Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include "builtin.h"
3#include "perf.h"
4
5#include "util/evsel.h"
6#include "util/evlist.h"
7#include "util/term.h"
8#include "util/util.h"
9#include "util/cache.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/parse-options.h>
16#include "util/trace-event.h"
17#include "util/debug.h"
18#include "util/tool.h"
19#include "util/stat.h"
20#include "util/top.h"
21#include "util/data.h"
22#include "util/ordered-events.h"
23
24#include <sys/prctl.h>
25#ifdef HAVE_TIMERFD_SUPPORT
26#include <sys/timerfd.h>
27#endif
28#include <sys/time.h>
29#include <sys/types.h>
30#include <sys/stat.h>
31#include <fcntl.h>
32
33#include <linux/kernel.h>
34#include <linux/time64.h>
35#include <errno.h>
36#include <inttypes.h>
37#include <poll.h>
38#include <termios.h>
39#include <semaphore.h>
40#include <signal.h>
41#include <math.h>
42
43static const char *get_filename_for_perf_kvm(void)
44{
45 const char *filename;
46
47 if (perf_host && !perf_guest)
48 filename = strdup("perf.data.host");
49 else if (!perf_host && perf_guest)
50 filename = strdup("perf.data.guest");
51 else
52 filename = strdup("perf.data.kvm");
53
54 return filename;
55}
56
57#ifdef HAVE_KVM_STAT_SUPPORT
58#include "util/kvm-stat.h"
59
60void exit_event_get_key(struct perf_evsel *evsel,
61 struct perf_sample *sample,
62 struct event_key *key)
63{
64 key->info = 0;
65 key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
66}
67
68bool kvm_exit_event(struct perf_evsel *evsel)
69{
70 return !strcmp(evsel->name, kvm_exit_trace);
71}
72
73bool exit_event_begin(struct perf_evsel *evsel,
74 struct perf_sample *sample, struct event_key *key)
75{
76 if (kvm_exit_event(evsel)) {
77 exit_event_get_key(evsel, sample, key);
78 return true;
79 }
80
81 return false;
82}
83
84bool kvm_entry_event(struct perf_evsel *evsel)
85{
86 return !strcmp(evsel->name, kvm_entry_trace);
87}
88
89bool exit_event_end(struct perf_evsel *evsel,
90 struct perf_sample *sample __maybe_unused,
91 struct event_key *key __maybe_unused)
92{
93 return kvm_entry_event(evsel);
94}
95
96static const char *get_exit_reason(struct perf_kvm_stat *kvm,
97 struct exit_reasons_table *tbl,
98 u64 exit_code)
99{
100 while (tbl->reason != NULL) {
101 if (tbl->exit_code == exit_code)
102 return tbl->reason;
103 tbl++;
104 }
105
106 pr_err("unknown kvm exit code:%lld on %s\n",
107 (unsigned long long)exit_code, kvm->exit_reasons_isa);
108 return "UNKNOWN";
109}
110
111void exit_event_decode_key(struct perf_kvm_stat *kvm,
112 struct event_key *key,
113 char *decode)
114{
115 const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
116 key->key);
117
118 scnprintf(decode, decode_str_len, "%s", exit_reason);
119}
120
121static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
122{
123 struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
124
125 for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
126 if (!strcmp(events_ops->name, kvm->report_event)) {
127 kvm->events_ops = events_ops->ops;
128 return true;
129 }
130 }
131
132 return false;
133}
134
135struct vcpu_event_record {
136 int vcpu_id;
137 u64 start_time;
138 struct kvm_event *last_event;
139};
140
141
142static void init_kvm_event_record(struct perf_kvm_stat *kvm)
143{
144 unsigned int i;
145
146 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
147 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
148}
149
150#ifdef HAVE_TIMERFD_SUPPORT
151static void clear_events_cache_stats(struct list_head *kvm_events_cache)
152{
153 struct list_head *head;
154 struct kvm_event *event;
155 unsigned int i;
156 int j;
157
158 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
159 head = &kvm_events_cache[i];
160 list_for_each_entry(event, head, hash_entry) {
161 /* reset stats for event */
162 event->total.time = 0;
163 init_stats(&event->total.stats);
164
165 for (j = 0; j < event->max_vcpu; ++j) {
166 event->vcpu[j].time = 0;
167 init_stats(&event->vcpu[j].stats);
168 }
169 }
170 }
171}
172#endif
173
174static int kvm_events_hash_fn(u64 key)
175{
176 return key & (EVENTS_CACHE_SIZE - 1);
177}
178
179static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
180{
181 int old_max_vcpu = event->max_vcpu;
182 void *prev;
183
184 if (vcpu_id < event->max_vcpu)
185 return true;
186
187 while (event->max_vcpu <= vcpu_id)
188 event->max_vcpu += DEFAULT_VCPU_NUM;
189
190 prev = event->vcpu;
191 event->vcpu = realloc(event->vcpu,
192 event->max_vcpu * sizeof(*event->vcpu));
193 if (!event->vcpu) {
194 free(prev);
195 pr_err("Not enough memory\n");
196 return false;
197 }
198
199 memset(event->vcpu + old_max_vcpu, 0,
200 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
201 return true;
202}
203
204static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
205{
206 struct kvm_event *event;
207
208 event = zalloc(sizeof(*event));
209 if (!event) {
210 pr_err("Not enough memory\n");
211 return NULL;
212 }
213
214 event->key = *key;
215 init_stats(&event->total.stats);
216 return event;
217}
218
219static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
220 struct event_key *key)
221{
222 struct kvm_event *event;
223 struct list_head *head;
224
225 BUG_ON(key->key == INVALID_KEY);
226
227 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
228 list_for_each_entry(event, head, hash_entry) {
229 if (event->key.key == key->key && event->key.info == key->info)
230 return event;
231 }
232
233 event = kvm_alloc_init_event(key);
234 if (!event)
235 return NULL;
236
237 list_add(&event->hash_entry, head);
238 return event;
239}
240
241static bool handle_begin_event(struct perf_kvm_stat *kvm,
242 struct vcpu_event_record *vcpu_record,
243 struct event_key *key, u64 timestamp)
244{
245 struct kvm_event *event = NULL;
246
247 if (key->key != INVALID_KEY)
248 event = find_create_kvm_event(kvm, key);
249
250 vcpu_record->last_event = event;
251 vcpu_record->start_time = timestamp;
252 return true;
253}
254
255static void
256kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
257{
258 kvm_stats->time += time_diff;
259 update_stats(&kvm_stats->stats, time_diff);
260}
261
262static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
263{
264 struct kvm_event_stats *kvm_stats = &event->total;
265
266 if (vcpu_id != -1)
267 kvm_stats = &event->vcpu[vcpu_id];
268
269 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
270 avg_stats(&kvm_stats->stats));
271}
272
273static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
274 u64 time_diff)
275{
276 if (vcpu_id == -1) {
277 kvm_update_event_stats(&event->total, time_diff);
278 return true;
279 }
280
281 if (!kvm_event_expand(event, vcpu_id))
282 return false;
283
284 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
285 return true;
286}
287
288static bool is_child_event(struct perf_kvm_stat *kvm,
289 struct perf_evsel *evsel,
290 struct perf_sample *sample,
291 struct event_key *key)
292{
293 struct child_event_ops *child_ops;
294
295 child_ops = kvm->events_ops->child_ops;
296
297 if (!child_ops)
298 return false;
299
300 for (; child_ops->name; child_ops++) {
301 if (!strcmp(evsel->name, child_ops->name)) {
302 child_ops->get_key(evsel, sample, key);
303 return true;
304 }
305 }
306
307 return false;
308}
309
310static bool handle_child_event(struct perf_kvm_stat *kvm,
311 struct vcpu_event_record *vcpu_record,
312 struct event_key *key,
313 struct perf_sample *sample __maybe_unused)
314{
315 struct kvm_event *event = NULL;
316
317 if (key->key != INVALID_KEY)
318 event = find_create_kvm_event(kvm, key);
319
320 vcpu_record->last_event = event;
321
322 return true;
323}
324
325static bool skip_event(const char *event)
326{
327 const char * const *skip_events;
328
329 for (skip_events = kvm_skip_events; *skip_events; skip_events++)
330 if (!strcmp(event, *skip_events))
331 return true;
332
333 return false;
334}
335
336static bool handle_end_event(struct perf_kvm_stat *kvm,
337 struct vcpu_event_record *vcpu_record,
338 struct event_key *key,
339 struct perf_sample *sample)
340{
341 struct kvm_event *event;
342 u64 time_begin, time_diff;
343 int vcpu;
344
345 if (kvm->trace_vcpu == -1)
346 vcpu = -1;
347 else
348 vcpu = vcpu_record->vcpu_id;
349
350 event = vcpu_record->last_event;
351 time_begin = vcpu_record->start_time;
352
353 /* The begin event is not caught. */
354 if (!time_begin)
355 return true;
356
357 /*
358 * In some case, the 'begin event' only records the start timestamp,
359 * the actual event is recognized in the 'end event' (e.g. mmio-event).
360 */
361
362 /* Both begin and end events did not get the key. */
363 if (!event && key->key == INVALID_KEY)
364 return true;
365
366 if (!event)
367 event = find_create_kvm_event(kvm, key);
368
369 if (!event)
370 return false;
371
372 vcpu_record->last_event = NULL;
373 vcpu_record->start_time = 0;
374
375 /* seems to happen once in a while during live mode */
376 if (sample->time < time_begin) {
377 pr_debug("End time before begin time; skipping event.\n");
378 return true;
379 }
380
381 time_diff = sample->time - time_begin;
382
383 if (kvm->duration && time_diff > kvm->duration) {
384 char decode[decode_str_len];
385
386 kvm->events_ops->decode_key(kvm, &event->key, decode);
387 if (!skip_event(decode)) {
388 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
389 sample->time, sample->pid, vcpu_record->vcpu_id,
390 decode, time_diff / NSEC_PER_USEC);
391 }
392 }
393
394 return update_kvm_event(event, vcpu, time_diff);
395}
396
397static
398struct vcpu_event_record *per_vcpu_record(struct thread *thread,
399 struct perf_evsel *evsel,
400 struct perf_sample *sample)
401{
402 /* Only kvm_entry records vcpu id. */
403 if (!thread__priv(thread) && kvm_entry_event(evsel)) {
404 struct vcpu_event_record *vcpu_record;
405
406 vcpu_record = zalloc(sizeof(*vcpu_record));
407 if (!vcpu_record) {
408 pr_err("%s: Not enough memory\n", __func__);
409 return NULL;
410 }
411
412 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
413 vcpu_id_str);
414 thread__set_priv(thread, vcpu_record);
415 }
416
417 return thread__priv(thread);
418}
419
420static bool handle_kvm_event(struct perf_kvm_stat *kvm,
421 struct thread *thread,
422 struct perf_evsel *evsel,
423 struct perf_sample *sample)
424{
425 struct vcpu_event_record *vcpu_record;
426 struct event_key key = { .key = INVALID_KEY,
427 .exit_reasons = kvm->exit_reasons };
428
429 vcpu_record = per_vcpu_record(thread, evsel, sample);
430 if (!vcpu_record)
431 return true;
432
433 /* only process events for vcpus user cares about */
434 if ((kvm->trace_vcpu != -1) &&
435 (kvm->trace_vcpu != vcpu_record->vcpu_id))
436 return true;
437
438 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
439 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
440
441 if (is_child_event(kvm, evsel, sample, &key))
442 return handle_child_event(kvm, vcpu_record, &key, sample);
443
444 if (kvm->events_ops->is_end_event(evsel, sample, &key))
445 return handle_end_event(kvm, vcpu_record, &key, sample);
446
447 return true;
448}
449
450#define GET_EVENT_KEY(func, field) \
451static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
452{ \
453 if (vcpu == -1) \
454 return event->total.field; \
455 \
456 if (vcpu >= event->max_vcpu) \
457 return 0; \
458 \
459 return event->vcpu[vcpu].field; \
460}
461
462#define COMPARE_EVENT_KEY(func, field) \
463GET_EVENT_KEY(func, field) \
464static int compare_kvm_event_ ## func(struct kvm_event *one, \
465 struct kvm_event *two, int vcpu)\
466{ \
467 return get_event_ ##func(one, vcpu) > \
468 get_event_ ##func(two, vcpu); \
469}
470
471GET_EVENT_KEY(time, time);
472COMPARE_EVENT_KEY(count, stats.n);
473COMPARE_EVENT_KEY(mean, stats.mean);
474GET_EVENT_KEY(max, stats.max);
475GET_EVENT_KEY(min, stats.min);
476
477#define DEF_SORT_NAME_KEY(name, compare_key) \
478 { #name, compare_kvm_event_ ## compare_key }
479
480static struct kvm_event_key keys[] = {
481 DEF_SORT_NAME_KEY(sample, count),
482 DEF_SORT_NAME_KEY(time, mean),
483 { NULL, NULL }
484};
485
486static bool select_key(struct perf_kvm_stat *kvm)
487{
488 int i;
489
490 for (i = 0; keys[i].name; i++) {
491 if (!strcmp(keys[i].name, kvm->sort_key)) {
492 kvm->compare = keys[i].key;
493 return true;
494 }
495 }
496
497 pr_err("Unknown compare key:%s\n", kvm->sort_key);
498 return false;
499}
500
501static void insert_to_result(struct rb_root *result, struct kvm_event *event,
502 key_cmp_fun bigger, int vcpu)
503{
504 struct rb_node **rb = &result->rb_node;
505 struct rb_node *parent = NULL;
506 struct kvm_event *p;
507
508 while (*rb) {
509 p = container_of(*rb, struct kvm_event, rb);
510 parent = *rb;
511
512 if (bigger(event, p, vcpu))
513 rb = &(*rb)->rb_left;
514 else
515 rb = &(*rb)->rb_right;
516 }
517
518 rb_link_node(&event->rb, parent, rb);
519 rb_insert_color(&event->rb, result);
520}
521
522static void
523update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
524{
525 int vcpu = kvm->trace_vcpu;
526
527 kvm->total_count += get_event_count(event, vcpu);
528 kvm->total_time += get_event_time(event, vcpu);
529}
530
531static bool event_is_valid(struct kvm_event *event, int vcpu)
532{
533 return !!get_event_count(event, vcpu);
534}
535
536static void sort_result(struct perf_kvm_stat *kvm)
537{
538 unsigned int i;
539 int vcpu = kvm->trace_vcpu;
540 struct kvm_event *event;
541
542 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
543 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
544 if (event_is_valid(event, vcpu)) {
545 update_total_count(kvm, event);
546 insert_to_result(&kvm->result, event,
547 kvm->compare, vcpu);
548 }
549 }
550 }
551}
552
553/* returns left most element of result, and erase it */
554static struct kvm_event *pop_from_result(struct rb_root *result)
555{
556 struct rb_node *node = rb_first(result);
557
558 if (!node)
559 return NULL;
560
561 rb_erase(node, result);
562 return container_of(node, struct kvm_event, rb);
563}
564
565static void print_vcpu_info(struct perf_kvm_stat *kvm)
566{
567 int vcpu = kvm->trace_vcpu;
568
569 pr_info("Analyze events for ");
570
571 if (kvm->opts.target.system_wide)
572 pr_info("all VMs, ");
573 else if (kvm->opts.target.pid)
574 pr_info("pid(s) %s, ", kvm->opts.target.pid);
575 else
576 pr_info("dazed and confused on what is monitored, ");
577
578 if (vcpu == -1)
579 pr_info("all VCPUs:\n\n");
580 else
581 pr_info("VCPU %d:\n\n", vcpu);
582}
583
584static void show_timeofday(void)
585{
586 char date[64];
587 struct timeval tv;
588 struct tm ltime;
589
590 gettimeofday(&tv, NULL);
591 if (localtime_r(&tv.tv_sec, <ime)) {
592 strftime(date, sizeof(date), "%H:%M:%S", <ime);
593 pr_info("%s.%06ld", date, tv.tv_usec);
594 } else
595 pr_info("00:00:00.000000");
596
597 return;
598}
599
600static void print_result(struct perf_kvm_stat *kvm)
601{
602 char decode[decode_str_len];
603 struct kvm_event *event;
604 int vcpu = kvm->trace_vcpu;
605
606 if (kvm->live) {
607 puts(CONSOLE_CLEAR);
608 show_timeofday();
609 }
610
611 pr_info("\n\n");
612 print_vcpu_info(kvm);
613 pr_info("%*s ", decode_str_len, kvm->events_ops->name);
614 pr_info("%10s ", "Samples");
615 pr_info("%9s ", "Samples%");
616
617 pr_info("%9s ", "Time%");
618 pr_info("%11s ", "Min Time");
619 pr_info("%11s ", "Max Time");
620 pr_info("%16s ", "Avg time");
621 pr_info("\n\n");
622
623 while ((event = pop_from_result(&kvm->result))) {
624 u64 ecount, etime, max, min;
625
626 ecount = get_event_count(event, vcpu);
627 etime = get_event_time(event, vcpu);
628 max = get_event_max(event, vcpu);
629 min = get_event_min(event, vcpu);
630
631 kvm->events_ops->decode_key(kvm, &event->key, decode);
632 pr_info("%*s ", decode_str_len, decode);
633 pr_info("%10llu ", (unsigned long long)ecount);
634 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
635 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
636 pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
637 pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
638 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
639 kvm_event_rel_stddev(vcpu, event));
640 pr_info("\n");
641 }
642
643 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
644 kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
645
646 if (kvm->lost_events)
647 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
648}
649
650#ifdef HAVE_TIMERFD_SUPPORT
651static int process_lost_event(struct perf_tool *tool,
652 union perf_event *event __maybe_unused,
653 struct perf_sample *sample __maybe_unused,
654 struct machine *machine __maybe_unused)
655{
656 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
657
658 kvm->lost_events++;
659 return 0;
660}
661#endif
662
663static bool skip_sample(struct perf_kvm_stat *kvm,
664 struct perf_sample *sample)
665{
666 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
667 return true;
668
669 return false;
670}
671
672static int process_sample_event(struct perf_tool *tool,
673 union perf_event *event,
674 struct perf_sample *sample,
675 struct perf_evsel *evsel,
676 struct machine *machine)
677{
678 int err = 0;
679 struct thread *thread;
680 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
681 tool);
682
683 if (skip_sample(kvm, sample))
684 return 0;
685
686 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
687 if (thread == NULL) {
688 pr_debug("problem processing %d event, skipping it.\n",
689 event->header.type);
690 return -1;
691 }
692
693 if (!handle_kvm_event(kvm, thread, evsel, sample))
694 err = -1;
695
696 thread__put(thread);
697 return err;
698}
699
700static int cpu_isa_config(struct perf_kvm_stat *kvm)
701{
702 char buf[64], *cpuid;
703 int err;
704
705 if (kvm->live) {
706 err = get_cpuid(buf, sizeof(buf));
707 if (err != 0) {
708 pr_err("Failed to look up CPU type\n");
709 return err;
710 }
711 cpuid = buf;
712 } else
713 cpuid = kvm->session->header.env.cpuid;
714
715 if (!cpuid) {
716 pr_err("Failed to look up CPU type\n");
717 return -EINVAL;
718 }
719
720 err = cpu_isa_init(kvm, cpuid);
721 if (err == -ENOTSUP)
722 pr_err("CPU %s is not supported.\n", cpuid);
723
724 return err;
725}
726
727static bool verify_vcpu(int vcpu)
728{
729 if (vcpu != -1 && vcpu < 0) {
730 pr_err("Invalid vcpu:%d.\n", vcpu);
731 return false;
732 }
733
734 return true;
735}
736
737#ifdef HAVE_TIMERFD_SUPPORT
738/* keeping the max events to a modest level to keep
739 * the processing of samples per mmap smooth.
740 */
741#define PERF_KVM__MAX_EVENTS_PER_MMAP 25
742
743static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
744 u64 *mmap_time)
745{
746 struct perf_evlist *evlist = kvm->evlist;
747 union perf_event *event;
748 struct perf_mmap *md;
749 u64 timestamp;
750 s64 n = 0;
751 int err;
752
753 *mmap_time = ULLONG_MAX;
754 md = &evlist->mmap[idx];
755 err = perf_mmap__read_init(md);
756 if (err < 0)
757 return (err == -EAGAIN) ? 0 : -1;
758
759 while ((event = perf_mmap__read_event(md)) != NULL) {
760 err = perf_evlist__parse_sample_timestamp(evlist, event, ×tamp);
761 if (err) {
762 perf_mmap__consume(md);
763 pr_err("Failed to parse sample\n");
764 return -1;
765 }
766
767 err = perf_session__queue_event(kvm->session, event, timestamp, 0);
768 /*
769 * FIXME: Here we can't consume the event, as perf_session__queue_event will
770 * point to it, and it'll get possibly overwritten by the kernel.
771 */
772 perf_mmap__consume(md);
773
774 if (err) {
775 pr_err("Failed to enqueue sample: %d\n", err);
776 return -1;
777 }
778
779 /* save time stamp of our first sample for this mmap */
780 if (n == 0)
781 *mmap_time = timestamp;
782
783 /* limit events per mmap handled all at once */
784 n++;
785 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
786 break;
787 }
788
789 perf_mmap__read_done(md);
790 return n;
791}
792
793static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
794{
795 int i, err, throttled = 0;
796 s64 n, ntotal = 0;
797 u64 flush_time = ULLONG_MAX, mmap_time;
798
799 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
800 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
801 if (n < 0)
802 return -1;
803
804 /* flush time is going to be the minimum of all the individual
805 * mmap times. Essentially, we flush all the samples queued up
806 * from the last pass under our minimal start time -- that leaves
807 * a very small race for samples to come in with a lower timestamp.
808 * The ioctl to return the perf_clock timestamp should close the
809 * race entirely.
810 */
811 if (mmap_time < flush_time)
812 flush_time = mmap_time;
813
814 ntotal += n;
815 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
816 throttled = 1;
817 }
818
819 /* flush queue after each round in which we processed events */
820 if (ntotal) {
821 struct ordered_events *oe = &kvm->session->ordered_events;
822
823 oe->next_flush = flush_time;
824 err = ordered_events__flush(oe, OE_FLUSH__ROUND);
825 if (err) {
826 if (kvm->lost_events)
827 pr_info("\nLost events: %" PRIu64 "\n\n",
828 kvm->lost_events);
829 return err;
830 }
831 }
832
833 return throttled;
834}
835
836static volatile int done;
837
838static void sig_handler(int sig __maybe_unused)
839{
840 done = 1;
841}
842
843static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
844{
845 struct itimerspec new_value;
846 int rc = -1;
847
848 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
849 if (kvm->timerfd < 0) {
850 pr_err("timerfd_create failed\n");
851 goto out;
852 }
853
854 new_value.it_value.tv_sec = kvm->display_time;
855 new_value.it_value.tv_nsec = 0;
856 new_value.it_interval.tv_sec = kvm->display_time;
857 new_value.it_interval.tv_nsec = 0;
858
859 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
860 pr_err("timerfd_settime failed: %d\n", errno);
861 close(kvm->timerfd);
862 goto out;
863 }
864
865 rc = 0;
866out:
867 return rc;
868}
869
870static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
871{
872 uint64_t c;
873 int rc;
874
875 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
876 if (rc < 0) {
877 if (errno == EAGAIN)
878 return 0;
879
880 pr_err("Failed to read timer fd: %d\n", errno);
881 return -1;
882 }
883
884 if (rc != sizeof(uint64_t)) {
885 pr_err("Error reading timer fd - invalid size returned\n");
886 return -1;
887 }
888
889 if (c != 1)
890 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
891
892 /* update display */
893 sort_result(kvm);
894 print_result(kvm);
895
896 /* reset counts */
897 clear_events_cache_stats(kvm->kvm_events_cache);
898 kvm->total_count = 0;
899 kvm->total_time = 0;
900 kvm->lost_events = 0;
901
902 return 0;
903}
904
905static int fd_set_nonblock(int fd)
906{
907 long arg = 0;
908
909 arg = fcntl(fd, F_GETFL);
910 if (arg < 0) {
911 pr_err("Failed to get current flags for fd %d\n", fd);
912 return -1;
913 }
914
915 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
916 pr_err("Failed to set non-block option on fd %d\n", fd);
917 return -1;
918 }
919
920 return 0;
921}
922
923static int perf_kvm__handle_stdin(void)
924{
925 int c;
926
927 c = getc(stdin);
928 if (c == 'q')
929 return 1;
930
931 return 0;
932}
933
934static int kvm_events_live_report(struct perf_kvm_stat *kvm)
935{
936 int nr_stdin, ret, err = -EINVAL;
937 struct termios save;
938
939 /* live flag must be set first */
940 kvm->live = true;
941
942 ret = cpu_isa_config(kvm);
943 if (ret < 0)
944 return ret;
945
946 if (!verify_vcpu(kvm->trace_vcpu) ||
947 !select_key(kvm) ||
948 !register_kvm_events_ops(kvm)) {
949 goto out;
950 }
951
952 set_term_quiet_input(&save);
953 init_kvm_event_record(kvm);
954
955 signal(SIGINT, sig_handler);
956 signal(SIGTERM, sig_handler);
957
958 /* add timer fd */
959 if (perf_kvm__timerfd_create(kvm) < 0) {
960 err = -1;
961 goto out;
962 }
963
964 if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
965 goto out;
966
967 nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
968 if (nr_stdin < 0)
969 goto out;
970
971 if (fd_set_nonblock(fileno(stdin)) != 0)
972 goto out;
973
974 /* everything is good - enable the events and process */
975 perf_evlist__enable(kvm->evlist);
976
977 while (!done) {
978 struct fdarray *fda = &kvm->evlist->pollfd;
979 int rc;
980
981 rc = perf_kvm__mmap_read(kvm);
982 if (rc < 0)
983 break;
984
985 err = perf_kvm__handle_timerfd(kvm);
986 if (err)
987 goto out;
988
989 if (fda->entries[nr_stdin].revents & POLLIN)
990 done = perf_kvm__handle_stdin();
991
992 if (!rc && !done)
993 err = fdarray__poll(fda, 100);
994 }
995
996 perf_evlist__disable(kvm->evlist);
997
998 if (err == 0) {
999 sort_result(kvm);
1000 print_result(kvm);
1001 }
1002
1003out:
1004 if (kvm->timerfd >= 0)
1005 close(kvm->timerfd);
1006
1007 tcsetattr(0, TCSAFLUSH, &save);
1008 return err;
1009}
1010
1011static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1012{
1013 int err, rc = -1;
1014 struct perf_evsel *pos;
1015 struct perf_evlist *evlist = kvm->evlist;
1016 char sbuf[STRERR_BUFSIZE];
1017
1018 perf_evlist__config(evlist, &kvm->opts, NULL);
1019
1020 /*
1021 * Note: exclude_{guest,host} do not apply here.
1022 * This command processes KVM tracepoints from host only
1023 */
1024 evlist__for_each_entry(evlist, pos) {
1025 struct perf_event_attr *attr = &pos->attr;
1026
1027 /* make sure these *are* set */
1028 perf_evsel__set_sample_bit(pos, TID);
1029 perf_evsel__set_sample_bit(pos, TIME);
1030 perf_evsel__set_sample_bit(pos, CPU);
1031 perf_evsel__set_sample_bit(pos, RAW);
1032 /* make sure these are *not*; want as small a sample as possible */
1033 perf_evsel__reset_sample_bit(pos, PERIOD);
1034 perf_evsel__reset_sample_bit(pos, IP);
1035 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1036 perf_evsel__reset_sample_bit(pos, ADDR);
1037 perf_evsel__reset_sample_bit(pos, READ);
1038 attr->mmap = 0;
1039 attr->comm = 0;
1040 attr->task = 0;
1041
1042 attr->sample_period = 1;
1043
1044 attr->watermark = 0;
1045 attr->wakeup_events = 1000;
1046
1047 /* will enable all once we are ready */
1048 attr->disabled = 1;
1049 }
1050
1051 err = perf_evlist__open(evlist);
1052 if (err < 0) {
1053 printf("Couldn't create the events: %s\n",
1054 str_error_r(errno, sbuf, sizeof(sbuf)));
1055 goto out;
1056 }
1057
1058 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1059 ui__error("Failed to mmap the events: %s\n",
1060 str_error_r(errno, sbuf, sizeof(sbuf)));
1061 perf_evlist__close(evlist);
1062 goto out;
1063 }
1064
1065 rc = 0;
1066
1067out:
1068 return rc;
1069}
1070#endif
1071
1072static int read_events(struct perf_kvm_stat *kvm)
1073{
1074 int ret;
1075
1076 struct perf_tool eops = {
1077 .sample = process_sample_event,
1078 .comm = perf_event__process_comm,
1079 .namespaces = perf_event__process_namespaces,
1080 .ordered_events = true,
1081 };
1082 struct perf_data file = {
1083 .file = {
1084 .path = kvm->file_name,
1085 },
1086 .mode = PERF_DATA_MODE_READ,
1087 .force = kvm->force,
1088 };
1089
1090 kvm->tool = eops;
1091 kvm->session = perf_session__new(&file, false, &kvm->tool);
1092 if (!kvm->session) {
1093 pr_err("Initializing perf session failed\n");
1094 return -1;
1095 }
1096
1097 symbol__init(&kvm->session->header.env);
1098
1099 if (!perf_session__has_traces(kvm->session, "kvm record")) {
1100 ret = -EINVAL;
1101 goto out_delete;
1102 }
1103
1104 /*
1105 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1106 * traced in the old kernel.
1107 */
1108 ret = cpu_isa_config(kvm);
1109 if (ret < 0)
1110 goto out_delete;
1111
1112 ret = perf_session__process_events(kvm->session);
1113
1114out_delete:
1115 perf_session__delete(kvm->session);
1116 return ret;
1117}
1118
1119static int parse_target_str(struct perf_kvm_stat *kvm)
1120{
1121 if (kvm->opts.target.pid) {
1122 kvm->pid_list = intlist__new(kvm->opts.target.pid);
1123 if (kvm->pid_list == NULL) {
1124 pr_err("Error parsing process id string\n");
1125 return -EINVAL;
1126 }
1127 }
1128
1129 return 0;
1130}
1131
1132static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1133{
1134 int ret = -EINVAL;
1135 int vcpu = kvm->trace_vcpu;
1136
1137 if (parse_target_str(kvm) != 0)
1138 goto exit;
1139
1140 if (!verify_vcpu(vcpu))
1141 goto exit;
1142
1143 if (!select_key(kvm))
1144 goto exit;
1145
1146 if (!register_kvm_events_ops(kvm))
1147 goto exit;
1148
1149 init_kvm_event_record(kvm);
1150 setup_pager();
1151
1152 ret = read_events(kvm);
1153 if (ret)
1154 goto exit;
1155
1156 sort_result(kvm);
1157 print_result(kvm);
1158
1159exit:
1160 return ret;
1161}
1162
1163#define STRDUP_FAIL_EXIT(s) \
1164 ({ char *_p; \
1165 _p = strdup(s); \
1166 if (!_p) \
1167 return -ENOMEM; \
1168 _p; \
1169 })
1170
1171int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1172{
1173 return 0;
1174}
1175
1176static int
1177kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1178{
1179 unsigned int rec_argc, i, j, events_tp_size;
1180 const char **rec_argv;
1181 const char * const record_args[] = {
1182 "record",
1183 "-R",
1184 "-m", "1024",
1185 "-c", "1",
1186 };
1187 const char * const kvm_stat_record_usage[] = {
1188 "perf kvm stat record [<options>]",
1189 NULL
1190 };
1191 const char * const *events_tp;
1192 int ret;
1193
1194 events_tp_size = 0;
1195 ret = setup_kvm_events_tp(kvm);
1196 if (ret < 0) {
1197 pr_err("Unable to setup the kvm tracepoints\n");
1198 return ret;
1199 }
1200
1201 for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1202 events_tp_size++;
1203
1204 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1205 2 * events_tp_size;
1206 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1207
1208 if (rec_argv == NULL)
1209 return -ENOMEM;
1210
1211 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1212 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1213
1214 for (j = 0; j < events_tp_size; j++) {
1215 rec_argv[i++] = "-e";
1216 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1217 }
1218
1219 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1220 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1221
1222 for (j = 1; j < (unsigned int)argc; j++, i++)
1223 rec_argv[i] = argv[j];
1224
1225 set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1226 set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1227 set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1228
1229 set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1230 set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1231 set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1232 set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1233 set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1234 set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1235 set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1236 set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1237 set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1238 set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1239 set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1240 set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1241 set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1242 set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1243 set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1244
1245 record_usage = kvm_stat_record_usage;
1246 return cmd_record(i, rec_argv);
1247}
1248
1249static int
1250kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1251{
1252 const struct option kvm_events_report_options[] = {
1253 OPT_STRING(0, "event", &kvm->report_event, "report event",
1254 "event for reporting: vmexit, "
1255 "mmio (x86 only), ioport (x86 only)"),
1256 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1257 "vcpu id to report"),
1258 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1259 "key for sorting: sample(sort by samples number)"
1260 " time (sort by avg time)"),
1261 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1262 "analyze events only for given process id(s)"),
1263 OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1264 OPT_END()
1265 };
1266
1267 const char * const kvm_events_report_usage[] = {
1268 "perf kvm stat report [<options>]",
1269 NULL
1270 };
1271
1272 if (argc) {
1273 argc = parse_options(argc, argv,
1274 kvm_events_report_options,
1275 kvm_events_report_usage, 0);
1276 if (argc)
1277 usage_with_options(kvm_events_report_usage,
1278 kvm_events_report_options);
1279 }
1280
1281 if (!kvm->opts.target.pid)
1282 kvm->opts.target.system_wide = true;
1283
1284 return kvm_events_report_vcpu(kvm);
1285}
1286
1287#ifdef HAVE_TIMERFD_SUPPORT
1288static struct perf_evlist *kvm_live_event_list(void)
1289{
1290 struct perf_evlist *evlist;
1291 char *tp, *name, *sys;
1292 int err = -1;
1293 const char * const *events_tp;
1294
1295 evlist = perf_evlist__new();
1296 if (evlist == NULL)
1297 return NULL;
1298
1299 for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1300
1301 tp = strdup(*events_tp);
1302 if (tp == NULL)
1303 goto out;
1304
1305 /* split tracepoint into subsystem and name */
1306 sys = tp;
1307 name = strchr(tp, ':');
1308 if (name == NULL) {
1309 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1310 *events_tp);
1311 free(tp);
1312 goto out;
1313 }
1314 *name = '\0';
1315 name++;
1316
1317 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1318 pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1319 free(tp);
1320 goto out;
1321 }
1322
1323 free(tp);
1324 }
1325
1326 err = 0;
1327
1328out:
1329 if (err) {
1330 perf_evlist__delete(evlist);
1331 evlist = NULL;
1332 }
1333
1334 return evlist;
1335}
1336
1337static int kvm_events_live(struct perf_kvm_stat *kvm,
1338 int argc, const char **argv)
1339{
1340 char errbuf[BUFSIZ];
1341 int err;
1342
1343 const struct option live_options[] = {
1344 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1345 "record events on existing process id"),
1346 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1347 "number of mmap data pages",
1348 perf_evlist__parse_mmap_pages),
1349 OPT_INCR('v', "verbose", &verbose,
1350 "be more verbose (show counter open errors, etc)"),
1351 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1352 "system-wide collection from all CPUs"),
1353 OPT_UINTEGER('d', "display", &kvm->display_time,
1354 "time in seconds between display updates"),
1355 OPT_STRING(0, "event", &kvm->report_event, "report event",
1356 "event for reporting: "
1357 "vmexit, mmio (x86 only), ioport (x86 only)"),
1358 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1359 "vcpu id to report"),
1360 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1361 "key for sorting: sample(sort by samples number)"
1362 " time (sort by avg time)"),
1363 OPT_U64(0, "duration", &kvm->duration,
1364 "show events other than"
1365 " HLT (x86 only) or Wait state (s390 only)"
1366 " that take longer than duration usecs"),
1367 OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1368 "per thread proc mmap processing timeout in ms"),
1369 OPT_END()
1370 };
1371 const char * const live_usage[] = {
1372 "perf kvm stat live [<options>]",
1373 NULL
1374 };
1375 struct perf_data data = {
1376 .mode = PERF_DATA_MODE_WRITE,
1377 };
1378
1379
1380 /* event handling */
1381 kvm->tool.sample = process_sample_event;
1382 kvm->tool.comm = perf_event__process_comm;
1383 kvm->tool.exit = perf_event__process_exit;
1384 kvm->tool.fork = perf_event__process_fork;
1385 kvm->tool.lost = process_lost_event;
1386 kvm->tool.namespaces = perf_event__process_namespaces;
1387 kvm->tool.ordered_events = true;
1388 perf_tool__fill_defaults(&kvm->tool);
1389
1390 /* set defaults */
1391 kvm->display_time = 1;
1392 kvm->opts.user_interval = 1;
1393 kvm->opts.mmap_pages = 512;
1394 kvm->opts.target.uses_mmap = false;
1395 kvm->opts.target.uid_str = NULL;
1396 kvm->opts.target.uid = UINT_MAX;
1397 kvm->opts.proc_map_timeout = 500;
1398
1399 symbol__init(NULL);
1400 disable_buildid_cache();
1401
1402 use_browser = 0;
1403
1404 if (argc) {
1405 argc = parse_options(argc, argv, live_options,
1406 live_usage, 0);
1407 if (argc)
1408 usage_with_options(live_usage, live_options);
1409 }
1410
1411 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1412
1413 /*
1414 * target related setups
1415 */
1416 err = target__validate(&kvm->opts.target);
1417 if (err) {
1418 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1419 ui__warning("%s", errbuf);
1420 }
1421
1422 if (target__none(&kvm->opts.target))
1423 kvm->opts.target.system_wide = true;
1424
1425
1426 /*
1427 * generate the event list
1428 */
1429 err = setup_kvm_events_tp(kvm);
1430 if (err < 0) {
1431 pr_err("Unable to setup the kvm tracepoints\n");
1432 return err;
1433 }
1434
1435 kvm->evlist = kvm_live_event_list();
1436 if (kvm->evlist == NULL) {
1437 err = -1;
1438 goto out;
1439 }
1440
1441 symbol_conf.nr_events = kvm->evlist->nr_entries;
1442
1443 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1444 usage_with_options(live_usage, live_options);
1445
1446 /*
1447 * perf session
1448 */
1449 kvm->session = perf_session__new(&data, false, &kvm->tool);
1450 if (kvm->session == NULL) {
1451 err = -1;
1452 goto out;
1453 }
1454 kvm->session->evlist = kvm->evlist;
1455 perf_session__set_id_hdr_size(kvm->session);
1456 ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1457 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1458 kvm->evlist->threads, false,
1459 kvm->opts.proc_map_timeout, 1);
1460 err = kvm_live_open_events(kvm);
1461 if (err)
1462 goto out;
1463
1464 err = kvm_events_live_report(kvm);
1465
1466out:
1467 perf_session__delete(kvm->session);
1468 kvm->session = NULL;
1469 perf_evlist__delete(kvm->evlist);
1470
1471 return err;
1472}
1473#endif
1474
1475static void print_kvm_stat_usage(void)
1476{
1477 printf("Usage: perf kvm stat <command>\n\n");
1478
1479 printf("# Available commands:\n");
1480 printf("\trecord: record kvm events\n");
1481 printf("\treport: report statistical data of kvm events\n");
1482 printf("\tlive: live reporting of statistical data of kvm events\n");
1483
1484 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1485}
1486
1487static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1488{
1489 struct perf_kvm_stat kvm = {
1490 .file_name = file_name,
1491
1492 .trace_vcpu = -1,
1493 .report_event = "vmexit",
1494 .sort_key = "sample",
1495
1496 };
1497
1498 if (argc == 1) {
1499 print_kvm_stat_usage();
1500 goto perf_stat;
1501 }
1502
1503 if (!strncmp(argv[1], "rec", 3))
1504 return kvm_events_record(&kvm, argc - 1, argv + 1);
1505
1506 if (!strncmp(argv[1], "rep", 3))
1507 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1508
1509#ifdef HAVE_TIMERFD_SUPPORT
1510 if (!strncmp(argv[1], "live", 4))
1511 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1512#endif
1513
1514perf_stat:
1515 return cmd_stat(argc, argv);
1516}
1517#endif /* HAVE_KVM_STAT_SUPPORT */
1518
1519static int __cmd_record(const char *file_name, int argc, const char **argv)
1520{
1521 int rec_argc, i = 0, j;
1522 const char **rec_argv;
1523
1524 rec_argc = argc + 2;
1525 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1526 rec_argv[i++] = strdup("record");
1527 rec_argv[i++] = strdup("-o");
1528 rec_argv[i++] = strdup(file_name);
1529 for (j = 1; j < argc; j++, i++)
1530 rec_argv[i] = argv[j];
1531
1532 BUG_ON(i != rec_argc);
1533
1534 return cmd_record(i, rec_argv);
1535}
1536
1537static int __cmd_report(const char *file_name, int argc, const char **argv)
1538{
1539 int rec_argc, i = 0, j;
1540 const char **rec_argv;
1541
1542 rec_argc = argc + 2;
1543 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1544 rec_argv[i++] = strdup("report");
1545 rec_argv[i++] = strdup("-i");
1546 rec_argv[i++] = strdup(file_name);
1547 for (j = 1; j < argc; j++, i++)
1548 rec_argv[i] = argv[j];
1549
1550 BUG_ON(i != rec_argc);
1551
1552 return cmd_report(i, rec_argv);
1553}
1554
1555static int
1556__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1557{
1558 int rec_argc, i = 0, j;
1559 const char **rec_argv;
1560
1561 rec_argc = argc + 2;
1562 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1563 rec_argv[i++] = strdup("buildid-list");
1564 rec_argv[i++] = strdup("-i");
1565 rec_argv[i++] = strdup(file_name);
1566 for (j = 1; j < argc; j++, i++)
1567 rec_argv[i] = argv[j];
1568
1569 BUG_ON(i != rec_argc);
1570
1571 return cmd_buildid_list(i, rec_argv);
1572}
1573
1574int cmd_kvm(int argc, const char **argv)
1575{
1576 const char *file_name = NULL;
1577 const struct option kvm_options[] = {
1578 OPT_STRING('i', "input", &file_name, "file",
1579 "Input file name"),
1580 OPT_STRING('o', "output", &file_name, "file",
1581 "Output file name"),
1582 OPT_BOOLEAN(0, "guest", &perf_guest,
1583 "Collect guest os data"),
1584 OPT_BOOLEAN(0, "host", &perf_host,
1585 "Collect host os data"),
1586 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1587 "guest mount directory under which every guest os"
1588 " instance has a subdir"),
1589 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1590 "file", "file saving guest os vmlinux"),
1591 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1592 "file", "file saving guest os /proc/kallsyms"),
1593 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1594 "file", "file saving guest os /proc/modules"),
1595 OPT_INCR('v', "verbose", &verbose,
1596 "be more verbose (show counter open errors, etc)"),
1597 OPT_END()
1598 };
1599
1600 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1601 "buildid-list", "stat", NULL };
1602 const char *kvm_usage[] = { NULL, NULL };
1603
1604 perf_host = 0;
1605 perf_guest = 1;
1606
1607 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1608 PARSE_OPT_STOP_AT_NON_OPTION);
1609 if (!argc)
1610 usage_with_options(kvm_usage, kvm_options);
1611
1612 if (!perf_host)
1613 perf_guest = 1;
1614
1615 if (!file_name) {
1616 file_name = get_filename_for_perf_kvm();
1617
1618 if (!file_name) {
1619 pr_err("Failed to allocate memory for filename\n");
1620 return -ENOMEM;
1621 }
1622 }
1623
1624 if (!strncmp(argv[0], "rec", 3))
1625 return __cmd_record(file_name, argc, argv);
1626 else if (!strncmp(argv[0], "rep", 3))
1627 return __cmd_report(file_name, argc, argv);
1628 else if (!strncmp(argv[0], "diff", 4))
1629 return cmd_diff(argc, argv);
1630 else if (!strncmp(argv[0], "top", 3))
1631 return cmd_top(argc, argv);
1632 else if (!strncmp(argv[0], "buildid-list", 12))
1633 return __cmd_buildid_list(file_name, argc, argv);
1634#ifdef HAVE_KVM_STAT_SUPPORT
1635 else if (!strncmp(argv[0], "stat", 4))
1636 return kvm_cmd_stat(file_name, argc, argv);
1637#endif
1638 else
1639 usage_with_options(kvm_usage, kvm_options);
1640
1641 return 0;
1642}
1#include "builtin.h"
2#include "perf.h"
3
4#include "util/evsel.h"
5#include "util/evlist.h"
6#include "util/util.h"
7#include "util/cache.h"
8#include "util/symbol.h"
9#include "util/thread.h"
10#include "util/header.h"
11#include "util/session.h"
12#include "util/intlist.h"
13#include "util/parse-options.h"
14#include "util/trace-event.h"
15#include "util/debug.h"
16#include <api/fs/debugfs.h>
17#include "util/tool.h"
18#include "util/stat.h"
19#include "util/top.h"
20#include "util/data.h"
21
22#include <sys/prctl.h>
23#ifdef HAVE_TIMERFD_SUPPORT
24#include <sys/timerfd.h>
25#endif
26
27#include <termios.h>
28#include <semaphore.h>
29#include <pthread.h>
30#include <math.h>
31
32#if defined(__i386__) || defined(__x86_64__)
33#include <asm/svm.h>
34#include <asm/vmx.h>
35#include <asm/kvm.h>
36
37struct event_key {
38 #define INVALID_KEY (~0ULL)
39 u64 key;
40 int info;
41};
42
43struct kvm_event_stats {
44 u64 time;
45 struct stats stats;
46};
47
48struct kvm_event {
49 struct list_head hash_entry;
50 struct rb_node rb;
51
52 struct event_key key;
53
54 struct kvm_event_stats total;
55
56 #define DEFAULT_VCPU_NUM 8
57 int max_vcpu;
58 struct kvm_event_stats *vcpu;
59};
60
61typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
62
63struct kvm_event_key {
64 const char *name;
65 key_cmp_fun key;
66};
67
68
69struct perf_kvm_stat;
70
71struct kvm_events_ops {
72 bool (*is_begin_event)(struct perf_evsel *evsel,
73 struct perf_sample *sample,
74 struct event_key *key);
75 bool (*is_end_event)(struct perf_evsel *evsel,
76 struct perf_sample *sample, struct event_key *key);
77 void (*decode_key)(struct perf_kvm_stat *kvm, struct event_key *key,
78 char decode[20]);
79 const char *name;
80};
81
82struct exit_reasons_table {
83 unsigned long exit_code;
84 const char *reason;
85};
86
87#define EVENTS_BITS 12
88#define EVENTS_CACHE_SIZE (1UL << EVENTS_BITS)
89
90struct perf_kvm_stat {
91 struct perf_tool tool;
92 struct record_opts opts;
93 struct perf_evlist *evlist;
94 struct perf_session *session;
95
96 const char *file_name;
97 const char *report_event;
98 const char *sort_key;
99 int trace_vcpu;
100
101 struct exit_reasons_table *exit_reasons;
102 int exit_reasons_size;
103 const char *exit_reasons_isa;
104
105 struct kvm_events_ops *events_ops;
106 key_cmp_fun compare;
107 struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
108
109 u64 total_time;
110 u64 total_count;
111 u64 lost_events;
112 u64 duration;
113
114 const char *pid_str;
115 struct intlist *pid_list;
116
117 struct rb_root result;
118
119 int timerfd;
120 unsigned int display_time;
121 bool live;
122};
123
124
125static void exit_event_get_key(struct perf_evsel *evsel,
126 struct perf_sample *sample,
127 struct event_key *key)
128{
129 key->info = 0;
130 key->key = perf_evsel__intval(evsel, sample, "exit_reason");
131}
132
133static bool kvm_exit_event(struct perf_evsel *evsel)
134{
135 return !strcmp(evsel->name, "kvm:kvm_exit");
136}
137
138static bool exit_event_begin(struct perf_evsel *evsel,
139 struct perf_sample *sample, struct event_key *key)
140{
141 if (kvm_exit_event(evsel)) {
142 exit_event_get_key(evsel, sample, key);
143 return true;
144 }
145
146 return false;
147}
148
149static bool kvm_entry_event(struct perf_evsel *evsel)
150{
151 return !strcmp(evsel->name, "kvm:kvm_entry");
152}
153
154static bool exit_event_end(struct perf_evsel *evsel,
155 struct perf_sample *sample __maybe_unused,
156 struct event_key *key __maybe_unused)
157{
158 return kvm_entry_event(evsel);
159}
160
161static struct exit_reasons_table vmx_exit_reasons[] = {
162 VMX_EXIT_REASONS
163};
164
165static struct exit_reasons_table svm_exit_reasons[] = {
166 SVM_EXIT_REASONS
167};
168
169static const char *get_exit_reason(struct perf_kvm_stat *kvm, u64 exit_code)
170{
171 int i = kvm->exit_reasons_size;
172 struct exit_reasons_table *tbl = kvm->exit_reasons;
173
174 while (i--) {
175 if (tbl->exit_code == exit_code)
176 return tbl->reason;
177 tbl++;
178 }
179
180 pr_err("unknown kvm exit code:%lld on %s\n",
181 (unsigned long long)exit_code, kvm->exit_reasons_isa);
182 return "UNKNOWN";
183}
184
185static void exit_event_decode_key(struct perf_kvm_stat *kvm,
186 struct event_key *key,
187 char decode[20])
188{
189 const char *exit_reason = get_exit_reason(kvm, key->key);
190
191 scnprintf(decode, 20, "%s", exit_reason);
192}
193
194static struct kvm_events_ops exit_events = {
195 .is_begin_event = exit_event_begin,
196 .is_end_event = exit_event_end,
197 .decode_key = exit_event_decode_key,
198 .name = "VM-EXIT"
199};
200
201/*
202 * For the mmio events, we treat:
203 * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
204 * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
205 */
206static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
207 struct event_key *key)
208{
209 key->key = perf_evsel__intval(evsel, sample, "gpa");
210 key->info = perf_evsel__intval(evsel, sample, "type");
211}
212
213#define KVM_TRACE_MMIO_READ_UNSATISFIED 0
214#define KVM_TRACE_MMIO_READ 1
215#define KVM_TRACE_MMIO_WRITE 2
216
217static bool mmio_event_begin(struct perf_evsel *evsel,
218 struct perf_sample *sample, struct event_key *key)
219{
220 /* MMIO read begin event in kernel. */
221 if (kvm_exit_event(evsel))
222 return true;
223
224 /* MMIO write begin event in kernel. */
225 if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
226 perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
227 mmio_event_get_key(evsel, sample, key);
228 return true;
229 }
230
231 return false;
232}
233
234static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
235 struct event_key *key)
236{
237 /* MMIO write end event in kernel. */
238 if (kvm_entry_event(evsel))
239 return true;
240
241 /* MMIO read end event in kernel.*/
242 if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
243 perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
244 mmio_event_get_key(evsel, sample, key);
245 return true;
246 }
247
248 return false;
249}
250
251static void mmio_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
252 struct event_key *key,
253 char decode[20])
254{
255 scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
256 key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
257}
258
259static struct kvm_events_ops mmio_events = {
260 .is_begin_event = mmio_event_begin,
261 .is_end_event = mmio_event_end,
262 .decode_key = mmio_event_decode_key,
263 .name = "MMIO Access"
264};
265
266 /* The time of emulation pio access is from kvm_pio to kvm_entry. */
267static void ioport_event_get_key(struct perf_evsel *evsel,
268 struct perf_sample *sample,
269 struct event_key *key)
270{
271 key->key = perf_evsel__intval(evsel, sample, "port");
272 key->info = perf_evsel__intval(evsel, sample, "rw");
273}
274
275static bool ioport_event_begin(struct perf_evsel *evsel,
276 struct perf_sample *sample,
277 struct event_key *key)
278{
279 if (!strcmp(evsel->name, "kvm:kvm_pio")) {
280 ioport_event_get_key(evsel, sample, key);
281 return true;
282 }
283
284 return false;
285}
286
287static bool ioport_event_end(struct perf_evsel *evsel,
288 struct perf_sample *sample __maybe_unused,
289 struct event_key *key __maybe_unused)
290{
291 return kvm_entry_event(evsel);
292}
293
294static void ioport_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
295 struct event_key *key,
296 char decode[20])
297{
298 scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
299 key->info ? "POUT" : "PIN");
300}
301
302static struct kvm_events_ops ioport_events = {
303 .is_begin_event = ioport_event_begin,
304 .is_end_event = ioport_event_end,
305 .decode_key = ioport_event_decode_key,
306 .name = "IO Port Access"
307};
308
309static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
310{
311 bool ret = true;
312
313 if (!strcmp(kvm->report_event, "vmexit"))
314 kvm->events_ops = &exit_events;
315 else if (!strcmp(kvm->report_event, "mmio"))
316 kvm->events_ops = &mmio_events;
317 else if (!strcmp(kvm->report_event, "ioport"))
318 kvm->events_ops = &ioport_events;
319 else {
320 pr_err("Unknown report event:%s\n", kvm->report_event);
321 ret = false;
322 }
323
324 return ret;
325}
326
327struct vcpu_event_record {
328 int vcpu_id;
329 u64 start_time;
330 struct kvm_event *last_event;
331};
332
333
334static void init_kvm_event_record(struct perf_kvm_stat *kvm)
335{
336 unsigned int i;
337
338 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
339 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
340}
341
342#ifdef HAVE_TIMERFD_SUPPORT
343static void clear_events_cache_stats(struct list_head *kvm_events_cache)
344{
345 struct list_head *head;
346 struct kvm_event *event;
347 unsigned int i;
348 int j;
349
350 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
351 head = &kvm_events_cache[i];
352 list_for_each_entry(event, head, hash_entry) {
353 /* reset stats for event */
354 event->total.time = 0;
355 init_stats(&event->total.stats);
356
357 for (j = 0; j < event->max_vcpu; ++j) {
358 event->vcpu[j].time = 0;
359 init_stats(&event->vcpu[j].stats);
360 }
361 }
362 }
363}
364#endif
365
366static int kvm_events_hash_fn(u64 key)
367{
368 return key & (EVENTS_CACHE_SIZE - 1);
369}
370
371static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
372{
373 int old_max_vcpu = event->max_vcpu;
374 void *prev;
375
376 if (vcpu_id < event->max_vcpu)
377 return true;
378
379 while (event->max_vcpu <= vcpu_id)
380 event->max_vcpu += DEFAULT_VCPU_NUM;
381
382 prev = event->vcpu;
383 event->vcpu = realloc(event->vcpu,
384 event->max_vcpu * sizeof(*event->vcpu));
385 if (!event->vcpu) {
386 free(prev);
387 pr_err("Not enough memory\n");
388 return false;
389 }
390
391 memset(event->vcpu + old_max_vcpu, 0,
392 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
393 return true;
394}
395
396static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
397{
398 struct kvm_event *event;
399
400 event = zalloc(sizeof(*event));
401 if (!event) {
402 pr_err("Not enough memory\n");
403 return NULL;
404 }
405
406 event->key = *key;
407 init_stats(&event->total.stats);
408 return event;
409}
410
411static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
412 struct event_key *key)
413{
414 struct kvm_event *event;
415 struct list_head *head;
416
417 BUG_ON(key->key == INVALID_KEY);
418
419 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
420 list_for_each_entry(event, head, hash_entry) {
421 if (event->key.key == key->key && event->key.info == key->info)
422 return event;
423 }
424
425 event = kvm_alloc_init_event(key);
426 if (!event)
427 return NULL;
428
429 list_add(&event->hash_entry, head);
430 return event;
431}
432
433static bool handle_begin_event(struct perf_kvm_stat *kvm,
434 struct vcpu_event_record *vcpu_record,
435 struct event_key *key, u64 timestamp)
436{
437 struct kvm_event *event = NULL;
438
439 if (key->key != INVALID_KEY)
440 event = find_create_kvm_event(kvm, key);
441
442 vcpu_record->last_event = event;
443 vcpu_record->start_time = timestamp;
444 return true;
445}
446
447static void
448kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
449{
450 kvm_stats->time += time_diff;
451 update_stats(&kvm_stats->stats, time_diff);
452}
453
454static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
455{
456 struct kvm_event_stats *kvm_stats = &event->total;
457
458 if (vcpu_id != -1)
459 kvm_stats = &event->vcpu[vcpu_id];
460
461 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
462 avg_stats(&kvm_stats->stats));
463}
464
465static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
466 u64 time_diff)
467{
468 if (vcpu_id == -1) {
469 kvm_update_event_stats(&event->total, time_diff);
470 return true;
471 }
472
473 if (!kvm_event_expand(event, vcpu_id))
474 return false;
475
476 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
477 return true;
478}
479
480static bool handle_end_event(struct perf_kvm_stat *kvm,
481 struct vcpu_event_record *vcpu_record,
482 struct event_key *key,
483 struct perf_sample *sample)
484{
485 struct kvm_event *event;
486 u64 time_begin, time_diff;
487 int vcpu;
488
489 if (kvm->trace_vcpu == -1)
490 vcpu = -1;
491 else
492 vcpu = vcpu_record->vcpu_id;
493
494 event = vcpu_record->last_event;
495 time_begin = vcpu_record->start_time;
496
497 /* The begin event is not caught. */
498 if (!time_begin)
499 return true;
500
501 /*
502 * In some case, the 'begin event' only records the start timestamp,
503 * the actual event is recognized in the 'end event' (e.g. mmio-event).
504 */
505
506 /* Both begin and end events did not get the key. */
507 if (!event && key->key == INVALID_KEY)
508 return true;
509
510 if (!event)
511 event = find_create_kvm_event(kvm, key);
512
513 if (!event)
514 return false;
515
516 vcpu_record->last_event = NULL;
517 vcpu_record->start_time = 0;
518
519 /* seems to happen once in a while during live mode */
520 if (sample->time < time_begin) {
521 pr_debug("End time before begin time; skipping event.\n");
522 return true;
523 }
524
525 time_diff = sample->time - time_begin;
526
527 if (kvm->duration && time_diff > kvm->duration) {
528 char decode[32];
529
530 kvm->events_ops->decode_key(kvm, &event->key, decode);
531 if (strcmp(decode, "HLT")) {
532 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
533 sample->time, sample->pid, vcpu_record->vcpu_id,
534 decode, time_diff/1000);
535 }
536 }
537
538 return update_kvm_event(event, vcpu, time_diff);
539}
540
541static
542struct vcpu_event_record *per_vcpu_record(struct thread *thread,
543 struct perf_evsel *evsel,
544 struct perf_sample *sample)
545{
546 /* Only kvm_entry records vcpu id. */
547 if (!thread->priv && kvm_entry_event(evsel)) {
548 struct vcpu_event_record *vcpu_record;
549
550 vcpu_record = zalloc(sizeof(*vcpu_record));
551 if (!vcpu_record) {
552 pr_err("%s: Not enough memory\n", __func__);
553 return NULL;
554 }
555
556 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
557 thread->priv = vcpu_record;
558 }
559
560 return thread->priv;
561}
562
563static bool handle_kvm_event(struct perf_kvm_stat *kvm,
564 struct thread *thread,
565 struct perf_evsel *evsel,
566 struct perf_sample *sample)
567{
568 struct vcpu_event_record *vcpu_record;
569 struct event_key key = {.key = INVALID_KEY};
570
571 vcpu_record = per_vcpu_record(thread, evsel, sample);
572 if (!vcpu_record)
573 return true;
574
575 /* only process events for vcpus user cares about */
576 if ((kvm->trace_vcpu != -1) &&
577 (kvm->trace_vcpu != vcpu_record->vcpu_id))
578 return true;
579
580 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
581 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
582
583 if (kvm->events_ops->is_end_event(evsel, sample, &key))
584 return handle_end_event(kvm, vcpu_record, &key, sample);
585
586 return true;
587}
588
589#define GET_EVENT_KEY(func, field) \
590static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
591{ \
592 if (vcpu == -1) \
593 return event->total.field; \
594 \
595 if (vcpu >= event->max_vcpu) \
596 return 0; \
597 \
598 return event->vcpu[vcpu].field; \
599}
600
601#define COMPARE_EVENT_KEY(func, field) \
602GET_EVENT_KEY(func, field) \
603static int compare_kvm_event_ ## func(struct kvm_event *one, \
604 struct kvm_event *two, int vcpu)\
605{ \
606 return get_event_ ##func(one, vcpu) > \
607 get_event_ ##func(two, vcpu); \
608}
609
610GET_EVENT_KEY(time, time);
611COMPARE_EVENT_KEY(count, stats.n);
612COMPARE_EVENT_KEY(mean, stats.mean);
613GET_EVENT_KEY(max, stats.max);
614GET_EVENT_KEY(min, stats.min);
615
616#define DEF_SORT_NAME_KEY(name, compare_key) \
617 { #name, compare_kvm_event_ ## compare_key }
618
619static struct kvm_event_key keys[] = {
620 DEF_SORT_NAME_KEY(sample, count),
621 DEF_SORT_NAME_KEY(time, mean),
622 { NULL, NULL }
623};
624
625static bool select_key(struct perf_kvm_stat *kvm)
626{
627 int i;
628
629 for (i = 0; keys[i].name; i++) {
630 if (!strcmp(keys[i].name, kvm->sort_key)) {
631 kvm->compare = keys[i].key;
632 return true;
633 }
634 }
635
636 pr_err("Unknown compare key:%s\n", kvm->sort_key);
637 return false;
638}
639
640static void insert_to_result(struct rb_root *result, struct kvm_event *event,
641 key_cmp_fun bigger, int vcpu)
642{
643 struct rb_node **rb = &result->rb_node;
644 struct rb_node *parent = NULL;
645 struct kvm_event *p;
646
647 while (*rb) {
648 p = container_of(*rb, struct kvm_event, rb);
649 parent = *rb;
650
651 if (bigger(event, p, vcpu))
652 rb = &(*rb)->rb_left;
653 else
654 rb = &(*rb)->rb_right;
655 }
656
657 rb_link_node(&event->rb, parent, rb);
658 rb_insert_color(&event->rb, result);
659}
660
661static void
662update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
663{
664 int vcpu = kvm->trace_vcpu;
665
666 kvm->total_count += get_event_count(event, vcpu);
667 kvm->total_time += get_event_time(event, vcpu);
668}
669
670static bool event_is_valid(struct kvm_event *event, int vcpu)
671{
672 return !!get_event_count(event, vcpu);
673}
674
675static void sort_result(struct perf_kvm_stat *kvm)
676{
677 unsigned int i;
678 int vcpu = kvm->trace_vcpu;
679 struct kvm_event *event;
680
681 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
682 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
683 if (event_is_valid(event, vcpu)) {
684 update_total_count(kvm, event);
685 insert_to_result(&kvm->result, event,
686 kvm->compare, vcpu);
687 }
688 }
689 }
690}
691
692/* returns left most element of result, and erase it */
693static struct kvm_event *pop_from_result(struct rb_root *result)
694{
695 struct rb_node *node = rb_first(result);
696
697 if (!node)
698 return NULL;
699
700 rb_erase(node, result);
701 return container_of(node, struct kvm_event, rb);
702}
703
704static void print_vcpu_info(struct perf_kvm_stat *kvm)
705{
706 int vcpu = kvm->trace_vcpu;
707
708 pr_info("Analyze events for ");
709
710 if (kvm->live) {
711 if (kvm->opts.target.system_wide)
712 pr_info("all VMs, ");
713 else if (kvm->opts.target.pid)
714 pr_info("pid(s) %s, ", kvm->opts.target.pid);
715 else
716 pr_info("dazed and confused on what is monitored, ");
717 }
718
719 if (vcpu == -1)
720 pr_info("all VCPUs:\n\n");
721 else
722 pr_info("VCPU %d:\n\n", vcpu);
723}
724
725static void show_timeofday(void)
726{
727 char date[64];
728 struct timeval tv;
729 struct tm ltime;
730
731 gettimeofday(&tv, NULL);
732 if (localtime_r(&tv.tv_sec, <ime)) {
733 strftime(date, sizeof(date), "%H:%M:%S", <ime);
734 pr_info("%s.%06ld", date, tv.tv_usec);
735 } else
736 pr_info("00:00:00.000000");
737
738 return;
739}
740
741static void print_result(struct perf_kvm_stat *kvm)
742{
743 char decode[20];
744 struct kvm_event *event;
745 int vcpu = kvm->trace_vcpu;
746
747 if (kvm->live) {
748 puts(CONSOLE_CLEAR);
749 show_timeofday();
750 }
751
752 pr_info("\n\n");
753 print_vcpu_info(kvm);
754 pr_info("%20s ", kvm->events_ops->name);
755 pr_info("%10s ", "Samples");
756 pr_info("%9s ", "Samples%");
757
758 pr_info("%9s ", "Time%");
759 pr_info("%10s ", "Min Time");
760 pr_info("%10s ", "Max Time");
761 pr_info("%16s ", "Avg time");
762 pr_info("\n\n");
763
764 while ((event = pop_from_result(&kvm->result))) {
765 u64 ecount, etime, max, min;
766
767 ecount = get_event_count(event, vcpu);
768 etime = get_event_time(event, vcpu);
769 max = get_event_max(event, vcpu);
770 min = get_event_min(event, vcpu);
771
772 kvm->events_ops->decode_key(kvm, &event->key, decode);
773 pr_info("%20s ", decode);
774 pr_info("%10llu ", (unsigned long long)ecount);
775 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
776 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
777 pr_info("%8" PRIu64 "us ", min / 1000);
778 pr_info("%8" PRIu64 "us ", max / 1000);
779 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
780 kvm_event_rel_stddev(vcpu, event));
781 pr_info("\n");
782 }
783
784 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
785 kvm->total_count, kvm->total_time / 1e3);
786
787 if (kvm->lost_events)
788 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
789}
790
791#ifdef HAVE_TIMERFD_SUPPORT
792static int process_lost_event(struct perf_tool *tool,
793 union perf_event *event __maybe_unused,
794 struct perf_sample *sample __maybe_unused,
795 struct machine *machine __maybe_unused)
796{
797 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
798
799 kvm->lost_events++;
800 return 0;
801}
802#endif
803
804static bool skip_sample(struct perf_kvm_stat *kvm,
805 struct perf_sample *sample)
806{
807 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
808 return true;
809
810 return false;
811}
812
813static int process_sample_event(struct perf_tool *tool,
814 union perf_event *event,
815 struct perf_sample *sample,
816 struct perf_evsel *evsel,
817 struct machine *machine)
818{
819 struct thread *thread;
820 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
821 tool);
822
823 if (skip_sample(kvm, sample))
824 return 0;
825
826 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
827 if (thread == NULL) {
828 pr_debug("problem processing %d event, skipping it.\n",
829 event->header.type);
830 return -1;
831 }
832
833 if (!handle_kvm_event(kvm, thread, evsel, sample))
834 return -1;
835
836 return 0;
837}
838
839static int cpu_isa_config(struct perf_kvm_stat *kvm)
840{
841 char buf[64], *cpuid;
842 int err, isa;
843
844 if (kvm->live) {
845 err = get_cpuid(buf, sizeof(buf));
846 if (err != 0) {
847 pr_err("Failed to look up CPU type (Intel or AMD)\n");
848 return err;
849 }
850 cpuid = buf;
851 } else
852 cpuid = kvm->session->header.env.cpuid;
853
854 if (strstr(cpuid, "Intel"))
855 isa = 1;
856 else if (strstr(cpuid, "AMD"))
857 isa = 0;
858 else {
859 pr_err("CPU %s is not supported.\n", cpuid);
860 return -ENOTSUP;
861 }
862
863 if (isa == 1) {
864 kvm->exit_reasons = vmx_exit_reasons;
865 kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
866 kvm->exit_reasons_isa = "VMX";
867 }
868
869 return 0;
870}
871
872static bool verify_vcpu(int vcpu)
873{
874 if (vcpu != -1 && vcpu < 0) {
875 pr_err("Invalid vcpu:%d.\n", vcpu);
876 return false;
877 }
878
879 return true;
880}
881
882#ifdef HAVE_TIMERFD_SUPPORT
883/* keeping the max events to a modest level to keep
884 * the processing of samples per mmap smooth.
885 */
886#define PERF_KVM__MAX_EVENTS_PER_MMAP 25
887
888static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
889 u64 *mmap_time)
890{
891 union perf_event *event;
892 struct perf_sample sample;
893 s64 n = 0;
894 int err;
895
896 *mmap_time = ULLONG_MAX;
897 while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
898 err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
899 if (err) {
900 perf_evlist__mmap_consume(kvm->evlist, idx);
901 pr_err("Failed to parse sample\n");
902 return -1;
903 }
904
905 err = perf_session_queue_event(kvm->session, event, &sample, 0);
906 /*
907 * FIXME: Here we can't consume the event, as perf_session_queue_event will
908 * point to it, and it'll get possibly overwritten by the kernel.
909 */
910 perf_evlist__mmap_consume(kvm->evlist, idx);
911
912 if (err) {
913 pr_err("Failed to enqueue sample: %d\n", err);
914 return -1;
915 }
916
917 /* save time stamp of our first sample for this mmap */
918 if (n == 0)
919 *mmap_time = sample.time;
920
921 /* limit events per mmap handled all at once */
922 n++;
923 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
924 break;
925 }
926
927 return n;
928}
929
930static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
931{
932 int i, err, throttled = 0;
933 s64 n, ntotal = 0;
934 u64 flush_time = ULLONG_MAX, mmap_time;
935
936 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
937 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
938 if (n < 0)
939 return -1;
940
941 /* flush time is going to be the minimum of all the individual
942 * mmap times. Essentially, we flush all the samples queued up
943 * from the last pass under our minimal start time -- that leaves
944 * a very small race for samples to come in with a lower timestamp.
945 * The ioctl to return the perf_clock timestamp should close the
946 * race entirely.
947 */
948 if (mmap_time < flush_time)
949 flush_time = mmap_time;
950
951 ntotal += n;
952 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
953 throttled = 1;
954 }
955
956 /* flush queue after each round in which we processed events */
957 if (ntotal) {
958 kvm->session->ordered_samples.next_flush = flush_time;
959 err = kvm->tool.finished_round(&kvm->tool, NULL, kvm->session);
960 if (err) {
961 if (kvm->lost_events)
962 pr_info("\nLost events: %" PRIu64 "\n\n",
963 kvm->lost_events);
964 return err;
965 }
966 }
967
968 return throttled;
969}
970
971static volatile int done;
972
973static void sig_handler(int sig __maybe_unused)
974{
975 done = 1;
976}
977
978static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
979{
980 struct itimerspec new_value;
981 int rc = -1;
982
983 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
984 if (kvm->timerfd < 0) {
985 pr_err("timerfd_create failed\n");
986 goto out;
987 }
988
989 new_value.it_value.tv_sec = kvm->display_time;
990 new_value.it_value.tv_nsec = 0;
991 new_value.it_interval.tv_sec = kvm->display_time;
992 new_value.it_interval.tv_nsec = 0;
993
994 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
995 pr_err("timerfd_settime failed: %d\n", errno);
996 close(kvm->timerfd);
997 goto out;
998 }
999
1000 rc = 0;
1001out:
1002 return rc;
1003}
1004
1005static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
1006{
1007 uint64_t c;
1008 int rc;
1009
1010 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
1011 if (rc < 0) {
1012 if (errno == EAGAIN)
1013 return 0;
1014
1015 pr_err("Failed to read timer fd: %d\n", errno);
1016 return -1;
1017 }
1018
1019 if (rc != sizeof(uint64_t)) {
1020 pr_err("Error reading timer fd - invalid size returned\n");
1021 return -1;
1022 }
1023
1024 if (c != 1)
1025 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
1026
1027 /* update display */
1028 sort_result(kvm);
1029 print_result(kvm);
1030
1031 /* reset counts */
1032 clear_events_cache_stats(kvm->kvm_events_cache);
1033 kvm->total_count = 0;
1034 kvm->total_time = 0;
1035 kvm->lost_events = 0;
1036
1037 return 0;
1038}
1039
1040static int fd_set_nonblock(int fd)
1041{
1042 long arg = 0;
1043
1044 arg = fcntl(fd, F_GETFL);
1045 if (arg < 0) {
1046 pr_err("Failed to get current flags for fd %d\n", fd);
1047 return -1;
1048 }
1049
1050 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
1051 pr_err("Failed to set non-block option on fd %d\n", fd);
1052 return -1;
1053 }
1054
1055 return 0;
1056}
1057
1058static
1059int perf_kvm__handle_stdin(struct termios *tc_now, struct termios *tc_save)
1060{
1061 int c;
1062
1063 tcsetattr(0, TCSANOW, tc_now);
1064 c = getc(stdin);
1065 tcsetattr(0, TCSAFLUSH, tc_save);
1066
1067 if (c == 'q')
1068 return 1;
1069
1070 return 0;
1071}
1072
1073static int kvm_events_live_report(struct perf_kvm_stat *kvm)
1074{
1075 struct pollfd *pollfds = NULL;
1076 int nr_fds, nr_stdin, ret, err = -EINVAL;
1077 struct termios tc, save;
1078
1079 /* live flag must be set first */
1080 kvm->live = true;
1081
1082 ret = cpu_isa_config(kvm);
1083 if (ret < 0)
1084 return ret;
1085
1086 if (!verify_vcpu(kvm->trace_vcpu) ||
1087 !select_key(kvm) ||
1088 !register_kvm_events_ops(kvm)) {
1089 goto out;
1090 }
1091
1092 init_kvm_event_record(kvm);
1093
1094 tcgetattr(0, &save);
1095 tc = save;
1096 tc.c_lflag &= ~(ICANON | ECHO);
1097 tc.c_cc[VMIN] = 0;
1098 tc.c_cc[VTIME] = 0;
1099
1100 signal(SIGINT, sig_handler);
1101 signal(SIGTERM, sig_handler);
1102
1103 /* copy pollfds -- need to add timerfd and stdin */
1104 nr_fds = kvm->evlist->nr_fds;
1105 pollfds = zalloc(sizeof(struct pollfd) * (nr_fds + 2));
1106 if (!pollfds) {
1107 err = -ENOMEM;
1108 goto out;
1109 }
1110 memcpy(pollfds, kvm->evlist->pollfd,
1111 sizeof(struct pollfd) * kvm->evlist->nr_fds);
1112
1113 /* add timer fd */
1114 if (perf_kvm__timerfd_create(kvm) < 0) {
1115 err = -1;
1116 goto out;
1117 }
1118
1119 pollfds[nr_fds].fd = kvm->timerfd;
1120 pollfds[nr_fds].events = POLLIN;
1121 nr_fds++;
1122
1123 pollfds[nr_fds].fd = fileno(stdin);
1124 pollfds[nr_fds].events = POLLIN;
1125 nr_stdin = nr_fds;
1126 nr_fds++;
1127 if (fd_set_nonblock(fileno(stdin)) != 0)
1128 goto out;
1129
1130 /* everything is good - enable the events and process */
1131 perf_evlist__enable(kvm->evlist);
1132
1133 while (!done) {
1134 int rc;
1135
1136 rc = perf_kvm__mmap_read(kvm);
1137 if (rc < 0)
1138 break;
1139
1140 err = perf_kvm__handle_timerfd(kvm);
1141 if (err)
1142 goto out;
1143
1144 if (pollfds[nr_stdin].revents & POLLIN)
1145 done = perf_kvm__handle_stdin(&tc, &save);
1146
1147 if (!rc && !done)
1148 err = poll(pollfds, nr_fds, 100);
1149 }
1150
1151 perf_evlist__disable(kvm->evlist);
1152
1153 if (err == 0) {
1154 sort_result(kvm);
1155 print_result(kvm);
1156 }
1157
1158out:
1159 if (kvm->timerfd >= 0)
1160 close(kvm->timerfd);
1161
1162 free(pollfds);
1163 return err;
1164}
1165
1166static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1167{
1168 int err, rc = -1;
1169 struct perf_evsel *pos;
1170 struct perf_evlist *evlist = kvm->evlist;
1171
1172 perf_evlist__config(evlist, &kvm->opts);
1173
1174 /*
1175 * Note: exclude_{guest,host} do not apply here.
1176 * This command processes KVM tracepoints from host only
1177 */
1178 evlist__for_each(evlist, pos) {
1179 struct perf_event_attr *attr = &pos->attr;
1180
1181 /* make sure these *are* set */
1182 perf_evsel__set_sample_bit(pos, TID);
1183 perf_evsel__set_sample_bit(pos, TIME);
1184 perf_evsel__set_sample_bit(pos, CPU);
1185 perf_evsel__set_sample_bit(pos, RAW);
1186 /* make sure these are *not*; want as small a sample as possible */
1187 perf_evsel__reset_sample_bit(pos, PERIOD);
1188 perf_evsel__reset_sample_bit(pos, IP);
1189 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1190 perf_evsel__reset_sample_bit(pos, ADDR);
1191 perf_evsel__reset_sample_bit(pos, READ);
1192 attr->mmap = 0;
1193 attr->comm = 0;
1194 attr->task = 0;
1195
1196 attr->sample_period = 1;
1197
1198 attr->watermark = 0;
1199 attr->wakeup_events = 1000;
1200
1201 /* will enable all once we are ready */
1202 attr->disabled = 1;
1203 }
1204
1205 err = perf_evlist__open(evlist);
1206 if (err < 0) {
1207 printf("Couldn't create the events: %s\n", strerror(errno));
1208 goto out;
1209 }
1210
1211 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1212 ui__error("Failed to mmap the events: %s\n", strerror(errno));
1213 perf_evlist__close(evlist);
1214 goto out;
1215 }
1216
1217 rc = 0;
1218
1219out:
1220 return rc;
1221}
1222#endif
1223
1224static int read_events(struct perf_kvm_stat *kvm)
1225{
1226 int ret;
1227
1228 struct perf_tool eops = {
1229 .sample = process_sample_event,
1230 .comm = perf_event__process_comm,
1231 .ordered_samples = true,
1232 };
1233 struct perf_data_file file = {
1234 .path = kvm->file_name,
1235 .mode = PERF_DATA_MODE_READ,
1236 };
1237
1238 kvm->tool = eops;
1239 kvm->session = perf_session__new(&file, false, &kvm->tool);
1240 if (!kvm->session) {
1241 pr_err("Initializing perf session failed\n");
1242 return -EINVAL;
1243 }
1244
1245 if (!perf_session__has_traces(kvm->session, "kvm record"))
1246 return -EINVAL;
1247
1248 /*
1249 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1250 * traced in the old kernel.
1251 */
1252 ret = cpu_isa_config(kvm);
1253 if (ret < 0)
1254 return ret;
1255
1256 return perf_session__process_events(kvm->session, &kvm->tool);
1257}
1258
1259static int parse_target_str(struct perf_kvm_stat *kvm)
1260{
1261 if (kvm->pid_str) {
1262 kvm->pid_list = intlist__new(kvm->pid_str);
1263 if (kvm->pid_list == NULL) {
1264 pr_err("Error parsing process id string\n");
1265 return -EINVAL;
1266 }
1267 }
1268
1269 return 0;
1270}
1271
1272static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1273{
1274 int ret = -EINVAL;
1275 int vcpu = kvm->trace_vcpu;
1276
1277 if (parse_target_str(kvm) != 0)
1278 goto exit;
1279
1280 if (!verify_vcpu(vcpu))
1281 goto exit;
1282
1283 if (!select_key(kvm))
1284 goto exit;
1285
1286 if (!register_kvm_events_ops(kvm))
1287 goto exit;
1288
1289 init_kvm_event_record(kvm);
1290 setup_pager();
1291
1292 ret = read_events(kvm);
1293 if (ret)
1294 goto exit;
1295
1296 sort_result(kvm);
1297 print_result(kvm);
1298
1299exit:
1300 return ret;
1301}
1302
1303static const char * const kvm_events_tp[] = {
1304 "kvm:kvm_entry",
1305 "kvm:kvm_exit",
1306 "kvm:kvm_mmio",
1307 "kvm:kvm_pio",
1308};
1309
1310#define STRDUP_FAIL_EXIT(s) \
1311 ({ char *_p; \
1312 _p = strdup(s); \
1313 if (!_p) \
1314 return -ENOMEM; \
1315 _p; \
1316 })
1317
1318static int
1319kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1320{
1321 unsigned int rec_argc, i, j;
1322 const char **rec_argv;
1323 const char * const record_args[] = {
1324 "record",
1325 "-R",
1326 "-m", "1024",
1327 "-c", "1",
1328 };
1329
1330 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1331 2 * ARRAY_SIZE(kvm_events_tp);
1332 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1333
1334 if (rec_argv == NULL)
1335 return -ENOMEM;
1336
1337 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1338 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1339
1340 for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1341 rec_argv[i++] = "-e";
1342 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1343 }
1344
1345 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1346 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1347
1348 for (j = 1; j < (unsigned int)argc; j++, i++)
1349 rec_argv[i] = argv[j];
1350
1351 return cmd_record(i, rec_argv, NULL);
1352}
1353
1354static int
1355kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1356{
1357 const struct option kvm_events_report_options[] = {
1358 OPT_STRING(0, "event", &kvm->report_event, "report event",
1359 "event for reporting: vmexit, mmio, ioport"),
1360 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1361 "vcpu id to report"),
1362 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1363 "key for sorting: sample(sort by samples number)"
1364 " time (sort by avg time)"),
1365 OPT_STRING('p', "pid", &kvm->pid_str, "pid",
1366 "analyze events only for given process id(s)"),
1367 OPT_END()
1368 };
1369
1370 const char * const kvm_events_report_usage[] = {
1371 "perf kvm stat report [<options>]",
1372 NULL
1373 };
1374
1375 symbol__init();
1376
1377 if (argc) {
1378 argc = parse_options(argc, argv,
1379 kvm_events_report_options,
1380 kvm_events_report_usage, 0);
1381 if (argc)
1382 usage_with_options(kvm_events_report_usage,
1383 kvm_events_report_options);
1384 }
1385
1386 return kvm_events_report_vcpu(kvm);
1387}
1388
1389#ifdef HAVE_TIMERFD_SUPPORT
1390static struct perf_evlist *kvm_live_event_list(void)
1391{
1392 struct perf_evlist *evlist;
1393 char *tp, *name, *sys;
1394 unsigned int j;
1395 int err = -1;
1396
1397 evlist = perf_evlist__new();
1398 if (evlist == NULL)
1399 return NULL;
1400
1401 for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1402
1403 tp = strdup(kvm_events_tp[j]);
1404 if (tp == NULL)
1405 goto out;
1406
1407 /* split tracepoint into subsystem and name */
1408 sys = tp;
1409 name = strchr(tp, ':');
1410 if (name == NULL) {
1411 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1412 kvm_events_tp[j]);
1413 free(tp);
1414 goto out;
1415 }
1416 *name = '\0';
1417 name++;
1418
1419 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1420 pr_err("Failed to add %s tracepoint to the list\n", kvm_events_tp[j]);
1421 free(tp);
1422 goto out;
1423 }
1424
1425 free(tp);
1426 }
1427
1428 err = 0;
1429
1430out:
1431 if (err) {
1432 perf_evlist__delete(evlist);
1433 evlist = NULL;
1434 }
1435
1436 return evlist;
1437}
1438
1439static int kvm_events_live(struct perf_kvm_stat *kvm,
1440 int argc, const char **argv)
1441{
1442 char errbuf[BUFSIZ];
1443 int err;
1444
1445 const struct option live_options[] = {
1446 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1447 "record events on existing process id"),
1448 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1449 "number of mmap data pages",
1450 perf_evlist__parse_mmap_pages),
1451 OPT_INCR('v', "verbose", &verbose,
1452 "be more verbose (show counter open errors, etc)"),
1453 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1454 "system-wide collection from all CPUs"),
1455 OPT_UINTEGER('d', "display", &kvm->display_time,
1456 "time in seconds between display updates"),
1457 OPT_STRING(0, "event", &kvm->report_event, "report event",
1458 "event for reporting: vmexit, mmio, ioport"),
1459 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1460 "vcpu id to report"),
1461 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1462 "key for sorting: sample(sort by samples number)"
1463 " time (sort by avg time)"),
1464 OPT_U64(0, "duration", &kvm->duration,
1465 "show events other than HALT that take longer than duration usecs"),
1466 OPT_END()
1467 };
1468 const char * const live_usage[] = {
1469 "perf kvm stat live [<options>]",
1470 NULL
1471 };
1472 struct perf_data_file file = {
1473 .mode = PERF_DATA_MODE_WRITE,
1474 };
1475
1476
1477 /* event handling */
1478 kvm->tool.sample = process_sample_event;
1479 kvm->tool.comm = perf_event__process_comm;
1480 kvm->tool.exit = perf_event__process_exit;
1481 kvm->tool.fork = perf_event__process_fork;
1482 kvm->tool.lost = process_lost_event;
1483 kvm->tool.ordered_samples = true;
1484 perf_tool__fill_defaults(&kvm->tool);
1485
1486 /* set defaults */
1487 kvm->display_time = 1;
1488 kvm->opts.user_interval = 1;
1489 kvm->opts.mmap_pages = 512;
1490 kvm->opts.target.uses_mmap = false;
1491 kvm->opts.target.uid_str = NULL;
1492 kvm->opts.target.uid = UINT_MAX;
1493
1494 symbol__init();
1495 disable_buildid_cache();
1496
1497 use_browser = 0;
1498 setup_browser(false);
1499
1500 if (argc) {
1501 argc = parse_options(argc, argv, live_options,
1502 live_usage, 0);
1503 if (argc)
1504 usage_with_options(live_usage, live_options);
1505 }
1506
1507 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1508
1509 /*
1510 * target related setups
1511 */
1512 err = target__validate(&kvm->opts.target);
1513 if (err) {
1514 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1515 ui__warning("%s", errbuf);
1516 }
1517
1518 if (target__none(&kvm->opts.target))
1519 kvm->opts.target.system_wide = true;
1520
1521
1522 /*
1523 * generate the event list
1524 */
1525 kvm->evlist = kvm_live_event_list();
1526 if (kvm->evlist == NULL) {
1527 err = -1;
1528 goto out;
1529 }
1530
1531 symbol_conf.nr_events = kvm->evlist->nr_entries;
1532
1533 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1534 usage_with_options(live_usage, live_options);
1535
1536 /*
1537 * perf session
1538 */
1539 kvm->session = perf_session__new(&file, false, &kvm->tool);
1540 if (kvm->session == NULL) {
1541 err = -ENOMEM;
1542 goto out;
1543 }
1544 kvm->session->evlist = kvm->evlist;
1545 perf_session__set_id_hdr_size(kvm->session);
1546 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1547 kvm->evlist->threads, false);
1548 err = kvm_live_open_events(kvm);
1549 if (err)
1550 goto out;
1551
1552 err = kvm_events_live_report(kvm);
1553
1554out:
1555 exit_browser(0);
1556
1557 if (kvm->session)
1558 perf_session__delete(kvm->session);
1559 kvm->session = NULL;
1560 if (kvm->evlist)
1561 perf_evlist__delete(kvm->evlist);
1562
1563 return err;
1564}
1565#endif
1566
1567static void print_kvm_stat_usage(void)
1568{
1569 printf("Usage: perf kvm stat <command>\n\n");
1570
1571 printf("# Available commands:\n");
1572 printf("\trecord: record kvm events\n");
1573 printf("\treport: report statistical data of kvm events\n");
1574 printf("\tlive: live reporting of statistical data of kvm events\n");
1575
1576 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1577}
1578
1579static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1580{
1581 struct perf_kvm_stat kvm = {
1582 .file_name = file_name,
1583
1584 .trace_vcpu = -1,
1585 .report_event = "vmexit",
1586 .sort_key = "sample",
1587
1588 .exit_reasons = svm_exit_reasons,
1589 .exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
1590 .exit_reasons_isa = "SVM",
1591 };
1592
1593 if (argc == 1) {
1594 print_kvm_stat_usage();
1595 goto perf_stat;
1596 }
1597
1598 if (!strncmp(argv[1], "rec", 3))
1599 return kvm_events_record(&kvm, argc - 1, argv + 1);
1600
1601 if (!strncmp(argv[1], "rep", 3))
1602 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1603
1604#ifdef HAVE_TIMERFD_SUPPORT
1605 if (!strncmp(argv[1], "live", 4))
1606 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1607#endif
1608
1609perf_stat:
1610 return cmd_stat(argc, argv, NULL);
1611}
1612#endif
1613
1614static int __cmd_record(const char *file_name, int argc, const char **argv)
1615{
1616 int rec_argc, i = 0, j;
1617 const char **rec_argv;
1618
1619 rec_argc = argc + 2;
1620 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1621 rec_argv[i++] = strdup("record");
1622 rec_argv[i++] = strdup("-o");
1623 rec_argv[i++] = strdup(file_name);
1624 for (j = 1; j < argc; j++, i++)
1625 rec_argv[i] = argv[j];
1626
1627 BUG_ON(i != rec_argc);
1628
1629 return cmd_record(i, rec_argv, NULL);
1630}
1631
1632static int __cmd_report(const char *file_name, int argc, const char **argv)
1633{
1634 int rec_argc, i = 0, j;
1635 const char **rec_argv;
1636
1637 rec_argc = argc + 2;
1638 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1639 rec_argv[i++] = strdup("report");
1640 rec_argv[i++] = strdup("-i");
1641 rec_argv[i++] = strdup(file_name);
1642 for (j = 1; j < argc; j++, i++)
1643 rec_argv[i] = argv[j];
1644
1645 BUG_ON(i != rec_argc);
1646
1647 return cmd_report(i, rec_argv, NULL);
1648}
1649
1650static int
1651__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1652{
1653 int rec_argc, i = 0, j;
1654 const char **rec_argv;
1655
1656 rec_argc = argc + 2;
1657 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1658 rec_argv[i++] = strdup("buildid-list");
1659 rec_argv[i++] = strdup("-i");
1660 rec_argv[i++] = strdup(file_name);
1661 for (j = 1; j < argc; j++, i++)
1662 rec_argv[i] = argv[j];
1663
1664 BUG_ON(i != rec_argc);
1665
1666 return cmd_buildid_list(i, rec_argv, NULL);
1667}
1668
1669int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1670{
1671 const char *file_name = NULL;
1672 const struct option kvm_options[] = {
1673 OPT_STRING('i', "input", &file_name, "file",
1674 "Input file name"),
1675 OPT_STRING('o', "output", &file_name, "file",
1676 "Output file name"),
1677 OPT_BOOLEAN(0, "guest", &perf_guest,
1678 "Collect guest os data"),
1679 OPT_BOOLEAN(0, "host", &perf_host,
1680 "Collect host os data"),
1681 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1682 "guest mount directory under which every guest os"
1683 " instance has a subdir"),
1684 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1685 "file", "file saving guest os vmlinux"),
1686 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1687 "file", "file saving guest os /proc/kallsyms"),
1688 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1689 "file", "file saving guest os /proc/modules"),
1690 OPT_INCR('v', "verbose", &verbose,
1691 "be more verbose (show counter open errors, etc)"),
1692 OPT_END()
1693 };
1694
1695 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1696 "buildid-list", "stat", NULL };
1697 const char *kvm_usage[] = { NULL, NULL };
1698
1699 perf_host = 0;
1700 perf_guest = 1;
1701
1702 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1703 PARSE_OPT_STOP_AT_NON_OPTION);
1704 if (!argc)
1705 usage_with_options(kvm_usage, kvm_options);
1706
1707 if (!perf_host)
1708 perf_guest = 1;
1709
1710 if (!file_name) {
1711 file_name = get_filename_for_perf_kvm();
1712
1713 if (!file_name) {
1714 pr_err("Failed to allocate memory for filename\n");
1715 return -ENOMEM;
1716 }
1717 }
1718
1719 if (!strncmp(argv[0], "rec", 3))
1720 return __cmd_record(file_name, argc, argv);
1721 else if (!strncmp(argv[0], "rep", 3))
1722 return __cmd_report(file_name, argc, argv);
1723 else if (!strncmp(argv[0], "diff", 4))
1724 return cmd_diff(argc, argv, NULL);
1725 else if (!strncmp(argv[0], "top", 3))
1726 return cmd_top(argc, argv, NULL);
1727 else if (!strncmp(argv[0], "buildid-list", 12))
1728 return __cmd_buildid_list(file_name, argc, argv);
1729#if defined(__i386__) || defined(__x86_64__)
1730 else if (!strncmp(argv[0], "stat", 4))
1731 return kvm_cmd_stat(file_name, argc, argv);
1732#endif
1733 else
1734 usage_with_options(kvm_usage, kvm_options);
1735
1736 return 0;
1737}