_exclude_guest(void)
434 {
435 int pt_mode;
436
437 if (sysfs__read_int("module/kvm_intel/parameters/pt_mode", &pt_mode))
438 pt_mode = 0;
439
440 return pt_mode == 1;
441 }
442
443 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
444 {
445 unsigned int val, last = 0, state = 1;
446 int p = 0;
447
448 str[0] = '\0';
449
450 for (val = 0; val <= 64; val++, valid >>= 1) {
451 if (valid & 1) {
452 last = val;
453 switch (state) {
454 case 0:
455 p += scnprintf(str + p, len - p, ",");
456 /* Fall through */
457 case 1:
458 p += scnprintf(str + p, len - p, "%u", val);
459 state = 2;
460 break;
461 case 2:
462 state = 3;
463 break;
464 case 3:
465 state = 4;
466 break;
467 default:
468 break;
469 }
470 } else {
471 switch (state) {
472 case 3:
473 p += scnprintf(str + p, len - p, ",%u", last);
474 state = 0;
475 break;
476 case 4:
477 p += scnprintf(str + p, len - p, "-%u", last);
478 state = 0;
479 break;
480 default:
481 break;
482 }
483 if (state != 1)
484 state = 0;
485 }
486 }
487 }
488
489 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu, int dirfd,
490 const char *caps, const char *name,
491 const char *supported, u64 config)
492 {
493 char valid_str[256];
494 unsigned int shift;
495 unsigned long long valid;
496 u64 bits;
497 int ok;
498
499 if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, caps, "%llx", &valid) != 1)
500 valid = 0;
501
502 if (supported &&
503 perf_pmu__scan_file_at(intel_pt_pmu, dirfd, supported, "%d", &ok) == 1 && !ok)
504 valid = 0;
505
506 valid |= 1;
507
508 bits = perf_pmu__format_bits(intel_pt_pmu, name);
509
510 config &= bits;
511
512 for (shift = 0; bits && !(bits & 1); shift++)
513 bits >>= 1;
514
515 config >>= shift;
516
517 if (config > 63)
518 goto out_err;
519
520 if (valid & (1 << config))
521 return 0;
522 out_err:
523 intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
524 pr_err("Invalid %s for %s. Valid values are: %s\n",
525 name, INTEL_PT_PMU_NAME, valid_str);
526 return -EINVAL;
527 }
528
529 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
530 struct evsel *evsel)
531 {
532 int err, dirfd;
533 char c;
534
535 if (!evsel)
536 return 0;
537
538 dirfd = perf_pmu__event_source_devices_fd();
539 if (dirfd < 0)
540 return dirfd;
541
542 /*
543 * If supported, force pass-through config term (pt=1) even if user
544 * sets pt=0, which avoids senseless kernel errors.
545 */
546 if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/pt", "%c", &c) == 1 &&
547 !(evsel->core.attr.config & 1)) {
548 pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
549 evsel->core.attr.config |= 1;
550 }
551
552 err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/cycle_thresholds",
553 "cyc_thresh", "caps/psb_cyc",
554 evsel->core.attr.config);
555 if (err)
556 goto out;
557
558 err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/mtc_periods",
559 "mtc_period", "caps/mtc",
560 evsel->core.attr.config);
561 if (err)
562 goto out;
563
564 err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/psb_periods",
565 "psb_period", "caps/psb_cyc",
566 evsel->core.attr.config);
567
568 out:
569 close(dirfd);
570 return err;
571 }
572
573 static void intel_pt_min_max_sample_sz(struct evlist *evlist,
574 size_t *min_sz, size_t *max_sz)
575 {
576 struct evsel *evsel;
577
578 evlist__for_each_entry(evlist, evsel) {
579 size_t sz = evsel->core.attr.aux_sample_size;
580
581 if (!sz)
582 continue;
583 if (min_sz && (sz < *min_sz || !*min_sz))
584 *min_sz = sz;
585 if (max_sz && sz > *max_sz)
586 *max_sz = sz;
587 }
588 }
589
590 /*
591 * Currently, there is not enough information to disambiguate different PEBS
592 * events, so only allow one.
593 */
594 static bool intel_pt_too_many_aux_output(struct evlist *evlist)
595 {
596 struct evsel *evsel;
597 int aux_output_cnt = 0;
598
599 evlist__for_each_entry(evlist, evsel)
600 aux_output_cnt += !!evsel->core.attr.aux_output;
601
602 if (aux_output_cnt > 1) {
603 pr_err(INTEL_PT_PMU_NAME " supports at most one event with aux-output\n");
604 return true;
605 }
606
607 return false;
608 }
609
610 static int intel_pt_recording_options(struct auxtrace_record *itr,
611 struct evlist *evlist,
612 struct record_opts *opts)
613 {
614 struct intel_pt_recording *ptr =
615 container_of(itr, struct intel_pt_recording, itr);
616 struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
617 bool have_timing_info, need_immediate = false;
618 struct evsel *evsel, *intel_pt_evsel = NULL;
619 const struct perf_cpu_map *cpus = evlist->core.user_requested_cpus;
620 bool privileged = perf_event_paranoid_check(-1);
621 u64 tsc_bit;
622 int err;
623
624 ptr->evlist = evlist;
625 ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
626
627 evlist__for_each_entry(evlist, evsel) {
628 if (evsel->core.attr.type == intel_pt_pmu->type) {
629 if (intel_pt_evsel) {
630 pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
631 return -EINVAL;
632 }
633 evsel->core.attr.freq = 0;
634 evsel->core.attr.sample_period = 1;
635 evsel->core.attr.exclude_guest = intel_pt_exclude_guest();
636 evsel->no_aux_samples = true;
637 evsel->needs_auxtrace_mmap = true;
638 intel_pt_evsel = evsel;
639 opts->full_auxtrace = true;
640 }
641 }
642
643 if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
644 pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
645 return -EINVAL;
646 }
647
648 if (opts->auxtrace_snapshot_mode && opts->auxtrace_sample_mode) {
649 pr_err("Snapshot mode (" INTEL_PT_PMU_NAME " PMU) and sample trace cannot be used together\n");
650 return -EINVAL;
651 }
652
653 if (opts->use_clockid) {
654 pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
655 return -EINVAL;
656 }
657
658 if (intel_pt_too_many_aux_output(evlist))
659 return -EINVAL;
660
661 if (!opts->full_auxtrace)
662 return 0;
663
664 if (opts->auxtrace_sample_mode)
665 evsel__set_config_if_unset(intel_pt_pmu, intel_pt_evsel,
666 "psb_period", 0);
667
668 err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
669 if (err)
670 return err;
671
672 /* Set default sizes for snapshot mode */
673 if (opts->auxtrace_snapshot_mode) {
674 size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
675
676 if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
677 if (privileged) {
678 opts->auxtrace_mmap_pages = MiB(4) / page_size;
679 } else {
680 opts->auxtrace_mmap_pages = KiB(128) / page_size;
681 if (opts->mmap_pages == UINT_MAX)
682 opts->mmap_pages = KiB(256) / page_size;
683 }
684 } else if (!opts->auxtrace_mmap_pages && !privileged &&
685 opts->mmap_pages == UINT_MAX) {
686 opts->mmap_pages = KiB(256) / page_size;
687 }
688 if (!opts->auxtrace_snapshot_size)
689 opts->auxtrace_snapshot_size =
690 opts->auxtrace_mmap_pages * (size_t)page_size;
691 if (!opts->auxtrace_mmap_pages) {
692 size_t sz = opts->auxtrace_snapshot_size;
693
694 sz = round_up(sz, page_size) / page_size;
695 opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
696 }
697 if (opts->auxtrace_snapshot_size >
698 opts->auxtrace_mmap_pages * (size_t)page_size) {
699 pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
700 opts->auxtrace_snapshot_size,
701 opts->auxtrace_mmap_pages * (size_t)page_size);
702 return -EINVAL;
703 }
704 if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
705 pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
706 return -EINVAL;
707 }
708 pr_debug2("Intel PT snapshot size: %zu\n",
709 opts->auxtrace_snapshot_size);
710 if (psb_period &&
711 opts->auxtrace_snapshot_size <= psb_period +
712 INTEL_PT_PSB_PERIOD_NEAR)
713 ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
714 opts->auxtrace_snapshot_size, psb_period);
715 }
716
717 /* Set default sizes for sample mode */
718 if (opts->auxtrace_sample_mode) {
719 size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
720 size_t min_sz = 0, max_sz = 0;
721
722 intel_pt_min_max_sample_sz(evlist, &min_sz, &max_sz);
723 if (!opts->auxtrace_mmap_pages && !privileged &&
724 opts->mmap_pages == UINT_MAX)
725 opts->mmap_pages = KiB(256) / page_size;
726 if (!opts->auxtrace_mmap_pages) {
727 size_t sz = round_up(max_sz, page_size) / page_size;
728
729 opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
730 }
731 if (max_sz > opts->auxtrace_mmap_pages * (size_t)page_size) {
732 pr_err("Sample size %zu must not be greater than AUX area tracing mmap size %zu\n",
733 max_sz,
734 opts->auxtrace_mmap_pages * (size_t)page_size);
735 return -EINVAL;
736 }
737 pr_debug2("Intel PT min. sample size: %zu max. sample size: %zu\n",
738 min_sz, max_sz);
739 if (psb_period &&
740 min_sz <= psb_period + INTEL_PT_PSB_PERIOD_NEAR)
741 ui__warning("Intel PT sample size (%zu) may be too small for PSB period (%zu)\n",
742 min_sz, psb_period);
743 }
744
745 /* Set default sizes for full trace mode */
746 if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
747 if (privileged) {
748 opts->auxtrace_mmap_pages = MiB(4) / page_size;
749 } else {
750 opts->auxtrace_mmap_pages = KiB(128) / page_size;
751 if (opts->mmap_pages == UINT_MAX)
752 opts->mmap_pages = KiB(256) / page_size;
753 }
754 }
755
756 /* Validate auxtrace_mmap_pages */
757 if (opts->auxtrace_mmap_pages) {
758 size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
759 size_t min_sz;
760
761 if (opts->auxtrace_snapshot_mode || opts->auxtrace_sample_mode)
762 min_sz = KiB(4);
763 else
764 min_sz = KiB(8);
765
766 if (sz < min_sz || !is_power_of_2(sz)) {
767 pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
768 min_sz / 1024);
769 return -EINVAL;
770 }
771 }
772
773 if (!opts->auxtrace_snapshot_mode && !opts->auxtrace_sample_mode) {
774 size_t aw = opts->auxtrace_mmap_pages * (size_t)page_size / 4;
775 u32 aux_watermark = aw > UINT_MAX ? UINT_MAX : aw;
776
777 intel_pt_evsel->core.attr.aux_watermark = aux_watermark;
778 }
779
780 intel_pt_parse_terms(intel_pt_pmu, "tsc", &tsc_bit);
781
782 if (opts->full_auxtrace && (intel_pt_evsel->core.attr.config & tsc_bit))
783 have_timing_info = true;
784 else
785 have_timing_info = false;
786
787 /*
788 * Per-cpu recording needs sched_switch events to distinguish different
789 * threads.
790 */
791 if (have_timing_info && !perf_cpu_map__is_any_cpu_or_is_empty(cpus) &&
792 !record_opts__no_switch_events(opts)) {
793 if (perf_can_record_switch_events()) {
794 bool cpu_wide = !target__none(&opts->target) &&
795 !target__has_task(&opts->target);
796
797 if (!cpu_wide && perf_can_record_cpu_wide()) {
798 struct evsel *switch_evsel;
799
800 switch_evsel = evlist__add_dummy_on_all_cpus(evlist);
801 if (!switch_evsel)
802 return -ENOMEM;
803
804 switch_evsel->core.attr.context_switch = 1;
805 switch_evsel->immediate = true;
806
807 evsel__set_sample_bit(switch_evsel, TID);
808 evsel__set_sample_bit(switch_evsel, TIME);
809 evsel__set_sample_bit(switch_evsel, CPU);
810 evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
811
812 opts->record_switch_events = false;
813 ptr->have_sched_switch = 3;
814 } else {
815 opts->record_switch_events = true;
816 need_immediate = true;
817 if (cpu_wide)
818 ptr->have_sched_switch = 3;
819 else
820 ptr->have_sched_switch = 2;
821 }
822 } else {
823 #ifdef HAVE_LIBTRACEEVENT
824 err = intel_pt_track_switches(evlist);
825 if (err == -EPERM)
826 pr_debug2("Unable to select sched:sched_switch\n");
827 else if (err)
828 return err;
829 else
830 ptr->have_sched_switch = 1;
831 #endif
832 }
833 }
834
835 if (have_timing_info && !intel_pt_evsel->core.attr.exclude_kernel &&
836 perf_can_record_text_poke_events() && perf_can_record_cpu_wide())
837 opts->text_poke = true;
838
839 if (intel_pt_evsel) {
840 /*
841 * To obtain the auxtrace buffer file descriptor, the auxtrace
842 * event must come first.
843 */
844 evlist__to_front(evlist, intel_pt_evsel);
845 /*
846 * In the case of per-cpu mmaps, we need the CPU on the
847 * AUX event.
848 */
849 if (!perf_cpu_map__is_any_cpu_or_is_empty(cpus))
850 evsel__set_sample_bit(intel_pt_evsel, CPU);
851 }
852
853 /* Add dummy event to keep tracking */
854 if (opts->full_auxtrace) {
855 bool need_system_wide_tracking;
856 struct evsel *tracking_evsel;
857
858 /*
859 * User space tasks can migrate between CPUs, so when tracing
860 * selected CPUs, sideband for all CPUs is still needed.
861 */
862 need_system_wide_tracking = opts->target.cpu_list &&
863 !intel_pt_evsel->core.attr.exclude_user;
864
865 tracking_evsel = evlist__add_aux_dummy(evlist, need_system_wide_tracking);
866 if (!tracking_evsel)
867 return -ENOMEM;
868
869 evlist__set_tracking_event(evlist, tracking_evsel);
870
871 if (need_immediate)
872 tracking_evsel->immediate = true;
873
874 /* In per-cpu case, always need the time of mmap events etc */
875 if (!perf_cpu_map__is_any_cpu_or_is_empty(cpus)) {
876 evsel__set_sample_bit(tracking_evsel, TIME);
877 /* And the CPU for switch events */
878 evsel__set_sample_bit(tracking_evsel, CPU);
879 }
880 evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
881 }
882
883 /*
884 * Warn the user when we do not have enough information to decode i.e.
885 * per-cpu with no sched_switch (except workload-only).
886 */
887 if (!ptr->have_sched_switch && !perf_cpu_map__is_any_cpu_or_is_empty(cpus) &&
888 !target__none(&opts->target) &&
889 !intel_pt_evsel->core.attr.exclude_user)
890 ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
891
892 return 0;
893 }
894
895 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
896 {
897 struct intel_pt_recording *ptr =
898 container_of(itr, struct intel_pt_recording, itr);
899 struct evsel *evsel;
900
901 evlist__for_each_entry(ptr->evlist, evsel) {
902 if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
903 return evsel__disable(evsel);
904 }
905 return -EINVAL;
906 }
907
908 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
909 {
910 struct intel_pt_recording *ptr =
911 container_of(itr, struct intel_pt_recording, itr);
912 struct evsel *evsel;
913
914 evlist__for_each_entry(ptr->evlist, evsel) {
915 if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
916 return evsel__enable(evsel);
917 }
918 return -EINVAL;
919 }
920
921 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
922 {
923 const size_t sz = sizeof(struct intel_pt_snapshot_ref);
924 int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
925 struct intel_pt_snapshot_ref *refs;
926
927 if (!new_cnt)
928 new_cnt = 16;
929
930 while (new_cnt <= idx)
931 new_cnt *= 2;
932
933 refs = calloc(new_cnt, sz);
934 if (!refs)
935 return -ENOMEM;
936
937 memcpy(refs, ptr->snapshot_refs, cnt * sz);
938
939 ptr->snapshot_refs = refs;
940 ptr->snapshot_ref_cnt = new_cnt;
941
942 return 0;
943 }
944
945 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
946 {
947 int i;
948
949 for (i = 0; i < ptr->snapshot_ref_cnt; i++)
950 zfree(&ptr->snapshot_refs[i].ref_buf);
951 zfree(&ptr->snapshot_refs);
952 }
953
954 static void intel_pt_recording_free(struct auxtrace_record *itr)
955 {
956 struct intel_pt_recording *ptr =
957 container_of(itr, struct intel_pt_recording, itr);
958
959 intel_pt_free_snapshot_refs(ptr);
960 free(ptr);
961 }
962
963 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
964 size_t snapshot_buf_size)
965 {
966 size_t ref_buf_size = ptr->snapshot_ref_buf_size;
967 void *ref_buf;
968
969 ref_buf = zalloc(ref_buf_size);
970 if (!ref_buf)
971 return -ENOMEM;
972
973 ptr->snapshot_refs[idx].ref_buf = ref_buf;
974 ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
975
976 return 0;
977 }
978
979 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
980 size_t snapshot_buf_size)
981 {
982 const size_t max_size = 256 * 1024;
983 size_t buf_size = 0, psb_period;
984
985 if (ptr->snapshot_size <= 64 * 1024)
986 return 0;
987
988 psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
989 if (psb_period)
990 buf_size = psb_period * 2;
991
992 if (!buf_size || buf_size > max_size)
993 buf_size = max_size;
994
995 if (buf_size >= snapshot_buf_size)
996 return 0;
997
998 if (buf_size >= ptr->snapshot_size / 2)
999 return 0;
1000
1001 return buf_size;
1002 }
1003
1004 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
1005 size_t snapshot_buf_size)
1006 {
1007 if (ptr->snapshot_init_done)
1008 return 0;
1009
1010 ptr->snapshot_init_done = true;
1011
1012 ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
1013 snapshot_buf_size);
1014
1015 return 0;
1016 }
1017
1018 /**
1019 * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
1020 * @buf1: first buffer
1021 * @compare_size: number of bytes to compare
1022 * @buf2: second buffer (a circular buffer)
1023 * @offs2: offset in second buffer
1024 * @buf2_size: size of second buffer
1025 *
1026 * The comparison allows for the possibility that the bytes to compare in the
1027 * circular buffer are not contiguous. It is assumed that @compare_size <=
1028 * @buf2_size. This function returns %false if the bytes are identical, %true
1029 * otherwise.
1030 */
1031 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
1032 void *buf2, size_t offs2, size_t buf2_size)
1033 {
1034 size_t end2 = offs2 + compare_size, part_size;
1035
1036 if (end2 <= buf2_size)
1037 return memcmp(buf1, buf2 + offs2, compare_size);
1038
1039 part_size = end2 - buf2_size;
1040 if (memcmp(buf1, buf2 + offs2, part_size))
1041 return true;
1042
1043 compare_size -= part_size;
1044
1045 return memcmp(buf1 + part_size, buf2, compare_size);
1046 }
1047
1048 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
1049 size_t ref_size, size_t buf_size,
1050 void *data, size_t head)
1051 {
1052 size_t ref_end = ref_offset + ref_size;
1053
1054 if (ref_end > buf_size) {
1055 if (head > ref_offset || head < ref_end - buf_size)
1056 return true;
1057 } else if (head > ref_offset && head < ref_end) {
1058 return true;
1059 }
1060
1061 return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
1062 buf_size);
1063 }
1064
1065 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
1066 void *data, size_t head)
1067 {
1068 if (head >= ref_size) {
1069 memcpy(ref_buf, data + head - ref_size, ref_size);
1070 } else {
1071 memcpy(ref_buf, data, head);
1072 ref_size -= head;
1073 memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
1074 }
1075 }
1076
1077 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
1078 struct auxtrace_mmap *mm, unsigned char *data,
1079 u64 head)
1080 {
1081 struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
1082 bool wrapped;
1083
1084 wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
1085 ptr->snapshot_ref_buf_size, mm->len,
1086 data, head);
1087
1088 intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
1089 data, head);
1090
1091 return wrapped;
1092 }
1093
1094 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
1095 {
1096 int i, a, b;
1097
1098 b = buf_size >> 3;
1099 a = b - 512;
1100 if (a < 0)
1101 a = 0;
1102
1103 for (i = a; i < b; i++) {
1104 if (data[i])
1105 return true;
1106 }
1107
1108 return false;
1109 }
1110
1111 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1112 struct auxtrace_mmap *mm, unsigned char *data,
1113 u64 *head, u64 *old)
1114 {
1115 struct intel_pt_recording *ptr =
1116 container_of(itr, struct intel_pt_recording, itr);
1117 bool wrapped;
1118 int err;
1119
1120 pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1121 __func__, idx, (size_t)*old, (size_t)*head);
1122
1123 err = intel_pt_snapshot_init(ptr, mm->len);
1124 if (err)
1125 goto out_err;
1126
1127 if (idx >= ptr->snapshot_ref_cnt) {
1128 err = intel_pt_alloc_snapshot_refs(ptr, idx);
1129 if (err)
1130 goto out_err;
1131 }
1132
1133 if (ptr->snapshot_ref_buf_size) {
1134 if (!ptr->snapshot_refs[idx].ref_buf) {
1135 err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1136 if (err)
1137 goto out_err;
1138 }
1139 wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1140 } else {
1141 wrapped = ptr->snapshot_refs[idx].wrapped;
1142 if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1143 ptr->snapshot_refs[idx].wrapped = true;
1144 wrapped = true;
1145 }
1146 }
1147
1148 /*
1149 * In full trace mode 'head' continually increases. However in snapshot
1150 * mode 'head' is an offset within the buffer. Here 'old' and 'head'
1151 * are adjusted to match the full trace case which expects that 'old' is
1152 * always less than 'head'.
1153 */
1154 if (wrapped) {
1155 *old = *head;
1156 *head += mm->len;
1157 } else {
1158 if (mm->mask)
1159 *old &= mm->mask;
1160 else
1161 *old %= mm->len;
1162 if (*old > *head)
1163 *head += mm->len;
1164 }
1165
1166 pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1167 __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1168
1169 return 0;
1170
1171 out_err:
1172 pr_err("%s: failed, error %d\n", __func__, err);
1173 return err;
1174 }
1175
1176 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1177 {
1178 return rdtsc();
1179 }
1180
1181 struct auxtrace_record *intel_pt_recording_init(int *err)
1182 {
1183 struct perf_pmu *intel_pt_pmu = perf_pmus__find(INTEL_PT_PMU_NAME);
1184 struct intel_pt_recording *ptr;
1185
1186 if (!intel_pt_pmu)
1187 return NULL;
1188
1189 if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1190 *err = -errno;
1191 return NULL;
1192 }
1193
1194 ptr = zalloc(sizeof(struct intel_pt_recording));
1195 if (!ptr) {
1196 *err = -ENOMEM;
1197 return NULL;
1198 }
1199
1200 ptr->intel_pt_pmu = intel_pt_pmu;
1201 ptr->itr.recording_options = intel_pt_recording_options;
1202 ptr->itr.info_priv_size = intel_pt_info_priv_size;
1203 ptr->itr.info_fill = intel_pt_info_fill;
1204 ptr->itr.free = intel_pt_recording_free;
1205 ptr->itr.snapshot_start = intel_pt_snapshot_start;
1206 ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1207 ptr->itr.find_snapshot = intel_pt_find_snapshot;
1208 ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1209 ptr->itr.reference = intel_pt_reference;
1210 ptr->itr.read_finish = auxtrace_record__read_finish;
1211 /*
1212 * Decoding starts at a PSB packet. Minimum PSB period is 2K so 4K
1213 * should give at least 1 PSB per sample.
1214 */
1215 ptr->itr.default_aux_sample_size = 4096;
1216 return &ptr->itr;
1217 }
Kernel DRM miscellaneous fixes and cross-tree changes