Loading...
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}
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 "util/util.h"
27#include "ui/browsers/hists.h"
28#include "ui/progress.h"
29#include "ui/ui.h"
30#include "util/string2.h"
31
32#include <sys/prctl.h>
33#ifdef HAVE_TIMERFD_SUPPORT
34#include <sys/timerfd.h>
35#endif
36#include <sys/time.h>
37#include <sys/types.h>
38#include <sys/stat.h>
39#include <fcntl.h>
40
41#include <linux/err.h>
42#include <linux/kernel.h>
43#include <linux/string.h>
44#include <linux/time64.h>
45#include <linux/zalloc.h>
46#include <errno.h>
47#include <inttypes.h>
48#include <poll.h>
49#include <termios.h>
50#include <semaphore.h>
51#include <signal.h>
52#include <math.h>
53#include <perf/mmap.h>
54
55#if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
56#define GET_EVENT_KEY(func, field) \
57static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
58{ \
59 if (vcpu == -1) \
60 return event->total.field; \
61 \
62 if (vcpu >= event->max_vcpu) \
63 return 0; \
64 \
65 return event->vcpu[vcpu].field; \
66}
67
68#define COMPARE_EVENT_KEY(func, field) \
69GET_EVENT_KEY(func, field) \
70static int64_t cmp_event_ ## func(struct kvm_event *one, \
71 struct kvm_event *two, int vcpu) \
72{ \
73 return get_event_ ##func(one, vcpu) - \
74 get_event_ ##func(two, vcpu); \
75}
76
77COMPARE_EVENT_KEY(time, time);
78COMPARE_EVENT_KEY(max, stats.max);
79COMPARE_EVENT_KEY(min, stats.min);
80COMPARE_EVENT_KEY(count, stats.n);
81COMPARE_EVENT_KEY(mean, stats.mean);
82
83struct kvm_hists {
84 struct hists hists;
85 struct perf_hpp_list list;
86};
87
88struct kvm_dimension {
89 const char *name;
90 const char *header;
91 int width;
92 int64_t (*cmp)(struct perf_hpp_fmt *fmt, struct hist_entry *left,
93 struct hist_entry *right);
94 int (*entry)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
95 struct hist_entry *he);
96};
97
98struct kvm_fmt {
99 struct perf_hpp_fmt fmt;
100 struct kvm_dimension *dim;
101};
102
103static struct kvm_hists kvm_hists;
104
105static int64_t ev_name_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
106 struct hist_entry *left,
107 struct hist_entry *right)
108{
109 /* Return opposite number for sorting in alphabetical order */
110 return -strcmp(left->kvm_info->name, right->kvm_info->name);
111}
112
113static int fmt_width(struct perf_hpp_fmt *fmt,
114 struct perf_hpp *hpp __maybe_unused,
115 struct hists *hists __maybe_unused);
116
117static int ev_name_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
118 struct hist_entry *he)
119{
120 int width = fmt_width(fmt, hpp, he->hists);
121
122 return scnprintf(hpp->buf, hpp->size, "%*s", width, he->kvm_info->name);
123}
124
125static struct kvm_dimension dim_event = {
126 .header = "Event name",
127 .name = "ev_name",
128 .cmp = ev_name_cmp,
129 .entry = ev_name_entry,
130 .width = 40,
131};
132
133#define EV_METRIC_CMP(metric) \
134static int64_t ev_cmp_##metric(struct perf_hpp_fmt *fmt __maybe_unused, \
135 struct hist_entry *left, \
136 struct hist_entry *right) \
137{ \
138 struct kvm_event *event_left; \
139 struct kvm_event *event_right; \
140 struct perf_kvm_stat *perf_kvm; \
141 \
142 event_left = container_of(left, struct kvm_event, he); \
143 event_right = container_of(right, struct kvm_event, he); \
144 \
145 perf_kvm = event_left->perf_kvm; \
146 return cmp_event_##metric(event_left, event_right, \
147 perf_kvm->trace_vcpu); \
148}
149
150EV_METRIC_CMP(time)
151EV_METRIC_CMP(count)
152EV_METRIC_CMP(max)
153EV_METRIC_CMP(min)
154EV_METRIC_CMP(mean)
155
156#define EV_METRIC_ENTRY(metric) \
157static int ev_entry_##metric(struct perf_hpp_fmt *fmt, \
158 struct perf_hpp *hpp, \
159 struct hist_entry *he) \
160{ \
161 struct kvm_event *event; \
162 int width = fmt_width(fmt, hpp, he->hists); \
163 struct perf_kvm_stat *perf_kvm; \
164 \
165 event = container_of(he, struct kvm_event, he); \
166 perf_kvm = event->perf_kvm; \
167 return scnprintf(hpp->buf, hpp->size, "%*lu", width, \
168 get_event_##metric(event, perf_kvm->trace_vcpu)); \
169}
170
171EV_METRIC_ENTRY(time)
172EV_METRIC_ENTRY(count)
173EV_METRIC_ENTRY(max)
174EV_METRIC_ENTRY(min)
175
176static struct kvm_dimension dim_time = {
177 .header = "Time (ns)",
178 .name = "time",
179 .cmp = ev_cmp_time,
180 .entry = ev_entry_time,
181 .width = 12,
182};
183
184static struct kvm_dimension dim_count = {
185 .header = "Samples",
186 .name = "sample",
187 .cmp = ev_cmp_count,
188 .entry = ev_entry_count,
189 .width = 12,
190};
191
192static struct kvm_dimension dim_max_time = {
193 .header = "Max Time (ns)",
194 .name = "max_t",
195 .cmp = ev_cmp_max,
196 .entry = ev_entry_max,
197 .width = 14,
198};
199
200static struct kvm_dimension dim_min_time = {
201 .header = "Min Time (ns)",
202 .name = "min_t",
203 .cmp = ev_cmp_min,
204 .entry = ev_entry_min,
205 .width = 14,
206};
207
208static int ev_entry_mean(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
209 struct hist_entry *he)
210{
211 struct kvm_event *event;
212 int width = fmt_width(fmt, hpp, he->hists);
213 struct perf_kvm_stat *perf_kvm;
214
215 event = container_of(he, struct kvm_event, he);
216 perf_kvm = event->perf_kvm;
217 return scnprintf(hpp->buf, hpp->size, "%*lu", width,
218 get_event_mean(event, perf_kvm->trace_vcpu));
219}
220
221static struct kvm_dimension dim_mean_time = {
222 .header = "Mean Time (ns)",
223 .name = "mean_t",
224 .cmp = ev_cmp_mean,
225 .entry = ev_entry_mean,
226 .width = 14,
227};
228
229#define PERC_STR(__s, __v) \
230({ \
231 scnprintf(__s, sizeof(__s), "%.2F%%", __v); \
232 __s; \
233})
234
235static double percent(u64 st, u64 tot)
236{
237 return tot ? 100. * (double) st / (double) tot : 0;
238}
239
240#define EV_METRIC_PERCENT(metric) \
241static int ev_percent_##metric(struct hist_entry *he) \
242{ \
243 struct kvm_event *event; \
244 struct perf_kvm_stat *perf_kvm; \
245 \
246 event = container_of(he, struct kvm_event, he); \
247 perf_kvm = event->perf_kvm; \
248 \
249 return percent(get_event_##metric(event, perf_kvm->trace_vcpu), \
250 perf_kvm->total_##metric); \
251}
252
253EV_METRIC_PERCENT(time)
254EV_METRIC_PERCENT(count)
255
256static int ev_entry_time_precent(struct perf_hpp_fmt *fmt,
257 struct perf_hpp *hpp,
258 struct hist_entry *he)
259{
260 int width = fmt_width(fmt, hpp, he->hists);
261 double per;
262 char buf[10];
263
264 per = ev_percent_time(he);
265 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
266}
267
268static int64_t
269ev_cmp_time_precent(struct perf_hpp_fmt *fmt __maybe_unused,
270 struct hist_entry *left, struct hist_entry *right)
271{
272 double per_left;
273 double per_right;
274
275 per_left = ev_percent_time(left);
276 per_right = ev_percent_time(right);
277
278 return per_left - per_right;
279}
280
281static struct kvm_dimension dim_time_percent = {
282 .header = "Time%",
283 .name = "percent_time",
284 .cmp = ev_cmp_time_precent,
285 .entry = ev_entry_time_precent,
286 .width = 12,
287};
288
289static int ev_entry_count_precent(struct perf_hpp_fmt *fmt,
290 struct perf_hpp *hpp,
291 struct hist_entry *he)
292{
293 int width = fmt_width(fmt, hpp, he->hists);
294 double per;
295 char buf[10];
296
297 per = ev_percent_count(he);
298 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
299}
300
301static int64_t
302ev_cmp_count_precent(struct perf_hpp_fmt *fmt __maybe_unused,
303 struct hist_entry *left, struct hist_entry *right)
304{
305 double per_left;
306 double per_right;
307
308 per_left = ev_percent_count(left);
309 per_right = ev_percent_count(right);
310
311 return per_left - per_right;
312}
313
314static struct kvm_dimension dim_count_percent = {
315 .header = "Sample%",
316 .name = "percent_sample",
317 .cmp = ev_cmp_count_precent,
318 .entry = ev_entry_count_precent,
319 .width = 12,
320};
321
322static struct kvm_dimension *dimensions[] = {
323 &dim_event,
324 &dim_time,
325 &dim_time_percent,
326 &dim_count,
327 &dim_count_percent,
328 &dim_max_time,
329 &dim_min_time,
330 &dim_mean_time,
331 NULL,
332};
333
334static int fmt_width(struct perf_hpp_fmt *fmt,
335 struct perf_hpp *hpp __maybe_unused,
336 struct hists *hists __maybe_unused)
337{
338 struct kvm_fmt *kvm_fmt;
339
340 kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
341 return kvm_fmt->dim->width;
342}
343
344static int fmt_header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
345 struct hists *hists, int line __maybe_unused,
346 int *span __maybe_unused)
347{
348 struct kvm_fmt *kvm_fmt;
349 struct kvm_dimension *dim;
350 int width = fmt_width(fmt, hpp, hists);
351
352 kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
353 dim = kvm_fmt->dim;
354
355 return scnprintf(hpp->buf, hpp->size, "%*s", width, dim->header);
356}
357
358static bool fmt_equal(struct perf_hpp_fmt *a, struct perf_hpp_fmt *b)
359{
360 struct kvm_fmt *kvm_fmt_a = container_of(a, struct kvm_fmt, fmt);
361 struct kvm_fmt *kvm_fmt_b = container_of(b, struct kvm_fmt, fmt);
362
363 return kvm_fmt_a->dim == kvm_fmt_b->dim;
364}
365
366static void fmt_free(struct perf_hpp_fmt *fmt)
367{
368 struct kvm_fmt *kvm_fmt;
369
370 kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
371 free(kvm_fmt);
372}
373
374static struct kvm_dimension *get_dimension(const char *name)
375{
376 unsigned int i;
377
378 for (i = 0; dimensions[i] != NULL; i++) {
379 if (!strcmp(dimensions[i]->name, name))
380 return dimensions[i];
381 }
382
383 return NULL;
384}
385
386static struct kvm_fmt *get_format(const char *name)
387{
388 struct kvm_dimension *dim = get_dimension(name);
389 struct kvm_fmt *kvm_fmt;
390 struct perf_hpp_fmt *fmt;
391
392 if (!dim)
393 return NULL;
394
395 kvm_fmt = zalloc(sizeof(*kvm_fmt));
396 if (!kvm_fmt)
397 return NULL;
398
399 kvm_fmt->dim = dim;
400
401 fmt = &kvm_fmt->fmt;
402 INIT_LIST_HEAD(&fmt->list);
403 INIT_LIST_HEAD(&fmt->sort_list);
404 fmt->cmp = dim->cmp;
405 fmt->sort = dim->cmp;
406 fmt->color = NULL;
407 fmt->entry = dim->entry;
408 fmt->header = fmt_header;
409 fmt->width = fmt_width;
410 fmt->collapse = dim->cmp;
411 fmt->equal = fmt_equal;
412 fmt->free = fmt_free;
413
414 return kvm_fmt;
415}
416
417static int kvm_hists__init_output(struct perf_hpp_list *hpp_list, char *name)
418{
419 struct kvm_fmt *kvm_fmt = get_format(name);
420
421 if (!kvm_fmt) {
422 pr_warning("Fail to find format for output field %s.\n", name);
423 return -EINVAL;
424 }
425
426 perf_hpp_list__column_register(hpp_list, &kvm_fmt->fmt);
427 return 0;
428}
429
430static int kvm_hists__init_sort(struct perf_hpp_list *hpp_list, char *name)
431{
432 struct kvm_fmt *kvm_fmt = get_format(name);
433
434 if (!kvm_fmt) {
435 pr_warning("Fail to find format for sorting %s.\n", name);
436 return -EINVAL;
437 }
438
439 perf_hpp_list__register_sort_field(hpp_list, &kvm_fmt->fmt);
440 return 0;
441}
442
443static int kvm_hpp_list__init(char *list,
444 struct perf_hpp_list *hpp_list,
445 int (*fn)(struct perf_hpp_list *hpp_list,
446 char *name))
447{
448 char *tmp, *tok;
449 int ret;
450
451 if (!list || !fn)
452 return 0;
453
454 for (tok = strtok_r(list, ", ", &tmp); tok;
455 tok = strtok_r(NULL, ", ", &tmp)) {
456 ret = fn(hpp_list, tok);
457 if (!ret)
458 continue;
459
460 /* Handle errors */
461 if (ret == -EINVAL)
462 pr_err("Invalid field key: '%s'", tok);
463 else if (ret == -ESRCH)
464 pr_err("Unknown field key: '%s'", tok);
465 else
466 pr_err("Fail to initialize for field key: '%s'", tok);
467
468 break;
469 }
470
471 return ret;
472}
473
474static int kvm_hpp_list__parse(struct perf_hpp_list *hpp_list,
475 const char *output_, const char *sort_)
476{
477 char *output = output_ ? strdup(output_) : NULL;
478 char *sort = sort_ ? strdup(sort_) : NULL;
479 int ret;
480
481 ret = kvm_hpp_list__init(output, hpp_list, kvm_hists__init_output);
482 if (ret)
483 goto out;
484
485 ret = kvm_hpp_list__init(sort, hpp_list, kvm_hists__init_sort);
486 if (ret)
487 goto out;
488
489 /* Copy sort keys to output fields */
490 perf_hpp__setup_output_field(hpp_list);
491
492 /* and then copy output fields to sort keys */
493 perf_hpp__append_sort_keys(hpp_list);
494out:
495 free(output);
496 free(sort);
497 return ret;
498}
499
500static int kvm_hists__init(void)
501{
502 kvm_hists.list.nr_header_lines = 1;
503 __hists__init(&kvm_hists.hists, &kvm_hists.list);
504 perf_hpp_list__init(&kvm_hists.list);
505 return kvm_hpp_list__parse(&kvm_hists.list, NULL, "ev_name");
506}
507
508static int kvm_hists__reinit(const char *output, const char *sort)
509{
510 perf_hpp__reset_output_field(&kvm_hists.list);
511 return kvm_hpp_list__parse(&kvm_hists.list, output, sort);
512}
513static void print_result(struct perf_kvm_stat *kvm);
514
515#ifdef HAVE_SLANG_SUPPORT
516static void kvm_browser__update_nr_entries(struct hist_browser *hb)
517{
518 struct rb_node *nd = rb_first_cached(&hb->hists->entries);
519 u64 nr_entries = 0;
520
521 for (; nd; nd = rb_next(nd)) {
522 struct hist_entry *he = rb_entry(nd, struct hist_entry,
523 rb_node);
524
525 if (!he->filtered)
526 nr_entries++;
527 }
528
529 hb->nr_non_filtered_entries = nr_entries;
530}
531
532static int kvm_browser__title(struct hist_browser *browser,
533 char *buf, size_t size)
534{
535 scnprintf(buf, size, "KVM event statistics (%lu entries)",
536 browser->nr_non_filtered_entries);
537 return 0;
538}
539
540static struct hist_browser*
541perf_kvm_browser__new(struct hists *hists)
542{
543 struct hist_browser *browser = hist_browser__new(hists);
544
545 if (browser)
546 browser->title = kvm_browser__title;
547
548 return browser;
549}
550
551static int kvm__hists_browse(struct hists *hists)
552{
553 struct hist_browser *browser;
554 int key = -1;
555
556 browser = perf_kvm_browser__new(hists);
557 if (browser == NULL)
558 return -1;
559
560 /* reset abort key so that it can get Ctrl-C as a key */
561 SLang_reset_tty();
562 SLang_init_tty(0, 0, 0);
563
564 kvm_browser__update_nr_entries(browser);
565
566 while (1) {
567 key = hist_browser__run(browser, "? - help", true, 0);
568
569 switch (key) {
570 case 'q':
571 goto out;
572 default:
573 break;
574 }
575 }
576
577out:
578 hist_browser__delete(browser);
579 return 0;
580}
581
582static void kvm_display(struct perf_kvm_stat *kvm)
583{
584 if (!use_browser)
585 print_result(kvm);
586 else
587 kvm__hists_browse(&kvm_hists.hists);
588}
589
590#else
591
592static void kvm_display(struct perf_kvm_stat *kvm)
593{
594 use_browser = 0;
595 print_result(kvm);
596}
597
598#endif /* HAVE_SLANG_SUPPORT */
599
600#endif // defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
601
602static const char *get_filename_for_perf_kvm(void)
603{
604 const char *filename;
605
606 if (perf_host && !perf_guest)
607 filename = strdup("perf.data.host");
608 else if (!perf_host && perf_guest)
609 filename = strdup("perf.data.guest");
610 else
611 filename = strdup("perf.data.kvm");
612
613 return filename;
614}
615
616#if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
617
618void exit_event_get_key(struct evsel *evsel,
619 struct perf_sample *sample,
620 struct event_key *key)
621{
622 key->info = 0;
623 key->key = evsel__intval(evsel, sample, kvm_exit_reason);
624}
625
626bool kvm_exit_event(struct evsel *evsel)
627{
628 return evsel__name_is(evsel, kvm_exit_trace);
629}
630
631bool exit_event_begin(struct evsel *evsel,
632 struct perf_sample *sample, struct event_key *key)
633{
634 if (kvm_exit_event(evsel)) {
635 exit_event_get_key(evsel, sample, key);
636 return true;
637 }
638
639 return false;
640}
641
642bool kvm_entry_event(struct evsel *evsel)
643{
644 return evsel__name_is(evsel, kvm_entry_trace);
645}
646
647bool exit_event_end(struct evsel *evsel,
648 struct perf_sample *sample __maybe_unused,
649 struct event_key *key __maybe_unused)
650{
651 return kvm_entry_event(evsel);
652}
653
654static const char *get_exit_reason(struct perf_kvm_stat *kvm,
655 struct exit_reasons_table *tbl,
656 u64 exit_code)
657{
658 while (tbl->reason != NULL) {
659 if (tbl->exit_code == exit_code)
660 return tbl->reason;
661 tbl++;
662 }
663
664 pr_err("unknown kvm exit code:%lld on %s\n",
665 (unsigned long long)exit_code, kvm->exit_reasons_isa);
666 return "UNKNOWN";
667}
668
669void exit_event_decode_key(struct perf_kvm_stat *kvm,
670 struct event_key *key,
671 char *decode)
672{
673 const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
674 key->key);
675
676 scnprintf(decode, KVM_EVENT_NAME_LEN, "%s", exit_reason);
677}
678
679static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
680{
681 struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
682
683 for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
684 if (!strcmp(events_ops->name, kvm->report_event)) {
685 kvm->events_ops = events_ops->ops;
686 return true;
687 }
688 }
689
690 return false;
691}
692
693struct vcpu_event_record {
694 int vcpu_id;
695 u64 start_time;
696 struct kvm_event *last_event;
697};
698
699#ifdef HAVE_TIMERFD_SUPPORT
700static void clear_events_cache_stats(void)
701{
702 struct rb_root_cached *root;
703 struct rb_node *nd;
704 struct kvm_event *event;
705 int i;
706
707 if (hists__has(&kvm_hists.hists, need_collapse))
708 root = &kvm_hists.hists.entries_collapsed;
709 else
710 root = kvm_hists.hists.entries_in;
711
712 for (nd = rb_first_cached(root); nd; nd = rb_next(nd)) {
713 struct hist_entry *he;
714
715 he = rb_entry(nd, struct hist_entry, rb_node_in);
716 event = container_of(he, struct kvm_event, he);
717
718 /* reset stats for event */
719 event->total.time = 0;
720 init_stats(&event->total.stats);
721
722 for (i = 0; i < event->max_vcpu; ++i) {
723 event->vcpu[i].time = 0;
724 init_stats(&event->vcpu[i].stats);
725 }
726 }
727}
728#endif
729
730static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
731{
732 int old_max_vcpu = event->max_vcpu;
733 void *prev;
734
735 if (vcpu_id < event->max_vcpu)
736 return true;
737
738 while (event->max_vcpu <= vcpu_id)
739 event->max_vcpu += DEFAULT_VCPU_NUM;
740
741 prev = event->vcpu;
742 event->vcpu = realloc(event->vcpu,
743 event->max_vcpu * sizeof(*event->vcpu));
744 if (!event->vcpu) {
745 free(prev);
746 pr_err("Not enough memory\n");
747 return false;
748 }
749
750 memset(event->vcpu + old_max_vcpu, 0,
751 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
752 return true;
753}
754
755static void *kvm_he_zalloc(size_t size)
756{
757 struct kvm_event *kvm_ev;
758
759 kvm_ev = zalloc(size + sizeof(*kvm_ev));
760 if (!kvm_ev)
761 return NULL;
762
763 init_stats(&kvm_ev->total.stats);
764 hists__inc_nr_samples(&kvm_hists.hists, 0);
765 return &kvm_ev->he;
766}
767
768static void kvm_he_free(void *he)
769{
770 struct kvm_event *kvm_ev;
771
772 kvm_ev = container_of(he, struct kvm_event, he);
773 free(kvm_ev);
774}
775
776static struct hist_entry_ops kvm_ev_entry_ops = {
777 .new = kvm_he_zalloc,
778 .free = kvm_he_free,
779};
780
781static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
782 struct event_key *key,
783 struct perf_sample *sample)
784{
785 struct kvm_event *event;
786 struct hist_entry *he;
787 struct kvm_info *ki;
788
789 BUG_ON(key->key == INVALID_KEY);
790
791 ki = kvm_info__new();
792 if (!ki) {
793 pr_err("Failed to allocate kvm info\n");
794 return NULL;
795 }
796
797 kvm->events_ops->decode_key(kvm, key, ki->name);
798 he = hists__add_entry_ops(&kvm_hists.hists, &kvm_ev_entry_ops,
799 &kvm->al, NULL, NULL, NULL, ki, sample, true);
800 if (he == NULL) {
801 pr_err("Failed to allocate hist entry\n");
802 free(ki);
803 return NULL;
804 }
805
806 event = container_of(he, struct kvm_event, he);
807 if (!event->perf_kvm) {
808 event->perf_kvm = kvm;
809 event->key = *key;
810 }
811
812 return event;
813}
814
815static bool handle_begin_event(struct perf_kvm_stat *kvm,
816 struct vcpu_event_record *vcpu_record,
817 struct event_key *key,
818 struct perf_sample *sample)
819{
820 struct kvm_event *event = NULL;
821
822 if (key->key != INVALID_KEY)
823 event = find_create_kvm_event(kvm, key, sample);
824
825 vcpu_record->last_event = event;
826 vcpu_record->start_time = sample->time;
827 return true;
828}
829
830static void
831kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
832{
833 kvm_stats->time += time_diff;
834 update_stats(&kvm_stats->stats, time_diff);
835}
836
837static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
838{
839 struct kvm_event_stats *kvm_stats = &event->total;
840
841 if (vcpu_id != -1)
842 kvm_stats = &event->vcpu[vcpu_id];
843
844 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
845 avg_stats(&kvm_stats->stats));
846}
847
848static bool update_kvm_event(struct perf_kvm_stat *kvm,
849 struct kvm_event *event, int vcpu_id,
850 u64 time_diff)
851{
852 /* Update overall statistics */
853 kvm->total_count++;
854 kvm->total_time += time_diff;
855
856 if (vcpu_id == -1) {
857 kvm_update_event_stats(&event->total, time_diff);
858 return true;
859 }
860
861 if (!kvm_event_expand(event, vcpu_id))
862 return false;
863
864 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
865 return true;
866}
867
868static bool is_child_event(struct perf_kvm_stat *kvm,
869 struct evsel *evsel,
870 struct perf_sample *sample,
871 struct event_key *key)
872{
873 struct child_event_ops *child_ops;
874
875 child_ops = kvm->events_ops->child_ops;
876
877 if (!child_ops)
878 return false;
879
880 for (; child_ops->name; child_ops++) {
881 if (evsel__name_is(evsel, child_ops->name)) {
882 child_ops->get_key(evsel, sample, key);
883 return true;
884 }
885 }
886
887 return false;
888}
889
890static bool handle_child_event(struct perf_kvm_stat *kvm,
891 struct vcpu_event_record *vcpu_record,
892 struct event_key *key,
893 struct perf_sample *sample)
894{
895 struct kvm_event *event = NULL;
896
897 if (key->key != INVALID_KEY)
898 event = find_create_kvm_event(kvm, key, sample);
899
900 vcpu_record->last_event = event;
901
902 return true;
903}
904
905static bool skip_event(const char *event)
906{
907 const char * const *skip_events;
908
909 for (skip_events = kvm_skip_events; *skip_events; skip_events++)
910 if (!strcmp(event, *skip_events))
911 return true;
912
913 return false;
914}
915
916static bool handle_end_event(struct perf_kvm_stat *kvm,
917 struct vcpu_event_record *vcpu_record,
918 struct event_key *key,
919 struct perf_sample *sample)
920{
921 struct kvm_event *event;
922 u64 time_begin, time_diff;
923 int vcpu;
924
925 if (kvm->trace_vcpu == -1)
926 vcpu = -1;
927 else
928 vcpu = vcpu_record->vcpu_id;
929
930 event = vcpu_record->last_event;
931 time_begin = vcpu_record->start_time;
932
933 /* The begin event is not caught. */
934 if (!time_begin)
935 return true;
936
937 /*
938 * In some case, the 'begin event' only records the start timestamp,
939 * the actual event is recognized in the 'end event' (e.g. mmio-event).
940 */
941
942 /* Both begin and end events did not get the key. */
943 if (!event && key->key == INVALID_KEY)
944 return true;
945
946 if (!event)
947 event = find_create_kvm_event(kvm, key, sample);
948
949 if (!event)
950 return false;
951
952 vcpu_record->last_event = NULL;
953 vcpu_record->start_time = 0;
954
955 /* seems to happen once in a while during live mode */
956 if (sample->time < time_begin) {
957 pr_debug("End time before begin time; skipping event.\n");
958 return true;
959 }
960
961 time_diff = sample->time - time_begin;
962
963 if (kvm->duration && time_diff > kvm->duration) {
964 char decode[KVM_EVENT_NAME_LEN];
965
966 kvm->events_ops->decode_key(kvm, &event->key, decode);
967 if (!skip_event(decode)) {
968 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
969 sample->time, sample->pid, vcpu_record->vcpu_id,
970 decode, time_diff / NSEC_PER_USEC);
971 }
972 }
973
974 return update_kvm_event(kvm, event, vcpu, time_diff);
975}
976
977static
978struct vcpu_event_record *per_vcpu_record(struct thread *thread,
979 struct evsel *evsel,
980 struct perf_sample *sample)
981{
982 /* Only kvm_entry records vcpu id. */
983 if (!thread__priv(thread) && kvm_entry_event(evsel)) {
984 struct vcpu_event_record *vcpu_record;
985
986 vcpu_record = zalloc(sizeof(*vcpu_record));
987 if (!vcpu_record) {
988 pr_err("%s: Not enough memory\n", __func__);
989 return NULL;
990 }
991
992 vcpu_record->vcpu_id = evsel__intval(evsel, sample, vcpu_id_str);
993 thread__set_priv(thread, vcpu_record);
994 }
995
996 return thread__priv(thread);
997}
998
999static bool handle_kvm_event(struct perf_kvm_stat *kvm,
1000 struct thread *thread,
1001 struct evsel *evsel,
1002 struct perf_sample *sample)
1003{
1004 struct vcpu_event_record *vcpu_record;
1005 struct event_key key = { .key = INVALID_KEY,
1006 .exit_reasons = kvm->exit_reasons };
1007
1008 vcpu_record = per_vcpu_record(thread, evsel, sample);
1009 if (!vcpu_record)
1010 return true;
1011
1012 /* only process events for vcpus user cares about */
1013 if ((kvm->trace_vcpu != -1) &&
1014 (kvm->trace_vcpu != vcpu_record->vcpu_id))
1015 return true;
1016
1017 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
1018 return handle_begin_event(kvm, vcpu_record, &key, sample);
1019
1020 if (is_child_event(kvm, evsel, sample, &key))
1021 return handle_child_event(kvm, vcpu_record, &key, sample);
1022
1023 if (kvm->events_ops->is_end_event(evsel, sample, &key))
1024 return handle_end_event(kvm, vcpu_record, &key, sample);
1025
1026 return true;
1027}
1028
1029static bool is_valid_key(struct perf_kvm_stat *kvm)
1030{
1031 static const char *key_array[] = {
1032 "ev_name", "sample", "time", "max_t", "min_t", "mean_t",
1033 };
1034 unsigned int i;
1035
1036 for (i = 0; i < ARRAY_SIZE(key_array); i++)
1037 if (!strcmp(key_array[i], kvm->sort_key))
1038 return true;
1039
1040 pr_err("Unsupported sort key: %s\n", kvm->sort_key);
1041 return false;
1042}
1043
1044static bool event_is_valid(struct kvm_event *event, int vcpu)
1045{
1046 return !!get_event_count(event, vcpu);
1047}
1048
1049static int filter_cb(struct hist_entry *he, void *arg __maybe_unused)
1050{
1051 struct kvm_event *event;
1052 struct perf_kvm_stat *perf_kvm;
1053
1054 event = container_of(he, struct kvm_event, he);
1055 perf_kvm = event->perf_kvm;
1056 if (!event_is_valid(event, perf_kvm->trace_vcpu))
1057 he->filtered = 1;
1058 else
1059 he->filtered = 0;
1060 return 0;
1061}
1062
1063static void sort_result(struct perf_kvm_stat *kvm)
1064{
1065 struct ui_progress prog;
1066 const char *output_columns = "ev_name,sample,percent_sample,"
1067 "time,percent_time,max_t,min_t,mean_t";
1068
1069 kvm_hists__reinit(output_columns, kvm->sort_key);
1070 ui_progress__init(&prog, kvm_hists.hists.nr_entries, "Sorting...");
1071 hists__collapse_resort(&kvm_hists.hists, NULL);
1072 hists__output_resort_cb(&kvm_hists.hists, NULL, filter_cb);
1073 ui_progress__finish();
1074}
1075
1076static void print_vcpu_info(struct perf_kvm_stat *kvm)
1077{
1078 int vcpu = kvm->trace_vcpu;
1079
1080 pr_info("Analyze events for ");
1081
1082 if (kvm->opts.target.system_wide)
1083 pr_info("all VMs, ");
1084 else if (kvm->opts.target.pid)
1085 pr_info("pid(s) %s, ", kvm->opts.target.pid);
1086 else
1087 pr_info("dazed and confused on what is monitored, ");
1088
1089 if (vcpu == -1)
1090 pr_info("all VCPUs:\n\n");
1091 else
1092 pr_info("VCPU %d:\n\n", vcpu);
1093}
1094
1095static void show_timeofday(void)
1096{
1097 char date[64];
1098 struct timeval tv;
1099 struct tm ltime;
1100
1101 gettimeofday(&tv, NULL);
1102 if (localtime_r(&tv.tv_sec, <ime)) {
1103 strftime(date, sizeof(date), "%H:%M:%S", <ime);
1104 pr_info("%s.%06ld", date, tv.tv_usec);
1105 } else
1106 pr_info("00:00:00.000000");
1107
1108 return;
1109}
1110
1111static void print_result(struct perf_kvm_stat *kvm)
1112{
1113 char decode[KVM_EVENT_NAME_LEN];
1114 struct kvm_event *event;
1115 int vcpu = kvm->trace_vcpu;
1116 struct rb_node *nd;
1117
1118 if (kvm->live) {
1119 puts(CONSOLE_CLEAR);
1120 show_timeofday();
1121 }
1122
1123 pr_info("\n\n");
1124 print_vcpu_info(kvm);
1125 pr_info("%*s ", KVM_EVENT_NAME_LEN, kvm->events_ops->name);
1126 pr_info("%10s ", "Samples");
1127 pr_info("%9s ", "Samples%");
1128
1129 pr_info("%9s ", "Time%");
1130 pr_info("%11s ", "Min Time");
1131 pr_info("%11s ", "Max Time");
1132 pr_info("%16s ", "Avg time");
1133 pr_info("\n\n");
1134
1135 for (nd = rb_first_cached(&kvm_hists.hists.entries); nd; nd = rb_next(nd)) {
1136 struct hist_entry *he;
1137 u64 ecount, etime, max, min;
1138
1139 he = rb_entry(nd, struct hist_entry, rb_node);
1140 if (he->filtered)
1141 continue;
1142
1143 event = container_of(he, struct kvm_event, he);
1144 ecount = get_event_count(event, vcpu);
1145 etime = get_event_time(event, vcpu);
1146 max = get_event_max(event, vcpu);
1147 min = get_event_min(event, vcpu);
1148
1149 kvm->events_ops->decode_key(kvm, &event->key, decode);
1150 pr_info("%*s ", KVM_EVENT_NAME_LEN, decode);
1151 pr_info("%10llu ", (unsigned long long)ecount);
1152 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
1153 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
1154 pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
1155 pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
1156 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
1157 kvm_event_rel_stddev(vcpu, event));
1158 pr_info("\n");
1159 }
1160
1161 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
1162 kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
1163
1164 if (kvm->lost_events)
1165 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
1166}
1167
1168#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1169static int process_lost_event(struct perf_tool *tool,
1170 union perf_event *event __maybe_unused,
1171 struct perf_sample *sample __maybe_unused,
1172 struct machine *machine __maybe_unused)
1173{
1174 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
1175
1176 kvm->lost_events++;
1177 return 0;
1178}
1179#endif
1180
1181static bool skip_sample(struct perf_kvm_stat *kvm,
1182 struct perf_sample *sample)
1183{
1184 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
1185 return true;
1186
1187 return false;
1188}
1189
1190static int process_sample_event(struct perf_tool *tool,
1191 union perf_event *event,
1192 struct perf_sample *sample,
1193 struct evsel *evsel,
1194 struct machine *machine)
1195{
1196 int err = 0;
1197 struct thread *thread;
1198 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
1199 tool);
1200
1201 if (skip_sample(kvm, sample))
1202 return 0;
1203
1204 if (machine__resolve(machine, &kvm->al, sample) < 0) {
1205 pr_warning("Fail to resolve address location, skip sample.\n");
1206 return 0;
1207 }
1208
1209 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
1210 if (thread == NULL) {
1211 pr_debug("problem processing %d event, skipping it.\n",
1212 event->header.type);
1213 return -1;
1214 }
1215
1216 if (!handle_kvm_event(kvm, thread, evsel, sample))
1217 err = -1;
1218
1219 thread__put(thread);
1220 return err;
1221}
1222
1223static int cpu_isa_config(struct perf_kvm_stat *kvm)
1224{
1225 char buf[128], *cpuid;
1226 int err;
1227
1228 if (kvm->live) {
1229 err = get_cpuid(buf, sizeof(buf));
1230 if (err != 0) {
1231 pr_err("Failed to look up CPU type: %s\n",
1232 str_error_r(err, buf, sizeof(buf)));
1233 return -err;
1234 }
1235 cpuid = buf;
1236 } else
1237 cpuid = kvm->session->header.env.cpuid;
1238
1239 if (!cpuid) {
1240 pr_err("Failed to look up CPU type\n");
1241 return -EINVAL;
1242 }
1243
1244 err = cpu_isa_init(kvm, cpuid);
1245 if (err == -ENOTSUP)
1246 pr_err("CPU %s is not supported.\n", cpuid);
1247
1248 return err;
1249}
1250
1251static bool verify_vcpu(int vcpu)
1252{
1253 if (vcpu != -1 && vcpu < 0) {
1254 pr_err("Invalid vcpu:%d.\n", vcpu);
1255 return false;
1256 }
1257
1258 return true;
1259}
1260
1261#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1262/* keeping the max events to a modest level to keep
1263 * the processing of samples per mmap smooth.
1264 */
1265#define PERF_KVM__MAX_EVENTS_PER_MMAP 25
1266
1267static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
1268 u64 *mmap_time)
1269{
1270 struct evlist *evlist = kvm->evlist;
1271 union perf_event *event;
1272 struct mmap *md;
1273 u64 timestamp;
1274 s64 n = 0;
1275 int err;
1276
1277 *mmap_time = ULLONG_MAX;
1278 md = &evlist->mmap[idx];
1279 err = perf_mmap__read_init(&md->core);
1280 if (err < 0)
1281 return (err == -EAGAIN) ? 0 : -1;
1282
1283 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
1284 err = evlist__parse_sample_timestamp(evlist, event, ×tamp);
1285 if (err) {
1286 perf_mmap__consume(&md->core);
1287 pr_err("Failed to parse sample\n");
1288 return -1;
1289 }
1290
1291 err = perf_session__queue_event(kvm->session, event, timestamp, 0, NULL);
1292 /*
1293 * FIXME: Here we can't consume the event, as perf_session__queue_event will
1294 * point to it, and it'll get possibly overwritten by the kernel.
1295 */
1296 perf_mmap__consume(&md->core);
1297
1298 if (err) {
1299 pr_err("Failed to enqueue sample: %d\n", err);
1300 return -1;
1301 }
1302
1303 /* save time stamp of our first sample for this mmap */
1304 if (n == 0)
1305 *mmap_time = timestamp;
1306
1307 /* limit events per mmap handled all at once */
1308 n++;
1309 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
1310 break;
1311 }
1312
1313 perf_mmap__read_done(&md->core);
1314 return n;
1315}
1316
1317static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
1318{
1319 int i, err, throttled = 0;
1320 s64 n, ntotal = 0;
1321 u64 flush_time = ULLONG_MAX, mmap_time;
1322
1323 for (i = 0; i < kvm->evlist->core.nr_mmaps; i++) {
1324 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
1325 if (n < 0)
1326 return -1;
1327
1328 /* flush time is going to be the minimum of all the individual
1329 * mmap times. Essentially, we flush all the samples queued up
1330 * from the last pass under our minimal start time -- that leaves
1331 * a very small race for samples to come in with a lower timestamp.
1332 * The ioctl to return the perf_clock timestamp should close the
1333 * race entirely.
1334 */
1335 if (mmap_time < flush_time)
1336 flush_time = mmap_time;
1337
1338 ntotal += n;
1339 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
1340 throttled = 1;
1341 }
1342
1343 /* flush queue after each round in which we processed events */
1344 if (ntotal) {
1345 struct ordered_events *oe = &kvm->session->ordered_events;
1346
1347 oe->next_flush = flush_time;
1348 err = ordered_events__flush(oe, OE_FLUSH__ROUND);
1349 if (err) {
1350 if (kvm->lost_events)
1351 pr_info("\nLost events: %" PRIu64 "\n\n",
1352 kvm->lost_events);
1353 return err;
1354 }
1355 }
1356
1357 return throttled;
1358}
1359
1360static volatile int done;
1361
1362static void sig_handler(int sig __maybe_unused)
1363{
1364 done = 1;
1365}
1366
1367static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
1368{
1369 struct itimerspec new_value;
1370 int rc = -1;
1371
1372 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
1373 if (kvm->timerfd < 0) {
1374 pr_err("timerfd_create failed\n");
1375 goto out;
1376 }
1377
1378 new_value.it_value.tv_sec = kvm->display_time;
1379 new_value.it_value.tv_nsec = 0;
1380 new_value.it_interval.tv_sec = kvm->display_time;
1381 new_value.it_interval.tv_nsec = 0;
1382
1383 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
1384 pr_err("timerfd_settime failed: %d\n", errno);
1385 close(kvm->timerfd);
1386 goto out;
1387 }
1388
1389 rc = 0;
1390out:
1391 return rc;
1392}
1393
1394static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
1395{
1396 uint64_t c;
1397 int rc;
1398
1399 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
1400 if (rc < 0) {
1401 if (errno == EAGAIN)
1402 return 0;
1403
1404 pr_err("Failed to read timer fd: %d\n", errno);
1405 return -1;
1406 }
1407
1408 if (rc != sizeof(uint64_t)) {
1409 pr_err("Error reading timer fd - invalid size returned\n");
1410 return -1;
1411 }
1412
1413 if (c != 1)
1414 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
1415
1416 /* update display */
1417 sort_result(kvm);
1418 print_result(kvm);
1419
1420 /* Reset sort list to "ev_name" */
1421 kvm_hists__reinit(NULL, "ev_name");
1422
1423 /* reset counts */
1424 clear_events_cache_stats();
1425 kvm->total_count = 0;
1426 kvm->total_time = 0;
1427 kvm->lost_events = 0;
1428
1429 return 0;
1430}
1431
1432static int fd_set_nonblock(int fd)
1433{
1434 long arg = 0;
1435
1436 arg = fcntl(fd, F_GETFL);
1437 if (arg < 0) {
1438 pr_err("Failed to get current flags for fd %d\n", fd);
1439 return -1;
1440 }
1441
1442 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
1443 pr_err("Failed to set non-block option on fd %d\n", fd);
1444 return -1;
1445 }
1446
1447 return 0;
1448}
1449
1450static int perf_kvm__handle_stdin(void)
1451{
1452 int c;
1453
1454 c = getc(stdin);
1455 if (c == 'q')
1456 return 1;
1457
1458 return 0;
1459}
1460
1461static int kvm_events_live_report(struct perf_kvm_stat *kvm)
1462{
1463 int nr_stdin, ret, err = -EINVAL;
1464 struct termios save;
1465
1466 /* live flag must be set first */
1467 kvm->live = true;
1468
1469 ret = cpu_isa_config(kvm);
1470 if (ret < 0)
1471 return ret;
1472
1473 if (!verify_vcpu(kvm->trace_vcpu) ||
1474 !is_valid_key(kvm) ||
1475 !register_kvm_events_ops(kvm)) {
1476 goto out;
1477 }
1478
1479 set_term_quiet_input(&save);
1480
1481 kvm_hists__init();
1482
1483 signal(SIGINT, sig_handler);
1484 signal(SIGTERM, sig_handler);
1485
1486 /* add timer fd */
1487 if (perf_kvm__timerfd_create(kvm) < 0) {
1488 err = -1;
1489 goto out;
1490 }
1491
1492 if (evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
1493 goto out;
1494
1495 nr_stdin = evlist__add_pollfd(kvm->evlist, fileno(stdin));
1496 if (nr_stdin < 0)
1497 goto out;
1498
1499 if (fd_set_nonblock(fileno(stdin)) != 0)
1500 goto out;
1501
1502 /* everything is good - enable the events and process */
1503 evlist__enable(kvm->evlist);
1504
1505 while (!done) {
1506 struct fdarray *fda = &kvm->evlist->core.pollfd;
1507 int rc;
1508
1509 rc = perf_kvm__mmap_read(kvm);
1510 if (rc < 0)
1511 break;
1512
1513 err = perf_kvm__handle_timerfd(kvm);
1514 if (err)
1515 goto out;
1516
1517 if (fda->entries[nr_stdin].revents & POLLIN)
1518 done = perf_kvm__handle_stdin();
1519
1520 if (!rc && !done)
1521 err = evlist__poll(kvm->evlist, 100);
1522 }
1523
1524 evlist__disable(kvm->evlist);
1525
1526 if (err == 0) {
1527 sort_result(kvm);
1528 print_result(kvm);
1529 }
1530
1531out:
1532 hists__delete_entries(&kvm_hists.hists);
1533
1534 if (kvm->timerfd >= 0)
1535 close(kvm->timerfd);
1536
1537 tcsetattr(0, TCSAFLUSH, &save);
1538 return err;
1539}
1540
1541static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1542{
1543 int err, rc = -1;
1544 struct evsel *pos;
1545 struct evlist *evlist = kvm->evlist;
1546 char sbuf[STRERR_BUFSIZE];
1547
1548 evlist__config(evlist, &kvm->opts, NULL);
1549
1550 /*
1551 * Note: exclude_{guest,host} do not apply here.
1552 * This command processes KVM tracepoints from host only
1553 */
1554 evlist__for_each_entry(evlist, pos) {
1555 struct perf_event_attr *attr = &pos->core.attr;
1556
1557 /* make sure these *are* set */
1558 evsel__set_sample_bit(pos, TID);
1559 evsel__set_sample_bit(pos, TIME);
1560 evsel__set_sample_bit(pos, CPU);
1561 evsel__set_sample_bit(pos, RAW);
1562 /* make sure these are *not*; want as small a sample as possible */
1563 evsel__reset_sample_bit(pos, PERIOD);
1564 evsel__reset_sample_bit(pos, IP);
1565 evsel__reset_sample_bit(pos, CALLCHAIN);
1566 evsel__reset_sample_bit(pos, ADDR);
1567 evsel__reset_sample_bit(pos, READ);
1568 attr->mmap = 0;
1569 attr->comm = 0;
1570 attr->task = 0;
1571
1572 attr->sample_period = 1;
1573
1574 attr->watermark = 0;
1575 attr->wakeup_events = 1000;
1576
1577 /* will enable all once we are ready */
1578 attr->disabled = 1;
1579 }
1580
1581 err = evlist__open(evlist);
1582 if (err < 0) {
1583 printf("Couldn't create the events: %s\n",
1584 str_error_r(errno, sbuf, sizeof(sbuf)));
1585 goto out;
1586 }
1587
1588 if (evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1589 ui__error("Failed to mmap the events: %s\n",
1590 str_error_r(errno, sbuf, sizeof(sbuf)));
1591 evlist__close(evlist);
1592 goto out;
1593 }
1594
1595 rc = 0;
1596
1597out:
1598 return rc;
1599}
1600#endif
1601
1602static int read_events(struct perf_kvm_stat *kvm)
1603{
1604 int ret;
1605
1606 struct perf_tool eops = {
1607 .sample = process_sample_event,
1608 .comm = perf_event__process_comm,
1609 .namespaces = perf_event__process_namespaces,
1610 .ordered_events = true,
1611 };
1612 struct perf_data file = {
1613 .path = kvm->file_name,
1614 .mode = PERF_DATA_MODE_READ,
1615 .force = kvm->force,
1616 };
1617
1618 kvm->tool = eops;
1619 kvm->session = perf_session__new(&file, &kvm->tool);
1620 if (IS_ERR(kvm->session)) {
1621 pr_err("Initializing perf session failed\n");
1622 return PTR_ERR(kvm->session);
1623 }
1624
1625 symbol__init(&kvm->session->header.env);
1626
1627 if (!perf_session__has_traces(kvm->session, "kvm record")) {
1628 ret = -EINVAL;
1629 goto out_delete;
1630 }
1631
1632 /*
1633 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1634 * traced in the old kernel.
1635 */
1636 ret = cpu_isa_config(kvm);
1637 if (ret < 0)
1638 goto out_delete;
1639
1640 ret = perf_session__process_events(kvm->session);
1641
1642out_delete:
1643 perf_session__delete(kvm->session);
1644 return ret;
1645}
1646
1647static int parse_target_str(struct perf_kvm_stat *kvm)
1648{
1649 if (kvm->opts.target.pid) {
1650 kvm->pid_list = intlist__new(kvm->opts.target.pid);
1651 if (kvm->pid_list == NULL) {
1652 pr_err("Error parsing process id string\n");
1653 return -EINVAL;
1654 }
1655 }
1656
1657 return 0;
1658}
1659
1660static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1661{
1662 int ret = -EINVAL;
1663 int vcpu = kvm->trace_vcpu;
1664
1665 if (parse_target_str(kvm) != 0)
1666 goto exit;
1667
1668 if (!verify_vcpu(vcpu))
1669 goto exit;
1670
1671 if (!is_valid_key(kvm))
1672 goto exit;
1673
1674 if (!register_kvm_events_ops(kvm))
1675 goto exit;
1676
1677 if (kvm->use_stdio) {
1678 use_browser = 0;
1679 setup_pager();
1680 } else {
1681 use_browser = 1;
1682 }
1683
1684 setup_browser(false);
1685
1686 kvm_hists__init();
1687
1688 ret = read_events(kvm);
1689 if (ret)
1690 goto exit;
1691
1692 sort_result(kvm);
1693 kvm_display(kvm);
1694
1695exit:
1696 hists__delete_entries(&kvm_hists.hists);
1697 return ret;
1698}
1699
1700#define STRDUP_FAIL_EXIT(s) \
1701 ({ char *_p; \
1702 _p = strdup(s); \
1703 if (!_p) \
1704 return -ENOMEM; \
1705 _p; \
1706 })
1707
1708int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1709{
1710 return 0;
1711}
1712
1713static int
1714kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1715{
1716 unsigned int rec_argc, i, j, events_tp_size;
1717 const char **rec_argv;
1718 const char * const record_args[] = {
1719 "record",
1720 "-R",
1721 "-m", "1024",
1722 "-c", "1",
1723 };
1724 const char * const kvm_stat_record_usage[] = {
1725 "perf kvm stat record [<options>]",
1726 NULL
1727 };
1728 const char * const *events_tp;
1729 int ret;
1730
1731 events_tp_size = 0;
1732 ret = setup_kvm_events_tp(kvm);
1733 if (ret < 0) {
1734 pr_err("Unable to setup the kvm tracepoints\n");
1735 return ret;
1736 }
1737
1738 for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1739 events_tp_size++;
1740
1741 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1742 2 * events_tp_size;
1743 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1744
1745 if (rec_argv == NULL)
1746 return -ENOMEM;
1747
1748 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1749 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1750
1751 for (j = 0; j < events_tp_size; j++) {
1752 rec_argv[i++] = "-e";
1753 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1754 }
1755
1756 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1757 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1758
1759 for (j = 1; j < (unsigned int)argc; j++, i++)
1760 rec_argv[i] = argv[j];
1761
1762 set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1763 set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1764 set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1765
1766 set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1767 set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1768 set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1769 set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1770 set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1771 set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1772 set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1773 set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1774 set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1775 set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1776 set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1777 set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1778 set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1779 set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1780 set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1781
1782 record_usage = kvm_stat_record_usage;
1783 return cmd_record(i, rec_argv);
1784}
1785
1786static int
1787kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1788{
1789 const struct option kvm_events_report_options[] = {
1790 OPT_STRING(0, "event", &kvm->report_event, "report event",
1791 "event for reporting: vmexit, "
1792 "mmio (x86 only), ioport (x86 only)"),
1793 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1794 "vcpu id to report"),
1795 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1796 "key for sorting: sample(sort by samples number)"
1797 " time (sort by avg time)"),
1798 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1799 "analyze events only for given process id(s)"),
1800 OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1801 OPT_BOOLEAN(0, "stdio", &kvm->use_stdio, "use the stdio interface"),
1802 OPT_END()
1803 };
1804
1805 const char * const kvm_events_report_usage[] = {
1806 "perf kvm stat report [<options>]",
1807 NULL
1808 };
1809
1810 if (argc) {
1811 argc = parse_options(argc, argv,
1812 kvm_events_report_options,
1813 kvm_events_report_usage, 0);
1814 if (argc)
1815 usage_with_options(kvm_events_report_usage,
1816 kvm_events_report_options);
1817 }
1818
1819#ifndef HAVE_SLANG_SUPPORT
1820 kvm->use_stdio = true;
1821#endif
1822
1823 if (!kvm->opts.target.pid)
1824 kvm->opts.target.system_wide = true;
1825
1826 return kvm_events_report_vcpu(kvm);
1827}
1828
1829#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1830static struct evlist *kvm_live_event_list(void)
1831{
1832 struct evlist *evlist;
1833 char *tp, *name, *sys;
1834 int err = -1;
1835 const char * const *events_tp;
1836
1837 evlist = evlist__new();
1838 if (evlist == NULL)
1839 return NULL;
1840
1841 for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1842
1843 tp = strdup(*events_tp);
1844 if (tp == NULL)
1845 goto out;
1846
1847 /* split tracepoint into subsystem and name */
1848 sys = tp;
1849 name = strchr(tp, ':');
1850 if (name == NULL) {
1851 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1852 *events_tp);
1853 free(tp);
1854 goto out;
1855 }
1856 *name = '\0';
1857 name++;
1858
1859 if (evlist__add_newtp(evlist, sys, name, NULL)) {
1860 pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1861 free(tp);
1862 goto out;
1863 }
1864
1865 free(tp);
1866 }
1867
1868 err = 0;
1869
1870out:
1871 if (err) {
1872 evlist__delete(evlist);
1873 evlist = NULL;
1874 }
1875
1876 return evlist;
1877}
1878
1879static int kvm_events_live(struct perf_kvm_stat *kvm,
1880 int argc, const char **argv)
1881{
1882 char errbuf[BUFSIZ];
1883 int err;
1884
1885 const struct option live_options[] = {
1886 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1887 "record events on existing process id"),
1888 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1889 "number of mmap data pages", evlist__parse_mmap_pages),
1890 OPT_INCR('v', "verbose", &verbose,
1891 "be more verbose (show counter open errors, etc)"),
1892 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1893 "system-wide collection from all CPUs"),
1894 OPT_UINTEGER('d', "display", &kvm->display_time,
1895 "time in seconds between display updates"),
1896 OPT_STRING(0, "event", &kvm->report_event, "report event",
1897 "event for reporting: "
1898 "vmexit, mmio (x86 only), ioport (x86 only)"),
1899 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1900 "vcpu id to report"),
1901 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1902 "key for sorting: sample(sort by samples number)"
1903 " time (sort by avg time)"),
1904 OPT_U64(0, "duration", &kvm->duration,
1905 "show events other than"
1906 " HLT (x86 only) or Wait state (s390 only)"
1907 " that take longer than duration usecs"),
1908 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1909 "per thread proc mmap processing timeout in ms"),
1910 OPT_END()
1911 };
1912 const char * const live_usage[] = {
1913 "perf kvm stat live [<options>]",
1914 NULL
1915 };
1916 struct perf_data data = {
1917 .mode = PERF_DATA_MODE_WRITE,
1918 };
1919
1920
1921 /* event handling */
1922 kvm->tool.sample = process_sample_event;
1923 kvm->tool.comm = perf_event__process_comm;
1924 kvm->tool.exit = perf_event__process_exit;
1925 kvm->tool.fork = perf_event__process_fork;
1926 kvm->tool.lost = process_lost_event;
1927 kvm->tool.namespaces = perf_event__process_namespaces;
1928 kvm->tool.ordered_events = true;
1929 perf_tool__fill_defaults(&kvm->tool);
1930
1931 /* set defaults */
1932 kvm->display_time = 1;
1933 kvm->opts.user_interval = 1;
1934 kvm->opts.mmap_pages = 512;
1935 kvm->opts.target.uses_mmap = false;
1936 kvm->opts.target.uid_str = NULL;
1937 kvm->opts.target.uid = UINT_MAX;
1938
1939 symbol__init(NULL);
1940 disable_buildid_cache();
1941
1942 use_browser = 0;
1943
1944 if (argc) {
1945 argc = parse_options(argc, argv, live_options,
1946 live_usage, 0);
1947 if (argc)
1948 usage_with_options(live_usage, live_options);
1949 }
1950
1951 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1952
1953 /*
1954 * target related setups
1955 */
1956 err = target__validate(&kvm->opts.target);
1957 if (err) {
1958 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1959 ui__warning("%s", errbuf);
1960 }
1961
1962 if (target__none(&kvm->opts.target))
1963 kvm->opts.target.system_wide = true;
1964
1965
1966 /*
1967 * generate the event list
1968 */
1969 err = setup_kvm_events_tp(kvm);
1970 if (err < 0) {
1971 pr_err("Unable to setup the kvm tracepoints\n");
1972 return err;
1973 }
1974
1975 kvm->evlist = kvm_live_event_list();
1976 if (kvm->evlist == NULL) {
1977 err = -1;
1978 goto out;
1979 }
1980
1981 if (evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1982 usage_with_options(live_usage, live_options);
1983
1984 /*
1985 * perf session
1986 */
1987 kvm->session = perf_session__new(&data, &kvm->tool);
1988 if (IS_ERR(kvm->session)) {
1989 err = PTR_ERR(kvm->session);
1990 goto out;
1991 }
1992 kvm->session->evlist = kvm->evlist;
1993 perf_session__set_id_hdr_size(kvm->session);
1994 ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1995 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1996 kvm->evlist->core.threads, true, false, 1);
1997 err = kvm_live_open_events(kvm);
1998 if (err)
1999 goto out;
2000
2001 err = kvm_events_live_report(kvm);
2002
2003out:
2004 perf_session__delete(kvm->session);
2005 kvm->session = NULL;
2006 evlist__delete(kvm->evlist);
2007
2008 return err;
2009}
2010#endif
2011
2012static void print_kvm_stat_usage(void)
2013{
2014 printf("Usage: perf kvm stat <command>\n\n");
2015
2016 printf("# Available commands:\n");
2017 printf("\trecord: record kvm events\n");
2018 printf("\treport: report statistical data of kvm events\n");
2019 printf("\tlive: live reporting of statistical data of kvm events\n");
2020
2021 printf("\nOtherwise, it is the alias of 'perf stat':\n");
2022}
2023
2024static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
2025{
2026 struct perf_kvm_stat kvm = {
2027 .file_name = file_name,
2028
2029 .trace_vcpu = -1,
2030 .report_event = "vmexit",
2031 .sort_key = "sample",
2032
2033 };
2034
2035 if (argc == 1) {
2036 print_kvm_stat_usage();
2037 goto perf_stat;
2038 }
2039
2040 if (strlen(argv[1]) > 2 && strstarts("record", argv[1]))
2041 return kvm_events_record(&kvm, argc - 1, argv + 1);
2042
2043 if (strlen(argv[1]) > 2 && strstarts("report", argv[1]))
2044 return kvm_events_report(&kvm, argc - 1 , argv + 1);
2045
2046#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
2047 if (!strncmp(argv[1], "live", 4))
2048 return kvm_events_live(&kvm, argc - 1 , argv + 1);
2049#endif
2050
2051perf_stat:
2052 return cmd_stat(argc, argv);
2053}
2054#endif /* HAVE_KVM_STAT_SUPPORT */
2055
2056int __weak kvm_add_default_arch_event(int *argc __maybe_unused,
2057 const char **argv __maybe_unused)
2058{
2059 return 0;
2060}
2061
2062static int __cmd_record(const char *file_name, int argc, const char **argv)
2063{
2064 int rec_argc, i = 0, j, ret;
2065 const char **rec_argv;
2066
2067 ret = kvm_add_default_arch_event(&argc, argv);
2068 if (ret)
2069 return -EINVAL;
2070
2071 rec_argc = argc + 2;
2072 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2073 rec_argv[i++] = strdup("record");
2074 rec_argv[i++] = strdup("-o");
2075 rec_argv[i++] = strdup(file_name);
2076 for (j = 1; j < argc; j++, i++)
2077 rec_argv[i] = argv[j];
2078
2079 BUG_ON(i != rec_argc);
2080
2081 return cmd_record(i, rec_argv);
2082}
2083
2084static int __cmd_report(const char *file_name, int argc, const char **argv)
2085{
2086 int rec_argc, i = 0, j;
2087 const char **rec_argv;
2088
2089 rec_argc = argc + 2;
2090 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2091 rec_argv[i++] = strdup("report");
2092 rec_argv[i++] = strdup("-i");
2093 rec_argv[i++] = strdup(file_name);
2094 for (j = 1; j < argc; j++, i++)
2095 rec_argv[i] = argv[j];
2096
2097 BUG_ON(i != rec_argc);
2098
2099 return cmd_report(i, rec_argv);
2100}
2101
2102static int
2103__cmd_buildid_list(const char *file_name, int argc, const char **argv)
2104{
2105 int rec_argc, i = 0, j;
2106 const char **rec_argv;
2107
2108 rec_argc = argc + 2;
2109 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2110 rec_argv[i++] = strdup("buildid-list");
2111 rec_argv[i++] = strdup("-i");
2112 rec_argv[i++] = strdup(file_name);
2113 for (j = 1; j < argc; j++, i++)
2114 rec_argv[i] = argv[j];
2115
2116 BUG_ON(i != rec_argc);
2117
2118 return cmd_buildid_list(i, rec_argv);
2119}
2120
2121int cmd_kvm(int argc, const char **argv)
2122{
2123 const char *file_name = NULL;
2124 const struct option kvm_options[] = {
2125 OPT_STRING('i', "input", &file_name, "file",
2126 "Input file name"),
2127 OPT_STRING('o', "output", &file_name, "file",
2128 "Output file name"),
2129 OPT_BOOLEAN(0, "guest", &perf_guest,
2130 "Collect guest os data"),
2131 OPT_BOOLEAN(0, "host", &perf_host,
2132 "Collect host os data"),
2133 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
2134 "guest mount directory under which every guest os"
2135 " instance has a subdir"),
2136 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
2137 "file", "file saving guest os vmlinux"),
2138 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
2139 "file", "file saving guest os /proc/kallsyms"),
2140 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
2141 "file", "file saving guest os /proc/modules"),
2142 OPT_BOOLEAN(0, "guest-code", &symbol_conf.guest_code,
2143 "Guest code can be found in hypervisor process"),
2144 OPT_INCR('v', "verbose", &verbose,
2145 "be more verbose (show counter open errors, etc)"),
2146 OPT_END()
2147 };
2148
2149 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
2150 "buildid-list", "stat", NULL };
2151 const char *kvm_usage[] = { NULL, NULL };
2152
2153 perf_host = 0;
2154 perf_guest = 1;
2155
2156 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
2157 PARSE_OPT_STOP_AT_NON_OPTION);
2158 if (!argc)
2159 usage_with_options(kvm_usage, kvm_options);
2160
2161 if (!perf_host)
2162 perf_guest = 1;
2163
2164 if (!file_name) {
2165 file_name = get_filename_for_perf_kvm();
2166
2167 if (!file_name) {
2168 pr_err("Failed to allocate memory for filename\n");
2169 return -ENOMEM;
2170 }
2171 }
2172
2173 if (strlen(argv[0]) > 2 && strstarts("record", argv[0]))
2174 return __cmd_record(file_name, argc, argv);
2175 else if (strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2176 return __cmd_report(file_name, argc, argv);
2177 else if (strlen(argv[0]) > 2 && strstarts("diff", argv[0]))
2178 return cmd_diff(argc, argv);
2179 else if (!strcmp(argv[0], "top"))
2180 return cmd_top(argc, argv);
2181 else if (strlen(argv[0]) > 2 && strstarts("buildid-list", argv[0]))
2182 return __cmd_buildid_list(file_name, argc, argv);
2183#if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
2184 else if (strlen(argv[0]) > 2 && strstarts("stat", argv[0]))
2185 return kvm_cmd_stat(file_name, argc, argv);
2186#endif
2187 else
2188 usage_with_options(kvm_usage, kvm_options);
2189
2190 return 0;
2191}