Linux 4.1.18
[linux/fpc-iii.git] / tools / perf / builtin-kvm.c
blob1f9338f6109cdbe79f8f08e510c00e4a171d1f82
1 #include "builtin.h"
2 #include "perf.h"
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 #include "util/ordered-events.h"
23 #include <sys/prctl.h>
24 #ifdef HAVE_TIMERFD_SUPPORT
25 #include <sys/timerfd.h>
26 #endif
28 #include <termios.h>
29 #include <semaphore.h>
30 #include <pthread.h>
31 #include <math.h>
33 #ifdef HAVE_KVM_STAT_SUPPORT
34 #include <asm/kvm_perf.h>
35 #include "util/kvm-stat.h"
37 void exit_event_get_key(struct perf_evsel *evsel,
38 struct perf_sample *sample,
39 struct event_key *key)
41 key->info = 0;
42 key->key = perf_evsel__intval(evsel, sample, KVM_EXIT_REASON);
45 bool kvm_exit_event(struct perf_evsel *evsel)
47 return !strcmp(evsel->name, KVM_EXIT_TRACE);
50 bool exit_event_begin(struct perf_evsel *evsel,
51 struct perf_sample *sample, struct event_key *key)
53 if (kvm_exit_event(evsel)) {
54 exit_event_get_key(evsel, sample, key);
55 return true;
58 return false;
61 bool kvm_entry_event(struct perf_evsel *evsel)
63 return !strcmp(evsel->name, KVM_ENTRY_TRACE);
66 bool exit_event_end(struct perf_evsel *evsel,
67 struct perf_sample *sample __maybe_unused,
68 struct event_key *key __maybe_unused)
70 return kvm_entry_event(evsel);
73 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
74 struct exit_reasons_table *tbl,
75 u64 exit_code)
77 while (tbl->reason != NULL) {
78 if (tbl->exit_code == exit_code)
79 return tbl->reason;
80 tbl++;
83 pr_err("unknown kvm exit code:%lld on %s\n",
84 (unsigned long long)exit_code, kvm->exit_reasons_isa);
85 return "UNKNOWN";
88 void exit_event_decode_key(struct perf_kvm_stat *kvm,
89 struct event_key *key,
90 char *decode)
92 const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
93 key->key);
95 scnprintf(decode, DECODE_STR_LEN, "%s", exit_reason);
98 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
100 struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
102 for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
103 if (!strcmp(events_ops->name, kvm->report_event)) {
104 kvm->events_ops = events_ops->ops;
105 return true;
109 return false;
112 struct vcpu_event_record {
113 int vcpu_id;
114 u64 start_time;
115 struct kvm_event *last_event;
119 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
121 unsigned int i;
123 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
124 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
127 #ifdef HAVE_TIMERFD_SUPPORT
128 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
130 struct list_head *head;
131 struct kvm_event *event;
132 unsigned int i;
133 int j;
135 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
136 head = &kvm_events_cache[i];
137 list_for_each_entry(event, head, hash_entry) {
138 /* reset stats for event */
139 event->total.time = 0;
140 init_stats(&event->total.stats);
142 for (j = 0; j < event->max_vcpu; ++j) {
143 event->vcpu[j].time = 0;
144 init_stats(&event->vcpu[j].stats);
149 #endif
151 static int kvm_events_hash_fn(u64 key)
153 return key & (EVENTS_CACHE_SIZE - 1);
156 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
158 int old_max_vcpu = event->max_vcpu;
159 void *prev;
161 if (vcpu_id < event->max_vcpu)
162 return true;
164 while (event->max_vcpu <= vcpu_id)
165 event->max_vcpu += DEFAULT_VCPU_NUM;
167 prev = event->vcpu;
168 event->vcpu = realloc(event->vcpu,
169 event->max_vcpu * sizeof(*event->vcpu));
170 if (!event->vcpu) {
171 free(prev);
172 pr_err("Not enough memory\n");
173 return false;
176 memset(event->vcpu + old_max_vcpu, 0,
177 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
178 return true;
181 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
183 struct kvm_event *event;
185 event = zalloc(sizeof(*event));
186 if (!event) {
187 pr_err("Not enough memory\n");
188 return NULL;
191 event->key = *key;
192 init_stats(&event->total.stats);
193 return event;
196 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
197 struct event_key *key)
199 struct kvm_event *event;
200 struct list_head *head;
202 BUG_ON(key->key == INVALID_KEY);
204 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
205 list_for_each_entry(event, head, hash_entry) {
206 if (event->key.key == key->key && event->key.info == key->info)
207 return event;
210 event = kvm_alloc_init_event(key);
211 if (!event)
212 return NULL;
214 list_add(&event->hash_entry, head);
215 return event;
218 static bool handle_begin_event(struct perf_kvm_stat *kvm,
219 struct vcpu_event_record *vcpu_record,
220 struct event_key *key, u64 timestamp)
222 struct kvm_event *event = NULL;
224 if (key->key != INVALID_KEY)
225 event = find_create_kvm_event(kvm, key);
227 vcpu_record->last_event = event;
228 vcpu_record->start_time = timestamp;
229 return true;
232 static void
233 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
235 kvm_stats->time += time_diff;
236 update_stats(&kvm_stats->stats, time_diff);
239 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
241 struct kvm_event_stats *kvm_stats = &event->total;
243 if (vcpu_id != -1)
244 kvm_stats = &event->vcpu[vcpu_id];
246 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
247 avg_stats(&kvm_stats->stats));
250 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
251 u64 time_diff)
253 if (vcpu_id == -1) {
254 kvm_update_event_stats(&event->total, time_diff);
255 return true;
258 if (!kvm_event_expand(event, vcpu_id))
259 return false;
261 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
262 return true;
265 static bool is_child_event(struct perf_kvm_stat *kvm,
266 struct perf_evsel *evsel,
267 struct perf_sample *sample,
268 struct event_key *key)
270 struct child_event_ops *child_ops;
272 child_ops = kvm->events_ops->child_ops;
274 if (!child_ops)
275 return false;
277 for (; child_ops->name; child_ops++) {
278 if (!strcmp(evsel->name, child_ops->name)) {
279 child_ops->get_key(evsel, sample, key);
280 return true;
284 return false;
287 static bool handle_child_event(struct perf_kvm_stat *kvm,
288 struct vcpu_event_record *vcpu_record,
289 struct event_key *key,
290 struct perf_sample *sample __maybe_unused)
292 struct kvm_event *event = NULL;
294 if (key->key != INVALID_KEY)
295 event = find_create_kvm_event(kvm, key);
297 vcpu_record->last_event = event;
299 return true;
302 static bool skip_event(const char *event)
304 const char * const *skip_events;
306 for (skip_events = kvm_skip_events; *skip_events; skip_events++)
307 if (!strcmp(event, *skip_events))
308 return true;
310 return false;
313 static bool handle_end_event(struct perf_kvm_stat *kvm,
314 struct vcpu_event_record *vcpu_record,
315 struct event_key *key,
316 struct perf_sample *sample)
318 struct kvm_event *event;
319 u64 time_begin, time_diff;
320 int vcpu;
322 if (kvm->trace_vcpu == -1)
323 vcpu = -1;
324 else
325 vcpu = vcpu_record->vcpu_id;
327 event = vcpu_record->last_event;
328 time_begin = vcpu_record->start_time;
330 /* The begin event is not caught. */
331 if (!time_begin)
332 return true;
335 * In some case, the 'begin event' only records the start timestamp,
336 * the actual event is recognized in the 'end event' (e.g. mmio-event).
339 /* Both begin and end events did not get the key. */
340 if (!event && key->key == INVALID_KEY)
341 return true;
343 if (!event)
344 event = find_create_kvm_event(kvm, key);
346 if (!event)
347 return false;
349 vcpu_record->last_event = NULL;
350 vcpu_record->start_time = 0;
352 /* seems to happen once in a while during live mode */
353 if (sample->time < time_begin) {
354 pr_debug("End time before begin time; skipping event.\n");
355 return true;
358 time_diff = sample->time - time_begin;
360 if (kvm->duration && time_diff > kvm->duration) {
361 char decode[DECODE_STR_LEN];
363 kvm->events_ops->decode_key(kvm, &event->key, decode);
364 if (!skip_event(decode)) {
365 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
366 sample->time, sample->pid, vcpu_record->vcpu_id,
367 decode, time_diff/1000);
371 return update_kvm_event(event, vcpu, time_diff);
374 static
375 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
376 struct perf_evsel *evsel,
377 struct perf_sample *sample)
379 /* Only kvm_entry records vcpu id. */
380 if (!thread__priv(thread) && kvm_entry_event(evsel)) {
381 struct vcpu_event_record *vcpu_record;
383 vcpu_record = zalloc(sizeof(*vcpu_record));
384 if (!vcpu_record) {
385 pr_err("%s: Not enough memory\n", __func__);
386 return NULL;
389 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, VCPU_ID);
390 thread__set_priv(thread, vcpu_record);
393 return thread__priv(thread);
396 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
397 struct thread *thread,
398 struct perf_evsel *evsel,
399 struct perf_sample *sample)
401 struct vcpu_event_record *vcpu_record;
402 struct event_key key = { .key = INVALID_KEY,
403 .exit_reasons = kvm->exit_reasons };
405 vcpu_record = per_vcpu_record(thread, evsel, sample);
406 if (!vcpu_record)
407 return true;
409 /* only process events for vcpus user cares about */
410 if ((kvm->trace_vcpu != -1) &&
411 (kvm->trace_vcpu != vcpu_record->vcpu_id))
412 return true;
414 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
415 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
417 if (is_child_event(kvm, evsel, sample, &key))
418 return handle_child_event(kvm, vcpu_record, &key, sample);
420 if (kvm->events_ops->is_end_event(evsel, sample, &key))
421 return handle_end_event(kvm, vcpu_record, &key, sample);
423 return true;
426 #define GET_EVENT_KEY(func, field) \
427 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
429 if (vcpu == -1) \
430 return event->total.field; \
432 if (vcpu >= event->max_vcpu) \
433 return 0; \
435 return event->vcpu[vcpu].field; \
438 #define COMPARE_EVENT_KEY(func, field) \
439 GET_EVENT_KEY(func, field) \
440 static int compare_kvm_event_ ## func(struct kvm_event *one, \
441 struct kvm_event *two, int vcpu)\
443 return get_event_ ##func(one, vcpu) > \
444 get_event_ ##func(two, vcpu); \
447 GET_EVENT_KEY(time, time);
448 COMPARE_EVENT_KEY(count, stats.n);
449 COMPARE_EVENT_KEY(mean, stats.mean);
450 GET_EVENT_KEY(max, stats.max);
451 GET_EVENT_KEY(min, stats.min);
453 #define DEF_SORT_NAME_KEY(name, compare_key) \
454 { #name, compare_kvm_event_ ## compare_key }
456 static struct kvm_event_key keys[] = {
457 DEF_SORT_NAME_KEY(sample, count),
458 DEF_SORT_NAME_KEY(time, mean),
459 { NULL, NULL }
462 static bool select_key(struct perf_kvm_stat *kvm)
464 int i;
466 for (i = 0; keys[i].name; i++) {
467 if (!strcmp(keys[i].name, kvm->sort_key)) {
468 kvm->compare = keys[i].key;
469 return true;
473 pr_err("Unknown compare key:%s\n", kvm->sort_key);
474 return false;
477 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
478 key_cmp_fun bigger, int vcpu)
480 struct rb_node **rb = &result->rb_node;
481 struct rb_node *parent = NULL;
482 struct kvm_event *p;
484 while (*rb) {
485 p = container_of(*rb, struct kvm_event, rb);
486 parent = *rb;
488 if (bigger(event, p, vcpu))
489 rb = &(*rb)->rb_left;
490 else
491 rb = &(*rb)->rb_right;
494 rb_link_node(&event->rb, parent, rb);
495 rb_insert_color(&event->rb, result);
498 static void
499 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
501 int vcpu = kvm->trace_vcpu;
503 kvm->total_count += get_event_count(event, vcpu);
504 kvm->total_time += get_event_time(event, vcpu);
507 static bool event_is_valid(struct kvm_event *event, int vcpu)
509 return !!get_event_count(event, vcpu);
512 static void sort_result(struct perf_kvm_stat *kvm)
514 unsigned int i;
515 int vcpu = kvm->trace_vcpu;
516 struct kvm_event *event;
518 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
519 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
520 if (event_is_valid(event, vcpu)) {
521 update_total_count(kvm, event);
522 insert_to_result(&kvm->result, event,
523 kvm->compare, vcpu);
529 /* returns left most element of result, and erase it */
530 static struct kvm_event *pop_from_result(struct rb_root *result)
532 struct rb_node *node = rb_first(result);
534 if (!node)
535 return NULL;
537 rb_erase(node, result);
538 return container_of(node, struct kvm_event, rb);
541 static void print_vcpu_info(struct perf_kvm_stat *kvm)
543 int vcpu = kvm->trace_vcpu;
545 pr_info("Analyze events for ");
547 if (kvm->opts.target.system_wide)
548 pr_info("all VMs, ");
549 else if (kvm->opts.target.pid)
550 pr_info("pid(s) %s, ", kvm->opts.target.pid);
551 else
552 pr_info("dazed and confused on what is monitored, ");
554 if (vcpu == -1)
555 pr_info("all VCPUs:\n\n");
556 else
557 pr_info("VCPU %d:\n\n", vcpu);
560 static void show_timeofday(void)
562 char date[64];
563 struct timeval tv;
564 struct tm ltime;
566 gettimeofday(&tv, NULL);
567 if (localtime_r(&tv.tv_sec, &ltime)) {
568 strftime(date, sizeof(date), "%H:%M:%S", &ltime);
569 pr_info("%s.%06ld", date, tv.tv_usec);
570 } else
571 pr_info("00:00:00.000000");
573 return;
576 static void print_result(struct perf_kvm_stat *kvm)
578 char decode[DECODE_STR_LEN];
579 struct kvm_event *event;
580 int vcpu = kvm->trace_vcpu;
582 if (kvm->live) {
583 puts(CONSOLE_CLEAR);
584 show_timeofday();
587 pr_info("\n\n");
588 print_vcpu_info(kvm);
589 pr_info("%*s ", DECODE_STR_LEN, kvm->events_ops->name);
590 pr_info("%10s ", "Samples");
591 pr_info("%9s ", "Samples%");
593 pr_info("%9s ", "Time%");
594 pr_info("%11s ", "Min Time");
595 pr_info("%11s ", "Max Time");
596 pr_info("%16s ", "Avg time");
597 pr_info("\n\n");
599 while ((event = pop_from_result(&kvm->result))) {
600 u64 ecount, etime, max, min;
602 ecount = get_event_count(event, vcpu);
603 etime = get_event_time(event, vcpu);
604 max = get_event_max(event, vcpu);
605 min = get_event_min(event, vcpu);
607 kvm->events_ops->decode_key(kvm, &event->key, decode);
608 pr_info("%*s ", DECODE_STR_LEN, decode);
609 pr_info("%10llu ", (unsigned long long)ecount);
610 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
611 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
612 pr_info("%9.2fus ", (double)min / 1e3);
613 pr_info("%9.2fus ", (double)max / 1e3);
614 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
615 kvm_event_rel_stddev(vcpu, event));
616 pr_info("\n");
619 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
620 kvm->total_count, kvm->total_time / 1e3);
622 if (kvm->lost_events)
623 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
626 #ifdef HAVE_TIMERFD_SUPPORT
627 static int process_lost_event(struct perf_tool *tool,
628 union perf_event *event __maybe_unused,
629 struct perf_sample *sample __maybe_unused,
630 struct machine *machine __maybe_unused)
632 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
634 kvm->lost_events++;
635 return 0;
637 #endif
639 static bool skip_sample(struct perf_kvm_stat *kvm,
640 struct perf_sample *sample)
642 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
643 return true;
645 return false;
648 static int process_sample_event(struct perf_tool *tool,
649 union perf_event *event,
650 struct perf_sample *sample,
651 struct perf_evsel *evsel,
652 struct machine *machine)
654 struct thread *thread;
655 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
656 tool);
658 if (skip_sample(kvm, sample))
659 return 0;
661 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
662 if (thread == NULL) {
663 pr_debug("problem processing %d event, skipping it.\n",
664 event->header.type);
665 return -1;
668 if (!handle_kvm_event(kvm, thread, evsel, sample))
669 return -1;
671 return 0;
674 static int cpu_isa_config(struct perf_kvm_stat *kvm)
676 char buf[64], *cpuid;
677 int err;
679 if (kvm->live) {
680 err = get_cpuid(buf, sizeof(buf));
681 if (err != 0) {
682 pr_err("Failed to look up CPU type\n");
683 return err;
685 cpuid = buf;
686 } else
687 cpuid = kvm->session->header.env.cpuid;
689 if (!cpuid) {
690 pr_err("Failed to look up CPU type\n");
691 return -EINVAL;
694 err = cpu_isa_init(kvm, cpuid);
695 if (err == -ENOTSUP)
696 pr_err("CPU %s is not supported.\n", cpuid);
698 return err;
701 static bool verify_vcpu(int vcpu)
703 if (vcpu != -1 && vcpu < 0) {
704 pr_err("Invalid vcpu:%d.\n", vcpu);
705 return false;
708 return true;
711 #ifdef HAVE_TIMERFD_SUPPORT
712 /* keeping the max events to a modest level to keep
713 * the processing of samples per mmap smooth.
715 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
717 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
718 u64 *mmap_time)
720 union perf_event *event;
721 struct perf_sample sample;
722 s64 n = 0;
723 int err;
725 *mmap_time = ULLONG_MAX;
726 while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
727 err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
728 if (err) {
729 perf_evlist__mmap_consume(kvm->evlist, idx);
730 pr_err("Failed to parse sample\n");
731 return -1;
734 err = perf_session__queue_event(kvm->session, event, &sample, 0);
736 * FIXME: Here we can't consume the event, as perf_session__queue_event will
737 * point to it, and it'll get possibly overwritten by the kernel.
739 perf_evlist__mmap_consume(kvm->evlist, idx);
741 if (err) {
742 pr_err("Failed to enqueue sample: %d\n", err);
743 return -1;
746 /* save time stamp of our first sample for this mmap */
747 if (n == 0)
748 *mmap_time = sample.time;
750 /* limit events per mmap handled all at once */
751 n++;
752 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
753 break;
756 return n;
759 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
761 int i, err, throttled = 0;
762 s64 n, ntotal = 0;
763 u64 flush_time = ULLONG_MAX, mmap_time;
765 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
766 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
767 if (n < 0)
768 return -1;
770 /* flush time is going to be the minimum of all the individual
771 * mmap times. Essentially, we flush all the samples queued up
772 * from the last pass under our minimal start time -- that leaves
773 * a very small race for samples to come in with a lower timestamp.
774 * The ioctl to return the perf_clock timestamp should close the
775 * race entirely.
777 if (mmap_time < flush_time)
778 flush_time = mmap_time;
780 ntotal += n;
781 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
782 throttled = 1;
785 /* flush queue after each round in which we processed events */
786 if (ntotal) {
787 struct ordered_events *oe = &kvm->session->ordered_events;
789 oe->next_flush = flush_time;
790 err = ordered_events__flush(oe, OE_FLUSH__ROUND);
791 if (err) {
792 if (kvm->lost_events)
793 pr_info("\nLost events: %" PRIu64 "\n\n",
794 kvm->lost_events);
795 return err;
799 return throttled;
802 static volatile int done;
804 static void sig_handler(int sig __maybe_unused)
806 done = 1;
809 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
811 struct itimerspec new_value;
812 int rc = -1;
814 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
815 if (kvm->timerfd < 0) {
816 pr_err("timerfd_create failed\n");
817 goto out;
820 new_value.it_value.tv_sec = kvm->display_time;
821 new_value.it_value.tv_nsec = 0;
822 new_value.it_interval.tv_sec = kvm->display_time;
823 new_value.it_interval.tv_nsec = 0;
825 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
826 pr_err("timerfd_settime failed: %d\n", errno);
827 close(kvm->timerfd);
828 goto out;
831 rc = 0;
832 out:
833 return rc;
836 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
838 uint64_t c;
839 int rc;
841 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
842 if (rc < 0) {
843 if (errno == EAGAIN)
844 return 0;
846 pr_err("Failed to read timer fd: %d\n", errno);
847 return -1;
850 if (rc != sizeof(uint64_t)) {
851 pr_err("Error reading timer fd - invalid size returned\n");
852 return -1;
855 if (c != 1)
856 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
858 /* update display */
859 sort_result(kvm);
860 print_result(kvm);
862 /* reset counts */
863 clear_events_cache_stats(kvm->kvm_events_cache);
864 kvm->total_count = 0;
865 kvm->total_time = 0;
866 kvm->lost_events = 0;
868 return 0;
871 static int fd_set_nonblock(int fd)
873 long arg = 0;
875 arg = fcntl(fd, F_GETFL);
876 if (arg < 0) {
877 pr_err("Failed to get current flags for fd %d\n", fd);
878 return -1;
881 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
882 pr_err("Failed to set non-block option on fd %d\n", fd);
883 return -1;
886 return 0;
889 static int perf_kvm__handle_stdin(void)
891 int c;
893 c = getc(stdin);
894 if (c == 'q')
895 return 1;
897 return 0;
900 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
902 int nr_stdin, ret, err = -EINVAL;
903 struct termios save;
905 /* live flag must be set first */
906 kvm->live = true;
908 ret = cpu_isa_config(kvm);
909 if (ret < 0)
910 return ret;
912 if (!verify_vcpu(kvm->trace_vcpu) ||
913 !select_key(kvm) ||
914 !register_kvm_events_ops(kvm)) {
915 goto out;
918 set_term_quiet_input(&save);
919 init_kvm_event_record(kvm);
921 signal(SIGINT, sig_handler);
922 signal(SIGTERM, sig_handler);
924 /* add timer fd */
925 if (perf_kvm__timerfd_create(kvm) < 0) {
926 err = -1;
927 goto out;
930 if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
931 goto out;
933 nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
934 if (nr_stdin < 0)
935 goto out;
937 if (fd_set_nonblock(fileno(stdin)) != 0)
938 goto out;
940 /* everything is good - enable the events and process */
941 perf_evlist__enable(kvm->evlist);
943 while (!done) {
944 struct fdarray *fda = &kvm->evlist->pollfd;
945 int rc;
947 rc = perf_kvm__mmap_read(kvm);
948 if (rc < 0)
949 break;
951 err = perf_kvm__handle_timerfd(kvm);
952 if (err)
953 goto out;
955 if (fda->entries[nr_stdin].revents & POLLIN)
956 done = perf_kvm__handle_stdin();
958 if (!rc && !done)
959 err = fdarray__poll(fda, 100);
962 perf_evlist__disable(kvm->evlist);
964 if (err == 0) {
965 sort_result(kvm);
966 print_result(kvm);
969 out:
970 if (kvm->timerfd >= 0)
971 close(kvm->timerfd);
973 tcsetattr(0, TCSAFLUSH, &save);
974 return err;
977 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
979 int err, rc = -1;
980 struct perf_evsel *pos;
981 struct perf_evlist *evlist = kvm->evlist;
982 char sbuf[STRERR_BUFSIZE];
984 perf_evlist__config(evlist, &kvm->opts);
987 * Note: exclude_{guest,host} do not apply here.
988 * This command processes KVM tracepoints from host only
990 evlist__for_each(evlist, pos) {
991 struct perf_event_attr *attr = &pos->attr;
993 /* make sure these *are* set */
994 perf_evsel__set_sample_bit(pos, TID);
995 perf_evsel__set_sample_bit(pos, TIME);
996 perf_evsel__set_sample_bit(pos, CPU);
997 perf_evsel__set_sample_bit(pos, RAW);
998 /* make sure these are *not*; want as small a sample as possible */
999 perf_evsel__reset_sample_bit(pos, PERIOD);
1000 perf_evsel__reset_sample_bit(pos, IP);
1001 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1002 perf_evsel__reset_sample_bit(pos, ADDR);
1003 perf_evsel__reset_sample_bit(pos, READ);
1004 attr->mmap = 0;
1005 attr->comm = 0;
1006 attr->task = 0;
1008 attr->sample_period = 1;
1010 attr->watermark = 0;
1011 attr->wakeup_events = 1000;
1013 /* will enable all once we are ready */
1014 attr->disabled = 1;
1017 err = perf_evlist__open(evlist);
1018 if (err < 0) {
1019 printf("Couldn't create the events: %s\n",
1020 strerror_r(errno, sbuf, sizeof(sbuf)));
1021 goto out;
1024 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1025 ui__error("Failed to mmap the events: %s\n",
1026 strerror_r(errno, sbuf, sizeof(sbuf)));
1027 perf_evlist__close(evlist);
1028 goto out;
1031 rc = 0;
1033 out:
1034 return rc;
1036 #endif
1038 static int read_events(struct perf_kvm_stat *kvm)
1040 int ret;
1042 struct perf_tool eops = {
1043 .sample = process_sample_event,
1044 .comm = perf_event__process_comm,
1045 .ordered_events = true,
1047 struct perf_data_file file = {
1048 .path = kvm->file_name,
1049 .mode = PERF_DATA_MODE_READ,
1050 .force = kvm->force,
1053 kvm->tool = eops;
1054 kvm->session = perf_session__new(&file, false, &kvm->tool);
1055 if (!kvm->session) {
1056 pr_err("Initializing perf session failed\n");
1057 return -1;
1060 symbol__init(&kvm->session->header.env);
1062 if (!perf_session__has_traces(kvm->session, "kvm record"))
1063 return -EINVAL;
1066 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1067 * traced in the old kernel.
1069 ret = cpu_isa_config(kvm);
1070 if (ret < 0)
1071 return ret;
1073 return perf_session__process_events(kvm->session);
1076 static int parse_target_str(struct perf_kvm_stat *kvm)
1078 if (kvm->opts.target.pid) {
1079 kvm->pid_list = intlist__new(kvm->opts.target.pid);
1080 if (kvm->pid_list == NULL) {
1081 pr_err("Error parsing process id string\n");
1082 return -EINVAL;
1086 return 0;
1089 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1091 int ret = -EINVAL;
1092 int vcpu = kvm->trace_vcpu;
1094 if (parse_target_str(kvm) != 0)
1095 goto exit;
1097 if (!verify_vcpu(vcpu))
1098 goto exit;
1100 if (!select_key(kvm))
1101 goto exit;
1103 if (!register_kvm_events_ops(kvm))
1104 goto exit;
1106 init_kvm_event_record(kvm);
1107 setup_pager();
1109 ret = read_events(kvm);
1110 if (ret)
1111 goto exit;
1113 sort_result(kvm);
1114 print_result(kvm);
1116 exit:
1117 return ret;
1120 #define STRDUP_FAIL_EXIT(s) \
1121 ({ char *_p; \
1122 _p = strdup(s); \
1123 if (!_p) \
1124 return -ENOMEM; \
1125 _p; \
1128 static int
1129 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1131 unsigned int rec_argc, i, j, events_tp_size;
1132 const char **rec_argv;
1133 const char * const record_args[] = {
1134 "record",
1135 "-R",
1136 "-m", "1024",
1137 "-c", "1",
1139 const char * const kvm_stat_record_usage[] = {
1140 "perf kvm stat record [<options>]",
1141 NULL
1143 const char * const *events_tp;
1144 events_tp_size = 0;
1146 for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1147 events_tp_size++;
1149 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1150 2 * events_tp_size;
1151 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1153 if (rec_argv == NULL)
1154 return -ENOMEM;
1156 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1157 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1159 for (j = 0; j < events_tp_size; j++) {
1160 rec_argv[i++] = "-e";
1161 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1164 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1165 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1167 for (j = 1; j < (unsigned int)argc; j++, i++)
1168 rec_argv[i] = argv[j];
1170 set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1171 set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1172 set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1174 set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1175 set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1176 set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1177 set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1178 set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1179 set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1180 set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1181 set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1182 set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1183 set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1184 set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1185 set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1186 set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1187 set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1188 set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1190 record_usage = kvm_stat_record_usage;
1191 return cmd_record(i, rec_argv, NULL);
1194 static int
1195 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1197 const struct option kvm_events_report_options[] = {
1198 OPT_STRING(0, "event", &kvm->report_event, "report event",
1199 "event for reporting: vmexit, "
1200 "mmio (x86 only), ioport (x86 only)"),
1201 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1202 "vcpu id to report"),
1203 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1204 "key for sorting: sample(sort by samples number)"
1205 " time (sort by avg time)"),
1206 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1207 "analyze events only for given process id(s)"),
1208 OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1209 OPT_END()
1212 const char * const kvm_events_report_usage[] = {
1213 "perf kvm stat report [<options>]",
1214 NULL
1217 if (argc) {
1218 argc = parse_options(argc, argv,
1219 kvm_events_report_options,
1220 kvm_events_report_usage, 0);
1221 if (argc)
1222 usage_with_options(kvm_events_report_usage,
1223 kvm_events_report_options);
1226 if (!kvm->opts.target.pid)
1227 kvm->opts.target.system_wide = true;
1229 return kvm_events_report_vcpu(kvm);
1232 #ifdef HAVE_TIMERFD_SUPPORT
1233 static struct perf_evlist *kvm_live_event_list(void)
1235 struct perf_evlist *evlist;
1236 char *tp, *name, *sys;
1237 int err = -1;
1238 const char * const *events_tp;
1240 evlist = perf_evlist__new();
1241 if (evlist == NULL)
1242 return NULL;
1244 for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1246 tp = strdup(*events_tp);
1247 if (tp == NULL)
1248 goto out;
1250 /* split tracepoint into subsystem and name */
1251 sys = tp;
1252 name = strchr(tp, ':');
1253 if (name == NULL) {
1254 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1255 *events_tp);
1256 free(tp);
1257 goto out;
1259 *name = '\0';
1260 name++;
1262 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1263 pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1264 free(tp);
1265 goto out;
1268 free(tp);
1271 err = 0;
1273 out:
1274 if (err) {
1275 perf_evlist__delete(evlist);
1276 evlist = NULL;
1279 return evlist;
1282 static int kvm_events_live(struct perf_kvm_stat *kvm,
1283 int argc, const char **argv)
1285 char errbuf[BUFSIZ];
1286 int err;
1288 const struct option live_options[] = {
1289 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1290 "record events on existing process id"),
1291 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1292 "number of mmap data pages",
1293 perf_evlist__parse_mmap_pages),
1294 OPT_INCR('v', "verbose", &verbose,
1295 "be more verbose (show counter open errors, etc)"),
1296 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1297 "system-wide collection from all CPUs"),
1298 OPT_UINTEGER('d', "display", &kvm->display_time,
1299 "time in seconds between display updates"),
1300 OPT_STRING(0, "event", &kvm->report_event, "report event",
1301 "event for reporting: "
1302 "vmexit, mmio (x86 only), ioport (x86 only)"),
1303 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1304 "vcpu id to report"),
1305 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1306 "key for sorting: sample(sort by samples number)"
1307 " time (sort by avg time)"),
1308 OPT_U64(0, "duration", &kvm->duration,
1309 "show events other than"
1310 " HLT (x86 only) or Wait state (s390 only)"
1311 " that take longer than duration usecs"),
1312 OPT_END()
1314 const char * const live_usage[] = {
1315 "perf kvm stat live [<options>]",
1316 NULL
1318 struct perf_data_file file = {
1319 .mode = PERF_DATA_MODE_WRITE,
1323 /* event handling */
1324 kvm->tool.sample = process_sample_event;
1325 kvm->tool.comm = perf_event__process_comm;
1326 kvm->tool.exit = perf_event__process_exit;
1327 kvm->tool.fork = perf_event__process_fork;
1328 kvm->tool.lost = process_lost_event;
1329 kvm->tool.ordered_events = true;
1330 perf_tool__fill_defaults(&kvm->tool);
1332 /* set defaults */
1333 kvm->display_time = 1;
1334 kvm->opts.user_interval = 1;
1335 kvm->opts.mmap_pages = 512;
1336 kvm->opts.target.uses_mmap = false;
1337 kvm->opts.target.uid_str = NULL;
1338 kvm->opts.target.uid = UINT_MAX;
1340 symbol__init(NULL);
1341 disable_buildid_cache();
1343 use_browser = 0;
1344 setup_browser(false);
1346 if (argc) {
1347 argc = parse_options(argc, argv, live_options,
1348 live_usage, 0);
1349 if (argc)
1350 usage_with_options(live_usage, live_options);
1353 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1356 * target related setups
1358 err = target__validate(&kvm->opts.target);
1359 if (err) {
1360 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1361 ui__warning("%s", errbuf);
1364 if (target__none(&kvm->opts.target))
1365 kvm->opts.target.system_wide = true;
1369 * generate the event list
1371 kvm->evlist = kvm_live_event_list();
1372 if (kvm->evlist == NULL) {
1373 err = -1;
1374 goto out;
1377 symbol_conf.nr_events = kvm->evlist->nr_entries;
1379 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1380 usage_with_options(live_usage, live_options);
1383 * perf session
1385 kvm->session = perf_session__new(&file, false, &kvm->tool);
1386 if (kvm->session == NULL) {
1387 err = -1;
1388 goto out;
1390 kvm->session->evlist = kvm->evlist;
1391 perf_session__set_id_hdr_size(kvm->session);
1392 ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1393 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1394 kvm->evlist->threads, false);
1395 err = kvm_live_open_events(kvm);
1396 if (err)
1397 goto out;
1399 err = kvm_events_live_report(kvm);
1401 out:
1402 exit_browser(0);
1404 if (kvm->session)
1405 perf_session__delete(kvm->session);
1406 kvm->session = NULL;
1407 if (kvm->evlist)
1408 perf_evlist__delete(kvm->evlist);
1410 return err;
1412 #endif
1414 static void print_kvm_stat_usage(void)
1416 printf("Usage: perf kvm stat <command>\n\n");
1418 printf("# Available commands:\n");
1419 printf("\trecord: record kvm events\n");
1420 printf("\treport: report statistical data of kvm events\n");
1421 printf("\tlive: live reporting of statistical data of kvm events\n");
1423 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1426 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1428 struct perf_kvm_stat kvm = {
1429 .file_name = file_name,
1431 .trace_vcpu = -1,
1432 .report_event = "vmexit",
1433 .sort_key = "sample",
1437 if (argc == 1) {
1438 print_kvm_stat_usage();
1439 goto perf_stat;
1442 if (!strncmp(argv[1], "rec", 3))
1443 return kvm_events_record(&kvm, argc - 1, argv + 1);
1445 if (!strncmp(argv[1], "rep", 3))
1446 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1448 #ifdef HAVE_TIMERFD_SUPPORT
1449 if (!strncmp(argv[1], "live", 4))
1450 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1451 #endif
1453 perf_stat:
1454 return cmd_stat(argc, argv, NULL);
1456 #endif /* HAVE_KVM_STAT_SUPPORT */
1458 static int __cmd_record(const char *file_name, int argc, const char **argv)
1460 int rec_argc, i = 0, j;
1461 const char **rec_argv;
1463 rec_argc = argc + 2;
1464 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1465 rec_argv[i++] = strdup("record");
1466 rec_argv[i++] = strdup("-o");
1467 rec_argv[i++] = strdup(file_name);
1468 for (j = 1; j < argc; j++, i++)
1469 rec_argv[i] = argv[j];
1471 BUG_ON(i != rec_argc);
1473 return cmd_record(i, rec_argv, NULL);
1476 static int __cmd_report(const char *file_name, int argc, const char **argv)
1478 int rec_argc, i = 0, j;
1479 const char **rec_argv;
1481 rec_argc = argc + 2;
1482 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1483 rec_argv[i++] = strdup("report");
1484 rec_argv[i++] = strdup("-i");
1485 rec_argv[i++] = strdup(file_name);
1486 for (j = 1; j < argc; j++, i++)
1487 rec_argv[i] = argv[j];
1489 BUG_ON(i != rec_argc);
1491 return cmd_report(i, rec_argv, NULL);
1494 static int
1495 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1497 int rec_argc, i = 0, j;
1498 const char **rec_argv;
1500 rec_argc = argc + 2;
1501 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1502 rec_argv[i++] = strdup("buildid-list");
1503 rec_argv[i++] = strdup("-i");
1504 rec_argv[i++] = strdup(file_name);
1505 for (j = 1; j < argc; j++, i++)
1506 rec_argv[i] = argv[j];
1508 BUG_ON(i != rec_argc);
1510 return cmd_buildid_list(i, rec_argv, NULL);
1513 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1515 const char *file_name = NULL;
1516 const struct option kvm_options[] = {
1517 OPT_STRING('i', "input", &file_name, "file",
1518 "Input file name"),
1519 OPT_STRING('o', "output", &file_name, "file",
1520 "Output file name"),
1521 OPT_BOOLEAN(0, "guest", &perf_guest,
1522 "Collect guest os data"),
1523 OPT_BOOLEAN(0, "host", &perf_host,
1524 "Collect host os data"),
1525 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1526 "guest mount directory under which every guest os"
1527 " instance has a subdir"),
1528 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1529 "file", "file saving guest os vmlinux"),
1530 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1531 "file", "file saving guest os /proc/kallsyms"),
1532 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1533 "file", "file saving guest os /proc/modules"),
1534 OPT_INCR('v', "verbose", &verbose,
1535 "be more verbose (show counter open errors, etc)"),
1536 OPT_END()
1539 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1540 "buildid-list", "stat", NULL };
1541 const char *kvm_usage[] = { NULL, NULL };
1543 perf_host = 0;
1544 perf_guest = 1;
1546 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1547 PARSE_OPT_STOP_AT_NON_OPTION);
1548 if (!argc)
1549 usage_with_options(kvm_usage, kvm_options);
1551 if (!perf_host)
1552 perf_guest = 1;
1554 if (!file_name) {
1555 file_name = get_filename_for_perf_kvm();
1557 if (!file_name) {
1558 pr_err("Failed to allocate memory for filename\n");
1559 return -ENOMEM;
1563 if (!strncmp(argv[0], "rec", 3))
1564 return __cmd_record(file_name, argc, argv);
1565 else if (!strncmp(argv[0], "rep", 3))
1566 return __cmd_report(file_name, argc, argv);
1567 else if (!strncmp(argv[0], "diff", 4))
1568 return cmd_diff(argc, argv, NULL);
1569 else if (!strncmp(argv[0], "top", 3))
1570 return cmd_top(argc, argv, NULL);
1571 else if (!strncmp(argv[0], "buildid-list", 12))
1572 return __cmd_buildid_list(file_name, argc, argv);
1573 #ifdef HAVE_KVM_STAT_SUPPORT
1574 else if (!strncmp(argv[0], "stat", 4))
1575 return kvm_cmd_stat(file_name, argc, argv);
1576 #endif
1577 else
1578 usage_with_options(kvm_usage, kvm_options);
1580 return 0;