Loading...
1#include "builtin.h"
2#include "perf.h"
3
4#include "util/evsel.h"
5#include "util/evlist.h"
6#include "util/util.h"
7#include "util/cache.h"
8#include "util/symbol.h"
9#include "util/thread.h"
10#include "util/header.h"
11#include "util/session.h"
12#include "util/intlist.h"
13#include <subcmd/parse-options.h>
14#include "util/trace-event.h"
15#include "util/debug.h"
16#include "util/tool.h"
17#include "util/stat.h"
18#include "util/top.h"
19#include "util/data.h"
20#include "util/ordered-events.h"
21
22#include <sys/prctl.h>
23#ifdef HAVE_TIMERFD_SUPPORT
24#include <sys/timerfd.h>
25#endif
26
27#include <linux/time64.h>
28#include <termios.h>
29#include <semaphore.h>
30#include <pthread.h>
31#include <math.h>
32
33#ifdef HAVE_KVM_STAT_SUPPORT
34#include "util/kvm-stat.h"
35
36void exit_event_get_key(struct perf_evsel *evsel,
37 struct perf_sample *sample,
38 struct event_key *key)
39{
40 key->info = 0;
41 key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
42}
43
44bool kvm_exit_event(struct perf_evsel *evsel)
45{
46 return !strcmp(evsel->name, kvm_exit_trace);
47}
48
49bool exit_event_begin(struct perf_evsel *evsel,
50 struct perf_sample *sample, struct event_key *key)
51{
52 if (kvm_exit_event(evsel)) {
53 exit_event_get_key(evsel, sample, key);
54 return true;
55 }
56
57 return false;
58}
59
60bool kvm_entry_event(struct perf_evsel *evsel)
61{
62 return !strcmp(evsel->name, kvm_entry_trace);
63}
64
65bool exit_event_end(struct perf_evsel *evsel,
66 struct perf_sample *sample __maybe_unused,
67 struct event_key *key __maybe_unused)
68{
69 return kvm_entry_event(evsel);
70}
71
72static const char *get_exit_reason(struct perf_kvm_stat *kvm,
73 struct exit_reasons_table *tbl,
74 u64 exit_code)
75{
76 while (tbl->reason != NULL) {
77 if (tbl->exit_code == exit_code)
78 return tbl->reason;
79 tbl++;
80 }
81
82 pr_err("unknown kvm exit code:%lld on %s\n",
83 (unsigned long long)exit_code, kvm->exit_reasons_isa);
84 return "UNKNOWN";
85}
86
87void exit_event_decode_key(struct perf_kvm_stat *kvm,
88 struct event_key *key,
89 char *decode)
90{
91 const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
92 key->key);
93
94 scnprintf(decode, decode_str_len, "%s", exit_reason);
95}
96
97static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
98{
99 struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
100
101 for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
102 if (!strcmp(events_ops->name, kvm->report_event)) {
103 kvm->events_ops = events_ops->ops;
104 return true;
105 }
106 }
107
108 return false;
109}
110
111struct vcpu_event_record {
112 int vcpu_id;
113 u64 start_time;
114 struct kvm_event *last_event;
115};
116
117
118static void init_kvm_event_record(struct perf_kvm_stat *kvm)
119{
120 unsigned int i;
121
122 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
123 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
124}
125
126#ifdef HAVE_TIMERFD_SUPPORT
127static void clear_events_cache_stats(struct list_head *kvm_events_cache)
128{
129 struct list_head *head;
130 struct kvm_event *event;
131 unsigned int i;
132 int j;
133
134 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
135 head = &kvm_events_cache[i];
136 list_for_each_entry(event, head, hash_entry) {
137 /* reset stats for event */
138 event->total.time = 0;
139 init_stats(&event->total.stats);
140
141 for (j = 0; j < event->max_vcpu; ++j) {
142 event->vcpu[j].time = 0;
143 init_stats(&event->vcpu[j].stats);
144 }
145 }
146 }
147}
148#endif
149
150static int kvm_events_hash_fn(u64 key)
151{
152 return key & (EVENTS_CACHE_SIZE - 1);
153}
154
155static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
156{
157 int old_max_vcpu = event->max_vcpu;
158 void *prev;
159
160 if (vcpu_id < event->max_vcpu)
161 return true;
162
163 while (event->max_vcpu <= vcpu_id)
164 event->max_vcpu += DEFAULT_VCPU_NUM;
165
166 prev = event->vcpu;
167 event->vcpu = realloc(event->vcpu,
168 event->max_vcpu * sizeof(*event->vcpu));
169 if (!event->vcpu) {
170 free(prev);
171 pr_err("Not enough memory\n");
172 return false;
173 }
174
175 memset(event->vcpu + old_max_vcpu, 0,
176 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
177 return true;
178}
179
180static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
181{
182 struct kvm_event *event;
183
184 event = zalloc(sizeof(*event));
185 if (!event) {
186 pr_err("Not enough memory\n");
187 return NULL;
188 }
189
190 event->key = *key;
191 init_stats(&event->total.stats);
192 return event;
193}
194
195static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
196 struct event_key *key)
197{
198 struct kvm_event *event;
199 struct list_head *head;
200
201 BUG_ON(key->key == INVALID_KEY);
202
203 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
204 list_for_each_entry(event, head, hash_entry) {
205 if (event->key.key == key->key && event->key.info == key->info)
206 return event;
207 }
208
209 event = kvm_alloc_init_event(key);
210 if (!event)
211 return NULL;
212
213 list_add(&event->hash_entry, head);
214 return event;
215}
216
217static bool handle_begin_event(struct perf_kvm_stat *kvm,
218 struct vcpu_event_record *vcpu_record,
219 struct event_key *key, u64 timestamp)
220{
221 struct kvm_event *event = NULL;
222
223 if (key->key != INVALID_KEY)
224 event = find_create_kvm_event(kvm, key);
225
226 vcpu_record->last_event = event;
227 vcpu_record->start_time = timestamp;
228 return true;
229}
230
231static void
232kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
233{
234 kvm_stats->time += time_diff;
235 update_stats(&kvm_stats->stats, time_diff);
236}
237
238static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
239{
240 struct kvm_event_stats *kvm_stats = &event->total;
241
242 if (vcpu_id != -1)
243 kvm_stats = &event->vcpu[vcpu_id];
244
245 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
246 avg_stats(&kvm_stats->stats));
247}
248
249static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
250 u64 time_diff)
251{
252 if (vcpu_id == -1) {
253 kvm_update_event_stats(&event->total, time_diff);
254 return true;
255 }
256
257 if (!kvm_event_expand(event, vcpu_id))
258 return false;
259
260 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
261 return true;
262}
263
264static bool is_child_event(struct perf_kvm_stat *kvm,
265 struct perf_evsel *evsel,
266 struct perf_sample *sample,
267 struct event_key *key)
268{
269 struct child_event_ops *child_ops;
270
271 child_ops = kvm->events_ops->child_ops;
272
273 if (!child_ops)
274 return false;
275
276 for (; child_ops->name; child_ops++) {
277 if (!strcmp(evsel->name, child_ops->name)) {
278 child_ops->get_key(evsel, sample, key);
279 return true;
280 }
281 }
282
283 return false;
284}
285
286static bool handle_child_event(struct perf_kvm_stat *kvm,
287 struct vcpu_event_record *vcpu_record,
288 struct event_key *key,
289 struct perf_sample *sample __maybe_unused)
290{
291 struct kvm_event *event = NULL;
292
293 if (key->key != INVALID_KEY)
294 event = find_create_kvm_event(kvm, key);
295
296 vcpu_record->last_event = event;
297
298 return true;
299}
300
301static bool skip_event(const char *event)
302{
303 const char * const *skip_events;
304
305 for (skip_events = kvm_skip_events; *skip_events; skip_events++)
306 if (!strcmp(event, *skip_events))
307 return true;
308
309 return false;
310}
311
312static bool handle_end_event(struct perf_kvm_stat *kvm,
313 struct vcpu_event_record *vcpu_record,
314 struct event_key *key,
315 struct perf_sample *sample)
316{
317 struct kvm_event *event;
318 u64 time_begin, time_diff;
319 int vcpu;
320
321 if (kvm->trace_vcpu == -1)
322 vcpu = -1;
323 else
324 vcpu = vcpu_record->vcpu_id;
325
326 event = vcpu_record->last_event;
327 time_begin = vcpu_record->start_time;
328
329 /* The begin event is not caught. */
330 if (!time_begin)
331 return true;
332
333 /*
334 * In some case, the 'begin event' only records the start timestamp,
335 * the actual event is recognized in the 'end event' (e.g. mmio-event).
336 */
337
338 /* Both begin and end events did not get the key. */
339 if (!event && key->key == INVALID_KEY)
340 return true;
341
342 if (!event)
343 event = find_create_kvm_event(kvm, key);
344
345 if (!event)
346 return false;
347
348 vcpu_record->last_event = NULL;
349 vcpu_record->start_time = 0;
350
351 /* seems to happen once in a while during live mode */
352 if (sample->time < time_begin) {
353 pr_debug("End time before begin time; skipping event.\n");
354 return true;
355 }
356
357 time_diff = sample->time - time_begin;
358
359 if (kvm->duration && time_diff > kvm->duration) {
360 char decode[decode_str_len];
361
362 kvm->events_ops->decode_key(kvm, &event->key, decode);
363 if (!skip_event(decode)) {
364 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
365 sample->time, sample->pid, vcpu_record->vcpu_id,
366 decode, time_diff / NSEC_PER_USEC);
367 }
368 }
369
370 return update_kvm_event(event, vcpu, time_diff);
371}
372
373static
374struct vcpu_event_record *per_vcpu_record(struct thread *thread,
375 struct perf_evsel *evsel,
376 struct perf_sample *sample)
377{
378 /* Only kvm_entry records vcpu id. */
379 if (!thread__priv(thread) && kvm_entry_event(evsel)) {
380 struct vcpu_event_record *vcpu_record;
381
382 vcpu_record = zalloc(sizeof(*vcpu_record));
383 if (!vcpu_record) {
384 pr_err("%s: Not enough memory\n", __func__);
385 return NULL;
386 }
387
388 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
389 vcpu_id_str);
390 thread__set_priv(thread, vcpu_record);
391 }
392
393 return thread__priv(thread);
394}
395
396static bool handle_kvm_event(struct perf_kvm_stat *kvm,
397 struct thread *thread,
398 struct perf_evsel *evsel,
399 struct perf_sample *sample)
400{
401 struct vcpu_event_record *vcpu_record;
402 struct event_key key = { .key = INVALID_KEY,
403 .exit_reasons = kvm->exit_reasons };
404
405 vcpu_record = per_vcpu_record(thread, evsel, sample);
406 if (!vcpu_record)
407 return true;
408
409 /* only process events for vcpus user cares about */
410 if ((kvm->trace_vcpu != -1) &&
411 (kvm->trace_vcpu != vcpu_record->vcpu_id))
412 return true;
413
414 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
415 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
416
417 if (is_child_event(kvm, evsel, sample, &key))
418 return handle_child_event(kvm, vcpu_record, &key, sample);
419
420 if (kvm->events_ops->is_end_event(evsel, sample, &key))
421 return handle_end_event(kvm, vcpu_record, &key, sample);
422
423 return true;
424}
425
426#define GET_EVENT_KEY(func, field) \
427static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
428{ \
429 if (vcpu == -1) \
430 return event->total.field; \
431 \
432 if (vcpu >= event->max_vcpu) \
433 return 0; \
434 \
435 return event->vcpu[vcpu].field; \
436}
437
438#define COMPARE_EVENT_KEY(func, field) \
439GET_EVENT_KEY(func, field) \
440static int compare_kvm_event_ ## func(struct kvm_event *one, \
441 struct kvm_event *two, int vcpu)\
442{ \
443 return get_event_ ##func(one, vcpu) > \
444 get_event_ ##func(two, vcpu); \
445}
446
447GET_EVENT_KEY(time, time);
448COMPARE_EVENT_KEY(count, stats.n);
449COMPARE_EVENT_KEY(mean, stats.mean);
450GET_EVENT_KEY(max, stats.max);
451GET_EVENT_KEY(min, stats.min);
452
453#define DEF_SORT_NAME_KEY(name, compare_key) \
454 { #name, compare_kvm_event_ ## compare_key }
455
456static struct kvm_event_key keys[] = {
457 DEF_SORT_NAME_KEY(sample, count),
458 DEF_SORT_NAME_KEY(time, mean),
459 { NULL, NULL }
460};
461
462static bool select_key(struct perf_kvm_stat *kvm)
463{
464 int i;
465
466 for (i = 0; keys[i].name; i++) {
467 if (!strcmp(keys[i].name, kvm->sort_key)) {
468 kvm->compare = keys[i].key;
469 return true;
470 }
471 }
472
473 pr_err("Unknown compare key:%s\n", kvm->sort_key);
474 return false;
475}
476
477static void insert_to_result(struct rb_root *result, struct kvm_event *event,
478 key_cmp_fun bigger, int vcpu)
479{
480 struct rb_node **rb = &result->rb_node;
481 struct rb_node *parent = NULL;
482 struct kvm_event *p;
483
484 while (*rb) {
485 p = container_of(*rb, struct kvm_event, rb);
486 parent = *rb;
487
488 if (bigger(event, p, vcpu))
489 rb = &(*rb)->rb_left;
490 else
491 rb = &(*rb)->rb_right;
492 }
493
494 rb_link_node(&event->rb, parent, rb);
495 rb_insert_color(&event->rb, result);
496}
497
498static void
499update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
500{
501 int vcpu = kvm->trace_vcpu;
502
503 kvm->total_count += get_event_count(event, vcpu);
504 kvm->total_time += get_event_time(event, vcpu);
505}
506
507static bool event_is_valid(struct kvm_event *event, int vcpu)
508{
509 return !!get_event_count(event, vcpu);
510}
511
512static void sort_result(struct perf_kvm_stat *kvm)
513{
514 unsigned int i;
515 int vcpu = kvm->trace_vcpu;
516 struct kvm_event *event;
517
518 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
519 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
520 if (event_is_valid(event, vcpu)) {
521 update_total_count(kvm, event);
522 insert_to_result(&kvm->result, event,
523 kvm->compare, vcpu);
524 }
525 }
526 }
527}
528
529/* returns left most element of result, and erase it */
530static struct kvm_event *pop_from_result(struct rb_root *result)
531{
532 struct rb_node *node = rb_first(result);
533
534 if (!node)
535 return NULL;
536
537 rb_erase(node, result);
538 return container_of(node, struct kvm_event, rb);
539}
540
541static void print_vcpu_info(struct perf_kvm_stat *kvm)
542{
543 int vcpu = kvm->trace_vcpu;
544
545 pr_info("Analyze events for ");
546
547 if (kvm->opts.target.system_wide)
548 pr_info("all VMs, ");
549 else if (kvm->opts.target.pid)
550 pr_info("pid(s) %s, ", kvm->opts.target.pid);
551 else
552 pr_info("dazed and confused on what is monitored, ");
553
554 if (vcpu == -1)
555 pr_info("all VCPUs:\n\n");
556 else
557 pr_info("VCPU %d:\n\n", vcpu);
558}
559
560static void show_timeofday(void)
561{
562 char date[64];
563 struct timeval tv;
564 struct tm ltime;
565
566 gettimeofday(&tv, NULL);
567 if (localtime_r(&tv.tv_sec, <ime)) {
568 strftime(date, sizeof(date), "%H:%M:%S", <ime);
569 pr_info("%s.%06ld", date, tv.tv_usec);
570 } else
571 pr_info("00:00:00.000000");
572
573 return;
574}
575
576static void print_result(struct perf_kvm_stat *kvm)
577{
578 char decode[decode_str_len];
579 struct kvm_event *event;
580 int vcpu = kvm->trace_vcpu;
581
582 if (kvm->live) {
583 puts(CONSOLE_CLEAR);
584 show_timeofday();
585 }
586
587 pr_info("\n\n");
588 print_vcpu_info(kvm);
589 pr_info("%*s ", decode_str_len, kvm->events_ops->name);
590 pr_info("%10s ", "Samples");
591 pr_info("%9s ", "Samples%");
592
593 pr_info("%9s ", "Time%");
594 pr_info("%11s ", "Min Time");
595 pr_info("%11s ", "Max Time");
596 pr_info("%16s ", "Avg time");
597 pr_info("\n\n");
598
599 while ((event = pop_from_result(&kvm->result))) {
600 u64 ecount, etime, max, min;
601
602 ecount = get_event_count(event, vcpu);
603 etime = get_event_time(event, vcpu);
604 max = get_event_max(event, vcpu);
605 min = get_event_min(event, vcpu);
606
607 kvm->events_ops->decode_key(kvm, &event->key, decode);
608 pr_info("%*s ", decode_str_len, decode);
609 pr_info("%10llu ", (unsigned long long)ecount);
610 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
611 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
612 pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
613 pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
614 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
615 kvm_event_rel_stddev(vcpu, event));
616 pr_info("\n");
617 }
618
619 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
620 kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
621
622 if (kvm->lost_events)
623 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
624}
625
626#ifdef HAVE_TIMERFD_SUPPORT
627static int process_lost_event(struct perf_tool *tool,
628 union perf_event *event __maybe_unused,
629 struct perf_sample *sample __maybe_unused,
630 struct machine *machine __maybe_unused)
631{
632 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
633
634 kvm->lost_events++;
635 return 0;
636}
637#endif
638
639static bool skip_sample(struct perf_kvm_stat *kvm,
640 struct perf_sample *sample)
641{
642 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
643 return true;
644
645 return false;
646}
647
648static int process_sample_event(struct perf_tool *tool,
649 union perf_event *event,
650 struct perf_sample *sample,
651 struct perf_evsel *evsel,
652 struct machine *machine)
653{
654 int err = 0;
655 struct thread *thread;
656 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
657 tool);
658
659 if (skip_sample(kvm, sample))
660 return 0;
661
662 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
663 if (thread == NULL) {
664 pr_debug("problem processing %d event, skipping it.\n",
665 event->header.type);
666 return -1;
667 }
668
669 if (!handle_kvm_event(kvm, thread, evsel, sample))
670 err = -1;
671
672 thread__put(thread);
673 return err;
674}
675
676static int cpu_isa_config(struct perf_kvm_stat *kvm)
677{
678 char buf[64], *cpuid;
679 int err;
680
681 if (kvm->live) {
682 err = get_cpuid(buf, sizeof(buf));
683 if (err != 0) {
684 pr_err("Failed to look up CPU type\n");
685 return err;
686 }
687 cpuid = buf;
688 } else
689 cpuid = kvm->session->header.env.cpuid;
690
691 if (!cpuid) {
692 pr_err("Failed to look up CPU type\n");
693 return -EINVAL;
694 }
695
696 err = cpu_isa_init(kvm, cpuid);
697 if (err == -ENOTSUP)
698 pr_err("CPU %s is not supported.\n", cpuid);
699
700 return err;
701}
702
703static bool verify_vcpu(int vcpu)
704{
705 if (vcpu != -1 && vcpu < 0) {
706 pr_err("Invalid vcpu:%d.\n", vcpu);
707 return false;
708 }
709
710 return true;
711}
712
713#ifdef HAVE_TIMERFD_SUPPORT
714/* keeping the max events to a modest level to keep
715 * the processing of samples per mmap smooth.
716 */
717#define PERF_KVM__MAX_EVENTS_PER_MMAP 25
718
719static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
720 u64 *mmap_time)
721{
722 union perf_event *event;
723 struct perf_sample sample;
724 s64 n = 0;
725 int err;
726
727 *mmap_time = ULLONG_MAX;
728 while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
729 err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
730 if (err) {
731 perf_evlist__mmap_consume(kvm->evlist, idx);
732 pr_err("Failed to parse sample\n");
733 return -1;
734 }
735
736 err = perf_session__queue_event(kvm->session, event, &sample, 0);
737 /*
738 * FIXME: Here we can't consume the event, as perf_session__queue_event will
739 * point to it, and it'll get possibly overwritten by the kernel.
740 */
741 perf_evlist__mmap_consume(kvm->evlist, idx);
742
743 if (err) {
744 pr_err("Failed to enqueue sample: %d\n", err);
745 return -1;
746 }
747
748 /* save time stamp of our first sample for this mmap */
749 if (n == 0)
750 *mmap_time = sample.time;
751
752 /* limit events per mmap handled all at once */
753 n++;
754 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
755 break;
756 }
757
758 return n;
759}
760
761static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
762{
763 int i, err, throttled = 0;
764 s64 n, ntotal = 0;
765 u64 flush_time = ULLONG_MAX, mmap_time;
766
767 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
768 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
769 if (n < 0)
770 return -1;
771
772 /* flush time is going to be the minimum of all the individual
773 * mmap times. Essentially, we flush all the samples queued up
774 * from the last pass under our minimal start time -- that leaves
775 * a very small race for samples to come in with a lower timestamp.
776 * The ioctl to return the perf_clock timestamp should close the
777 * race entirely.
778 */
779 if (mmap_time < flush_time)
780 flush_time = mmap_time;
781
782 ntotal += n;
783 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
784 throttled = 1;
785 }
786
787 /* flush queue after each round in which we processed events */
788 if (ntotal) {
789 struct ordered_events *oe = &kvm->session->ordered_events;
790
791 oe->next_flush = flush_time;
792 err = ordered_events__flush(oe, OE_FLUSH__ROUND);
793 if (err) {
794 if (kvm->lost_events)
795 pr_info("\nLost events: %" PRIu64 "\n\n",
796 kvm->lost_events);
797 return err;
798 }
799 }
800
801 return throttled;
802}
803
804static volatile int done;
805
806static void sig_handler(int sig __maybe_unused)
807{
808 done = 1;
809}
810
811static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
812{
813 struct itimerspec new_value;
814 int rc = -1;
815
816 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
817 if (kvm->timerfd < 0) {
818 pr_err("timerfd_create failed\n");
819 goto out;
820 }
821
822 new_value.it_value.tv_sec = kvm->display_time;
823 new_value.it_value.tv_nsec = 0;
824 new_value.it_interval.tv_sec = kvm->display_time;
825 new_value.it_interval.tv_nsec = 0;
826
827 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
828 pr_err("timerfd_settime failed: %d\n", errno);
829 close(kvm->timerfd);
830 goto out;
831 }
832
833 rc = 0;
834out:
835 return rc;
836}
837
838static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
839{
840 uint64_t c;
841 int rc;
842
843 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
844 if (rc < 0) {
845 if (errno == EAGAIN)
846 return 0;
847
848 pr_err("Failed to read timer fd: %d\n", errno);
849 return -1;
850 }
851
852 if (rc != sizeof(uint64_t)) {
853 pr_err("Error reading timer fd - invalid size returned\n");
854 return -1;
855 }
856
857 if (c != 1)
858 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
859
860 /* update display */
861 sort_result(kvm);
862 print_result(kvm);
863
864 /* reset counts */
865 clear_events_cache_stats(kvm->kvm_events_cache);
866 kvm->total_count = 0;
867 kvm->total_time = 0;
868 kvm->lost_events = 0;
869
870 return 0;
871}
872
873static int fd_set_nonblock(int fd)
874{
875 long arg = 0;
876
877 arg = fcntl(fd, F_GETFL);
878 if (arg < 0) {
879 pr_err("Failed to get current flags for fd %d\n", fd);
880 return -1;
881 }
882
883 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
884 pr_err("Failed to set non-block option on fd %d\n", fd);
885 return -1;
886 }
887
888 return 0;
889}
890
891static int perf_kvm__handle_stdin(void)
892{
893 int c;
894
895 c = getc(stdin);
896 if (c == 'q')
897 return 1;
898
899 return 0;
900}
901
902static int kvm_events_live_report(struct perf_kvm_stat *kvm)
903{
904 int nr_stdin, ret, err = -EINVAL;
905 struct termios save;
906
907 /* live flag must be set first */
908 kvm->live = true;
909
910 ret = cpu_isa_config(kvm);
911 if (ret < 0)
912 return ret;
913
914 if (!verify_vcpu(kvm->trace_vcpu) ||
915 !select_key(kvm) ||
916 !register_kvm_events_ops(kvm)) {
917 goto out;
918 }
919
920 set_term_quiet_input(&save);
921 init_kvm_event_record(kvm);
922
923 signal(SIGINT, sig_handler);
924 signal(SIGTERM, sig_handler);
925
926 /* add timer fd */
927 if (perf_kvm__timerfd_create(kvm) < 0) {
928 err = -1;
929 goto out;
930 }
931
932 if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
933 goto out;
934
935 nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
936 if (nr_stdin < 0)
937 goto out;
938
939 if (fd_set_nonblock(fileno(stdin)) != 0)
940 goto out;
941
942 /* everything is good - enable the events and process */
943 perf_evlist__enable(kvm->evlist);
944
945 while (!done) {
946 struct fdarray *fda = &kvm->evlist->pollfd;
947 int rc;
948
949 rc = perf_kvm__mmap_read(kvm);
950 if (rc < 0)
951 break;
952
953 err = perf_kvm__handle_timerfd(kvm);
954 if (err)
955 goto out;
956
957 if (fda->entries[nr_stdin].revents & POLLIN)
958 done = perf_kvm__handle_stdin();
959
960 if (!rc && !done)
961 err = fdarray__poll(fda, 100);
962 }
963
964 perf_evlist__disable(kvm->evlist);
965
966 if (err == 0) {
967 sort_result(kvm);
968 print_result(kvm);
969 }
970
971out:
972 if (kvm->timerfd >= 0)
973 close(kvm->timerfd);
974
975 tcsetattr(0, TCSAFLUSH, &save);
976 return err;
977}
978
979static int kvm_live_open_events(struct perf_kvm_stat *kvm)
980{
981 int err, rc = -1;
982 struct perf_evsel *pos;
983 struct perf_evlist *evlist = kvm->evlist;
984 char sbuf[STRERR_BUFSIZE];
985
986 perf_evlist__config(evlist, &kvm->opts, NULL);
987
988 /*
989 * Note: exclude_{guest,host} do not apply here.
990 * This command processes KVM tracepoints from host only
991 */
992 evlist__for_each_entry(evlist, pos) {
993 struct perf_event_attr *attr = &pos->attr;
994
995 /* make sure these *are* set */
996 perf_evsel__set_sample_bit(pos, TID);
997 perf_evsel__set_sample_bit(pos, TIME);
998 perf_evsel__set_sample_bit(pos, CPU);
999 perf_evsel__set_sample_bit(pos, RAW);
1000 /* make sure these are *not*; want as small a sample as possible */
1001 perf_evsel__reset_sample_bit(pos, PERIOD);
1002 perf_evsel__reset_sample_bit(pos, IP);
1003 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1004 perf_evsel__reset_sample_bit(pos, ADDR);
1005 perf_evsel__reset_sample_bit(pos, READ);
1006 attr->mmap = 0;
1007 attr->comm = 0;
1008 attr->task = 0;
1009
1010 attr->sample_period = 1;
1011
1012 attr->watermark = 0;
1013 attr->wakeup_events = 1000;
1014
1015 /* will enable all once we are ready */
1016 attr->disabled = 1;
1017 }
1018
1019 err = perf_evlist__open(evlist);
1020 if (err < 0) {
1021 printf("Couldn't create the events: %s\n",
1022 str_error_r(errno, sbuf, sizeof(sbuf)));
1023 goto out;
1024 }
1025
1026 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1027 ui__error("Failed to mmap the events: %s\n",
1028 str_error_r(errno, sbuf, sizeof(sbuf)));
1029 perf_evlist__close(evlist);
1030 goto out;
1031 }
1032
1033 rc = 0;
1034
1035out:
1036 return rc;
1037}
1038#endif
1039
1040static int read_events(struct perf_kvm_stat *kvm)
1041{
1042 int ret;
1043
1044 struct perf_tool eops = {
1045 .sample = process_sample_event,
1046 .comm = perf_event__process_comm,
1047 .ordered_events = true,
1048 };
1049 struct perf_data_file file = {
1050 .path = kvm->file_name,
1051 .mode = PERF_DATA_MODE_READ,
1052 .force = kvm->force,
1053 };
1054
1055 kvm->tool = eops;
1056 kvm->session = perf_session__new(&file, false, &kvm->tool);
1057 if (!kvm->session) {
1058 pr_err("Initializing perf session failed\n");
1059 return -1;
1060 }
1061
1062 symbol__init(&kvm->session->header.env);
1063
1064 if (!perf_session__has_traces(kvm->session, "kvm record")) {
1065 ret = -EINVAL;
1066 goto out_delete;
1067 }
1068
1069 /*
1070 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1071 * traced in the old kernel.
1072 */
1073 ret = cpu_isa_config(kvm);
1074 if (ret < 0)
1075 goto out_delete;
1076
1077 ret = perf_session__process_events(kvm->session);
1078
1079out_delete:
1080 perf_session__delete(kvm->session);
1081 return ret;
1082}
1083
1084static int parse_target_str(struct perf_kvm_stat *kvm)
1085{
1086 if (kvm->opts.target.pid) {
1087 kvm->pid_list = intlist__new(kvm->opts.target.pid);
1088 if (kvm->pid_list == NULL) {
1089 pr_err("Error parsing process id string\n");
1090 return -EINVAL;
1091 }
1092 }
1093
1094 return 0;
1095}
1096
1097static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1098{
1099 int ret = -EINVAL;
1100 int vcpu = kvm->trace_vcpu;
1101
1102 if (parse_target_str(kvm) != 0)
1103 goto exit;
1104
1105 if (!verify_vcpu(vcpu))
1106 goto exit;
1107
1108 if (!select_key(kvm))
1109 goto exit;
1110
1111 if (!register_kvm_events_ops(kvm))
1112 goto exit;
1113
1114 init_kvm_event_record(kvm);
1115 setup_pager();
1116
1117 ret = read_events(kvm);
1118 if (ret)
1119 goto exit;
1120
1121 sort_result(kvm);
1122 print_result(kvm);
1123
1124exit:
1125 return ret;
1126}
1127
1128#define STRDUP_FAIL_EXIT(s) \
1129 ({ char *_p; \
1130 _p = strdup(s); \
1131 if (!_p) \
1132 return -ENOMEM; \
1133 _p; \
1134 })
1135
1136int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1137{
1138 return 0;
1139}
1140
1141static int
1142kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1143{
1144 unsigned int rec_argc, i, j, events_tp_size;
1145 const char **rec_argv;
1146 const char * const record_args[] = {
1147 "record",
1148 "-R",
1149 "-m", "1024",
1150 "-c", "1",
1151 };
1152 const char * const kvm_stat_record_usage[] = {
1153 "perf kvm stat record [<options>]",
1154 NULL
1155 };
1156 const char * const *events_tp;
1157 int ret;
1158
1159 events_tp_size = 0;
1160 ret = setup_kvm_events_tp(kvm);
1161 if (ret < 0) {
1162 pr_err("Unable to setup the kvm tracepoints\n");
1163 return ret;
1164 }
1165
1166 for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1167 events_tp_size++;
1168
1169 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1170 2 * events_tp_size;
1171 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1172
1173 if (rec_argv == NULL)
1174 return -ENOMEM;
1175
1176 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1177 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1178
1179 for (j = 0; j < events_tp_size; j++) {
1180 rec_argv[i++] = "-e";
1181 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1182 }
1183
1184 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1185 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1186
1187 for (j = 1; j < (unsigned int)argc; j++, i++)
1188 rec_argv[i] = argv[j];
1189
1190 set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1191 set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1192 set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1193
1194 set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1195 set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1196 set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1197 set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1198 set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1199 set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1200 set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1201 set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1202 set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1203 set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1204 set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1205 set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1206 set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1207 set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1208 set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1209
1210 record_usage = kvm_stat_record_usage;
1211 return cmd_record(i, rec_argv, NULL);
1212}
1213
1214static int
1215kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1216{
1217 const struct option kvm_events_report_options[] = {
1218 OPT_STRING(0, "event", &kvm->report_event, "report event",
1219 "event for reporting: vmexit, "
1220 "mmio (x86 only), ioport (x86 only)"),
1221 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1222 "vcpu id to report"),
1223 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1224 "key for sorting: sample(sort by samples number)"
1225 " time (sort by avg time)"),
1226 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1227 "analyze events only for given process id(s)"),
1228 OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1229 OPT_END()
1230 };
1231
1232 const char * const kvm_events_report_usage[] = {
1233 "perf kvm stat report [<options>]",
1234 NULL
1235 };
1236
1237 if (argc) {
1238 argc = parse_options(argc, argv,
1239 kvm_events_report_options,
1240 kvm_events_report_usage, 0);
1241 if (argc)
1242 usage_with_options(kvm_events_report_usage,
1243 kvm_events_report_options);
1244 }
1245
1246 if (!kvm->opts.target.pid)
1247 kvm->opts.target.system_wide = true;
1248
1249 return kvm_events_report_vcpu(kvm);
1250}
1251
1252#ifdef HAVE_TIMERFD_SUPPORT
1253static struct perf_evlist *kvm_live_event_list(void)
1254{
1255 struct perf_evlist *evlist;
1256 char *tp, *name, *sys;
1257 int err = -1;
1258 const char * const *events_tp;
1259
1260 evlist = perf_evlist__new();
1261 if (evlist == NULL)
1262 return NULL;
1263
1264 for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1265
1266 tp = strdup(*events_tp);
1267 if (tp == NULL)
1268 goto out;
1269
1270 /* split tracepoint into subsystem and name */
1271 sys = tp;
1272 name = strchr(tp, ':');
1273 if (name == NULL) {
1274 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1275 *events_tp);
1276 free(tp);
1277 goto out;
1278 }
1279 *name = '\0';
1280 name++;
1281
1282 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1283 pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1284 free(tp);
1285 goto out;
1286 }
1287
1288 free(tp);
1289 }
1290
1291 err = 0;
1292
1293out:
1294 if (err) {
1295 perf_evlist__delete(evlist);
1296 evlist = NULL;
1297 }
1298
1299 return evlist;
1300}
1301
1302static int kvm_events_live(struct perf_kvm_stat *kvm,
1303 int argc, const char **argv)
1304{
1305 char errbuf[BUFSIZ];
1306 int err;
1307
1308 const struct option live_options[] = {
1309 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1310 "record events on existing process id"),
1311 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1312 "number of mmap data pages",
1313 perf_evlist__parse_mmap_pages),
1314 OPT_INCR('v', "verbose", &verbose,
1315 "be more verbose (show counter open errors, etc)"),
1316 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1317 "system-wide collection from all CPUs"),
1318 OPT_UINTEGER('d', "display", &kvm->display_time,
1319 "time in seconds between display updates"),
1320 OPT_STRING(0, "event", &kvm->report_event, "report event",
1321 "event for reporting: "
1322 "vmexit, mmio (x86 only), ioport (x86 only)"),
1323 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1324 "vcpu id to report"),
1325 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1326 "key for sorting: sample(sort by samples number)"
1327 " time (sort by avg time)"),
1328 OPT_U64(0, "duration", &kvm->duration,
1329 "show events other than"
1330 " HLT (x86 only) or Wait state (s390 only)"
1331 " that take longer than duration usecs"),
1332 OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1333 "per thread proc mmap processing timeout in ms"),
1334 OPT_END()
1335 };
1336 const char * const live_usage[] = {
1337 "perf kvm stat live [<options>]",
1338 NULL
1339 };
1340 struct perf_data_file file = {
1341 .mode = PERF_DATA_MODE_WRITE,
1342 };
1343
1344
1345 /* event handling */
1346 kvm->tool.sample = process_sample_event;
1347 kvm->tool.comm = perf_event__process_comm;
1348 kvm->tool.exit = perf_event__process_exit;
1349 kvm->tool.fork = perf_event__process_fork;
1350 kvm->tool.lost = process_lost_event;
1351 kvm->tool.ordered_events = true;
1352 perf_tool__fill_defaults(&kvm->tool);
1353
1354 /* set defaults */
1355 kvm->display_time = 1;
1356 kvm->opts.user_interval = 1;
1357 kvm->opts.mmap_pages = 512;
1358 kvm->opts.target.uses_mmap = false;
1359 kvm->opts.target.uid_str = NULL;
1360 kvm->opts.target.uid = UINT_MAX;
1361 kvm->opts.proc_map_timeout = 500;
1362
1363 symbol__init(NULL);
1364 disable_buildid_cache();
1365
1366 use_browser = 0;
1367
1368 if (argc) {
1369 argc = parse_options(argc, argv, live_options,
1370 live_usage, 0);
1371 if (argc)
1372 usage_with_options(live_usage, live_options);
1373 }
1374
1375 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1376
1377 /*
1378 * target related setups
1379 */
1380 err = target__validate(&kvm->opts.target);
1381 if (err) {
1382 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1383 ui__warning("%s", errbuf);
1384 }
1385
1386 if (target__none(&kvm->opts.target))
1387 kvm->opts.target.system_wide = true;
1388
1389
1390 /*
1391 * generate the event list
1392 */
1393 err = setup_kvm_events_tp(kvm);
1394 if (err < 0) {
1395 pr_err("Unable to setup the kvm tracepoints\n");
1396 return err;
1397 }
1398
1399 kvm->evlist = kvm_live_event_list();
1400 if (kvm->evlist == NULL) {
1401 err = -1;
1402 goto out;
1403 }
1404
1405 symbol_conf.nr_events = kvm->evlist->nr_entries;
1406
1407 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1408 usage_with_options(live_usage, live_options);
1409
1410 /*
1411 * perf session
1412 */
1413 kvm->session = perf_session__new(&file, false, &kvm->tool);
1414 if (kvm->session == NULL) {
1415 err = -1;
1416 goto out;
1417 }
1418 kvm->session->evlist = kvm->evlist;
1419 perf_session__set_id_hdr_size(kvm->session);
1420 ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1421 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1422 kvm->evlist->threads, false, kvm->opts.proc_map_timeout);
1423 err = kvm_live_open_events(kvm);
1424 if (err)
1425 goto out;
1426
1427 err = kvm_events_live_report(kvm);
1428
1429out:
1430 perf_session__delete(kvm->session);
1431 kvm->session = NULL;
1432 perf_evlist__delete(kvm->evlist);
1433
1434 return err;
1435}
1436#endif
1437
1438static void print_kvm_stat_usage(void)
1439{
1440 printf("Usage: perf kvm stat <command>\n\n");
1441
1442 printf("# Available commands:\n");
1443 printf("\trecord: record kvm events\n");
1444 printf("\treport: report statistical data of kvm events\n");
1445 printf("\tlive: live reporting of statistical data of kvm events\n");
1446
1447 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1448}
1449
1450static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1451{
1452 struct perf_kvm_stat kvm = {
1453 .file_name = file_name,
1454
1455 .trace_vcpu = -1,
1456 .report_event = "vmexit",
1457 .sort_key = "sample",
1458
1459 };
1460
1461 if (argc == 1) {
1462 print_kvm_stat_usage();
1463 goto perf_stat;
1464 }
1465
1466 if (!strncmp(argv[1], "rec", 3))
1467 return kvm_events_record(&kvm, argc - 1, argv + 1);
1468
1469 if (!strncmp(argv[1], "rep", 3))
1470 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1471
1472#ifdef HAVE_TIMERFD_SUPPORT
1473 if (!strncmp(argv[1], "live", 4))
1474 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1475#endif
1476
1477perf_stat:
1478 return cmd_stat(argc, argv, NULL);
1479}
1480#endif /* HAVE_KVM_STAT_SUPPORT */
1481
1482static int __cmd_record(const char *file_name, int argc, const char **argv)
1483{
1484 int rec_argc, i = 0, j;
1485 const char **rec_argv;
1486
1487 rec_argc = argc + 2;
1488 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1489 rec_argv[i++] = strdup("record");
1490 rec_argv[i++] = strdup("-o");
1491 rec_argv[i++] = strdup(file_name);
1492 for (j = 1; j < argc; j++, i++)
1493 rec_argv[i] = argv[j];
1494
1495 BUG_ON(i != rec_argc);
1496
1497 return cmd_record(i, rec_argv, NULL);
1498}
1499
1500static int __cmd_report(const char *file_name, int argc, const char **argv)
1501{
1502 int rec_argc, i = 0, j;
1503 const char **rec_argv;
1504
1505 rec_argc = argc + 2;
1506 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1507 rec_argv[i++] = strdup("report");
1508 rec_argv[i++] = strdup("-i");
1509 rec_argv[i++] = strdup(file_name);
1510 for (j = 1; j < argc; j++, i++)
1511 rec_argv[i] = argv[j];
1512
1513 BUG_ON(i != rec_argc);
1514
1515 return cmd_report(i, rec_argv, NULL);
1516}
1517
1518static int
1519__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1520{
1521 int rec_argc, i = 0, j;
1522 const char **rec_argv;
1523
1524 rec_argc = argc + 2;
1525 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1526 rec_argv[i++] = strdup("buildid-list");
1527 rec_argv[i++] = strdup("-i");
1528 rec_argv[i++] = strdup(file_name);
1529 for (j = 1; j < argc; j++, i++)
1530 rec_argv[i] = argv[j];
1531
1532 BUG_ON(i != rec_argc);
1533
1534 return cmd_buildid_list(i, rec_argv, NULL);
1535}
1536
1537int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1538{
1539 const char *file_name = NULL;
1540 const struct option kvm_options[] = {
1541 OPT_STRING('i', "input", &file_name, "file",
1542 "Input file name"),
1543 OPT_STRING('o', "output", &file_name, "file",
1544 "Output file name"),
1545 OPT_BOOLEAN(0, "guest", &perf_guest,
1546 "Collect guest os data"),
1547 OPT_BOOLEAN(0, "host", &perf_host,
1548 "Collect host os data"),
1549 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1550 "guest mount directory under which every guest os"
1551 " instance has a subdir"),
1552 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1553 "file", "file saving guest os vmlinux"),
1554 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1555 "file", "file saving guest os /proc/kallsyms"),
1556 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1557 "file", "file saving guest os /proc/modules"),
1558 OPT_INCR('v', "verbose", &verbose,
1559 "be more verbose (show counter open errors, etc)"),
1560 OPT_END()
1561 };
1562
1563 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1564 "buildid-list", "stat", NULL };
1565 const char *kvm_usage[] = { NULL, NULL };
1566
1567 perf_host = 0;
1568 perf_guest = 1;
1569
1570 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1571 PARSE_OPT_STOP_AT_NON_OPTION);
1572 if (!argc)
1573 usage_with_options(kvm_usage, kvm_options);
1574
1575 if (!perf_host)
1576 perf_guest = 1;
1577
1578 if (!file_name) {
1579 file_name = get_filename_for_perf_kvm();
1580
1581 if (!file_name) {
1582 pr_err("Failed to allocate memory for filename\n");
1583 return -ENOMEM;
1584 }
1585 }
1586
1587 if (!strncmp(argv[0], "rec", 3))
1588 return __cmd_record(file_name, argc, argv);
1589 else if (!strncmp(argv[0], "rep", 3))
1590 return __cmd_report(file_name, argc, argv);
1591 else if (!strncmp(argv[0], "diff", 4))
1592 return cmd_diff(argc, argv, NULL);
1593 else if (!strncmp(argv[0], "top", 3))
1594 return cmd_top(argc, argv, NULL);
1595 else if (!strncmp(argv[0], "buildid-list", 12))
1596 return __cmd_buildid_list(file_name, argc, argv);
1597#ifdef HAVE_KVM_STAT_SUPPORT
1598 else if (!strncmp(argv[0], "stat", 4))
1599 return kvm_cmd_stat(file_name, argc, argv);
1600#endif
1601 else
1602 usage_with_options(kvm_usage, kvm_options);
1603
1604 return 0;
1605}
1#include "builtin.h"
2#include "perf.h"
3
4#include "util/evsel.h"
5#include "util/evlist.h"
6#include "util/util.h"
7#include "util/cache.h"
8#include "util/symbol.h"
9#include "util/thread.h"
10#include "util/header.h"
11#include "util/session.h"
12#include "util/intlist.h"
13#include "util/parse-options.h"
14#include "util/trace-event.h"
15#include "util/debug.h"
16#include <api/fs/debugfs.h>
17#include "util/tool.h"
18#include "util/stat.h"
19#include "util/top.h"
20#include "util/data.h"
21
22#include <sys/prctl.h>
23#ifdef HAVE_TIMERFD_SUPPORT
24#include <sys/timerfd.h>
25#endif
26
27#include <termios.h>
28#include <semaphore.h>
29#include <pthread.h>
30#include <math.h>
31
32#if defined(__i386__) || defined(__x86_64__)
33#include <asm/svm.h>
34#include <asm/vmx.h>
35#include <asm/kvm.h>
36
37struct event_key {
38 #define INVALID_KEY (~0ULL)
39 u64 key;
40 int info;
41};
42
43struct kvm_event_stats {
44 u64 time;
45 struct stats stats;
46};
47
48struct kvm_event {
49 struct list_head hash_entry;
50 struct rb_node rb;
51
52 struct event_key key;
53
54 struct kvm_event_stats total;
55
56 #define DEFAULT_VCPU_NUM 8
57 int max_vcpu;
58 struct kvm_event_stats *vcpu;
59};
60
61typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
62
63struct kvm_event_key {
64 const char *name;
65 key_cmp_fun key;
66};
67
68
69struct perf_kvm_stat;
70
71struct kvm_events_ops {
72 bool (*is_begin_event)(struct perf_evsel *evsel,
73 struct perf_sample *sample,
74 struct event_key *key);
75 bool (*is_end_event)(struct perf_evsel *evsel,
76 struct perf_sample *sample, struct event_key *key);
77 void (*decode_key)(struct perf_kvm_stat *kvm, struct event_key *key,
78 char decode[20]);
79 const char *name;
80};
81
82struct exit_reasons_table {
83 unsigned long exit_code;
84 const char *reason;
85};
86
87#define EVENTS_BITS 12
88#define EVENTS_CACHE_SIZE (1UL << EVENTS_BITS)
89
90struct perf_kvm_stat {
91 struct perf_tool tool;
92 struct record_opts opts;
93 struct perf_evlist *evlist;
94 struct perf_session *session;
95
96 const char *file_name;
97 const char *report_event;
98 const char *sort_key;
99 int trace_vcpu;
100
101 struct exit_reasons_table *exit_reasons;
102 int exit_reasons_size;
103 const char *exit_reasons_isa;
104
105 struct kvm_events_ops *events_ops;
106 key_cmp_fun compare;
107 struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
108
109 u64 total_time;
110 u64 total_count;
111 u64 lost_events;
112 u64 duration;
113
114 const char *pid_str;
115 struct intlist *pid_list;
116
117 struct rb_root result;
118
119 int timerfd;
120 unsigned int display_time;
121 bool live;
122};
123
124
125static void exit_event_get_key(struct perf_evsel *evsel,
126 struct perf_sample *sample,
127 struct event_key *key)
128{
129 key->info = 0;
130 key->key = perf_evsel__intval(evsel, sample, "exit_reason");
131}
132
133static bool kvm_exit_event(struct perf_evsel *evsel)
134{
135 return !strcmp(evsel->name, "kvm:kvm_exit");
136}
137
138static bool exit_event_begin(struct perf_evsel *evsel,
139 struct perf_sample *sample, struct event_key *key)
140{
141 if (kvm_exit_event(evsel)) {
142 exit_event_get_key(evsel, sample, key);
143 return true;
144 }
145
146 return false;
147}
148
149static bool kvm_entry_event(struct perf_evsel *evsel)
150{
151 return !strcmp(evsel->name, "kvm:kvm_entry");
152}
153
154static bool exit_event_end(struct perf_evsel *evsel,
155 struct perf_sample *sample __maybe_unused,
156 struct event_key *key __maybe_unused)
157{
158 return kvm_entry_event(evsel);
159}
160
161static struct exit_reasons_table vmx_exit_reasons[] = {
162 VMX_EXIT_REASONS
163};
164
165static struct exit_reasons_table svm_exit_reasons[] = {
166 SVM_EXIT_REASONS
167};
168
169static const char *get_exit_reason(struct perf_kvm_stat *kvm, u64 exit_code)
170{
171 int i = kvm->exit_reasons_size;
172 struct exit_reasons_table *tbl = kvm->exit_reasons;
173
174 while (i--) {
175 if (tbl->exit_code == exit_code)
176 return tbl->reason;
177 tbl++;
178 }
179
180 pr_err("unknown kvm exit code:%lld on %s\n",
181 (unsigned long long)exit_code, kvm->exit_reasons_isa);
182 return "UNKNOWN";
183}
184
185static void exit_event_decode_key(struct perf_kvm_stat *kvm,
186 struct event_key *key,
187 char decode[20])
188{
189 const char *exit_reason = get_exit_reason(kvm, key->key);
190
191 scnprintf(decode, 20, "%s", exit_reason);
192}
193
194static struct kvm_events_ops exit_events = {
195 .is_begin_event = exit_event_begin,
196 .is_end_event = exit_event_end,
197 .decode_key = exit_event_decode_key,
198 .name = "VM-EXIT"
199};
200
201/*
202 * For the mmio events, we treat:
203 * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
204 * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
205 */
206static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
207 struct event_key *key)
208{
209 key->key = perf_evsel__intval(evsel, sample, "gpa");
210 key->info = perf_evsel__intval(evsel, sample, "type");
211}
212
213#define KVM_TRACE_MMIO_READ_UNSATISFIED 0
214#define KVM_TRACE_MMIO_READ 1
215#define KVM_TRACE_MMIO_WRITE 2
216
217static bool mmio_event_begin(struct perf_evsel *evsel,
218 struct perf_sample *sample, struct event_key *key)
219{
220 /* MMIO read begin event in kernel. */
221 if (kvm_exit_event(evsel))
222 return true;
223
224 /* MMIO write begin event in kernel. */
225 if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
226 perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
227 mmio_event_get_key(evsel, sample, key);
228 return true;
229 }
230
231 return false;
232}
233
234static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
235 struct event_key *key)
236{
237 /* MMIO write end event in kernel. */
238 if (kvm_entry_event(evsel))
239 return true;
240
241 /* MMIO read end event in kernel.*/
242 if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
243 perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
244 mmio_event_get_key(evsel, sample, key);
245 return true;
246 }
247
248 return false;
249}
250
251static void mmio_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
252 struct event_key *key,
253 char decode[20])
254{
255 scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
256 key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
257}
258
259static struct kvm_events_ops mmio_events = {
260 .is_begin_event = mmio_event_begin,
261 .is_end_event = mmio_event_end,
262 .decode_key = mmio_event_decode_key,
263 .name = "MMIO Access"
264};
265
266 /* The time of emulation pio access is from kvm_pio to kvm_entry. */
267static void ioport_event_get_key(struct perf_evsel *evsel,
268 struct perf_sample *sample,
269 struct event_key *key)
270{
271 key->key = perf_evsel__intval(evsel, sample, "port");
272 key->info = perf_evsel__intval(evsel, sample, "rw");
273}
274
275static bool ioport_event_begin(struct perf_evsel *evsel,
276 struct perf_sample *sample,
277 struct event_key *key)
278{
279 if (!strcmp(evsel->name, "kvm:kvm_pio")) {
280 ioport_event_get_key(evsel, sample, key);
281 return true;
282 }
283
284 return false;
285}
286
287static bool ioport_event_end(struct perf_evsel *evsel,
288 struct perf_sample *sample __maybe_unused,
289 struct event_key *key __maybe_unused)
290{
291 return kvm_entry_event(evsel);
292}
293
294static void ioport_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
295 struct event_key *key,
296 char decode[20])
297{
298 scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
299 key->info ? "POUT" : "PIN");
300}
301
302static struct kvm_events_ops ioport_events = {
303 .is_begin_event = ioport_event_begin,
304 .is_end_event = ioport_event_end,
305 .decode_key = ioport_event_decode_key,
306 .name = "IO Port Access"
307};
308
309static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
310{
311 bool ret = true;
312
313 if (!strcmp(kvm->report_event, "vmexit"))
314 kvm->events_ops = &exit_events;
315 else if (!strcmp(kvm->report_event, "mmio"))
316 kvm->events_ops = &mmio_events;
317 else if (!strcmp(kvm->report_event, "ioport"))
318 kvm->events_ops = &ioport_events;
319 else {
320 pr_err("Unknown report event:%s\n", kvm->report_event);
321 ret = false;
322 }
323
324 return ret;
325}
326
327struct vcpu_event_record {
328 int vcpu_id;
329 u64 start_time;
330 struct kvm_event *last_event;
331};
332
333
334static void init_kvm_event_record(struct perf_kvm_stat *kvm)
335{
336 unsigned int i;
337
338 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
339 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
340}
341
342#ifdef HAVE_TIMERFD_SUPPORT
343static void clear_events_cache_stats(struct list_head *kvm_events_cache)
344{
345 struct list_head *head;
346 struct kvm_event *event;
347 unsigned int i;
348 int j;
349
350 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
351 head = &kvm_events_cache[i];
352 list_for_each_entry(event, head, hash_entry) {
353 /* reset stats for event */
354 event->total.time = 0;
355 init_stats(&event->total.stats);
356
357 for (j = 0; j < event->max_vcpu; ++j) {
358 event->vcpu[j].time = 0;
359 init_stats(&event->vcpu[j].stats);
360 }
361 }
362 }
363}
364#endif
365
366static int kvm_events_hash_fn(u64 key)
367{
368 return key & (EVENTS_CACHE_SIZE - 1);
369}
370
371static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
372{
373 int old_max_vcpu = event->max_vcpu;
374 void *prev;
375
376 if (vcpu_id < event->max_vcpu)
377 return true;
378
379 while (event->max_vcpu <= vcpu_id)
380 event->max_vcpu += DEFAULT_VCPU_NUM;
381
382 prev = event->vcpu;
383 event->vcpu = realloc(event->vcpu,
384 event->max_vcpu * sizeof(*event->vcpu));
385 if (!event->vcpu) {
386 free(prev);
387 pr_err("Not enough memory\n");
388 return false;
389 }
390
391 memset(event->vcpu + old_max_vcpu, 0,
392 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
393 return true;
394}
395
396static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
397{
398 struct kvm_event *event;
399
400 event = zalloc(sizeof(*event));
401 if (!event) {
402 pr_err("Not enough memory\n");
403 return NULL;
404 }
405
406 event->key = *key;
407 init_stats(&event->total.stats);
408 return event;
409}
410
411static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
412 struct event_key *key)
413{
414 struct kvm_event *event;
415 struct list_head *head;
416
417 BUG_ON(key->key == INVALID_KEY);
418
419 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
420 list_for_each_entry(event, head, hash_entry) {
421 if (event->key.key == key->key && event->key.info == key->info)
422 return event;
423 }
424
425 event = kvm_alloc_init_event(key);
426 if (!event)
427 return NULL;
428
429 list_add(&event->hash_entry, head);
430 return event;
431}
432
433static bool handle_begin_event(struct perf_kvm_stat *kvm,
434 struct vcpu_event_record *vcpu_record,
435 struct event_key *key, u64 timestamp)
436{
437 struct kvm_event *event = NULL;
438
439 if (key->key != INVALID_KEY)
440 event = find_create_kvm_event(kvm, key);
441
442 vcpu_record->last_event = event;
443 vcpu_record->start_time = timestamp;
444 return true;
445}
446
447static void
448kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
449{
450 kvm_stats->time += time_diff;
451 update_stats(&kvm_stats->stats, time_diff);
452}
453
454static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
455{
456 struct kvm_event_stats *kvm_stats = &event->total;
457
458 if (vcpu_id != -1)
459 kvm_stats = &event->vcpu[vcpu_id];
460
461 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
462 avg_stats(&kvm_stats->stats));
463}
464
465static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
466 u64 time_diff)
467{
468 if (vcpu_id == -1) {
469 kvm_update_event_stats(&event->total, time_diff);
470 return true;
471 }
472
473 if (!kvm_event_expand(event, vcpu_id))
474 return false;
475
476 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
477 return true;
478}
479
480static bool handle_end_event(struct perf_kvm_stat *kvm,
481 struct vcpu_event_record *vcpu_record,
482 struct event_key *key,
483 struct perf_sample *sample)
484{
485 struct kvm_event *event;
486 u64 time_begin, time_diff;
487 int vcpu;
488
489 if (kvm->trace_vcpu == -1)
490 vcpu = -1;
491 else
492 vcpu = vcpu_record->vcpu_id;
493
494 event = vcpu_record->last_event;
495 time_begin = vcpu_record->start_time;
496
497 /* The begin event is not caught. */
498 if (!time_begin)
499 return true;
500
501 /*
502 * In some case, the 'begin event' only records the start timestamp,
503 * the actual event is recognized in the 'end event' (e.g. mmio-event).
504 */
505
506 /* Both begin and end events did not get the key. */
507 if (!event && key->key == INVALID_KEY)
508 return true;
509
510 if (!event)
511 event = find_create_kvm_event(kvm, key);
512
513 if (!event)
514 return false;
515
516 vcpu_record->last_event = NULL;
517 vcpu_record->start_time = 0;
518
519 /* seems to happen once in a while during live mode */
520 if (sample->time < time_begin) {
521 pr_debug("End time before begin time; skipping event.\n");
522 return true;
523 }
524
525 time_diff = sample->time - time_begin;
526
527 if (kvm->duration && time_diff > kvm->duration) {
528 char decode[32];
529
530 kvm->events_ops->decode_key(kvm, &event->key, decode);
531 if (strcmp(decode, "HLT")) {
532 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
533 sample->time, sample->pid, vcpu_record->vcpu_id,
534 decode, time_diff/1000);
535 }
536 }
537
538 return update_kvm_event(event, vcpu, time_diff);
539}
540
541static
542struct vcpu_event_record *per_vcpu_record(struct thread *thread,
543 struct perf_evsel *evsel,
544 struct perf_sample *sample)
545{
546 /* Only kvm_entry records vcpu id. */
547 if (!thread->priv && kvm_entry_event(evsel)) {
548 struct vcpu_event_record *vcpu_record;
549
550 vcpu_record = zalloc(sizeof(*vcpu_record));
551 if (!vcpu_record) {
552 pr_err("%s: Not enough memory\n", __func__);
553 return NULL;
554 }
555
556 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
557 thread->priv = vcpu_record;
558 }
559
560 return thread->priv;
561}
562
563static bool handle_kvm_event(struct perf_kvm_stat *kvm,
564 struct thread *thread,
565 struct perf_evsel *evsel,
566 struct perf_sample *sample)
567{
568 struct vcpu_event_record *vcpu_record;
569 struct event_key key = {.key = INVALID_KEY};
570
571 vcpu_record = per_vcpu_record(thread, evsel, sample);
572 if (!vcpu_record)
573 return true;
574
575 /* only process events for vcpus user cares about */
576 if ((kvm->trace_vcpu != -1) &&
577 (kvm->trace_vcpu != vcpu_record->vcpu_id))
578 return true;
579
580 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
581 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
582
583 if (kvm->events_ops->is_end_event(evsel, sample, &key))
584 return handle_end_event(kvm, vcpu_record, &key, sample);
585
586 return true;
587}
588
589#define GET_EVENT_KEY(func, field) \
590static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
591{ \
592 if (vcpu == -1) \
593 return event->total.field; \
594 \
595 if (vcpu >= event->max_vcpu) \
596 return 0; \
597 \
598 return event->vcpu[vcpu].field; \
599}
600
601#define COMPARE_EVENT_KEY(func, field) \
602GET_EVENT_KEY(func, field) \
603static int compare_kvm_event_ ## func(struct kvm_event *one, \
604 struct kvm_event *two, int vcpu)\
605{ \
606 return get_event_ ##func(one, vcpu) > \
607 get_event_ ##func(two, vcpu); \
608}
609
610GET_EVENT_KEY(time, time);
611COMPARE_EVENT_KEY(count, stats.n);
612COMPARE_EVENT_KEY(mean, stats.mean);
613GET_EVENT_KEY(max, stats.max);
614GET_EVENT_KEY(min, stats.min);
615
616#define DEF_SORT_NAME_KEY(name, compare_key) \
617 { #name, compare_kvm_event_ ## compare_key }
618
619static struct kvm_event_key keys[] = {
620 DEF_SORT_NAME_KEY(sample, count),
621 DEF_SORT_NAME_KEY(time, mean),
622 { NULL, NULL }
623};
624
625static bool select_key(struct perf_kvm_stat *kvm)
626{
627 int i;
628
629 for (i = 0; keys[i].name; i++) {
630 if (!strcmp(keys[i].name, kvm->sort_key)) {
631 kvm->compare = keys[i].key;
632 return true;
633 }
634 }
635
636 pr_err("Unknown compare key:%s\n", kvm->sort_key);
637 return false;
638}
639
640static void insert_to_result(struct rb_root *result, struct kvm_event *event,
641 key_cmp_fun bigger, int vcpu)
642{
643 struct rb_node **rb = &result->rb_node;
644 struct rb_node *parent = NULL;
645 struct kvm_event *p;
646
647 while (*rb) {
648 p = container_of(*rb, struct kvm_event, rb);
649 parent = *rb;
650
651 if (bigger(event, p, vcpu))
652 rb = &(*rb)->rb_left;
653 else
654 rb = &(*rb)->rb_right;
655 }
656
657 rb_link_node(&event->rb, parent, rb);
658 rb_insert_color(&event->rb, result);
659}
660
661static void
662update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
663{
664 int vcpu = kvm->trace_vcpu;
665
666 kvm->total_count += get_event_count(event, vcpu);
667 kvm->total_time += get_event_time(event, vcpu);
668}
669
670static bool event_is_valid(struct kvm_event *event, int vcpu)
671{
672 return !!get_event_count(event, vcpu);
673}
674
675static void sort_result(struct perf_kvm_stat *kvm)
676{
677 unsigned int i;
678 int vcpu = kvm->trace_vcpu;
679 struct kvm_event *event;
680
681 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
682 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
683 if (event_is_valid(event, vcpu)) {
684 update_total_count(kvm, event);
685 insert_to_result(&kvm->result, event,
686 kvm->compare, vcpu);
687 }
688 }
689 }
690}
691
692/* returns left most element of result, and erase it */
693static struct kvm_event *pop_from_result(struct rb_root *result)
694{
695 struct rb_node *node = rb_first(result);
696
697 if (!node)
698 return NULL;
699
700 rb_erase(node, result);
701 return container_of(node, struct kvm_event, rb);
702}
703
704static void print_vcpu_info(struct perf_kvm_stat *kvm)
705{
706 int vcpu = kvm->trace_vcpu;
707
708 pr_info("Analyze events for ");
709
710 if (kvm->live) {
711 if (kvm->opts.target.system_wide)
712 pr_info("all VMs, ");
713 else if (kvm->opts.target.pid)
714 pr_info("pid(s) %s, ", kvm->opts.target.pid);
715 else
716 pr_info("dazed and confused on what is monitored, ");
717 }
718
719 if (vcpu == -1)
720 pr_info("all VCPUs:\n\n");
721 else
722 pr_info("VCPU %d:\n\n", vcpu);
723}
724
725static void show_timeofday(void)
726{
727 char date[64];
728 struct timeval tv;
729 struct tm ltime;
730
731 gettimeofday(&tv, NULL);
732 if (localtime_r(&tv.tv_sec, <ime)) {
733 strftime(date, sizeof(date), "%H:%M:%S", <ime);
734 pr_info("%s.%06ld", date, tv.tv_usec);
735 } else
736 pr_info("00:00:00.000000");
737
738 return;
739}
740
741static void print_result(struct perf_kvm_stat *kvm)
742{
743 char decode[20];
744 struct kvm_event *event;
745 int vcpu = kvm->trace_vcpu;
746
747 if (kvm->live) {
748 puts(CONSOLE_CLEAR);
749 show_timeofday();
750 }
751
752 pr_info("\n\n");
753 print_vcpu_info(kvm);
754 pr_info("%20s ", kvm->events_ops->name);
755 pr_info("%10s ", "Samples");
756 pr_info("%9s ", "Samples%");
757
758 pr_info("%9s ", "Time%");
759 pr_info("%10s ", "Min Time");
760 pr_info("%10s ", "Max Time");
761 pr_info("%16s ", "Avg time");
762 pr_info("\n\n");
763
764 while ((event = pop_from_result(&kvm->result))) {
765 u64 ecount, etime, max, min;
766
767 ecount = get_event_count(event, vcpu);
768 etime = get_event_time(event, vcpu);
769 max = get_event_max(event, vcpu);
770 min = get_event_min(event, vcpu);
771
772 kvm->events_ops->decode_key(kvm, &event->key, decode);
773 pr_info("%20s ", decode);
774 pr_info("%10llu ", (unsigned long long)ecount);
775 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
776 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
777 pr_info("%8" PRIu64 "us ", min / 1000);
778 pr_info("%8" PRIu64 "us ", max / 1000);
779 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
780 kvm_event_rel_stddev(vcpu, event));
781 pr_info("\n");
782 }
783
784 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
785 kvm->total_count, kvm->total_time / 1e3);
786
787 if (kvm->lost_events)
788 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
789}
790
791#ifdef HAVE_TIMERFD_SUPPORT
792static int process_lost_event(struct perf_tool *tool,
793 union perf_event *event __maybe_unused,
794 struct perf_sample *sample __maybe_unused,
795 struct machine *machine __maybe_unused)
796{
797 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
798
799 kvm->lost_events++;
800 return 0;
801}
802#endif
803
804static bool skip_sample(struct perf_kvm_stat *kvm,
805 struct perf_sample *sample)
806{
807 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
808 return true;
809
810 return false;
811}
812
813static int process_sample_event(struct perf_tool *tool,
814 union perf_event *event,
815 struct perf_sample *sample,
816 struct perf_evsel *evsel,
817 struct machine *machine)
818{
819 struct thread *thread;
820 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
821 tool);
822
823 if (skip_sample(kvm, sample))
824 return 0;
825
826 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
827 if (thread == NULL) {
828 pr_debug("problem processing %d event, skipping it.\n",
829 event->header.type);
830 return -1;
831 }
832
833 if (!handle_kvm_event(kvm, thread, evsel, sample))
834 return -1;
835
836 return 0;
837}
838
839static int cpu_isa_config(struct perf_kvm_stat *kvm)
840{
841 char buf[64], *cpuid;
842 int err, isa;
843
844 if (kvm->live) {
845 err = get_cpuid(buf, sizeof(buf));
846 if (err != 0) {
847 pr_err("Failed to look up CPU type (Intel or AMD)\n");
848 return err;
849 }
850 cpuid = buf;
851 } else
852 cpuid = kvm->session->header.env.cpuid;
853
854 if (strstr(cpuid, "Intel"))
855 isa = 1;
856 else if (strstr(cpuid, "AMD"))
857 isa = 0;
858 else {
859 pr_err("CPU %s is not supported.\n", cpuid);
860 return -ENOTSUP;
861 }
862
863 if (isa == 1) {
864 kvm->exit_reasons = vmx_exit_reasons;
865 kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
866 kvm->exit_reasons_isa = "VMX";
867 }
868
869 return 0;
870}
871
872static bool verify_vcpu(int vcpu)
873{
874 if (vcpu != -1 && vcpu < 0) {
875 pr_err("Invalid vcpu:%d.\n", vcpu);
876 return false;
877 }
878
879 return true;
880}
881
882#ifdef HAVE_TIMERFD_SUPPORT
883/* keeping the max events to a modest level to keep
884 * the processing of samples per mmap smooth.
885 */
886#define PERF_KVM__MAX_EVENTS_PER_MMAP 25
887
888static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
889 u64 *mmap_time)
890{
891 union perf_event *event;
892 struct perf_sample sample;
893 s64 n = 0;
894 int err;
895
896 *mmap_time = ULLONG_MAX;
897 while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
898 err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
899 if (err) {
900 perf_evlist__mmap_consume(kvm->evlist, idx);
901 pr_err("Failed to parse sample\n");
902 return -1;
903 }
904
905 err = perf_session_queue_event(kvm->session, event, &sample, 0);
906 /*
907 * FIXME: Here we can't consume the event, as perf_session_queue_event will
908 * point to it, and it'll get possibly overwritten by the kernel.
909 */
910 perf_evlist__mmap_consume(kvm->evlist, idx);
911
912 if (err) {
913 pr_err("Failed to enqueue sample: %d\n", err);
914 return -1;
915 }
916
917 /* save time stamp of our first sample for this mmap */
918 if (n == 0)
919 *mmap_time = sample.time;
920
921 /* limit events per mmap handled all at once */
922 n++;
923 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
924 break;
925 }
926
927 return n;
928}
929
930static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
931{
932 int i, err, throttled = 0;
933 s64 n, ntotal = 0;
934 u64 flush_time = ULLONG_MAX, mmap_time;
935
936 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
937 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
938 if (n < 0)
939 return -1;
940
941 /* flush time is going to be the minimum of all the individual
942 * mmap times. Essentially, we flush all the samples queued up
943 * from the last pass under our minimal start time -- that leaves
944 * a very small race for samples to come in with a lower timestamp.
945 * The ioctl to return the perf_clock timestamp should close the
946 * race entirely.
947 */
948 if (mmap_time < flush_time)
949 flush_time = mmap_time;
950
951 ntotal += n;
952 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
953 throttled = 1;
954 }
955
956 /* flush queue after each round in which we processed events */
957 if (ntotal) {
958 kvm->session->ordered_samples.next_flush = flush_time;
959 err = kvm->tool.finished_round(&kvm->tool, NULL, kvm->session);
960 if (err) {
961 if (kvm->lost_events)
962 pr_info("\nLost events: %" PRIu64 "\n\n",
963 kvm->lost_events);
964 return err;
965 }
966 }
967
968 return throttled;
969}
970
971static volatile int done;
972
973static void sig_handler(int sig __maybe_unused)
974{
975 done = 1;
976}
977
978static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
979{
980 struct itimerspec new_value;
981 int rc = -1;
982
983 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
984 if (kvm->timerfd < 0) {
985 pr_err("timerfd_create failed\n");
986 goto out;
987 }
988
989 new_value.it_value.tv_sec = kvm->display_time;
990 new_value.it_value.tv_nsec = 0;
991 new_value.it_interval.tv_sec = kvm->display_time;
992 new_value.it_interval.tv_nsec = 0;
993
994 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
995 pr_err("timerfd_settime failed: %d\n", errno);
996 close(kvm->timerfd);
997 goto out;
998 }
999
1000 rc = 0;
1001out:
1002 return rc;
1003}
1004
1005static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
1006{
1007 uint64_t c;
1008 int rc;
1009
1010 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
1011 if (rc < 0) {
1012 if (errno == EAGAIN)
1013 return 0;
1014
1015 pr_err("Failed to read timer fd: %d\n", errno);
1016 return -1;
1017 }
1018
1019 if (rc != sizeof(uint64_t)) {
1020 pr_err("Error reading timer fd - invalid size returned\n");
1021 return -1;
1022 }
1023
1024 if (c != 1)
1025 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
1026
1027 /* update display */
1028 sort_result(kvm);
1029 print_result(kvm);
1030
1031 /* reset counts */
1032 clear_events_cache_stats(kvm->kvm_events_cache);
1033 kvm->total_count = 0;
1034 kvm->total_time = 0;
1035 kvm->lost_events = 0;
1036
1037 return 0;
1038}
1039
1040static int fd_set_nonblock(int fd)
1041{
1042 long arg = 0;
1043
1044 arg = fcntl(fd, F_GETFL);
1045 if (arg < 0) {
1046 pr_err("Failed to get current flags for fd %d\n", fd);
1047 return -1;
1048 }
1049
1050 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
1051 pr_err("Failed to set non-block option on fd %d\n", fd);
1052 return -1;
1053 }
1054
1055 return 0;
1056}
1057
1058static
1059int perf_kvm__handle_stdin(struct termios *tc_now, struct termios *tc_save)
1060{
1061 int c;
1062
1063 tcsetattr(0, TCSANOW, tc_now);
1064 c = getc(stdin);
1065 tcsetattr(0, TCSAFLUSH, tc_save);
1066
1067 if (c == 'q')
1068 return 1;
1069
1070 return 0;
1071}
1072
1073static int kvm_events_live_report(struct perf_kvm_stat *kvm)
1074{
1075 struct pollfd *pollfds = NULL;
1076 int nr_fds, nr_stdin, ret, err = -EINVAL;
1077 struct termios tc, save;
1078
1079 /* live flag must be set first */
1080 kvm->live = true;
1081
1082 ret = cpu_isa_config(kvm);
1083 if (ret < 0)
1084 return ret;
1085
1086 if (!verify_vcpu(kvm->trace_vcpu) ||
1087 !select_key(kvm) ||
1088 !register_kvm_events_ops(kvm)) {
1089 goto out;
1090 }
1091
1092 init_kvm_event_record(kvm);
1093
1094 tcgetattr(0, &save);
1095 tc = save;
1096 tc.c_lflag &= ~(ICANON | ECHO);
1097 tc.c_cc[VMIN] = 0;
1098 tc.c_cc[VTIME] = 0;
1099
1100 signal(SIGINT, sig_handler);
1101 signal(SIGTERM, sig_handler);
1102
1103 /* copy pollfds -- need to add timerfd and stdin */
1104 nr_fds = kvm->evlist->nr_fds;
1105 pollfds = zalloc(sizeof(struct pollfd) * (nr_fds + 2));
1106 if (!pollfds) {
1107 err = -ENOMEM;
1108 goto out;
1109 }
1110 memcpy(pollfds, kvm->evlist->pollfd,
1111 sizeof(struct pollfd) * kvm->evlist->nr_fds);
1112
1113 /* add timer fd */
1114 if (perf_kvm__timerfd_create(kvm) < 0) {
1115 err = -1;
1116 goto out;
1117 }
1118
1119 pollfds[nr_fds].fd = kvm->timerfd;
1120 pollfds[nr_fds].events = POLLIN;
1121 nr_fds++;
1122
1123 pollfds[nr_fds].fd = fileno(stdin);
1124 pollfds[nr_fds].events = POLLIN;
1125 nr_stdin = nr_fds;
1126 nr_fds++;
1127 if (fd_set_nonblock(fileno(stdin)) != 0)
1128 goto out;
1129
1130 /* everything is good - enable the events and process */
1131 perf_evlist__enable(kvm->evlist);
1132
1133 while (!done) {
1134 int rc;
1135
1136 rc = perf_kvm__mmap_read(kvm);
1137 if (rc < 0)
1138 break;
1139
1140 err = perf_kvm__handle_timerfd(kvm);
1141 if (err)
1142 goto out;
1143
1144 if (pollfds[nr_stdin].revents & POLLIN)
1145 done = perf_kvm__handle_stdin(&tc, &save);
1146
1147 if (!rc && !done)
1148 err = poll(pollfds, nr_fds, 100);
1149 }
1150
1151 perf_evlist__disable(kvm->evlist);
1152
1153 if (err == 0) {
1154 sort_result(kvm);
1155 print_result(kvm);
1156 }
1157
1158out:
1159 if (kvm->timerfd >= 0)
1160 close(kvm->timerfd);
1161
1162 free(pollfds);
1163 return err;
1164}
1165
1166static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1167{
1168 int err, rc = -1;
1169 struct perf_evsel *pos;
1170 struct perf_evlist *evlist = kvm->evlist;
1171
1172 perf_evlist__config(evlist, &kvm->opts);
1173
1174 /*
1175 * Note: exclude_{guest,host} do not apply here.
1176 * This command processes KVM tracepoints from host only
1177 */
1178 evlist__for_each(evlist, pos) {
1179 struct perf_event_attr *attr = &pos->attr;
1180
1181 /* make sure these *are* set */
1182 perf_evsel__set_sample_bit(pos, TID);
1183 perf_evsel__set_sample_bit(pos, TIME);
1184 perf_evsel__set_sample_bit(pos, CPU);
1185 perf_evsel__set_sample_bit(pos, RAW);
1186 /* make sure these are *not*; want as small a sample as possible */
1187 perf_evsel__reset_sample_bit(pos, PERIOD);
1188 perf_evsel__reset_sample_bit(pos, IP);
1189 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1190 perf_evsel__reset_sample_bit(pos, ADDR);
1191 perf_evsel__reset_sample_bit(pos, READ);
1192 attr->mmap = 0;
1193 attr->comm = 0;
1194 attr->task = 0;
1195
1196 attr->sample_period = 1;
1197
1198 attr->watermark = 0;
1199 attr->wakeup_events = 1000;
1200
1201 /* will enable all once we are ready */
1202 attr->disabled = 1;
1203 }
1204
1205 err = perf_evlist__open(evlist);
1206 if (err < 0) {
1207 printf("Couldn't create the events: %s\n", strerror(errno));
1208 goto out;
1209 }
1210
1211 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1212 ui__error("Failed to mmap the events: %s\n", strerror(errno));
1213 perf_evlist__close(evlist);
1214 goto out;
1215 }
1216
1217 rc = 0;
1218
1219out:
1220 return rc;
1221}
1222#endif
1223
1224static int read_events(struct perf_kvm_stat *kvm)
1225{
1226 int ret;
1227
1228 struct perf_tool eops = {
1229 .sample = process_sample_event,
1230 .comm = perf_event__process_comm,
1231 .ordered_samples = true,
1232 };
1233 struct perf_data_file file = {
1234 .path = kvm->file_name,
1235 .mode = PERF_DATA_MODE_READ,
1236 };
1237
1238 kvm->tool = eops;
1239 kvm->session = perf_session__new(&file, false, &kvm->tool);
1240 if (!kvm->session) {
1241 pr_err("Initializing perf session failed\n");
1242 return -EINVAL;
1243 }
1244
1245 if (!perf_session__has_traces(kvm->session, "kvm record"))
1246 return -EINVAL;
1247
1248 /*
1249 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1250 * traced in the old kernel.
1251 */
1252 ret = cpu_isa_config(kvm);
1253 if (ret < 0)
1254 return ret;
1255
1256 return perf_session__process_events(kvm->session, &kvm->tool);
1257}
1258
1259static int parse_target_str(struct perf_kvm_stat *kvm)
1260{
1261 if (kvm->pid_str) {
1262 kvm->pid_list = intlist__new(kvm->pid_str);
1263 if (kvm->pid_list == NULL) {
1264 pr_err("Error parsing process id string\n");
1265 return -EINVAL;
1266 }
1267 }
1268
1269 return 0;
1270}
1271
1272static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1273{
1274 int ret = -EINVAL;
1275 int vcpu = kvm->trace_vcpu;
1276
1277 if (parse_target_str(kvm) != 0)
1278 goto exit;
1279
1280 if (!verify_vcpu(vcpu))
1281 goto exit;
1282
1283 if (!select_key(kvm))
1284 goto exit;
1285
1286 if (!register_kvm_events_ops(kvm))
1287 goto exit;
1288
1289 init_kvm_event_record(kvm);
1290 setup_pager();
1291
1292 ret = read_events(kvm);
1293 if (ret)
1294 goto exit;
1295
1296 sort_result(kvm);
1297 print_result(kvm);
1298
1299exit:
1300 return ret;
1301}
1302
1303static const char * const kvm_events_tp[] = {
1304 "kvm:kvm_entry",
1305 "kvm:kvm_exit",
1306 "kvm:kvm_mmio",
1307 "kvm:kvm_pio",
1308};
1309
1310#define STRDUP_FAIL_EXIT(s) \
1311 ({ char *_p; \
1312 _p = strdup(s); \
1313 if (!_p) \
1314 return -ENOMEM; \
1315 _p; \
1316 })
1317
1318static int
1319kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1320{
1321 unsigned int rec_argc, i, j;
1322 const char **rec_argv;
1323 const char * const record_args[] = {
1324 "record",
1325 "-R",
1326 "-m", "1024",
1327 "-c", "1",
1328 };
1329
1330 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1331 2 * ARRAY_SIZE(kvm_events_tp);
1332 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1333
1334 if (rec_argv == NULL)
1335 return -ENOMEM;
1336
1337 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1338 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1339
1340 for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1341 rec_argv[i++] = "-e";
1342 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1343 }
1344
1345 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1346 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1347
1348 for (j = 1; j < (unsigned int)argc; j++, i++)
1349 rec_argv[i] = argv[j];
1350
1351 return cmd_record(i, rec_argv, NULL);
1352}
1353
1354static int
1355kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1356{
1357 const struct option kvm_events_report_options[] = {
1358 OPT_STRING(0, "event", &kvm->report_event, "report event",
1359 "event for reporting: vmexit, mmio, ioport"),
1360 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1361 "vcpu id to report"),
1362 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1363 "key for sorting: sample(sort by samples number)"
1364 " time (sort by avg time)"),
1365 OPT_STRING('p', "pid", &kvm->pid_str, "pid",
1366 "analyze events only for given process id(s)"),
1367 OPT_END()
1368 };
1369
1370 const char * const kvm_events_report_usage[] = {
1371 "perf kvm stat report [<options>]",
1372 NULL
1373 };
1374
1375 symbol__init();
1376
1377 if (argc) {
1378 argc = parse_options(argc, argv,
1379 kvm_events_report_options,
1380 kvm_events_report_usage, 0);
1381 if (argc)
1382 usage_with_options(kvm_events_report_usage,
1383 kvm_events_report_options);
1384 }
1385
1386 return kvm_events_report_vcpu(kvm);
1387}
1388
1389#ifdef HAVE_TIMERFD_SUPPORT
1390static struct perf_evlist *kvm_live_event_list(void)
1391{
1392 struct perf_evlist *evlist;
1393 char *tp, *name, *sys;
1394 unsigned int j;
1395 int err = -1;
1396
1397 evlist = perf_evlist__new();
1398 if (evlist == NULL)
1399 return NULL;
1400
1401 for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1402
1403 tp = strdup(kvm_events_tp[j]);
1404 if (tp == NULL)
1405 goto out;
1406
1407 /* split tracepoint into subsystem and name */
1408 sys = tp;
1409 name = strchr(tp, ':');
1410 if (name == NULL) {
1411 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1412 kvm_events_tp[j]);
1413 free(tp);
1414 goto out;
1415 }
1416 *name = '\0';
1417 name++;
1418
1419 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1420 pr_err("Failed to add %s tracepoint to the list\n", kvm_events_tp[j]);
1421 free(tp);
1422 goto out;
1423 }
1424
1425 free(tp);
1426 }
1427
1428 err = 0;
1429
1430out:
1431 if (err) {
1432 perf_evlist__delete(evlist);
1433 evlist = NULL;
1434 }
1435
1436 return evlist;
1437}
1438
1439static int kvm_events_live(struct perf_kvm_stat *kvm,
1440 int argc, const char **argv)
1441{
1442 char errbuf[BUFSIZ];
1443 int err;
1444
1445 const struct option live_options[] = {
1446 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1447 "record events on existing process id"),
1448 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1449 "number of mmap data pages",
1450 perf_evlist__parse_mmap_pages),
1451 OPT_INCR('v', "verbose", &verbose,
1452 "be more verbose (show counter open errors, etc)"),
1453 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1454 "system-wide collection from all CPUs"),
1455 OPT_UINTEGER('d', "display", &kvm->display_time,
1456 "time in seconds between display updates"),
1457 OPT_STRING(0, "event", &kvm->report_event, "report event",
1458 "event for reporting: vmexit, mmio, ioport"),
1459 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1460 "vcpu id to report"),
1461 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1462 "key for sorting: sample(sort by samples number)"
1463 " time (sort by avg time)"),
1464 OPT_U64(0, "duration", &kvm->duration,
1465 "show events other than HALT that take longer than duration usecs"),
1466 OPT_END()
1467 };
1468 const char * const live_usage[] = {
1469 "perf kvm stat live [<options>]",
1470 NULL
1471 };
1472 struct perf_data_file file = {
1473 .mode = PERF_DATA_MODE_WRITE,
1474 };
1475
1476
1477 /* event handling */
1478 kvm->tool.sample = process_sample_event;
1479 kvm->tool.comm = perf_event__process_comm;
1480 kvm->tool.exit = perf_event__process_exit;
1481 kvm->tool.fork = perf_event__process_fork;
1482 kvm->tool.lost = process_lost_event;
1483 kvm->tool.ordered_samples = true;
1484 perf_tool__fill_defaults(&kvm->tool);
1485
1486 /* set defaults */
1487 kvm->display_time = 1;
1488 kvm->opts.user_interval = 1;
1489 kvm->opts.mmap_pages = 512;
1490 kvm->opts.target.uses_mmap = false;
1491 kvm->opts.target.uid_str = NULL;
1492 kvm->opts.target.uid = UINT_MAX;
1493
1494 symbol__init();
1495 disable_buildid_cache();
1496
1497 use_browser = 0;
1498 setup_browser(false);
1499
1500 if (argc) {
1501 argc = parse_options(argc, argv, live_options,
1502 live_usage, 0);
1503 if (argc)
1504 usage_with_options(live_usage, live_options);
1505 }
1506
1507 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1508
1509 /*
1510 * target related setups
1511 */
1512 err = target__validate(&kvm->opts.target);
1513 if (err) {
1514 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1515 ui__warning("%s", errbuf);
1516 }
1517
1518 if (target__none(&kvm->opts.target))
1519 kvm->opts.target.system_wide = true;
1520
1521
1522 /*
1523 * generate the event list
1524 */
1525 kvm->evlist = kvm_live_event_list();
1526 if (kvm->evlist == NULL) {
1527 err = -1;
1528 goto out;
1529 }
1530
1531 symbol_conf.nr_events = kvm->evlist->nr_entries;
1532
1533 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1534 usage_with_options(live_usage, live_options);
1535
1536 /*
1537 * perf session
1538 */
1539 kvm->session = perf_session__new(&file, false, &kvm->tool);
1540 if (kvm->session == NULL) {
1541 err = -ENOMEM;
1542 goto out;
1543 }
1544 kvm->session->evlist = kvm->evlist;
1545 perf_session__set_id_hdr_size(kvm->session);
1546 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1547 kvm->evlist->threads, false);
1548 err = kvm_live_open_events(kvm);
1549 if (err)
1550 goto out;
1551
1552 err = kvm_events_live_report(kvm);
1553
1554out:
1555 exit_browser(0);
1556
1557 if (kvm->session)
1558 perf_session__delete(kvm->session);
1559 kvm->session = NULL;
1560 if (kvm->evlist)
1561 perf_evlist__delete(kvm->evlist);
1562
1563 return err;
1564}
1565#endif
1566
1567static void print_kvm_stat_usage(void)
1568{
1569 printf("Usage: perf kvm stat <command>\n\n");
1570
1571 printf("# Available commands:\n");
1572 printf("\trecord: record kvm events\n");
1573 printf("\treport: report statistical data of kvm events\n");
1574 printf("\tlive: live reporting of statistical data of kvm events\n");
1575
1576 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1577}
1578
1579static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1580{
1581 struct perf_kvm_stat kvm = {
1582 .file_name = file_name,
1583
1584 .trace_vcpu = -1,
1585 .report_event = "vmexit",
1586 .sort_key = "sample",
1587
1588 .exit_reasons = svm_exit_reasons,
1589 .exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
1590 .exit_reasons_isa = "SVM",
1591 };
1592
1593 if (argc == 1) {
1594 print_kvm_stat_usage();
1595 goto perf_stat;
1596 }
1597
1598 if (!strncmp(argv[1], "rec", 3))
1599 return kvm_events_record(&kvm, argc - 1, argv + 1);
1600
1601 if (!strncmp(argv[1], "rep", 3))
1602 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1603
1604#ifdef HAVE_TIMERFD_SUPPORT
1605 if (!strncmp(argv[1], "live", 4))
1606 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1607#endif
1608
1609perf_stat:
1610 return cmd_stat(argc, argv, NULL);
1611}
1612#endif
1613
1614static int __cmd_record(const char *file_name, int argc, const char **argv)
1615{
1616 int rec_argc, i = 0, j;
1617 const char **rec_argv;
1618
1619 rec_argc = argc + 2;
1620 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1621 rec_argv[i++] = strdup("record");
1622 rec_argv[i++] = strdup("-o");
1623 rec_argv[i++] = strdup(file_name);
1624 for (j = 1; j < argc; j++, i++)
1625 rec_argv[i] = argv[j];
1626
1627 BUG_ON(i != rec_argc);
1628
1629 return cmd_record(i, rec_argv, NULL);
1630}
1631
1632static int __cmd_report(const char *file_name, int argc, const char **argv)
1633{
1634 int rec_argc, i = 0, j;
1635 const char **rec_argv;
1636
1637 rec_argc = argc + 2;
1638 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1639 rec_argv[i++] = strdup("report");
1640 rec_argv[i++] = strdup("-i");
1641 rec_argv[i++] = strdup(file_name);
1642 for (j = 1; j < argc; j++, i++)
1643 rec_argv[i] = argv[j];
1644
1645 BUG_ON(i != rec_argc);
1646
1647 return cmd_report(i, rec_argv, NULL);
1648}
1649
1650static int
1651__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1652{
1653 int rec_argc, i = 0, j;
1654 const char **rec_argv;
1655
1656 rec_argc = argc + 2;
1657 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1658 rec_argv[i++] = strdup("buildid-list");
1659 rec_argv[i++] = strdup("-i");
1660 rec_argv[i++] = strdup(file_name);
1661 for (j = 1; j < argc; j++, i++)
1662 rec_argv[i] = argv[j];
1663
1664 BUG_ON(i != rec_argc);
1665
1666 return cmd_buildid_list(i, rec_argv, NULL);
1667}
1668
1669int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1670{
1671 const char *file_name = NULL;
1672 const struct option kvm_options[] = {
1673 OPT_STRING('i', "input", &file_name, "file",
1674 "Input file name"),
1675 OPT_STRING('o', "output", &file_name, "file",
1676 "Output file name"),
1677 OPT_BOOLEAN(0, "guest", &perf_guest,
1678 "Collect guest os data"),
1679 OPT_BOOLEAN(0, "host", &perf_host,
1680 "Collect host os data"),
1681 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1682 "guest mount directory under which every guest os"
1683 " instance has a subdir"),
1684 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1685 "file", "file saving guest os vmlinux"),
1686 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1687 "file", "file saving guest os /proc/kallsyms"),
1688 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1689 "file", "file saving guest os /proc/modules"),
1690 OPT_INCR('v', "verbose", &verbose,
1691 "be more verbose (show counter open errors, etc)"),
1692 OPT_END()
1693 };
1694
1695 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1696 "buildid-list", "stat", NULL };
1697 const char *kvm_usage[] = { NULL, NULL };
1698
1699 perf_host = 0;
1700 perf_guest = 1;
1701
1702 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1703 PARSE_OPT_STOP_AT_NON_OPTION);
1704 if (!argc)
1705 usage_with_options(kvm_usage, kvm_options);
1706
1707 if (!perf_host)
1708 perf_guest = 1;
1709
1710 if (!file_name) {
1711 file_name = get_filename_for_perf_kvm();
1712
1713 if (!file_name) {
1714 pr_err("Failed to allocate memory for filename\n");
1715 return -ENOMEM;
1716 }
1717 }
1718
1719 if (!strncmp(argv[0], "rec", 3))
1720 return __cmd_record(file_name, argc, argv);
1721 else if (!strncmp(argv[0], "rep", 3))
1722 return __cmd_report(file_name, argc, argv);
1723 else if (!strncmp(argv[0], "diff", 4))
1724 return cmd_diff(argc, argv, NULL);
1725 else if (!strncmp(argv[0], "top", 3))
1726 return cmd_top(argc, argv, NULL);
1727 else if (!strncmp(argv[0], "buildid-list", 12))
1728 return __cmd_buildid_list(file_name, argc, argv);
1729#if defined(__i386__) || defined(__x86_64__)
1730 else if (!strncmp(argv[0], "stat", 4))
1731 return kvm_cmd_stat(file_name, argc, argv);
1732#endif
1733 else
1734 usage_with_options(kvm_usage, kvm_options);
1735
1736 return 0;
1737}