4 #include "util/evsel.h"
5 #include "util/evlist.h"
8 #include "util/cache.h"
9 #include "util/symbol.h"
10 #include "util/thread.h"
11 #include "util/header.h"
12 #include "util/session.h"
13 #include "util/intlist.h"
14 #include <subcmd/parse-options.h>
15 #include "util/trace-event.h"
16 #include "util/debug.h"
17 #include "util/tool.h"
18 #include "util/stat.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>
29 #include <linux/kernel.h>
30 #include <linux/time64.h>
35 #include <semaphore.h>
39 static const char *get_filename_for_perf_kvm(void)
43 if (perf_host && !perf_guest)
44 filename = strdup("perf.data.host");
45 else if (!perf_host && perf_guest)
46 filename = strdup("perf.data.guest");
48 filename = strdup("perf.data.kvm");
53 #ifdef HAVE_KVM_STAT_SUPPORT
54 #include "util/kvm-stat.h"
56 void exit_event_get_key(struct perf_evsel *evsel,
57 struct perf_sample *sample,
58 struct event_key *key)
61 key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
64 bool kvm_exit_event(struct perf_evsel *evsel)
66 return !strcmp(evsel->name, kvm_exit_trace);
69 bool exit_event_begin(struct perf_evsel *evsel,
70 struct perf_sample *sample, struct event_key *key)
72 if (kvm_exit_event(evsel)) {
73 exit_event_get_key(evsel, sample, key);
80 bool kvm_entry_event(struct perf_evsel *evsel)
82 return !strcmp(evsel->name, kvm_entry_trace);
85 bool exit_event_end(struct perf_evsel *evsel,
86 struct perf_sample *sample __maybe_unused,
87 struct event_key *key __maybe_unused)
89 return kvm_entry_event(evsel);
92 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
93 struct exit_reasons_table *tbl,
96 while (tbl->reason != NULL) {
97 if (tbl->exit_code == exit_code)
102 pr_err("unknown kvm exit code:%lld on %s\n",
103 (unsigned long long)exit_code, kvm->exit_reasons_isa);
107 void exit_event_decode_key(struct perf_kvm_stat *kvm,
108 struct event_key *key,
111 const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
114 scnprintf(decode, decode_str_len, "%s", exit_reason);
117 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
119 struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
121 for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
122 if (!strcmp(events_ops->name, kvm->report_event)) {
123 kvm->events_ops = events_ops->ops;
131 struct vcpu_event_record {
134 struct kvm_event *last_event;
138 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
142 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
143 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
146 #ifdef HAVE_TIMERFD_SUPPORT
147 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
149 struct list_head *head;
150 struct kvm_event *event;
154 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
155 head = &kvm_events_cache[i];
156 list_for_each_entry(event, head, hash_entry) {
157 /* reset stats for event */
158 event->total.time = 0;
159 init_stats(&event->total.stats);
161 for (j = 0; j < event->max_vcpu; ++j) {
162 event->vcpu[j].time = 0;
163 init_stats(&event->vcpu[j].stats);
170 static int kvm_events_hash_fn(u64 key)
172 return key & (EVENTS_CACHE_SIZE - 1);
175 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
177 int old_max_vcpu = event->max_vcpu;
180 if (vcpu_id < event->max_vcpu)
183 while (event->max_vcpu <= vcpu_id)
184 event->max_vcpu += DEFAULT_VCPU_NUM;
187 event->vcpu = realloc(event->vcpu,
188 event->max_vcpu * sizeof(*event->vcpu));
191 pr_err("Not enough memory\n");
195 memset(event->vcpu + old_max_vcpu, 0,
196 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
200 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
202 struct kvm_event *event;
204 event = zalloc(sizeof(*event));
206 pr_err("Not enough memory\n");
211 init_stats(&event->total.stats);
215 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
216 struct event_key *key)
218 struct kvm_event *event;
219 struct list_head *head;
221 BUG_ON(key->key == INVALID_KEY);
223 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
224 list_for_each_entry(event, head, hash_entry) {
225 if (event->key.key == key->key && event->key.info == key->info)
229 event = kvm_alloc_init_event(key);
233 list_add(&event->hash_entry, head);
237 static bool handle_begin_event(struct perf_kvm_stat *kvm,
238 struct vcpu_event_record *vcpu_record,
239 struct event_key *key, u64 timestamp)
241 struct kvm_event *event = NULL;
243 if (key->key != INVALID_KEY)
244 event = find_create_kvm_event(kvm, key);
246 vcpu_record->last_event = event;
247 vcpu_record->start_time = timestamp;
252 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
254 kvm_stats->time += time_diff;
255 update_stats(&kvm_stats->stats, time_diff);
258 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
260 struct kvm_event_stats *kvm_stats = &event->total;
263 kvm_stats = &event->vcpu[vcpu_id];
265 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
266 avg_stats(&kvm_stats->stats));
269 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
273 kvm_update_event_stats(&event->total, time_diff);
277 if (!kvm_event_expand(event, vcpu_id))
280 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
284 static bool is_child_event(struct perf_kvm_stat *kvm,
285 struct perf_evsel *evsel,
286 struct perf_sample *sample,
287 struct event_key *key)
289 struct child_event_ops *child_ops;
291 child_ops = kvm->events_ops->child_ops;
296 for (; child_ops->name; child_ops++) {
297 if (!strcmp(evsel->name, child_ops->name)) {
298 child_ops->get_key(evsel, sample, key);
306 static bool handle_child_event(struct perf_kvm_stat *kvm,
307 struct vcpu_event_record *vcpu_record,
308 struct event_key *key,
309 struct perf_sample *sample __maybe_unused)
311 struct kvm_event *event = NULL;
313 if (key->key != INVALID_KEY)
314 event = find_create_kvm_event(kvm, key);
316 vcpu_record->last_event = event;
321 static bool skip_event(const char *event)
323 const char * const *skip_events;
325 for (skip_events = kvm_skip_events; *skip_events; skip_events++)
326 if (!strcmp(event, *skip_events))
332 static bool handle_end_event(struct perf_kvm_stat *kvm,
333 struct vcpu_event_record *vcpu_record,
334 struct event_key *key,
335 struct perf_sample *sample)
337 struct kvm_event *event;
338 u64 time_begin, time_diff;
341 if (kvm->trace_vcpu == -1)
344 vcpu = vcpu_record->vcpu_id;
346 event = vcpu_record->last_event;
347 time_begin = vcpu_record->start_time;
349 /* The begin event is not caught. */
354 * In some case, the 'begin event' only records the start timestamp,
355 * the actual event is recognized in the 'end event' (e.g. mmio-event).
358 /* Both begin and end events did not get the key. */
359 if (!event && key->key == INVALID_KEY)
363 event = find_create_kvm_event(kvm, key);
368 vcpu_record->last_event = NULL;
369 vcpu_record->start_time = 0;
371 /* seems to happen once in a while during live mode */
372 if (sample->time < time_begin) {
373 pr_debug("End time before begin time; skipping event.\n");
377 time_diff = sample->time - time_begin;
379 if (kvm->duration && time_diff > kvm->duration) {
380 char decode[decode_str_len];
382 kvm->events_ops->decode_key(kvm, &event->key, decode);
383 if (!skip_event(decode)) {
384 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
385 sample->time, sample->pid, vcpu_record->vcpu_id,
386 decode, time_diff / NSEC_PER_USEC);
390 return update_kvm_event(event, vcpu, time_diff);
394 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
395 struct perf_evsel *evsel,
396 struct perf_sample *sample)
398 /* Only kvm_entry records vcpu id. */
399 if (!thread__priv(thread) && kvm_entry_event(evsel)) {
400 struct vcpu_event_record *vcpu_record;
402 vcpu_record = zalloc(sizeof(*vcpu_record));
404 pr_err("%s: Not enough memory\n", __func__);
408 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
410 thread__set_priv(thread, vcpu_record);
413 return thread__priv(thread);
416 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
417 struct thread *thread,
418 struct perf_evsel *evsel,
419 struct perf_sample *sample)
421 struct vcpu_event_record *vcpu_record;
422 struct event_key key = { .key = INVALID_KEY,
423 .exit_reasons = kvm->exit_reasons };
425 vcpu_record = per_vcpu_record(thread, evsel, sample);
429 /* only process events for vcpus user cares about */
430 if ((kvm->trace_vcpu != -1) &&
431 (kvm->trace_vcpu != vcpu_record->vcpu_id))
434 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
435 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
437 if (is_child_event(kvm, evsel, sample, &key))
438 return handle_child_event(kvm, vcpu_record, &key, sample);
440 if (kvm->events_ops->is_end_event(evsel, sample, &key))
441 return handle_end_event(kvm, vcpu_record, &key, sample);
446 #define GET_EVENT_KEY(func, field) \
447 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
450 return event->total.field; \
452 if (vcpu >= event->max_vcpu) \
455 return event->vcpu[vcpu].field; \
458 #define COMPARE_EVENT_KEY(func, field) \
459 GET_EVENT_KEY(func, field) \
460 static int compare_kvm_event_ ## func(struct kvm_event *one, \
461 struct kvm_event *two, int vcpu)\
463 return get_event_ ##func(one, vcpu) > \
464 get_event_ ##func(two, vcpu); \
467 GET_EVENT_KEY(time, time);
468 COMPARE_EVENT_KEY(count, stats.n);
469 COMPARE_EVENT_KEY(mean, stats.mean);
470 GET_EVENT_KEY(max, stats.max);
471 GET_EVENT_KEY(min, stats.min);
473 #define DEF_SORT_NAME_KEY(name, compare_key) \
474 { #name, compare_kvm_event_ ## compare_key }
476 static struct kvm_event_key keys[] = {
477 DEF_SORT_NAME_KEY(sample, count),
478 DEF_SORT_NAME_KEY(time, mean),
482 static bool select_key(struct perf_kvm_stat *kvm)
486 for (i = 0; keys[i].name; i++) {
487 if (!strcmp(keys[i].name, kvm->sort_key)) {
488 kvm->compare = keys[i].key;
493 pr_err("Unknown compare key:%s\n", kvm->sort_key);
497 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
498 key_cmp_fun bigger, int vcpu)
500 struct rb_node **rb = &result->rb_node;
501 struct rb_node *parent = NULL;
505 p = container_of(*rb, struct kvm_event, rb);
508 if (bigger(event, p, vcpu))
509 rb = &(*rb)->rb_left;
511 rb = &(*rb)->rb_right;
514 rb_link_node(&event->rb, parent, rb);
515 rb_insert_color(&event->rb, result);
519 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
521 int vcpu = kvm->trace_vcpu;
523 kvm->total_count += get_event_count(event, vcpu);
524 kvm->total_time += get_event_time(event, vcpu);
527 static bool event_is_valid(struct kvm_event *event, int vcpu)
529 return !!get_event_count(event, vcpu);
532 static void sort_result(struct perf_kvm_stat *kvm)
535 int vcpu = kvm->trace_vcpu;
536 struct kvm_event *event;
538 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
539 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
540 if (event_is_valid(event, vcpu)) {
541 update_total_count(kvm, event);
542 insert_to_result(&kvm->result, event,
549 /* returns left most element of result, and erase it */
550 static struct kvm_event *pop_from_result(struct rb_root *result)
552 struct rb_node *node = rb_first(result);
557 rb_erase(node, result);
558 return container_of(node, struct kvm_event, rb);
561 static void print_vcpu_info(struct perf_kvm_stat *kvm)
563 int vcpu = kvm->trace_vcpu;
565 pr_info("Analyze events for ");
567 if (kvm->opts.target.system_wide)
568 pr_info("all VMs, ");
569 else if (kvm->opts.target.pid)
570 pr_info("pid(s) %s, ", kvm->opts.target.pid);
572 pr_info("dazed and confused on what is monitored, ");
575 pr_info("all VCPUs:\n\n");
577 pr_info("VCPU %d:\n\n", vcpu);
580 static void show_timeofday(void)
586 gettimeofday(&tv, NULL);
587 if (localtime_r(&tv.tv_sec, <ime)) {
588 strftime(date, sizeof(date), "%H:%M:%S", <ime);
589 pr_info("%s.%06ld", date, tv.tv_usec);
591 pr_info("00:00:00.000000");
596 static void print_result(struct perf_kvm_stat *kvm)
598 char decode[decode_str_len];
599 struct kvm_event *event;
600 int vcpu = kvm->trace_vcpu;
608 print_vcpu_info(kvm);
609 pr_info("%*s ", decode_str_len, kvm->events_ops->name);
610 pr_info("%10s ", "Samples");
611 pr_info("%9s ", "Samples%");
613 pr_info("%9s ", "Time%");
614 pr_info("%11s ", "Min Time");
615 pr_info("%11s ", "Max Time");
616 pr_info("%16s ", "Avg time");
619 while ((event = pop_from_result(&kvm->result))) {
620 u64 ecount, etime, max, min;
622 ecount = get_event_count(event, vcpu);
623 etime = get_event_time(event, vcpu);
624 max = get_event_max(event, vcpu);
625 min = get_event_min(event, vcpu);
627 kvm->events_ops->decode_key(kvm, &event->key, decode);
628 pr_info("%*s ", decode_str_len, decode);
629 pr_info("%10llu ", (unsigned long long)ecount);
630 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
631 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
632 pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
633 pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
634 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
635 kvm_event_rel_stddev(vcpu, event));
639 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
640 kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
642 if (kvm->lost_events)
643 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
646 #ifdef HAVE_TIMERFD_SUPPORT
647 static int process_lost_event(struct perf_tool *tool,
648 union perf_event *event __maybe_unused,
649 struct perf_sample *sample __maybe_unused,
650 struct machine *machine __maybe_unused)
652 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
659 static bool skip_sample(struct perf_kvm_stat *kvm,
660 struct perf_sample *sample)
662 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
668 static int process_sample_event(struct perf_tool *tool,
669 union perf_event *event,
670 struct perf_sample *sample,
671 struct perf_evsel *evsel,
672 struct machine *machine)
675 struct thread *thread;
676 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
679 if (skip_sample(kvm, sample))
682 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
683 if (thread == NULL) {
684 pr_debug("problem processing %d event, skipping it.\n",
689 if (!handle_kvm_event(kvm, thread, evsel, sample))
696 static int cpu_isa_config(struct perf_kvm_stat *kvm)
698 char buf[64], *cpuid;
702 err = get_cpuid(buf, sizeof(buf));
704 pr_err("Failed to look up CPU type\n");
709 cpuid = kvm->session->header.env.cpuid;
712 pr_err("Failed to look up CPU type\n");
716 err = cpu_isa_init(kvm, cpuid);
718 pr_err("CPU %s is not supported.\n", cpuid);
723 static bool verify_vcpu(int vcpu)
725 if (vcpu != -1 && vcpu < 0) {
726 pr_err("Invalid vcpu:%d.\n", vcpu);
733 #ifdef HAVE_TIMERFD_SUPPORT
734 /* keeping the max events to a modest level to keep
735 * the processing of samples per mmap smooth.
737 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
739 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
742 union perf_event *event;
743 struct perf_sample sample;
747 *mmap_time = ULLONG_MAX;
748 while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
749 err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
751 perf_evlist__mmap_consume(kvm->evlist, idx);
752 pr_err("Failed to parse sample\n");
756 err = perf_session__queue_event(kvm->session, event, &sample, 0);
758 * FIXME: Here we can't consume the event, as perf_session__queue_event will
759 * point to it, and it'll get possibly overwritten by the kernel.
761 perf_evlist__mmap_consume(kvm->evlist, idx);
764 pr_err("Failed to enqueue sample: %d\n", err);
768 /* save time stamp of our first sample for this mmap */
770 *mmap_time = sample.time;
772 /* limit events per mmap handled all at once */
774 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
781 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
783 int i, err, throttled = 0;
785 u64 flush_time = ULLONG_MAX, mmap_time;
787 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
788 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
792 /* flush time is going to be the minimum of all the individual
793 * mmap times. Essentially, we flush all the samples queued up
794 * from the last pass under our minimal start time -- that leaves
795 * a very small race for samples to come in with a lower timestamp.
796 * The ioctl to return the perf_clock timestamp should close the
799 if (mmap_time < flush_time)
800 flush_time = mmap_time;
803 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
807 /* flush queue after each round in which we processed events */
809 struct ordered_events *oe = &kvm->session->ordered_events;
811 oe->next_flush = flush_time;
812 err = ordered_events__flush(oe, OE_FLUSH__ROUND);
814 if (kvm->lost_events)
815 pr_info("\nLost events: %" PRIu64 "\n\n",
824 static volatile int done;
826 static void sig_handler(int sig __maybe_unused)
831 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
833 struct itimerspec new_value;
836 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
837 if (kvm->timerfd < 0) {
838 pr_err("timerfd_create failed\n");
842 new_value.it_value.tv_sec = kvm->display_time;
843 new_value.it_value.tv_nsec = 0;
844 new_value.it_interval.tv_sec = kvm->display_time;
845 new_value.it_interval.tv_nsec = 0;
847 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
848 pr_err("timerfd_settime failed: %d\n", errno);
858 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
863 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
868 pr_err("Failed to read timer fd: %d\n", errno);
872 if (rc != sizeof(uint64_t)) {
873 pr_err("Error reading timer fd - invalid size returned\n");
878 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
885 clear_events_cache_stats(kvm->kvm_events_cache);
886 kvm->total_count = 0;
888 kvm->lost_events = 0;
893 static int fd_set_nonblock(int fd)
897 arg = fcntl(fd, F_GETFL);
899 pr_err("Failed to get current flags for fd %d\n", fd);
903 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
904 pr_err("Failed to set non-block option on fd %d\n", fd);
911 static int perf_kvm__handle_stdin(void)
922 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
924 int nr_stdin, ret, err = -EINVAL;
927 /* live flag must be set first */
930 ret = cpu_isa_config(kvm);
934 if (!verify_vcpu(kvm->trace_vcpu) ||
936 !register_kvm_events_ops(kvm)) {
940 set_term_quiet_input(&save);
941 init_kvm_event_record(kvm);
943 signal(SIGINT, sig_handler);
944 signal(SIGTERM, sig_handler);
947 if (perf_kvm__timerfd_create(kvm) < 0) {
952 if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
955 nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
959 if (fd_set_nonblock(fileno(stdin)) != 0)
962 /* everything is good - enable the events and process */
963 perf_evlist__enable(kvm->evlist);
966 struct fdarray *fda = &kvm->evlist->pollfd;
969 rc = perf_kvm__mmap_read(kvm);
973 err = perf_kvm__handle_timerfd(kvm);
977 if (fda->entries[nr_stdin].revents & POLLIN)
978 done = perf_kvm__handle_stdin();
981 err = fdarray__poll(fda, 100);
984 perf_evlist__disable(kvm->evlist);
992 if (kvm->timerfd >= 0)
995 tcsetattr(0, TCSAFLUSH, &save);
999 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1002 struct perf_evsel *pos;
1003 struct perf_evlist *evlist = kvm->evlist;
1004 char sbuf[STRERR_BUFSIZE];
1006 perf_evlist__config(evlist, &kvm->opts, NULL);
1009 * Note: exclude_{guest,host} do not apply here.
1010 * This command processes KVM tracepoints from host only
1012 evlist__for_each_entry(evlist, pos) {
1013 struct perf_event_attr *attr = &pos->attr;
1015 /* make sure these *are* set */
1016 perf_evsel__set_sample_bit(pos, TID);
1017 perf_evsel__set_sample_bit(pos, TIME);
1018 perf_evsel__set_sample_bit(pos, CPU);
1019 perf_evsel__set_sample_bit(pos, RAW);
1020 /* make sure these are *not*; want as small a sample as possible */
1021 perf_evsel__reset_sample_bit(pos, PERIOD);
1022 perf_evsel__reset_sample_bit(pos, IP);
1023 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1024 perf_evsel__reset_sample_bit(pos, ADDR);
1025 perf_evsel__reset_sample_bit(pos, READ);
1030 attr->sample_period = 1;
1032 attr->watermark = 0;
1033 attr->wakeup_events = 1000;
1035 /* will enable all once we are ready */
1039 err = perf_evlist__open(evlist);
1041 printf("Couldn't create the events: %s\n",
1042 str_error_r(errno, sbuf, sizeof(sbuf)));
1046 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1047 ui__error("Failed to mmap the events: %s\n",
1048 str_error_r(errno, sbuf, sizeof(sbuf)));
1049 perf_evlist__close(evlist);
1060 static int read_events(struct perf_kvm_stat *kvm)
1064 struct perf_tool eops = {
1065 .sample = process_sample_event,
1066 .comm = perf_event__process_comm,
1067 .namespaces = perf_event__process_namespaces,
1068 .ordered_events = true,
1070 struct perf_data file = {
1071 .path = kvm->file_name,
1072 .mode = PERF_DATA_MODE_READ,
1073 .force = kvm->force,
1077 kvm->session = perf_session__new(&file, false, &kvm->tool);
1078 if (!kvm->session) {
1079 pr_err("Initializing perf session failed\n");
1083 symbol__init(&kvm->session->header.env);
1085 if (!perf_session__has_traces(kvm->session, "kvm record")) {
1091 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1092 * traced in the old kernel.
1094 ret = cpu_isa_config(kvm);
1098 ret = perf_session__process_events(kvm->session);
1101 perf_session__delete(kvm->session);
1105 static int parse_target_str(struct perf_kvm_stat *kvm)
1107 if (kvm->opts.target.pid) {
1108 kvm->pid_list = intlist__new(kvm->opts.target.pid);
1109 if (kvm->pid_list == NULL) {
1110 pr_err("Error parsing process id string\n");
1118 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1121 int vcpu = kvm->trace_vcpu;
1123 if (parse_target_str(kvm) != 0)
1126 if (!verify_vcpu(vcpu))
1129 if (!select_key(kvm))
1132 if (!register_kvm_events_ops(kvm))
1135 init_kvm_event_record(kvm);
1138 ret = read_events(kvm);
1149 #define STRDUP_FAIL_EXIT(s) \
1157 int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1163 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1165 unsigned int rec_argc, i, j, events_tp_size;
1166 const char **rec_argv;
1167 const char * const record_args[] = {
1173 const char * const kvm_stat_record_usage[] = {
1174 "perf kvm stat record [<options>]",
1177 const char * const *events_tp;
1181 ret = setup_kvm_events_tp(kvm);
1183 pr_err("Unable to setup the kvm tracepoints\n");
1187 for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1190 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1192 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1194 if (rec_argv == NULL)
1197 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1198 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1200 for (j = 0; j < events_tp_size; j++) {
1201 rec_argv[i++] = "-e";
1202 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1205 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1206 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1208 for (j = 1; j < (unsigned int)argc; j++, i++)
1209 rec_argv[i] = argv[j];
1211 set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1212 set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1213 set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1215 set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1216 set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1217 set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1218 set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1219 set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1220 set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1221 set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1222 set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1223 set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1224 set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1225 set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1226 set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1227 set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1228 set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1229 set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1231 record_usage = kvm_stat_record_usage;
1232 return cmd_record(i, rec_argv);
1236 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1238 const struct option kvm_events_report_options[] = {
1239 OPT_STRING(0, "event", &kvm->report_event, "report event",
1240 "event for reporting: vmexit, "
1241 "mmio (x86 only), ioport (x86 only)"),
1242 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1243 "vcpu id to report"),
1244 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1245 "key for sorting: sample(sort by samples number)"
1246 " time (sort by avg time)"),
1247 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1248 "analyze events only for given process id(s)"),
1249 OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1253 const char * const kvm_events_report_usage[] = {
1254 "perf kvm stat report [<options>]",
1259 argc = parse_options(argc, argv,
1260 kvm_events_report_options,
1261 kvm_events_report_usage, 0);
1263 usage_with_options(kvm_events_report_usage,
1264 kvm_events_report_options);
1267 if (!kvm->opts.target.pid)
1268 kvm->opts.target.system_wide = true;
1270 return kvm_events_report_vcpu(kvm);
1273 #ifdef HAVE_TIMERFD_SUPPORT
1274 static struct perf_evlist *kvm_live_event_list(void)
1276 struct perf_evlist *evlist;
1277 char *tp, *name, *sys;
1279 const char * const *events_tp;
1281 evlist = perf_evlist__new();
1285 for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1287 tp = strdup(*events_tp);
1291 /* split tracepoint into subsystem and name */
1293 name = strchr(tp, ':');
1295 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1303 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1304 pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1316 perf_evlist__delete(evlist);
1323 static int kvm_events_live(struct perf_kvm_stat *kvm,
1324 int argc, const char **argv)
1326 char errbuf[BUFSIZ];
1329 const struct option live_options[] = {
1330 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1331 "record events on existing process id"),
1332 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1333 "number of mmap data pages",
1334 perf_evlist__parse_mmap_pages),
1335 OPT_INCR('v', "verbose", &verbose,
1336 "be more verbose (show counter open errors, etc)"),
1337 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1338 "system-wide collection from all CPUs"),
1339 OPT_UINTEGER('d', "display", &kvm->display_time,
1340 "time in seconds between display updates"),
1341 OPT_STRING(0, "event", &kvm->report_event, "report event",
1342 "event for reporting: "
1343 "vmexit, mmio (x86 only), ioport (x86 only)"),
1344 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1345 "vcpu id to report"),
1346 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1347 "key for sorting: sample(sort by samples number)"
1348 " time (sort by avg time)"),
1349 OPT_U64(0, "duration", &kvm->duration,
1350 "show events other than"
1351 " HLT (x86 only) or Wait state (s390 only)"
1352 " that take longer than duration usecs"),
1353 OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1354 "per thread proc mmap processing timeout in ms"),
1357 const char * const live_usage[] = {
1358 "perf kvm stat live [<options>]",
1361 struct perf_data data = {
1362 .mode = PERF_DATA_MODE_WRITE,
1366 /* event handling */
1367 kvm->tool.sample = process_sample_event;
1368 kvm->tool.comm = perf_event__process_comm;
1369 kvm->tool.exit = perf_event__process_exit;
1370 kvm->tool.fork = perf_event__process_fork;
1371 kvm->tool.lost = process_lost_event;
1372 kvm->tool.namespaces = perf_event__process_namespaces;
1373 kvm->tool.ordered_events = true;
1374 perf_tool__fill_defaults(&kvm->tool);
1377 kvm->display_time = 1;
1378 kvm->opts.user_interval = 1;
1379 kvm->opts.mmap_pages = 512;
1380 kvm->opts.target.uses_mmap = false;
1381 kvm->opts.target.uid_str = NULL;
1382 kvm->opts.target.uid = UINT_MAX;
1383 kvm->opts.proc_map_timeout = 500;
1386 disable_buildid_cache();
1391 argc = parse_options(argc, argv, live_options,
1394 usage_with_options(live_usage, live_options);
1397 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1400 * target related setups
1402 err = target__validate(&kvm->opts.target);
1404 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1405 ui__warning("%s", errbuf);
1408 if (target__none(&kvm->opts.target))
1409 kvm->opts.target.system_wide = true;
1413 * generate the event list
1415 err = setup_kvm_events_tp(kvm);
1417 pr_err("Unable to setup the kvm tracepoints\n");
1421 kvm->evlist = kvm_live_event_list();
1422 if (kvm->evlist == NULL) {
1427 symbol_conf.nr_events = kvm->evlist->nr_entries;
1429 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1430 usage_with_options(live_usage, live_options);
1435 kvm->session = perf_session__new(&data, false, &kvm->tool);
1436 if (kvm->session == NULL) {
1440 kvm->session->evlist = kvm->evlist;
1441 perf_session__set_id_hdr_size(kvm->session);
1442 ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1443 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1444 kvm->evlist->threads, false,
1445 kvm->opts.proc_map_timeout, 1);
1446 err = kvm_live_open_events(kvm);
1450 err = kvm_events_live_report(kvm);
1453 perf_session__delete(kvm->session);
1454 kvm->session = NULL;
1455 perf_evlist__delete(kvm->evlist);
1461 static void print_kvm_stat_usage(void)
1463 printf("Usage: perf kvm stat <command>\n\n");
1465 printf("# Available commands:\n");
1466 printf("\trecord: record kvm events\n");
1467 printf("\treport: report statistical data of kvm events\n");
1468 printf("\tlive: live reporting of statistical data of kvm events\n");
1470 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1473 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1475 struct perf_kvm_stat kvm = {
1476 .file_name = file_name,
1479 .report_event = "vmexit",
1480 .sort_key = "sample",
1485 print_kvm_stat_usage();
1489 if (!strncmp(argv[1], "rec", 3))
1490 return kvm_events_record(&kvm, argc - 1, argv + 1);
1492 if (!strncmp(argv[1], "rep", 3))
1493 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1495 #ifdef HAVE_TIMERFD_SUPPORT
1496 if (!strncmp(argv[1], "live", 4))
1497 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1501 return cmd_stat(argc, argv);
1503 #endif /* HAVE_KVM_STAT_SUPPORT */
1505 static int __cmd_record(const char *file_name, int argc, const char **argv)
1507 int rec_argc, i = 0, j;
1508 const char **rec_argv;
1510 rec_argc = argc + 2;
1511 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1512 rec_argv[i++] = strdup("record");
1513 rec_argv[i++] = strdup("-o");
1514 rec_argv[i++] = strdup(file_name);
1515 for (j = 1; j < argc; j++, i++)
1516 rec_argv[i] = argv[j];
1518 BUG_ON(i != rec_argc);
1520 return cmd_record(i, rec_argv);
1523 static int __cmd_report(const char *file_name, int argc, const char **argv)
1525 int rec_argc, i = 0, j;
1526 const char **rec_argv;
1528 rec_argc = argc + 2;
1529 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1530 rec_argv[i++] = strdup("report");
1531 rec_argv[i++] = strdup("-i");
1532 rec_argv[i++] = strdup(file_name);
1533 for (j = 1; j < argc; j++, i++)
1534 rec_argv[i] = argv[j];
1536 BUG_ON(i != rec_argc);
1538 return cmd_report(i, rec_argv);
1542 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1544 int rec_argc, i = 0, j;
1545 const char **rec_argv;
1547 rec_argc = argc + 2;
1548 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1549 rec_argv[i++] = strdup("buildid-list");
1550 rec_argv[i++] = strdup("-i");
1551 rec_argv[i++] = strdup(file_name);
1552 for (j = 1; j < argc; j++, i++)
1553 rec_argv[i] = argv[j];
1555 BUG_ON(i != rec_argc);
1557 return cmd_buildid_list(i, rec_argv);
1560 int cmd_kvm(int argc, const char **argv)
1562 const char *file_name = NULL;
1563 const struct option kvm_options[] = {
1564 OPT_STRING('i', "input", &file_name, "file",
1566 OPT_STRING('o', "output", &file_name, "file",
1567 "Output file name"),
1568 OPT_BOOLEAN(0, "guest", &perf_guest,
1569 "Collect guest os data"),
1570 OPT_BOOLEAN(0, "host", &perf_host,
1571 "Collect host os data"),
1572 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1573 "guest mount directory under which every guest os"
1574 " instance has a subdir"),
1575 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1576 "file", "file saving guest os vmlinux"),
1577 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1578 "file", "file saving guest os /proc/kallsyms"),
1579 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1580 "file", "file saving guest os /proc/modules"),
1581 OPT_INCR('v', "verbose", &verbose,
1582 "be more verbose (show counter open errors, etc)"),
1586 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1587 "buildid-list", "stat", NULL };
1588 const char *kvm_usage[] = { NULL, NULL };
1593 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1594 PARSE_OPT_STOP_AT_NON_OPTION);
1596 usage_with_options(kvm_usage, kvm_options);
1602 file_name = get_filename_for_perf_kvm();
1605 pr_err("Failed to allocate memory for filename\n");
1610 if (!strncmp(argv[0], "rec", 3))
1611 return __cmd_record(file_name, argc, argv);
1612 else if (!strncmp(argv[0], "rep", 3))
1613 return __cmd_report(file_name, argc, argv);
1614 else if (!strncmp(argv[0], "diff", 4))
1615 return cmd_diff(argc, argv);
1616 else if (!strncmp(argv[0], "top", 3))
1617 return cmd_top(argc, argv);
1618 else if (!strncmp(argv[0], "buildid-list", 12))
1619 return __cmd_buildid_list(file_name, argc, argv);
1620 #ifdef HAVE_KVM_STAT_SUPPORT
1621 else if (!strncmp(argv[0], "stat", 4))
1622 return kvm_cmd_stat(file_name, argc, argv);
1625 usage_with_options(kvm_usage, kvm_options);