Merge tag 'ccn/fixes-for-4.8-v2' of git://git.linaro.org/people/pawel.moll/linux...
[cascardo/linux.git] / tools / perf / arch / x86 / util / intel-pt.c
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_entry(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, need_immediate = false;
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_entry(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                                 need_immediate = true;
659                                 if (cpu_wide)
660                                         ptr->have_sched_switch = 3;
661                                 else
662                                         ptr->have_sched_switch = 2;
663                         }
664                 } else {
665                         err = intel_pt_track_switches(evlist);
666                         if (err == -EPERM)
667                                 pr_debug2("Unable to select sched:sched_switch\n");
668                         else if (err)
669                                 return err;
670                         else
671                                 ptr->have_sched_switch = 1;
672                 }
673         }
674
675         if (intel_pt_evsel) {
676                 /*
677                  * To obtain the auxtrace buffer file descriptor, the auxtrace
678                  * event must come first.
679                  */
680                 perf_evlist__to_front(evlist, intel_pt_evsel);
681                 /*
682                  * In the case of per-cpu mmaps, we need the CPU on the
683                  * AUX event.
684                  */
685                 if (!cpu_map__empty(cpus))
686                         perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
687         }
688
689         /* Add dummy event to keep tracking */
690         if (opts->full_auxtrace) {
691                 struct perf_evsel *tracking_evsel;
692
693                 err = parse_events(evlist, "dummy:u", NULL);
694                 if (err)
695                         return err;
696
697                 tracking_evsel = perf_evlist__last(evlist);
698
699                 perf_evlist__set_tracking_event(evlist, tracking_evsel);
700
701                 tracking_evsel->attr.freq = 0;
702                 tracking_evsel->attr.sample_period = 1;
703
704                 if (need_immediate)
705                         tracking_evsel->immediate = true;
706
707                 /* In per-cpu case, always need the time of mmap events etc */
708                 if (!cpu_map__empty(cpus)) {
709                         perf_evsel__set_sample_bit(tracking_evsel, TIME);
710                         /* And the CPU for switch events */
711                         perf_evsel__set_sample_bit(tracking_evsel, CPU);
712                 }
713         }
714
715         /*
716          * Warn the user when we do not have enough information to decode i.e.
717          * per-cpu with no sched_switch (except workload-only).
718          */
719         if (!ptr->have_sched_switch && !cpu_map__empty(cpus) &&
720             !target__none(&opts->target))
721                 ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
722
723         return 0;
724 }
725
726 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
727 {
728         struct intel_pt_recording *ptr =
729                         container_of(itr, struct intel_pt_recording, itr);
730         struct perf_evsel *evsel;
731
732         evlist__for_each_entry(ptr->evlist, evsel) {
733                 if (evsel->attr.type == ptr->intel_pt_pmu->type)
734                         return perf_evsel__disable(evsel);
735         }
736         return -EINVAL;
737 }
738
739 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
740 {
741         struct intel_pt_recording *ptr =
742                         container_of(itr, struct intel_pt_recording, itr);
743         struct perf_evsel *evsel;
744
745         evlist__for_each_entry(ptr->evlist, evsel) {
746                 if (evsel->attr.type == ptr->intel_pt_pmu->type)
747                         return perf_evsel__enable(evsel);
748         }
749         return -EINVAL;
750 }
751
752 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
753 {
754         const size_t sz = sizeof(struct intel_pt_snapshot_ref);
755         int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
756         struct intel_pt_snapshot_ref *refs;
757
758         if (!new_cnt)
759                 new_cnt = 16;
760
761         while (new_cnt <= idx)
762                 new_cnt *= 2;
763
764         refs = calloc(new_cnt, sz);
765         if (!refs)
766                 return -ENOMEM;
767
768         memcpy(refs, ptr->snapshot_refs, cnt * sz);
769
770         ptr->snapshot_refs = refs;
771         ptr->snapshot_ref_cnt = new_cnt;
772
773         return 0;
774 }
775
776 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
777 {
778         int i;
779
780         for (i = 0; i < ptr->snapshot_ref_cnt; i++)
781                 zfree(&ptr->snapshot_refs[i].ref_buf);
782         zfree(&ptr->snapshot_refs);
783 }
784
785 static void intel_pt_recording_free(struct auxtrace_record *itr)
786 {
787         struct intel_pt_recording *ptr =
788                         container_of(itr, struct intel_pt_recording, itr);
789
790         intel_pt_free_snapshot_refs(ptr);
791         free(ptr);
792 }
793
794 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
795                                        size_t snapshot_buf_size)
796 {
797         size_t ref_buf_size = ptr->snapshot_ref_buf_size;
798         void *ref_buf;
799
800         ref_buf = zalloc(ref_buf_size);
801         if (!ref_buf)
802                 return -ENOMEM;
803
804         ptr->snapshot_refs[idx].ref_buf = ref_buf;
805         ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
806
807         return 0;
808 }
809
810 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
811                                              size_t snapshot_buf_size)
812 {
813         const size_t max_size = 256 * 1024;
814         size_t buf_size = 0, psb_period;
815
816         if (ptr->snapshot_size <= 64 * 1024)
817                 return 0;
818
819         psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
820         if (psb_period)
821                 buf_size = psb_period * 2;
822
823         if (!buf_size || buf_size > max_size)
824                 buf_size = max_size;
825
826         if (buf_size >= snapshot_buf_size)
827                 return 0;
828
829         if (buf_size >= ptr->snapshot_size / 2)
830                 return 0;
831
832         return buf_size;
833 }
834
835 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
836                                   size_t snapshot_buf_size)
837 {
838         if (ptr->snapshot_init_done)
839                 return 0;
840
841         ptr->snapshot_init_done = true;
842
843         ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
844                                                         snapshot_buf_size);
845
846         return 0;
847 }
848
849 /**
850  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
851  * @buf1: first buffer
852  * @compare_size: number of bytes to compare
853  * @buf2: second buffer (a circular buffer)
854  * @offs2: offset in second buffer
855  * @buf2_size: size of second buffer
856  *
857  * The comparison allows for the possibility that the bytes to compare in the
858  * circular buffer are not contiguous.  It is assumed that @compare_size <=
859  * @buf2_size.  This function returns %false if the bytes are identical, %true
860  * otherwise.
861  */
862 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
863                                      void *buf2, size_t offs2, size_t buf2_size)
864 {
865         size_t end2 = offs2 + compare_size, part_size;
866
867         if (end2 <= buf2_size)
868                 return memcmp(buf1, buf2 + offs2, compare_size);
869
870         part_size = end2 - buf2_size;
871         if (memcmp(buf1, buf2 + offs2, part_size))
872                 return true;
873
874         compare_size -= part_size;
875
876         return memcmp(buf1 + part_size, buf2, compare_size);
877 }
878
879 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
880                                  size_t ref_size, size_t buf_size,
881                                  void *data, size_t head)
882 {
883         size_t ref_end = ref_offset + ref_size;
884
885         if (ref_end > buf_size) {
886                 if (head > ref_offset || head < ref_end - buf_size)
887                         return true;
888         } else if (head > ref_offset && head < ref_end) {
889                 return true;
890         }
891
892         return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
893                                         buf_size);
894 }
895
896 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
897                               void *data, size_t head)
898 {
899         if (head >= ref_size) {
900                 memcpy(ref_buf, data + head - ref_size, ref_size);
901         } else {
902                 memcpy(ref_buf, data, head);
903                 ref_size -= head;
904                 memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
905         }
906 }
907
908 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
909                              struct auxtrace_mmap *mm, unsigned char *data,
910                              u64 head)
911 {
912         struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
913         bool wrapped;
914
915         wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
916                                        ptr->snapshot_ref_buf_size, mm->len,
917                                        data, head);
918
919         intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
920                           data, head);
921
922         return wrapped;
923 }
924
925 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
926 {
927         int i, a, b;
928
929         b = buf_size >> 3;
930         a = b - 512;
931         if (a < 0)
932                 a = 0;
933
934         for (i = a; i < b; i++) {
935                 if (data[i])
936                         return true;
937         }
938
939         return false;
940 }
941
942 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
943                                   struct auxtrace_mmap *mm, unsigned char *data,
944                                   u64 *head, u64 *old)
945 {
946         struct intel_pt_recording *ptr =
947                         container_of(itr, struct intel_pt_recording, itr);
948         bool wrapped;
949         int err;
950
951         pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
952                   __func__, idx, (size_t)*old, (size_t)*head);
953
954         err = intel_pt_snapshot_init(ptr, mm->len);
955         if (err)
956                 goto out_err;
957
958         if (idx >= ptr->snapshot_ref_cnt) {
959                 err = intel_pt_alloc_snapshot_refs(ptr, idx);
960                 if (err)
961                         goto out_err;
962         }
963
964         if (ptr->snapshot_ref_buf_size) {
965                 if (!ptr->snapshot_refs[idx].ref_buf) {
966                         err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
967                         if (err)
968                                 goto out_err;
969                 }
970                 wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
971         } else {
972                 wrapped = ptr->snapshot_refs[idx].wrapped;
973                 if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
974                         ptr->snapshot_refs[idx].wrapped = true;
975                         wrapped = true;
976                 }
977         }
978
979         /*
980          * In full trace mode 'head' continually increases.  However in snapshot
981          * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
982          * are adjusted to match the full trace case which expects that 'old' is
983          * always less than 'head'.
984          */
985         if (wrapped) {
986                 *old = *head;
987                 *head += mm->len;
988         } else {
989                 if (mm->mask)
990                         *old &= mm->mask;
991                 else
992                         *old %= mm->len;
993                 if (*old > *head)
994                         *head += mm->len;
995         }
996
997         pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
998                   __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
999
1000         return 0;
1001
1002 out_err:
1003         pr_err("%s: failed, error %d\n", __func__, err);
1004         return err;
1005 }
1006
1007 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1008 {
1009         return rdtsc();
1010 }
1011
1012 static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1013 {
1014         struct intel_pt_recording *ptr =
1015                         container_of(itr, struct intel_pt_recording, itr);
1016         struct perf_evsel *evsel;
1017
1018         evlist__for_each_entry(ptr->evlist, evsel) {
1019                 if (evsel->attr.type == ptr->intel_pt_pmu->type)
1020                         return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1021                                                              idx);
1022         }
1023         return -EINVAL;
1024 }
1025
1026 struct auxtrace_record *intel_pt_recording_init(int *err)
1027 {
1028         struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1029         struct intel_pt_recording *ptr;
1030
1031         if (!intel_pt_pmu)
1032                 return NULL;
1033
1034         if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1035                 *err = -errno;
1036                 return NULL;
1037         }
1038
1039         ptr = zalloc(sizeof(struct intel_pt_recording));
1040         if (!ptr) {
1041                 *err = -ENOMEM;
1042                 return NULL;
1043         }
1044
1045         ptr->intel_pt_pmu = intel_pt_pmu;
1046         ptr->itr.recording_options = intel_pt_recording_options;
1047         ptr->itr.info_priv_size = intel_pt_info_priv_size;
1048         ptr->itr.info_fill = intel_pt_info_fill;
1049         ptr->itr.free = intel_pt_recording_free;
1050         ptr->itr.snapshot_start = intel_pt_snapshot_start;
1051         ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1052         ptr->itr.find_snapshot = intel_pt_find_snapshot;
1053         ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1054         ptr->itr.reference = intel_pt_reference;
1055         ptr->itr.read_finish = intel_pt_read_finish;
1056         return &ptr->itr;
1057 }