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// 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}