1 // SPDX-License-Identifier: GPL-2.0
5 #include "util/evsel.h"
6 #include "util/evlist.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"
21 #include "util/data.h"
22 #include "util/ordered-events.h"
24 #include <sys/prctl.h>
25 #ifdef HAVE_TIMERFD_SUPPORT
26 #include <sys/timerfd.h>
29 #include <sys/types.h>
33 #include <linux/kernel.h>
34 #include <linux/time64.h>
39 #include <semaphore.h>
43 static const char *get_filename_for_perf_kvm(void)
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");
52 filename = strdup("perf.data.kvm");
57 #ifdef HAVE_KVM_STAT_SUPPORT
58 #include "util/kvm-stat.h"
60 void exit_event_get_key(struct perf_evsel *evsel,
61 struct perf_sample *sample,
62 struct event_key *key)
65 key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
68 bool kvm_exit_event(struct perf_evsel *evsel)
70 return !strcmp(evsel->name, kvm_exit_trace);
73 bool exit_event_begin(struct perf_evsel *evsel,
74 struct perf_sample *sample, struct event_key *key)
76 if (kvm_exit_event(evsel)) {
77 exit_event_get_key(evsel, sample, key);
84 bool kvm_entry_event(struct perf_evsel *evsel)
86 return !strcmp(evsel->name, kvm_entry_trace);
89 bool exit_event_end(struct perf_evsel *evsel,
90 struct perf_sample *sample __maybe_unused,
91 struct event_key *key __maybe_unused)
93 return kvm_entry_event(evsel);
96 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
97 struct exit_reasons_table *tbl,
100 while (tbl->reason != NULL) {
101 if (tbl->exit_code == exit_code)
106 pr_err("unknown kvm exit code:%lld on %s\n",
107 (unsigned long long)exit_code, kvm->exit_reasons_isa);
111 void exit_event_decode_key(struct perf_kvm_stat *kvm,
112 struct event_key *key,
115 const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
118 scnprintf(decode, decode_str_len, "%s", exit_reason);
121 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
123 struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
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;
135 struct vcpu_event_record {
138 struct kvm_event *last_event;
142 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
146 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
147 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
150 #ifdef HAVE_TIMERFD_SUPPORT
151 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
153 struct list_head *head;
154 struct kvm_event *event;
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);
165 for (j = 0; j < event->max_vcpu; ++j) {
166 event->vcpu[j].time = 0;
167 init_stats(&event->vcpu[j].stats);
174 static int kvm_events_hash_fn(u64 key)
176 return key & (EVENTS_CACHE_SIZE - 1);
179 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
181 int old_max_vcpu = event->max_vcpu;
184 if (vcpu_id < event->max_vcpu)
187 while (event->max_vcpu <= vcpu_id)
188 event->max_vcpu += DEFAULT_VCPU_NUM;
191 event->vcpu = realloc(event->vcpu,
192 event->max_vcpu * sizeof(*event->vcpu));
195 pr_err("Not enough memory\n");
199 memset(event->vcpu + old_max_vcpu, 0,
200 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
204 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
206 struct kvm_event *event;
208 event = zalloc(sizeof(*event));
210 pr_err("Not enough memory\n");
215 init_stats(&event->total.stats);
219 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
220 struct event_key *key)
222 struct kvm_event *event;
223 struct list_head *head;
225 BUG_ON(key->key == INVALID_KEY);
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)
233 event = kvm_alloc_init_event(key);
237 list_add(&event->hash_entry, head);
241 static bool handle_begin_event(struct perf_kvm_stat *kvm,
242 struct vcpu_event_record *vcpu_record,
243 struct event_key *key, u64 timestamp)
245 struct kvm_event *event = NULL;
247 if (key->key != INVALID_KEY)
248 event = find_create_kvm_event(kvm, key);
250 vcpu_record->last_event = event;
251 vcpu_record->start_time = timestamp;
256 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
258 kvm_stats->time += time_diff;
259 update_stats(&kvm_stats->stats, time_diff);
262 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
264 struct kvm_event_stats *kvm_stats = &event->total;
267 kvm_stats = &event->vcpu[vcpu_id];
269 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
270 avg_stats(&kvm_stats->stats));
273 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
277 kvm_update_event_stats(&event->total, time_diff);
281 if (!kvm_event_expand(event, vcpu_id))
284 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
288 static bool is_child_event(struct perf_kvm_stat *kvm,
289 struct perf_evsel *evsel,
290 struct perf_sample *sample,
291 struct event_key *key)
293 struct child_event_ops *child_ops;
295 child_ops = kvm->events_ops->child_ops;
300 for (; child_ops->name; child_ops++) {
301 if (!strcmp(evsel->name, child_ops->name)) {
302 child_ops->get_key(evsel, sample, key);
310 static 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)
315 struct kvm_event *event = NULL;
317 if (key->key != INVALID_KEY)
318 event = find_create_kvm_event(kvm, key);
320 vcpu_record->last_event = event;
325 static bool skip_event(const char *event)
327 const char * const *skip_events;
329 for (skip_events = kvm_skip_events; *skip_events; skip_events++)
330 if (!strcmp(event, *skip_events))
336 static 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)
341 struct kvm_event *event;
342 u64 time_begin, time_diff;
345 if (kvm->trace_vcpu == -1)
348 vcpu = vcpu_record->vcpu_id;
350 event = vcpu_record->last_event;
351 time_begin = vcpu_record->start_time;
353 /* The begin event is not caught. */
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).
362 /* Both begin and end events did not get the key. */
363 if (!event && key->key == INVALID_KEY)
367 event = find_create_kvm_event(kvm, key);
372 vcpu_record->last_event = NULL;
373 vcpu_record->start_time = 0;
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");
381 time_diff = sample->time - time_begin;
383 if (kvm->duration && time_diff > kvm->duration) {
384 char decode[decode_str_len];
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);
394 return update_kvm_event(event, vcpu, time_diff);
398 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
399 struct perf_evsel *evsel,
400 struct perf_sample *sample)
402 /* Only kvm_entry records vcpu id. */
403 if (!thread__priv(thread) && kvm_entry_event(evsel)) {
404 struct vcpu_event_record *vcpu_record;
406 vcpu_record = zalloc(sizeof(*vcpu_record));
408 pr_err("%s: Not enough memory\n", __func__);
412 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
414 thread__set_priv(thread, vcpu_record);
417 return thread__priv(thread);
420 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
421 struct thread *thread,
422 struct perf_evsel *evsel,
423 struct perf_sample *sample)
425 struct vcpu_event_record *vcpu_record;
426 struct event_key key = { .key = INVALID_KEY,
427 .exit_reasons = kvm->exit_reasons };
429 vcpu_record = per_vcpu_record(thread, evsel, sample);
433 /* only process events for vcpus user cares about */
434 if ((kvm->trace_vcpu != -1) &&
435 (kvm->trace_vcpu != vcpu_record->vcpu_id))
438 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
439 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
441 if (is_child_event(kvm, evsel, sample, &key))
442 return handle_child_event(kvm, vcpu_record, &key, sample);
444 if (kvm->events_ops->is_end_event(evsel, sample, &key))
445 return handle_end_event(kvm, vcpu_record, &key, sample);
450 #define GET_EVENT_KEY(func, field) \
451 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
454 return event->total.field; \
456 if (vcpu >= event->max_vcpu) \
459 return event->vcpu[vcpu].field; \
462 #define COMPARE_EVENT_KEY(func, field) \
463 GET_EVENT_KEY(func, field) \
464 static int compare_kvm_event_ ## func(struct kvm_event *one, \
465 struct kvm_event *two, int vcpu)\
467 return get_event_ ##func(one, vcpu) > \
468 get_event_ ##func(two, vcpu); \
471 GET_EVENT_KEY(time, time);
472 COMPARE_EVENT_KEY(count, stats.n);
473 COMPARE_EVENT_KEY(mean, stats.mean);
474 GET_EVENT_KEY(max, stats.max);
475 GET_EVENT_KEY(min, stats.min);
477 #define DEF_SORT_NAME_KEY(name, compare_key) \
478 { #name, compare_kvm_event_ ## compare_key }
480 static struct kvm_event_key keys[] = {
481 DEF_SORT_NAME_KEY(sample, count),
482 DEF_SORT_NAME_KEY(time, mean),
486 static bool select_key(struct perf_kvm_stat *kvm)
490 for (i = 0; keys[i].name; i++) {
491 if (!strcmp(keys[i].name, kvm->sort_key)) {
492 kvm->compare = keys[i].key;
497 pr_err("Unknown compare key:%s\n", kvm->sort_key);
501 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
502 key_cmp_fun bigger, int vcpu)
504 struct rb_node **rb = &result->rb_node;
505 struct rb_node *parent = NULL;
509 p = container_of(*rb, struct kvm_event, rb);
512 if (bigger(event, p, vcpu))
513 rb = &(*rb)->rb_left;
515 rb = &(*rb)->rb_right;
518 rb_link_node(&event->rb, parent, rb);
519 rb_insert_color(&event->rb, result);
523 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
525 int vcpu = kvm->trace_vcpu;
527 kvm->total_count += get_event_count(event, vcpu);
528 kvm->total_time += get_event_time(event, vcpu);
531 static bool event_is_valid(struct kvm_event *event, int vcpu)
533 return !!get_event_count(event, vcpu);
536 static void sort_result(struct perf_kvm_stat *kvm)
539 int vcpu = kvm->trace_vcpu;
540 struct kvm_event *event;
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,
553 /* returns left most element of result, and erase it */
554 static struct kvm_event *pop_from_result(struct rb_root *result)
556 struct rb_node *node = rb_first(result);
561 rb_erase(node, result);
562 return container_of(node, struct kvm_event, rb);
565 static void print_vcpu_info(struct perf_kvm_stat *kvm)
567 int vcpu = kvm->trace_vcpu;
569 pr_info("Analyze events for ");
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);
576 pr_info("dazed and confused on what is monitored, ");
579 pr_info("all VCPUs:\n\n");
581 pr_info("VCPU %d:\n\n", vcpu);
584 static void show_timeofday(void)
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);
595 pr_info("00:00:00.000000");
600 static void print_result(struct perf_kvm_stat *kvm)
602 char decode[decode_str_len];
603 struct kvm_event *event;
604 int vcpu = kvm->trace_vcpu;
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%");
617 pr_info("%9s ", "Time%");
618 pr_info("%11s ", "Min Time");
619 pr_info("%11s ", "Max Time");
620 pr_info("%16s ", "Avg time");
623 while ((event = pop_from_result(&kvm->result))) {
624 u64 ecount, etime, max, min;
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);
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));
643 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
644 kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
646 if (kvm->lost_events)
647 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
650 #ifdef HAVE_TIMERFD_SUPPORT
651 static 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)
656 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
663 static bool skip_sample(struct perf_kvm_stat *kvm,
664 struct perf_sample *sample)
666 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
672 static 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)
679 struct thread *thread;
680 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
683 if (skip_sample(kvm, sample))
686 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
687 if (thread == NULL) {
688 pr_debug("problem processing %d event, skipping it.\n",
693 if (!handle_kvm_event(kvm, thread, evsel, sample))
700 static int cpu_isa_config(struct perf_kvm_stat *kvm)
702 char buf[128], *cpuid;
706 err = get_cpuid(buf, sizeof(buf));
708 pr_err("Failed to look up CPU type: %s\n",
709 str_error_r(err, buf, sizeof(buf)));
714 cpuid = kvm->session->header.env.cpuid;
717 pr_err("Failed to look up CPU type\n");
721 err = cpu_isa_init(kvm, cpuid);
723 pr_err("CPU %s is not supported.\n", cpuid);
728 static bool verify_vcpu(int vcpu)
730 if (vcpu != -1 && vcpu < 0) {
731 pr_err("Invalid vcpu:%d.\n", vcpu);
738 #ifdef HAVE_TIMERFD_SUPPORT
739 /* keeping the max events to a modest level to keep
740 * the processing of samples per mmap smooth.
742 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
744 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
747 struct perf_evlist *evlist = kvm->evlist;
748 union perf_event *event;
749 struct perf_mmap *md;
754 *mmap_time = ULLONG_MAX;
755 md = &evlist->mmap[idx];
756 err = perf_mmap__read_init(md);
758 return (err == -EAGAIN) ? 0 : -1;
760 while ((event = perf_mmap__read_event(md)) != NULL) {
761 err = perf_evlist__parse_sample_timestamp(evlist, event, ×tamp);
763 perf_mmap__consume(md);
764 pr_err("Failed to parse sample\n");
768 err = perf_session__queue_event(kvm->session, event, timestamp, 0);
770 * FIXME: Here we can't consume the event, as perf_session__queue_event will
771 * point to it, and it'll get possibly overwritten by the kernel.
773 perf_mmap__consume(md);
776 pr_err("Failed to enqueue sample: %d\n", err);
780 /* save time stamp of our first sample for this mmap */
782 *mmap_time = timestamp;
784 /* limit events per mmap handled all at once */
786 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
790 perf_mmap__read_done(md);
794 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
796 int i, err, throttled = 0;
798 u64 flush_time = ULLONG_MAX, mmap_time;
800 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
801 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
805 /* flush time is going to be the minimum of all the individual
806 * mmap times. Essentially, we flush all the samples queued up
807 * from the last pass under our minimal start time -- that leaves
808 * a very small race for samples to come in with a lower timestamp.
809 * The ioctl to return the perf_clock timestamp should close the
812 if (mmap_time < flush_time)
813 flush_time = mmap_time;
816 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
820 /* flush queue after each round in which we processed events */
822 struct ordered_events *oe = &kvm->session->ordered_events;
824 oe->next_flush = flush_time;
825 err = ordered_events__flush(oe, OE_FLUSH__ROUND);
827 if (kvm->lost_events)
828 pr_info("\nLost events: %" PRIu64 "\n\n",
837 static volatile int done;
839 static void sig_handler(int sig __maybe_unused)
844 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
846 struct itimerspec new_value;
849 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
850 if (kvm->timerfd < 0) {
851 pr_err("timerfd_create failed\n");
855 new_value.it_value.tv_sec = kvm->display_time;
856 new_value.it_value.tv_nsec = 0;
857 new_value.it_interval.tv_sec = kvm->display_time;
858 new_value.it_interval.tv_nsec = 0;
860 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
861 pr_err("timerfd_settime failed: %d\n", errno);
871 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
876 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
881 pr_err("Failed to read timer fd: %d\n", errno);
885 if (rc != sizeof(uint64_t)) {
886 pr_err("Error reading timer fd - invalid size returned\n");
891 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
898 clear_events_cache_stats(kvm->kvm_events_cache);
899 kvm->total_count = 0;
901 kvm->lost_events = 0;
906 static int fd_set_nonblock(int fd)
910 arg = fcntl(fd, F_GETFL);
912 pr_err("Failed to get current flags for fd %d\n", fd);
916 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
917 pr_err("Failed to set non-block option on fd %d\n", fd);
924 static int perf_kvm__handle_stdin(void)
935 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
937 int nr_stdin, ret, err = -EINVAL;
940 /* live flag must be set first */
943 ret = cpu_isa_config(kvm);
947 if (!verify_vcpu(kvm->trace_vcpu) ||
949 !register_kvm_events_ops(kvm)) {
953 set_term_quiet_input(&save);
954 init_kvm_event_record(kvm);
956 signal(SIGINT, sig_handler);
957 signal(SIGTERM, sig_handler);
960 if (perf_kvm__timerfd_create(kvm) < 0) {
965 if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
968 nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
972 if (fd_set_nonblock(fileno(stdin)) != 0)
975 /* everything is good - enable the events and process */
976 perf_evlist__enable(kvm->evlist);
979 struct fdarray *fda = &kvm->evlist->pollfd;
982 rc = perf_kvm__mmap_read(kvm);
986 err = perf_kvm__handle_timerfd(kvm);
990 if (fda->entries[nr_stdin].revents & POLLIN)
991 done = perf_kvm__handle_stdin();
994 err = fdarray__poll(fda, 100);
997 perf_evlist__disable(kvm->evlist);
1005 if (kvm->timerfd >= 0)
1006 close(kvm->timerfd);
1008 tcsetattr(0, TCSAFLUSH, &save);
1012 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1015 struct perf_evsel *pos;
1016 struct perf_evlist *evlist = kvm->evlist;
1017 char sbuf[STRERR_BUFSIZE];
1019 perf_evlist__config(evlist, &kvm->opts, NULL);
1022 * Note: exclude_{guest,host} do not apply here.
1023 * This command processes KVM tracepoints from host only
1025 evlist__for_each_entry(evlist, pos) {
1026 struct perf_event_attr *attr = &pos->attr;
1028 /* make sure these *are* set */
1029 perf_evsel__set_sample_bit(pos, TID);
1030 perf_evsel__set_sample_bit(pos, TIME);
1031 perf_evsel__set_sample_bit(pos, CPU);
1032 perf_evsel__set_sample_bit(pos, RAW);
1033 /* make sure these are *not*; want as small a sample as possible */
1034 perf_evsel__reset_sample_bit(pos, PERIOD);
1035 perf_evsel__reset_sample_bit(pos, IP);
1036 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1037 perf_evsel__reset_sample_bit(pos, ADDR);
1038 perf_evsel__reset_sample_bit(pos, READ);
1043 attr->sample_period = 1;
1045 attr->watermark = 0;
1046 attr->wakeup_events = 1000;
1048 /* will enable all once we are ready */
1052 err = perf_evlist__open(evlist);
1054 printf("Couldn't create the events: %s\n",
1055 str_error_r(errno, sbuf, sizeof(sbuf)));
1059 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1060 ui__error("Failed to mmap the events: %s\n",
1061 str_error_r(errno, sbuf, sizeof(sbuf)));
1062 perf_evlist__close(evlist);
1073 static int read_events(struct perf_kvm_stat *kvm)
1077 struct perf_tool eops = {
1078 .sample = process_sample_event,
1079 .comm = perf_event__process_comm,
1080 .namespaces = perf_event__process_namespaces,
1081 .ordered_events = true,
1083 struct perf_data file = {
1085 .path = kvm->file_name,
1087 .mode = PERF_DATA_MODE_READ,
1088 .force = kvm->force,
1092 kvm->session = perf_session__new(&file, false, &kvm->tool);
1093 if (!kvm->session) {
1094 pr_err("Initializing perf session failed\n");
1098 symbol__init(&kvm->session->header.env);
1100 if (!perf_session__has_traces(kvm->session, "kvm record")) {
1106 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1107 * traced in the old kernel.
1109 ret = cpu_isa_config(kvm);
1113 ret = perf_session__process_events(kvm->session);
1116 perf_session__delete(kvm->session);
1120 static int parse_target_str(struct perf_kvm_stat *kvm)
1122 if (kvm->opts.target.pid) {
1123 kvm->pid_list = intlist__new(kvm->opts.target.pid);
1124 if (kvm->pid_list == NULL) {
1125 pr_err("Error parsing process id string\n");
1133 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1136 int vcpu = kvm->trace_vcpu;
1138 if (parse_target_str(kvm) != 0)
1141 if (!verify_vcpu(vcpu))
1144 if (!select_key(kvm))
1147 if (!register_kvm_events_ops(kvm))
1150 init_kvm_event_record(kvm);
1153 ret = read_events(kvm);
1164 #define STRDUP_FAIL_EXIT(s) \
1172 int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1178 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1180 unsigned int rec_argc, i, j, events_tp_size;
1181 const char **rec_argv;
1182 const char * const record_args[] = {
1188 const char * const kvm_stat_record_usage[] = {
1189 "perf kvm stat record [<options>]",
1192 const char * const *events_tp;
1196 ret = setup_kvm_events_tp(kvm);
1198 pr_err("Unable to setup the kvm tracepoints\n");
1202 for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1205 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1207 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1209 if (rec_argv == NULL)
1212 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1213 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1215 for (j = 0; j < events_tp_size; j++) {
1216 rec_argv[i++] = "-e";
1217 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1220 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1221 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1223 for (j = 1; j < (unsigned int)argc; j++, i++)
1224 rec_argv[i] = argv[j];
1226 set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1227 set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1228 set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1230 set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1231 set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1232 set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1233 set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1234 set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1235 set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1236 set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1237 set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1238 set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1239 set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1240 set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1241 set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1242 set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1243 set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1244 set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1246 record_usage = kvm_stat_record_usage;
1247 return cmd_record(i, rec_argv);
1251 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1253 const struct option kvm_events_report_options[] = {
1254 OPT_STRING(0, "event", &kvm->report_event, "report event",
1255 "event for reporting: vmexit, "
1256 "mmio (x86 only), ioport (x86 only)"),
1257 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1258 "vcpu id to report"),
1259 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1260 "key for sorting: sample(sort by samples number)"
1261 " time (sort by avg time)"),
1262 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1263 "analyze events only for given process id(s)"),
1264 OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1268 const char * const kvm_events_report_usage[] = {
1269 "perf kvm stat report [<options>]",
1274 argc = parse_options(argc, argv,
1275 kvm_events_report_options,
1276 kvm_events_report_usage, 0);
1278 usage_with_options(kvm_events_report_usage,
1279 kvm_events_report_options);
1282 if (!kvm->opts.target.pid)
1283 kvm->opts.target.system_wide = true;
1285 return kvm_events_report_vcpu(kvm);
1288 #ifdef HAVE_TIMERFD_SUPPORT
1289 static struct perf_evlist *kvm_live_event_list(void)
1291 struct perf_evlist *evlist;
1292 char *tp, *name, *sys;
1294 const char * const *events_tp;
1296 evlist = perf_evlist__new();
1300 for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1302 tp = strdup(*events_tp);
1306 /* split tracepoint into subsystem and name */
1308 name = strchr(tp, ':');
1310 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1318 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1319 pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1331 perf_evlist__delete(evlist);
1338 static int kvm_events_live(struct perf_kvm_stat *kvm,
1339 int argc, const char **argv)
1341 char errbuf[BUFSIZ];
1344 const struct option live_options[] = {
1345 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1346 "record events on existing process id"),
1347 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1348 "number of mmap data pages",
1349 perf_evlist__parse_mmap_pages),
1350 OPT_INCR('v', "verbose", &verbose,
1351 "be more verbose (show counter open errors, etc)"),
1352 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1353 "system-wide collection from all CPUs"),
1354 OPT_UINTEGER('d', "display", &kvm->display_time,
1355 "time in seconds between display updates"),
1356 OPT_STRING(0, "event", &kvm->report_event, "report event",
1357 "event for reporting: "
1358 "vmexit, mmio (x86 only), ioport (x86 only)"),
1359 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1360 "vcpu id to report"),
1361 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1362 "key for sorting: sample(sort by samples number)"
1363 " time (sort by avg time)"),
1364 OPT_U64(0, "duration", &kvm->duration,
1365 "show events other than"
1366 " HLT (x86 only) or Wait state (s390 only)"
1367 " that take longer than duration usecs"),
1368 OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1369 "per thread proc mmap processing timeout in ms"),
1372 const char * const live_usage[] = {
1373 "perf kvm stat live [<options>]",
1376 struct perf_data data = {
1377 .mode = PERF_DATA_MODE_WRITE,
1381 /* event handling */
1382 kvm->tool.sample = process_sample_event;
1383 kvm->tool.comm = perf_event__process_comm;
1384 kvm->tool.exit = perf_event__process_exit;
1385 kvm->tool.fork = perf_event__process_fork;
1386 kvm->tool.lost = process_lost_event;
1387 kvm->tool.namespaces = perf_event__process_namespaces;
1388 kvm->tool.ordered_events = true;
1389 perf_tool__fill_defaults(&kvm->tool);
1392 kvm->display_time = 1;
1393 kvm->opts.user_interval = 1;
1394 kvm->opts.mmap_pages = 512;
1395 kvm->opts.target.uses_mmap = false;
1396 kvm->opts.target.uid_str = NULL;
1397 kvm->opts.target.uid = UINT_MAX;
1398 kvm->opts.proc_map_timeout = 500;
1401 disable_buildid_cache();
1406 argc = parse_options(argc, argv, live_options,
1409 usage_with_options(live_usage, live_options);
1412 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1415 * target related setups
1417 err = target__validate(&kvm->opts.target);
1419 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1420 ui__warning("%s", errbuf);
1423 if (target__none(&kvm->opts.target))
1424 kvm->opts.target.system_wide = true;
1428 * generate the event list
1430 err = setup_kvm_events_tp(kvm);
1432 pr_err("Unable to setup the kvm tracepoints\n");
1436 kvm->evlist = kvm_live_event_list();
1437 if (kvm->evlist == NULL) {
1442 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1443 usage_with_options(live_usage, live_options);
1448 kvm->session = perf_session__new(&data, false, &kvm->tool);
1449 if (kvm->session == NULL) {
1453 kvm->session->evlist = kvm->evlist;
1454 perf_session__set_id_hdr_size(kvm->session);
1455 ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1456 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1457 kvm->evlist->threads, false,
1458 kvm->opts.proc_map_timeout, 1);
1459 err = kvm_live_open_events(kvm);
1463 err = kvm_events_live_report(kvm);
1466 perf_session__delete(kvm->session);
1467 kvm->session = NULL;
1468 perf_evlist__delete(kvm->evlist);
1474 static void print_kvm_stat_usage(void)
1476 printf("Usage: perf kvm stat <command>\n\n");
1478 printf("# Available commands:\n");
1479 printf("\trecord: record kvm events\n");
1480 printf("\treport: report statistical data of kvm events\n");
1481 printf("\tlive: live reporting of statistical data of kvm events\n");
1483 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1486 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1488 struct perf_kvm_stat kvm = {
1489 .file_name = file_name,
1492 .report_event = "vmexit",
1493 .sort_key = "sample",
1498 print_kvm_stat_usage();
1502 if (!strncmp(argv[1], "rec", 3))
1503 return kvm_events_record(&kvm, argc - 1, argv + 1);
1505 if (!strncmp(argv[1], "rep", 3))
1506 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1508 #ifdef HAVE_TIMERFD_SUPPORT
1509 if (!strncmp(argv[1], "live", 4))
1510 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1514 return cmd_stat(argc, argv);
1516 #endif /* HAVE_KVM_STAT_SUPPORT */
1518 static int __cmd_record(const char *file_name, int argc, const char **argv)
1520 int rec_argc, i = 0, j;
1521 const char **rec_argv;
1523 rec_argc = argc + 2;
1524 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1525 rec_argv[i++] = strdup("record");
1526 rec_argv[i++] = strdup("-o");
1527 rec_argv[i++] = strdup(file_name);
1528 for (j = 1; j < argc; j++, i++)
1529 rec_argv[i] = argv[j];
1531 BUG_ON(i != rec_argc);
1533 return cmd_record(i, rec_argv);
1536 static int __cmd_report(const char *file_name, int argc, const char **argv)
1538 int rec_argc, i = 0, j;
1539 const char **rec_argv;
1541 rec_argc = argc + 2;
1542 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1543 rec_argv[i++] = strdup("report");
1544 rec_argv[i++] = strdup("-i");
1545 rec_argv[i++] = strdup(file_name);
1546 for (j = 1; j < argc; j++, i++)
1547 rec_argv[i] = argv[j];
1549 BUG_ON(i != rec_argc);
1551 return cmd_report(i, rec_argv);
1555 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1557 int rec_argc, i = 0, j;
1558 const char **rec_argv;
1560 rec_argc = argc + 2;
1561 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1562 rec_argv[i++] = strdup("buildid-list");
1563 rec_argv[i++] = strdup("-i");
1564 rec_argv[i++] = strdup(file_name);
1565 for (j = 1; j < argc; j++, i++)
1566 rec_argv[i] = argv[j];
1568 BUG_ON(i != rec_argc);
1570 return cmd_buildid_list(i, rec_argv);
1573 int cmd_kvm(int argc, const char **argv)
1575 const char *file_name = NULL;
1576 const struct option kvm_options[] = {
1577 OPT_STRING('i', "input", &file_name, "file",
1579 OPT_STRING('o', "output", &file_name, "file",
1580 "Output file name"),
1581 OPT_BOOLEAN(0, "guest", &perf_guest,
1582 "Collect guest os data"),
1583 OPT_BOOLEAN(0, "host", &perf_host,
1584 "Collect host os data"),
1585 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1586 "guest mount directory under which every guest os"
1587 " instance has a subdir"),
1588 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1589 "file", "file saving guest os vmlinux"),
1590 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1591 "file", "file saving guest os /proc/kallsyms"),
1592 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1593 "file", "file saving guest os /proc/modules"),
1594 OPT_INCR('v', "verbose", &verbose,
1595 "be more verbose (show counter open errors, etc)"),
1599 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1600 "buildid-list", "stat", NULL };
1601 const char *kvm_usage[] = { NULL, NULL };
1606 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1607 PARSE_OPT_STOP_AT_NON_OPTION);
1609 usage_with_options(kvm_usage, kvm_options);
1615 file_name = get_filename_for_perf_kvm();
1618 pr_err("Failed to allocate memory for filename\n");
1623 if (!strncmp(argv[0], "rec", 3))
1624 return __cmd_record(file_name, argc, argv);
1625 else if (!strncmp(argv[0], "rep", 3))
1626 return __cmd_report(file_name, argc, argv);
1627 else if (!strncmp(argv[0], "diff", 4))
1628 return cmd_diff(argc, argv);
1629 else if (!strncmp(argv[0], "top", 3))
1630 return cmd_top(argc, argv);
1631 else if (!strncmp(argv[0], "buildid-list", 12))
1632 return __cmd_buildid_list(file_name, argc, argv);
1633 #ifdef HAVE_KVM_STAT_SUPPORT
1634 else if (!strncmp(argv[0], "stat", 4))
1635 return kvm_cmd_stat(file_name, argc, argv);
1638 usage_with_options(kvm_usage, kvm_options);