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 "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}
1// SPDX-License-Identifier: GPL-2.0
2#include "builtin.h"
3#include "perf.h"
4
5#include "util/build-id.h"
6#include "util/evsel.h"
7#include "util/evlist.h"
8#include "util/mmap.h"
9#include "util/term.h"
10#include "util/symbol.h"
11#include "util/thread.h"
12#include "util/header.h"
13#include "util/session.h"
14#include "util/intlist.h"
15#include <subcmd/pager.h>
16#include <subcmd/parse-options.h>
17#include "util/trace-event.h"
18#include "util/debug.h"
19#include "util/tool.h"
20#include "util/stat.h"
21#include "util/synthetic-events.h"
22#include "util/top.h"
23#include "util/data.h"
24#include "util/ordered-events.h"
25#include "util/kvm-stat.h"
26#include "ui/ui.h"
27
28#include <sys/prctl.h>
29#ifdef HAVE_TIMERFD_SUPPORT
30#include <sys/timerfd.h>
31#endif
32#include <sys/time.h>
33#include <sys/types.h>
34#include <sys/stat.h>
35#include <fcntl.h>
36
37#include <linux/err.h>
38#include <linux/kernel.h>
39#include <linux/string.h>
40#include <linux/time64.h>
41#include <linux/zalloc.h>
42#include <errno.h>
43#include <inttypes.h>
44#include <poll.h>
45#include <termios.h>
46#include <semaphore.h>
47#include <signal.h>
48#include <math.h>
49
50static const char *get_filename_for_perf_kvm(void)
51{
52 const char *filename;
53
54 if (perf_host && !perf_guest)
55 filename = strdup("perf.data.host");
56 else if (!perf_host && perf_guest)
57 filename = strdup("perf.data.guest");
58 else
59 filename = strdup("perf.data.kvm");
60
61 return filename;
62}
63
64#ifdef HAVE_KVM_STAT_SUPPORT
65
66void exit_event_get_key(struct evsel *evsel,
67 struct perf_sample *sample,
68 struct event_key *key)
69{
70 key->info = 0;
71 key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
72}
73
74bool kvm_exit_event(struct evsel *evsel)
75{
76 return !strcmp(evsel->name, kvm_exit_trace);
77}
78
79bool exit_event_begin(struct evsel *evsel,
80 struct perf_sample *sample, struct event_key *key)
81{
82 if (kvm_exit_event(evsel)) {
83 exit_event_get_key(evsel, sample, key);
84 return true;
85 }
86
87 return false;
88}
89
90bool kvm_entry_event(struct evsel *evsel)
91{
92 return !strcmp(evsel->name, kvm_entry_trace);
93}
94
95bool exit_event_end(struct evsel *evsel,
96 struct perf_sample *sample __maybe_unused,
97 struct event_key *key __maybe_unused)
98{
99 return kvm_entry_event(evsel);
100}
101
102static const char *get_exit_reason(struct perf_kvm_stat *kvm,
103 struct exit_reasons_table *tbl,
104 u64 exit_code)
105{
106 while (tbl->reason != NULL) {
107 if (tbl->exit_code == exit_code)
108 return tbl->reason;
109 tbl++;
110 }
111
112 pr_err("unknown kvm exit code:%lld on %s\n",
113 (unsigned long long)exit_code, kvm->exit_reasons_isa);
114 return "UNKNOWN";
115}
116
117void exit_event_decode_key(struct perf_kvm_stat *kvm,
118 struct event_key *key,
119 char *decode)
120{
121 const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
122 key->key);
123
124 scnprintf(decode, decode_str_len, "%s", exit_reason);
125}
126
127static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
128{
129 struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
130
131 for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
132 if (!strcmp(events_ops->name, kvm->report_event)) {
133 kvm->events_ops = events_ops->ops;
134 return true;
135 }
136 }
137
138 return false;
139}
140
141struct vcpu_event_record {
142 int vcpu_id;
143 u64 start_time;
144 struct kvm_event *last_event;
145};
146
147
148static void init_kvm_event_record(struct perf_kvm_stat *kvm)
149{
150 unsigned int i;
151
152 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
153 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
154}
155
156#ifdef HAVE_TIMERFD_SUPPORT
157static void clear_events_cache_stats(struct list_head *kvm_events_cache)
158{
159 struct list_head *head;
160 struct kvm_event *event;
161 unsigned int i;
162 int j;
163
164 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
165 head = &kvm_events_cache[i];
166 list_for_each_entry(event, head, hash_entry) {
167 /* reset stats for event */
168 event->total.time = 0;
169 init_stats(&event->total.stats);
170
171 for (j = 0; j < event->max_vcpu; ++j) {
172 event->vcpu[j].time = 0;
173 init_stats(&event->vcpu[j].stats);
174 }
175 }
176 }
177}
178#endif
179
180static int kvm_events_hash_fn(u64 key)
181{
182 return key & (EVENTS_CACHE_SIZE - 1);
183}
184
185static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
186{
187 int old_max_vcpu = event->max_vcpu;
188 void *prev;
189
190 if (vcpu_id < event->max_vcpu)
191 return true;
192
193 while (event->max_vcpu <= vcpu_id)
194 event->max_vcpu += DEFAULT_VCPU_NUM;
195
196 prev = event->vcpu;
197 event->vcpu = realloc(event->vcpu,
198 event->max_vcpu * sizeof(*event->vcpu));
199 if (!event->vcpu) {
200 free(prev);
201 pr_err("Not enough memory\n");
202 return false;
203 }
204
205 memset(event->vcpu + old_max_vcpu, 0,
206 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
207 return true;
208}
209
210static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
211{
212 struct kvm_event *event;
213
214 event = zalloc(sizeof(*event));
215 if (!event) {
216 pr_err("Not enough memory\n");
217 return NULL;
218 }
219
220 event->key = *key;
221 init_stats(&event->total.stats);
222 return event;
223}
224
225static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
226 struct event_key *key)
227{
228 struct kvm_event *event;
229 struct list_head *head;
230
231 BUG_ON(key->key == INVALID_KEY);
232
233 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
234 list_for_each_entry(event, head, hash_entry) {
235 if (event->key.key == key->key && event->key.info == key->info)
236 return event;
237 }
238
239 event = kvm_alloc_init_event(key);
240 if (!event)
241 return NULL;
242
243 list_add(&event->hash_entry, head);
244 return event;
245}
246
247static bool handle_begin_event(struct perf_kvm_stat *kvm,
248 struct vcpu_event_record *vcpu_record,
249 struct event_key *key, u64 timestamp)
250{
251 struct kvm_event *event = NULL;
252
253 if (key->key != INVALID_KEY)
254 event = find_create_kvm_event(kvm, key);
255
256 vcpu_record->last_event = event;
257 vcpu_record->start_time = timestamp;
258 return true;
259}
260
261static void
262kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
263{
264 kvm_stats->time += time_diff;
265 update_stats(&kvm_stats->stats, time_diff);
266}
267
268static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
269{
270 struct kvm_event_stats *kvm_stats = &event->total;
271
272 if (vcpu_id != -1)
273 kvm_stats = &event->vcpu[vcpu_id];
274
275 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
276 avg_stats(&kvm_stats->stats));
277}
278
279static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
280 u64 time_diff)
281{
282 if (vcpu_id == -1) {
283 kvm_update_event_stats(&event->total, time_diff);
284 return true;
285 }
286
287 if (!kvm_event_expand(event, vcpu_id))
288 return false;
289
290 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
291 return true;
292}
293
294static bool is_child_event(struct perf_kvm_stat *kvm,
295 struct evsel *evsel,
296 struct perf_sample *sample,
297 struct event_key *key)
298{
299 struct child_event_ops *child_ops;
300
301 child_ops = kvm->events_ops->child_ops;
302
303 if (!child_ops)
304 return false;
305
306 for (; child_ops->name; child_ops++) {
307 if (!strcmp(evsel->name, child_ops->name)) {
308 child_ops->get_key(evsel, sample, key);
309 return true;
310 }
311 }
312
313 return false;
314}
315
316static bool handle_child_event(struct perf_kvm_stat *kvm,
317 struct vcpu_event_record *vcpu_record,
318 struct event_key *key,
319 struct perf_sample *sample __maybe_unused)
320{
321 struct kvm_event *event = NULL;
322
323 if (key->key != INVALID_KEY)
324 event = find_create_kvm_event(kvm, key);
325
326 vcpu_record->last_event = event;
327
328 return true;
329}
330
331static bool skip_event(const char *event)
332{
333 const char * const *skip_events;
334
335 for (skip_events = kvm_skip_events; *skip_events; skip_events++)
336 if (!strcmp(event, *skip_events))
337 return true;
338
339 return false;
340}
341
342static bool handle_end_event(struct perf_kvm_stat *kvm,
343 struct vcpu_event_record *vcpu_record,
344 struct event_key *key,
345 struct perf_sample *sample)
346{
347 struct kvm_event *event;
348 u64 time_begin, time_diff;
349 int vcpu;
350
351 if (kvm->trace_vcpu == -1)
352 vcpu = -1;
353 else
354 vcpu = vcpu_record->vcpu_id;
355
356 event = vcpu_record->last_event;
357 time_begin = vcpu_record->start_time;
358
359 /* The begin event is not caught. */
360 if (!time_begin)
361 return true;
362
363 /*
364 * In some case, the 'begin event' only records the start timestamp,
365 * the actual event is recognized in the 'end event' (e.g. mmio-event).
366 */
367
368 /* Both begin and end events did not get the key. */
369 if (!event && key->key == INVALID_KEY)
370 return true;
371
372 if (!event)
373 event = find_create_kvm_event(kvm, key);
374
375 if (!event)
376 return false;
377
378 vcpu_record->last_event = NULL;
379 vcpu_record->start_time = 0;
380
381 /* seems to happen once in a while during live mode */
382 if (sample->time < time_begin) {
383 pr_debug("End time before begin time; skipping event.\n");
384 return true;
385 }
386
387 time_diff = sample->time - time_begin;
388
389 if (kvm->duration && time_diff > kvm->duration) {
390 char decode[decode_str_len];
391
392 kvm->events_ops->decode_key(kvm, &event->key, decode);
393 if (!skip_event(decode)) {
394 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
395 sample->time, sample->pid, vcpu_record->vcpu_id,
396 decode, time_diff / NSEC_PER_USEC);
397 }
398 }
399
400 return update_kvm_event(event, vcpu, time_diff);
401}
402
403static
404struct vcpu_event_record *per_vcpu_record(struct thread *thread,
405 struct evsel *evsel,
406 struct perf_sample *sample)
407{
408 /* Only kvm_entry records vcpu id. */
409 if (!thread__priv(thread) && kvm_entry_event(evsel)) {
410 struct vcpu_event_record *vcpu_record;
411
412 vcpu_record = zalloc(sizeof(*vcpu_record));
413 if (!vcpu_record) {
414 pr_err("%s: Not enough memory\n", __func__);
415 return NULL;
416 }
417
418 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
419 vcpu_id_str);
420 thread__set_priv(thread, vcpu_record);
421 }
422
423 return thread__priv(thread);
424}
425
426static bool handle_kvm_event(struct perf_kvm_stat *kvm,
427 struct thread *thread,
428 struct evsel *evsel,
429 struct perf_sample *sample)
430{
431 struct vcpu_event_record *vcpu_record;
432 struct event_key key = { .key = INVALID_KEY,
433 .exit_reasons = kvm->exit_reasons };
434
435 vcpu_record = per_vcpu_record(thread, evsel, sample);
436 if (!vcpu_record)
437 return true;
438
439 /* only process events for vcpus user cares about */
440 if ((kvm->trace_vcpu != -1) &&
441 (kvm->trace_vcpu != vcpu_record->vcpu_id))
442 return true;
443
444 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
445 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
446
447 if (is_child_event(kvm, evsel, sample, &key))
448 return handle_child_event(kvm, vcpu_record, &key, sample);
449
450 if (kvm->events_ops->is_end_event(evsel, sample, &key))
451 return handle_end_event(kvm, vcpu_record, &key, sample);
452
453 return true;
454}
455
456#define GET_EVENT_KEY(func, field) \
457static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
458{ \
459 if (vcpu == -1) \
460 return event->total.field; \
461 \
462 if (vcpu >= event->max_vcpu) \
463 return 0; \
464 \
465 return event->vcpu[vcpu].field; \
466}
467
468#define COMPARE_EVENT_KEY(func, field) \
469GET_EVENT_KEY(func, field) \
470static int compare_kvm_event_ ## func(struct kvm_event *one, \
471 struct kvm_event *two, int vcpu)\
472{ \
473 return get_event_ ##func(one, vcpu) > \
474 get_event_ ##func(two, vcpu); \
475}
476
477GET_EVENT_KEY(time, time);
478COMPARE_EVENT_KEY(count, stats.n);
479COMPARE_EVENT_KEY(mean, stats.mean);
480GET_EVENT_KEY(max, stats.max);
481GET_EVENT_KEY(min, stats.min);
482
483#define DEF_SORT_NAME_KEY(name, compare_key) \
484 { #name, compare_kvm_event_ ## compare_key }
485
486static struct kvm_event_key keys[] = {
487 DEF_SORT_NAME_KEY(sample, count),
488 DEF_SORT_NAME_KEY(time, mean),
489 { NULL, NULL }
490};
491
492static bool select_key(struct perf_kvm_stat *kvm)
493{
494 int i;
495
496 for (i = 0; keys[i].name; i++) {
497 if (!strcmp(keys[i].name, kvm->sort_key)) {
498 kvm->compare = keys[i].key;
499 return true;
500 }
501 }
502
503 pr_err("Unknown compare key:%s\n", kvm->sort_key);
504 return false;
505}
506
507static void insert_to_result(struct rb_root *result, struct kvm_event *event,
508 key_cmp_fun bigger, int vcpu)
509{
510 struct rb_node **rb = &result->rb_node;
511 struct rb_node *parent = NULL;
512 struct kvm_event *p;
513
514 while (*rb) {
515 p = container_of(*rb, struct kvm_event, rb);
516 parent = *rb;
517
518 if (bigger(event, p, vcpu))
519 rb = &(*rb)->rb_left;
520 else
521 rb = &(*rb)->rb_right;
522 }
523
524 rb_link_node(&event->rb, parent, rb);
525 rb_insert_color(&event->rb, result);
526}
527
528static void
529update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
530{
531 int vcpu = kvm->trace_vcpu;
532
533 kvm->total_count += get_event_count(event, vcpu);
534 kvm->total_time += get_event_time(event, vcpu);
535}
536
537static bool event_is_valid(struct kvm_event *event, int vcpu)
538{
539 return !!get_event_count(event, vcpu);
540}
541
542static void sort_result(struct perf_kvm_stat *kvm)
543{
544 unsigned int i;
545 int vcpu = kvm->trace_vcpu;
546 struct kvm_event *event;
547
548 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
549 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
550 if (event_is_valid(event, vcpu)) {
551 update_total_count(kvm, event);
552 insert_to_result(&kvm->result, event,
553 kvm->compare, vcpu);
554 }
555 }
556 }
557}
558
559/* returns left most element of result, and erase it */
560static struct kvm_event *pop_from_result(struct rb_root *result)
561{
562 struct rb_node *node = rb_first(result);
563
564 if (!node)
565 return NULL;
566
567 rb_erase(node, result);
568 return container_of(node, struct kvm_event, rb);
569}
570
571static void print_vcpu_info(struct perf_kvm_stat *kvm)
572{
573 int vcpu = kvm->trace_vcpu;
574
575 pr_info("Analyze events for ");
576
577 if (kvm->opts.target.system_wide)
578 pr_info("all VMs, ");
579 else if (kvm->opts.target.pid)
580 pr_info("pid(s) %s, ", kvm->opts.target.pid);
581 else
582 pr_info("dazed and confused on what is monitored, ");
583
584 if (vcpu == -1)
585 pr_info("all VCPUs:\n\n");
586 else
587 pr_info("VCPU %d:\n\n", vcpu);
588}
589
590static void show_timeofday(void)
591{
592 char date[64];
593 struct timeval tv;
594 struct tm ltime;
595
596 gettimeofday(&tv, NULL);
597 if (localtime_r(&tv.tv_sec, <ime)) {
598 strftime(date, sizeof(date), "%H:%M:%S", <ime);
599 pr_info("%s.%06ld", date, tv.tv_usec);
600 } else
601 pr_info("00:00:00.000000");
602
603 return;
604}
605
606static void print_result(struct perf_kvm_stat *kvm)
607{
608 char decode[decode_str_len];
609 struct kvm_event *event;
610 int vcpu = kvm->trace_vcpu;
611
612 if (kvm->live) {
613 puts(CONSOLE_CLEAR);
614 show_timeofday();
615 }
616
617 pr_info("\n\n");
618 print_vcpu_info(kvm);
619 pr_info("%*s ", decode_str_len, kvm->events_ops->name);
620 pr_info("%10s ", "Samples");
621 pr_info("%9s ", "Samples%");
622
623 pr_info("%9s ", "Time%");
624 pr_info("%11s ", "Min Time");
625 pr_info("%11s ", "Max Time");
626 pr_info("%16s ", "Avg time");
627 pr_info("\n\n");
628
629 while ((event = pop_from_result(&kvm->result))) {
630 u64 ecount, etime, max, min;
631
632 ecount = get_event_count(event, vcpu);
633 etime = get_event_time(event, vcpu);
634 max = get_event_max(event, vcpu);
635 min = get_event_min(event, vcpu);
636
637 kvm->events_ops->decode_key(kvm, &event->key, decode);
638 pr_info("%*s ", decode_str_len, decode);
639 pr_info("%10llu ", (unsigned long long)ecount);
640 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
641 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
642 pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
643 pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
644 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
645 kvm_event_rel_stddev(vcpu, event));
646 pr_info("\n");
647 }
648
649 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
650 kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
651
652 if (kvm->lost_events)
653 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
654}
655
656#ifdef HAVE_TIMERFD_SUPPORT
657static int process_lost_event(struct perf_tool *tool,
658 union perf_event *event __maybe_unused,
659 struct perf_sample *sample __maybe_unused,
660 struct machine *machine __maybe_unused)
661{
662 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
663
664 kvm->lost_events++;
665 return 0;
666}
667#endif
668
669static bool skip_sample(struct perf_kvm_stat *kvm,
670 struct perf_sample *sample)
671{
672 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
673 return true;
674
675 return false;
676}
677
678static int process_sample_event(struct perf_tool *tool,
679 union perf_event *event,
680 struct perf_sample *sample,
681 struct evsel *evsel,
682 struct machine *machine)
683{
684 int err = 0;
685 struct thread *thread;
686 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
687 tool);
688
689 if (skip_sample(kvm, sample))
690 return 0;
691
692 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
693 if (thread == NULL) {
694 pr_debug("problem processing %d event, skipping it.\n",
695 event->header.type);
696 return -1;
697 }
698
699 if (!handle_kvm_event(kvm, thread, evsel, sample))
700 err = -1;
701
702 thread__put(thread);
703 return err;
704}
705
706static int cpu_isa_config(struct perf_kvm_stat *kvm)
707{
708 char buf[128], *cpuid;
709 int err;
710
711 if (kvm->live) {
712 err = get_cpuid(buf, sizeof(buf));
713 if (err != 0) {
714 pr_err("Failed to look up CPU type: %s\n",
715 str_error_r(err, buf, sizeof(buf)));
716 return -err;
717 }
718 cpuid = buf;
719 } else
720 cpuid = kvm->session->header.env.cpuid;
721
722 if (!cpuid) {
723 pr_err("Failed to look up CPU type\n");
724 return -EINVAL;
725 }
726
727 err = cpu_isa_init(kvm, cpuid);
728 if (err == -ENOTSUP)
729 pr_err("CPU %s is not supported.\n", cpuid);
730
731 return err;
732}
733
734static bool verify_vcpu(int vcpu)
735{
736 if (vcpu != -1 && vcpu < 0) {
737 pr_err("Invalid vcpu:%d.\n", vcpu);
738 return false;
739 }
740
741 return true;
742}
743
744#ifdef HAVE_TIMERFD_SUPPORT
745/* keeping the max events to a modest level to keep
746 * the processing of samples per mmap smooth.
747 */
748#define PERF_KVM__MAX_EVENTS_PER_MMAP 25
749
750static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
751 u64 *mmap_time)
752{
753 struct evlist *evlist = kvm->evlist;
754 union perf_event *event;
755 struct mmap *md;
756 u64 timestamp;
757 s64 n = 0;
758 int err;
759
760 *mmap_time = ULLONG_MAX;
761 md = &evlist->mmap[idx];
762 err = perf_mmap__read_init(md);
763 if (err < 0)
764 return (err == -EAGAIN) ? 0 : -1;
765
766 while ((event = perf_mmap__read_event(md)) != NULL) {
767 err = perf_evlist__parse_sample_timestamp(evlist, event, ×tamp);
768 if (err) {
769 perf_mmap__consume(md);
770 pr_err("Failed to parse sample\n");
771 return -1;
772 }
773
774 err = perf_session__queue_event(kvm->session, event, timestamp, 0);
775 /*
776 * FIXME: Here we can't consume the event, as perf_session__queue_event will
777 * point to it, and it'll get possibly overwritten by the kernel.
778 */
779 perf_mmap__consume(md);
780
781 if (err) {
782 pr_err("Failed to enqueue sample: %d\n", err);
783 return -1;
784 }
785
786 /* save time stamp of our first sample for this mmap */
787 if (n == 0)
788 *mmap_time = timestamp;
789
790 /* limit events per mmap handled all at once */
791 n++;
792 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
793 break;
794 }
795
796 perf_mmap__read_done(md);
797 return n;
798}
799
800static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
801{
802 int i, err, throttled = 0;
803 s64 n, ntotal = 0;
804 u64 flush_time = ULLONG_MAX, mmap_time;
805
806 for (i = 0; i < kvm->evlist->core.nr_mmaps; i++) {
807 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
808 if (n < 0)
809 return -1;
810
811 /* flush time is going to be the minimum of all the individual
812 * mmap times. Essentially, we flush all the samples queued up
813 * from the last pass under our minimal start time -- that leaves
814 * a very small race for samples to come in with a lower timestamp.
815 * The ioctl to return the perf_clock timestamp should close the
816 * race entirely.
817 */
818 if (mmap_time < flush_time)
819 flush_time = mmap_time;
820
821 ntotal += n;
822 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
823 throttled = 1;
824 }
825
826 /* flush queue after each round in which we processed events */
827 if (ntotal) {
828 struct ordered_events *oe = &kvm->session->ordered_events;
829
830 oe->next_flush = flush_time;
831 err = ordered_events__flush(oe, OE_FLUSH__ROUND);
832 if (err) {
833 if (kvm->lost_events)
834 pr_info("\nLost events: %" PRIu64 "\n\n",
835 kvm->lost_events);
836 return err;
837 }
838 }
839
840 return throttled;
841}
842
843static volatile int done;
844
845static void sig_handler(int sig __maybe_unused)
846{
847 done = 1;
848}
849
850static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
851{
852 struct itimerspec new_value;
853 int rc = -1;
854
855 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
856 if (kvm->timerfd < 0) {
857 pr_err("timerfd_create failed\n");
858 goto out;
859 }
860
861 new_value.it_value.tv_sec = kvm->display_time;
862 new_value.it_value.tv_nsec = 0;
863 new_value.it_interval.tv_sec = kvm->display_time;
864 new_value.it_interval.tv_nsec = 0;
865
866 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
867 pr_err("timerfd_settime failed: %d\n", errno);
868 close(kvm->timerfd);
869 goto out;
870 }
871
872 rc = 0;
873out:
874 return rc;
875}
876
877static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
878{
879 uint64_t c;
880 int rc;
881
882 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
883 if (rc < 0) {
884 if (errno == EAGAIN)
885 return 0;
886
887 pr_err("Failed to read timer fd: %d\n", errno);
888 return -1;
889 }
890
891 if (rc != sizeof(uint64_t)) {
892 pr_err("Error reading timer fd - invalid size returned\n");
893 return -1;
894 }
895
896 if (c != 1)
897 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
898
899 /* update display */
900 sort_result(kvm);
901 print_result(kvm);
902
903 /* reset counts */
904 clear_events_cache_stats(kvm->kvm_events_cache);
905 kvm->total_count = 0;
906 kvm->total_time = 0;
907 kvm->lost_events = 0;
908
909 return 0;
910}
911
912static int fd_set_nonblock(int fd)
913{
914 long arg = 0;
915
916 arg = fcntl(fd, F_GETFL);
917 if (arg < 0) {
918 pr_err("Failed to get current flags for fd %d\n", fd);
919 return -1;
920 }
921
922 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
923 pr_err("Failed to set non-block option on fd %d\n", fd);
924 return -1;
925 }
926
927 return 0;
928}
929
930static int perf_kvm__handle_stdin(void)
931{
932 int c;
933
934 c = getc(stdin);
935 if (c == 'q')
936 return 1;
937
938 return 0;
939}
940
941static int kvm_events_live_report(struct perf_kvm_stat *kvm)
942{
943 int nr_stdin, ret, err = -EINVAL;
944 struct termios save;
945
946 /* live flag must be set first */
947 kvm->live = true;
948
949 ret = cpu_isa_config(kvm);
950 if (ret < 0)
951 return ret;
952
953 if (!verify_vcpu(kvm->trace_vcpu) ||
954 !select_key(kvm) ||
955 !register_kvm_events_ops(kvm)) {
956 goto out;
957 }
958
959 set_term_quiet_input(&save);
960 init_kvm_event_record(kvm);
961
962 signal(SIGINT, sig_handler);
963 signal(SIGTERM, sig_handler);
964
965 /* add timer fd */
966 if (perf_kvm__timerfd_create(kvm) < 0) {
967 err = -1;
968 goto out;
969 }
970
971 if (evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
972 goto out;
973
974 nr_stdin = evlist__add_pollfd(kvm->evlist, fileno(stdin));
975 if (nr_stdin < 0)
976 goto out;
977
978 if (fd_set_nonblock(fileno(stdin)) != 0)
979 goto out;
980
981 /* everything is good - enable the events and process */
982 evlist__enable(kvm->evlist);
983
984 while (!done) {
985 struct fdarray *fda = &kvm->evlist->core.pollfd;
986 int rc;
987
988 rc = perf_kvm__mmap_read(kvm);
989 if (rc < 0)
990 break;
991
992 err = perf_kvm__handle_timerfd(kvm);
993 if (err)
994 goto out;
995
996 if (fda->entries[nr_stdin].revents & POLLIN)
997 done = perf_kvm__handle_stdin();
998
999 if (!rc && !done)
1000 err = fdarray__poll(fda, 100);
1001 }
1002
1003 evlist__disable(kvm->evlist);
1004
1005 if (err == 0) {
1006 sort_result(kvm);
1007 print_result(kvm);
1008 }
1009
1010out:
1011 if (kvm->timerfd >= 0)
1012 close(kvm->timerfd);
1013
1014 tcsetattr(0, TCSAFLUSH, &save);
1015 return err;
1016}
1017
1018static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1019{
1020 int err, rc = -1;
1021 struct evsel *pos;
1022 struct evlist *evlist = kvm->evlist;
1023 char sbuf[STRERR_BUFSIZE];
1024
1025 perf_evlist__config(evlist, &kvm->opts, NULL);
1026
1027 /*
1028 * Note: exclude_{guest,host} do not apply here.
1029 * This command processes KVM tracepoints from host only
1030 */
1031 evlist__for_each_entry(evlist, pos) {
1032 struct perf_event_attr *attr = &pos->core.attr;
1033
1034 /* make sure these *are* set */
1035 perf_evsel__set_sample_bit(pos, TID);
1036 perf_evsel__set_sample_bit(pos, TIME);
1037 perf_evsel__set_sample_bit(pos, CPU);
1038 perf_evsel__set_sample_bit(pos, RAW);
1039 /* make sure these are *not*; want as small a sample as possible */
1040 perf_evsel__reset_sample_bit(pos, PERIOD);
1041 perf_evsel__reset_sample_bit(pos, IP);
1042 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1043 perf_evsel__reset_sample_bit(pos, ADDR);
1044 perf_evsel__reset_sample_bit(pos, READ);
1045 attr->mmap = 0;
1046 attr->comm = 0;
1047 attr->task = 0;
1048
1049 attr->sample_period = 1;
1050
1051 attr->watermark = 0;
1052 attr->wakeup_events = 1000;
1053
1054 /* will enable all once we are ready */
1055 attr->disabled = 1;
1056 }
1057
1058 err = evlist__open(evlist);
1059 if (err < 0) {
1060 printf("Couldn't create the events: %s\n",
1061 str_error_r(errno, sbuf, sizeof(sbuf)));
1062 goto out;
1063 }
1064
1065 if (evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1066 ui__error("Failed to mmap the events: %s\n",
1067 str_error_r(errno, sbuf, sizeof(sbuf)));
1068 evlist__close(evlist);
1069 goto out;
1070 }
1071
1072 rc = 0;
1073
1074out:
1075 return rc;
1076}
1077#endif
1078
1079static int read_events(struct perf_kvm_stat *kvm)
1080{
1081 int ret;
1082
1083 struct perf_tool eops = {
1084 .sample = process_sample_event,
1085 .comm = perf_event__process_comm,
1086 .namespaces = perf_event__process_namespaces,
1087 .ordered_events = true,
1088 };
1089 struct perf_data file = {
1090 .path = kvm->file_name,
1091 .mode = PERF_DATA_MODE_READ,
1092 .force = kvm->force,
1093 };
1094
1095 kvm->tool = eops;
1096 kvm->session = perf_session__new(&file, false, &kvm->tool);
1097 if (IS_ERR(kvm->session)) {
1098 pr_err("Initializing perf session failed\n");
1099 return PTR_ERR(kvm->session);
1100 }
1101
1102 symbol__init(&kvm->session->header.env);
1103
1104 if (!perf_session__has_traces(kvm->session, "kvm record")) {
1105 ret = -EINVAL;
1106 goto out_delete;
1107 }
1108
1109 /*
1110 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1111 * traced in the old kernel.
1112 */
1113 ret = cpu_isa_config(kvm);
1114 if (ret < 0)
1115 goto out_delete;
1116
1117 ret = perf_session__process_events(kvm->session);
1118
1119out_delete:
1120 perf_session__delete(kvm->session);
1121 return ret;
1122}
1123
1124static int parse_target_str(struct perf_kvm_stat *kvm)
1125{
1126 if (kvm->opts.target.pid) {
1127 kvm->pid_list = intlist__new(kvm->opts.target.pid);
1128 if (kvm->pid_list == NULL) {
1129 pr_err("Error parsing process id string\n");
1130 return -EINVAL;
1131 }
1132 }
1133
1134 return 0;
1135}
1136
1137static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1138{
1139 int ret = -EINVAL;
1140 int vcpu = kvm->trace_vcpu;
1141
1142 if (parse_target_str(kvm) != 0)
1143 goto exit;
1144
1145 if (!verify_vcpu(vcpu))
1146 goto exit;
1147
1148 if (!select_key(kvm))
1149 goto exit;
1150
1151 if (!register_kvm_events_ops(kvm))
1152 goto exit;
1153
1154 init_kvm_event_record(kvm);
1155 setup_pager();
1156
1157 ret = read_events(kvm);
1158 if (ret)
1159 goto exit;
1160
1161 sort_result(kvm);
1162 print_result(kvm);
1163
1164exit:
1165 return ret;
1166}
1167
1168#define STRDUP_FAIL_EXIT(s) \
1169 ({ char *_p; \
1170 _p = strdup(s); \
1171 if (!_p) \
1172 return -ENOMEM; \
1173 _p; \
1174 })
1175
1176int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1177{
1178 return 0;
1179}
1180
1181static int
1182kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1183{
1184 unsigned int rec_argc, i, j, events_tp_size;
1185 const char **rec_argv;
1186 const char * const record_args[] = {
1187 "record",
1188 "-R",
1189 "-m", "1024",
1190 "-c", "1",
1191 };
1192 const char * const kvm_stat_record_usage[] = {
1193 "perf kvm stat record [<options>]",
1194 NULL
1195 };
1196 const char * const *events_tp;
1197 int ret;
1198
1199 events_tp_size = 0;
1200 ret = setup_kvm_events_tp(kvm);
1201 if (ret < 0) {
1202 pr_err("Unable to setup the kvm tracepoints\n");
1203 return ret;
1204 }
1205
1206 for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1207 events_tp_size++;
1208
1209 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1210 2 * events_tp_size;
1211 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1212
1213 if (rec_argv == NULL)
1214 return -ENOMEM;
1215
1216 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1217 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1218
1219 for (j = 0; j < events_tp_size; j++) {
1220 rec_argv[i++] = "-e";
1221 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1222 }
1223
1224 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1225 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1226
1227 for (j = 1; j < (unsigned int)argc; j++, i++)
1228 rec_argv[i] = argv[j];
1229
1230 set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1231 set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1232 set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1233
1234 set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1235 set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1236 set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1237 set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1238 set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1239 set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1240 set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1241 set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1242 set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1243 set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1244 set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1245 set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1246 set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1247 set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1248 set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1249
1250 record_usage = kvm_stat_record_usage;
1251 return cmd_record(i, rec_argv);
1252}
1253
1254static int
1255kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1256{
1257 const struct option kvm_events_report_options[] = {
1258 OPT_STRING(0, "event", &kvm->report_event, "report event",
1259 "event for reporting: vmexit, "
1260 "mmio (x86 only), ioport (x86 only)"),
1261 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1262 "vcpu id to report"),
1263 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1264 "key for sorting: sample(sort by samples number)"
1265 " time (sort by avg time)"),
1266 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1267 "analyze events only for given process id(s)"),
1268 OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1269 OPT_END()
1270 };
1271
1272 const char * const kvm_events_report_usage[] = {
1273 "perf kvm stat report [<options>]",
1274 NULL
1275 };
1276
1277 if (argc) {
1278 argc = parse_options(argc, argv,
1279 kvm_events_report_options,
1280 kvm_events_report_usage, 0);
1281 if (argc)
1282 usage_with_options(kvm_events_report_usage,
1283 kvm_events_report_options);
1284 }
1285
1286 if (!kvm->opts.target.pid)
1287 kvm->opts.target.system_wide = true;
1288
1289 return kvm_events_report_vcpu(kvm);
1290}
1291
1292#ifdef HAVE_TIMERFD_SUPPORT
1293static struct evlist *kvm_live_event_list(void)
1294{
1295 struct evlist *evlist;
1296 char *tp, *name, *sys;
1297 int err = -1;
1298 const char * const *events_tp;
1299
1300 evlist = evlist__new();
1301 if (evlist == NULL)
1302 return NULL;
1303
1304 for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1305
1306 tp = strdup(*events_tp);
1307 if (tp == NULL)
1308 goto out;
1309
1310 /* split tracepoint into subsystem and name */
1311 sys = tp;
1312 name = strchr(tp, ':');
1313 if (name == NULL) {
1314 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1315 *events_tp);
1316 free(tp);
1317 goto out;
1318 }
1319 *name = '\0';
1320 name++;
1321
1322 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1323 pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1324 free(tp);
1325 goto out;
1326 }
1327
1328 free(tp);
1329 }
1330
1331 err = 0;
1332
1333out:
1334 if (err) {
1335 evlist__delete(evlist);
1336 evlist = NULL;
1337 }
1338
1339 return evlist;
1340}
1341
1342static int kvm_events_live(struct perf_kvm_stat *kvm,
1343 int argc, const char **argv)
1344{
1345 char errbuf[BUFSIZ];
1346 int err;
1347
1348 const struct option live_options[] = {
1349 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1350 "record events on existing process id"),
1351 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1352 "number of mmap data pages",
1353 perf_evlist__parse_mmap_pages),
1354 OPT_INCR('v', "verbose", &verbose,
1355 "be more verbose (show counter open errors, etc)"),
1356 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1357 "system-wide collection from all CPUs"),
1358 OPT_UINTEGER('d', "display", &kvm->display_time,
1359 "time in seconds between display updates"),
1360 OPT_STRING(0, "event", &kvm->report_event, "report event",
1361 "event for reporting: "
1362 "vmexit, mmio (x86 only), ioport (x86 only)"),
1363 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1364 "vcpu id to report"),
1365 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1366 "key for sorting: sample(sort by samples number)"
1367 " time (sort by avg time)"),
1368 OPT_U64(0, "duration", &kvm->duration,
1369 "show events other than"
1370 " HLT (x86 only) or Wait state (s390 only)"
1371 " that take longer than duration usecs"),
1372 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1373 "per thread proc mmap processing timeout in ms"),
1374 OPT_END()
1375 };
1376 const char * const live_usage[] = {
1377 "perf kvm stat live [<options>]",
1378 NULL
1379 };
1380 struct perf_data data = {
1381 .mode = PERF_DATA_MODE_WRITE,
1382 };
1383
1384
1385 /* event handling */
1386 kvm->tool.sample = process_sample_event;
1387 kvm->tool.comm = perf_event__process_comm;
1388 kvm->tool.exit = perf_event__process_exit;
1389 kvm->tool.fork = perf_event__process_fork;
1390 kvm->tool.lost = process_lost_event;
1391 kvm->tool.namespaces = perf_event__process_namespaces;
1392 kvm->tool.ordered_events = true;
1393 perf_tool__fill_defaults(&kvm->tool);
1394
1395 /* set defaults */
1396 kvm->display_time = 1;
1397 kvm->opts.user_interval = 1;
1398 kvm->opts.mmap_pages = 512;
1399 kvm->opts.target.uses_mmap = false;
1400 kvm->opts.target.uid_str = NULL;
1401 kvm->opts.target.uid = UINT_MAX;
1402
1403 symbol__init(NULL);
1404 disable_buildid_cache();
1405
1406 use_browser = 0;
1407
1408 if (argc) {
1409 argc = parse_options(argc, argv, live_options,
1410 live_usage, 0);
1411 if (argc)
1412 usage_with_options(live_usage, live_options);
1413 }
1414
1415 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1416
1417 /*
1418 * target related setups
1419 */
1420 err = target__validate(&kvm->opts.target);
1421 if (err) {
1422 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1423 ui__warning("%s", errbuf);
1424 }
1425
1426 if (target__none(&kvm->opts.target))
1427 kvm->opts.target.system_wide = true;
1428
1429
1430 /*
1431 * generate the event list
1432 */
1433 err = setup_kvm_events_tp(kvm);
1434 if (err < 0) {
1435 pr_err("Unable to setup the kvm tracepoints\n");
1436 return err;
1437 }
1438
1439 kvm->evlist = kvm_live_event_list();
1440 if (kvm->evlist == NULL) {
1441 err = -1;
1442 goto out;
1443 }
1444
1445 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1446 usage_with_options(live_usage, live_options);
1447
1448 /*
1449 * perf session
1450 */
1451 kvm->session = perf_session__new(&data, false, &kvm->tool);
1452 if (IS_ERR(kvm->session)) {
1453 err = PTR_ERR(kvm->session);
1454 goto out;
1455 }
1456 kvm->session->evlist = kvm->evlist;
1457 perf_session__set_id_hdr_size(kvm->session);
1458 ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1459 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1460 kvm->evlist->core.threads, false, 1);
1461 err = kvm_live_open_events(kvm);
1462 if (err)
1463 goto out;
1464
1465 err = kvm_events_live_report(kvm);
1466
1467out:
1468 perf_session__delete(kvm->session);
1469 kvm->session = NULL;
1470 evlist__delete(kvm->evlist);
1471
1472 return err;
1473}
1474#endif
1475
1476static void print_kvm_stat_usage(void)
1477{
1478 printf("Usage: perf kvm stat <command>\n\n");
1479
1480 printf("# Available commands:\n");
1481 printf("\trecord: record kvm events\n");
1482 printf("\treport: report statistical data of kvm events\n");
1483 printf("\tlive: live reporting of statistical data of kvm events\n");
1484
1485 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1486}
1487
1488static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1489{
1490 struct perf_kvm_stat kvm = {
1491 .file_name = file_name,
1492
1493 .trace_vcpu = -1,
1494 .report_event = "vmexit",
1495 .sort_key = "sample",
1496
1497 };
1498
1499 if (argc == 1) {
1500 print_kvm_stat_usage();
1501 goto perf_stat;
1502 }
1503
1504 if (!strncmp(argv[1], "rec", 3))
1505 return kvm_events_record(&kvm, argc - 1, argv + 1);
1506
1507 if (!strncmp(argv[1], "rep", 3))
1508 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1509
1510#ifdef HAVE_TIMERFD_SUPPORT
1511 if (!strncmp(argv[1], "live", 4))
1512 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1513#endif
1514
1515perf_stat:
1516 return cmd_stat(argc, argv);
1517}
1518#endif /* HAVE_KVM_STAT_SUPPORT */
1519
1520int __weak kvm_add_default_arch_event(int *argc __maybe_unused,
1521 const char **argv __maybe_unused)
1522{
1523 return 0;
1524}
1525
1526static int __cmd_record(const char *file_name, int argc, const char **argv)
1527{
1528 int rec_argc, i = 0, j, ret;
1529 const char **rec_argv;
1530
1531 ret = kvm_add_default_arch_event(&argc, argv);
1532 if (ret)
1533 return -EINVAL;
1534
1535 rec_argc = argc + 2;
1536 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1537 rec_argv[i++] = strdup("record");
1538 rec_argv[i++] = strdup("-o");
1539 rec_argv[i++] = strdup(file_name);
1540 for (j = 1; j < argc; j++, i++)
1541 rec_argv[i] = argv[j];
1542
1543 BUG_ON(i != rec_argc);
1544
1545 return cmd_record(i, rec_argv);
1546}
1547
1548static int __cmd_report(const char *file_name, int argc, const char **argv)
1549{
1550 int rec_argc, i = 0, j;
1551 const char **rec_argv;
1552
1553 rec_argc = argc + 2;
1554 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1555 rec_argv[i++] = strdup("report");
1556 rec_argv[i++] = strdup("-i");
1557 rec_argv[i++] = strdup(file_name);
1558 for (j = 1; j < argc; j++, i++)
1559 rec_argv[i] = argv[j];
1560
1561 BUG_ON(i != rec_argc);
1562
1563 return cmd_report(i, rec_argv);
1564}
1565
1566static int
1567__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1568{
1569 int rec_argc, i = 0, j;
1570 const char **rec_argv;
1571
1572 rec_argc = argc + 2;
1573 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1574 rec_argv[i++] = strdup("buildid-list");
1575 rec_argv[i++] = strdup("-i");
1576 rec_argv[i++] = strdup(file_name);
1577 for (j = 1; j < argc; j++, i++)
1578 rec_argv[i] = argv[j];
1579
1580 BUG_ON(i != rec_argc);
1581
1582 return cmd_buildid_list(i, rec_argv);
1583}
1584
1585int cmd_kvm(int argc, const char **argv)
1586{
1587 const char *file_name = NULL;
1588 const struct option kvm_options[] = {
1589 OPT_STRING('i', "input", &file_name, "file",
1590 "Input file name"),
1591 OPT_STRING('o', "output", &file_name, "file",
1592 "Output file name"),
1593 OPT_BOOLEAN(0, "guest", &perf_guest,
1594 "Collect guest os data"),
1595 OPT_BOOLEAN(0, "host", &perf_host,
1596 "Collect host os data"),
1597 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1598 "guest mount directory under which every guest os"
1599 " instance has a subdir"),
1600 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1601 "file", "file saving guest os vmlinux"),
1602 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1603 "file", "file saving guest os /proc/kallsyms"),
1604 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1605 "file", "file saving guest os /proc/modules"),
1606 OPT_INCR('v', "verbose", &verbose,
1607 "be more verbose (show counter open errors, etc)"),
1608 OPT_END()
1609 };
1610
1611 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1612 "buildid-list", "stat", NULL };
1613 const char *kvm_usage[] = { NULL, NULL };
1614
1615 perf_host = 0;
1616 perf_guest = 1;
1617
1618 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1619 PARSE_OPT_STOP_AT_NON_OPTION);
1620 if (!argc)
1621 usage_with_options(kvm_usage, kvm_options);
1622
1623 if (!perf_host)
1624 perf_guest = 1;
1625
1626 if (!file_name) {
1627 file_name = get_filename_for_perf_kvm();
1628
1629 if (!file_name) {
1630 pr_err("Failed to allocate memory for filename\n");
1631 return -ENOMEM;
1632 }
1633 }
1634
1635 if (!strncmp(argv[0], "rec", 3))
1636 return __cmd_record(file_name, argc, argv);
1637 else if (!strncmp(argv[0], "rep", 3))
1638 return __cmd_report(file_name, argc, argv);
1639 else if (!strncmp(argv[0], "diff", 4))
1640 return cmd_diff(argc, argv);
1641 else if (!strncmp(argv[0], "top", 3))
1642 return cmd_top(argc, argv);
1643 else if (!strncmp(argv[0], "buildid-list", 12))
1644 return __cmd_buildid_list(file_name, argc, argv);
1645#ifdef HAVE_KVM_STAT_SUPPORT
1646 else if (!strncmp(argv[0], "stat", 4))
1647 return kvm_cmd_stat(file_name, argc, argv);
1648#endif
1649 else
1650 usage_with_options(kvm_usage, kvm_options);
1651
1652 return 0;
1653}