4 #include "util/evsel.h"
5 #include "util/evlist.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"
20 #include "util/data.h"
22 #include <sys/prctl.h>
23 #ifdef HAVE_TIMERFD_SUPPORT
24 #include <sys/timerfd.h>
28 #include <semaphore.h>
32 #if defined(__i386__) || defined(__x86_64__)
38 #define INVALID_KEY (~0ULL)
43 struct kvm_event_stats {
49 struct list_head hash_entry;
54 struct kvm_event_stats total;
56 #define DEFAULT_VCPU_NUM 8
58 struct kvm_event_stats *vcpu;
61 typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
63 struct kvm_event_key {
71 struct kvm_events_ops {
72 bool (*is_begin_event)(struct perf_evsel *evsel,
73 struct perf_sample *sample,
74 struct event_key *key);
75 bool (*is_end_event)(struct perf_evsel *evsel,
76 struct perf_sample *sample, struct event_key *key);
77 void (*decode_key)(struct perf_kvm_stat *kvm, struct event_key *key,
82 struct exit_reasons_table {
83 unsigned long exit_code;
87 #define EVENTS_BITS 12
88 #define EVENTS_CACHE_SIZE (1UL << EVENTS_BITS)
90 struct perf_kvm_stat {
91 struct perf_tool tool;
92 struct record_opts opts;
93 struct perf_evlist *evlist;
94 struct perf_session *session;
96 const char *file_name;
97 const char *report_event;
101 struct exit_reasons_table *exit_reasons;
102 int exit_reasons_size;
103 const char *exit_reasons_isa;
105 struct kvm_events_ops *events_ops;
107 struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
115 struct intlist *pid_list;
117 struct rb_root result;
120 unsigned int display_time;
125 static void exit_event_get_key(struct perf_evsel *evsel,
126 struct perf_sample *sample,
127 struct event_key *key)
130 key->key = perf_evsel__intval(evsel, sample, "exit_reason");
133 static bool kvm_exit_event(struct perf_evsel *evsel)
135 return !strcmp(evsel->name, "kvm:kvm_exit");
138 static bool exit_event_begin(struct perf_evsel *evsel,
139 struct perf_sample *sample, struct event_key *key)
141 if (kvm_exit_event(evsel)) {
142 exit_event_get_key(evsel, sample, key);
149 static bool kvm_entry_event(struct perf_evsel *evsel)
151 return !strcmp(evsel->name, "kvm:kvm_entry");
154 static bool exit_event_end(struct perf_evsel *evsel,
155 struct perf_sample *sample __maybe_unused,
156 struct event_key *key __maybe_unused)
158 return kvm_entry_event(evsel);
161 static struct exit_reasons_table vmx_exit_reasons[] = {
165 static struct exit_reasons_table svm_exit_reasons[] = {
169 static const char *get_exit_reason(struct perf_kvm_stat *kvm, u64 exit_code)
171 int i = kvm->exit_reasons_size;
172 struct exit_reasons_table *tbl = kvm->exit_reasons;
175 if (tbl->exit_code == exit_code)
180 pr_err("unknown kvm exit code:%lld on %s\n",
181 (unsigned long long)exit_code, kvm->exit_reasons_isa);
185 static void exit_event_decode_key(struct perf_kvm_stat *kvm,
186 struct event_key *key,
189 const char *exit_reason = get_exit_reason(kvm, key->key);
191 scnprintf(decode, 20, "%s", exit_reason);
194 static struct kvm_events_ops exit_events = {
195 .is_begin_event = exit_event_begin,
196 .is_end_event = exit_event_end,
197 .decode_key = exit_event_decode_key,
202 * For the mmio events, we treat:
203 * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
204 * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
206 static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
207 struct event_key *key)
209 key->key = perf_evsel__intval(evsel, sample, "gpa");
210 key->info = perf_evsel__intval(evsel, sample, "type");
213 #define KVM_TRACE_MMIO_READ_UNSATISFIED 0
214 #define KVM_TRACE_MMIO_READ 1
215 #define KVM_TRACE_MMIO_WRITE 2
217 static bool mmio_event_begin(struct perf_evsel *evsel,
218 struct perf_sample *sample, struct event_key *key)
220 /* MMIO read begin event in kernel. */
221 if (kvm_exit_event(evsel))
224 /* MMIO write begin event in kernel. */
225 if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
226 perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
227 mmio_event_get_key(evsel, sample, key);
234 static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
235 struct event_key *key)
237 /* MMIO write end event in kernel. */
238 if (kvm_entry_event(evsel))
241 /* MMIO read end event in kernel.*/
242 if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
243 perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
244 mmio_event_get_key(evsel, sample, key);
251 static void mmio_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
252 struct event_key *key,
255 scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
256 key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
259 static struct kvm_events_ops mmio_events = {
260 .is_begin_event = mmio_event_begin,
261 .is_end_event = mmio_event_end,
262 .decode_key = mmio_event_decode_key,
263 .name = "MMIO Access"
266 /* The time of emulation pio access is from kvm_pio to kvm_entry. */
267 static void ioport_event_get_key(struct perf_evsel *evsel,
268 struct perf_sample *sample,
269 struct event_key *key)
271 key->key = perf_evsel__intval(evsel, sample, "port");
272 key->info = perf_evsel__intval(evsel, sample, "rw");
275 static bool ioport_event_begin(struct perf_evsel *evsel,
276 struct perf_sample *sample,
277 struct event_key *key)
279 if (!strcmp(evsel->name, "kvm:kvm_pio")) {
280 ioport_event_get_key(evsel, sample, key);
287 static bool ioport_event_end(struct perf_evsel *evsel,
288 struct perf_sample *sample __maybe_unused,
289 struct event_key *key __maybe_unused)
291 return kvm_entry_event(evsel);
294 static void ioport_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
295 struct event_key *key,
298 scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
299 key->info ? "POUT" : "PIN");
302 static struct kvm_events_ops ioport_events = {
303 .is_begin_event = ioport_event_begin,
304 .is_end_event = ioport_event_end,
305 .decode_key = ioport_event_decode_key,
306 .name = "IO Port Access"
309 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
313 if (!strcmp(kvm->report_event, "vmexit"))
314 kvm->events_ops = &exit_events;
315 else if (!strcmp(kvm->report_event, "mmio"))
316 kvm->events_ops = &mmio_events;
317 else if (!strcmp(kvm->report_event, "ioport"))
318 kvm->events_ops = &ioport_events;
320 pr_err("Unknown report event:%s\n", kvm->report_event);
327 struct vcpu_event_record {
330 struct kvm_event *last_event;
334 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
338 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
339 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
342 #ifdef HAVE_TIMERFD_SUPPORT
343 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
345 struct list_head *head;
346 struct kvm_event *event;
350 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
351 head = &kvm_events_cache[i];
352 list_for_each_entry(event, head, hash_entry) {
353 /* reset stats for event */
354 event->total.time = 0;
355 init_stats(&event->total.stats);
357 for (j = 0; j < event->max_vcpu; ++j) {
358 event->vcpu[j].time = 0;
359 init_stats(&event->vcpu[j].stats);
366 static int kvm_events_hash_fn(u64 key)
368 return key & (EVENTS_CACHE_SIZE - 1);
371 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
373 int old_max_vcpu = event->max_vcpu;
376 if (vcpu_id < event->max_vcpu)
379 while (event->max_vcpu <= vcpu_id)
380 event->max_vcpu += DEFAULT_VCPU_NUM;
383 event->vcpu = realloc(event->vcpu,
384 event->max_vcpu * sizeof(*event->vcpu));
387 pr_err("Not enough memory\n");
391 memset(event->vcpu + old_max_vcpu, 0,
392 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
396 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
398 struct kvm_event *event;
400 event = zalloc(sizeof(*event));
402 pr_err("Not enough memory\n");
407 init_stats(&event->total.stats);
411 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
412 struct event_key *key)
414 struct kvm_event *event;
415 struct list_head *head;
417 BUG_ON(key->key == INVALID_KEY);
419 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
420 list_for_each_entry(event, head, hash_entry) {
421 if (event->key.key == key->key && event->key.info == key->info)
425 event = kvm_alloc_init_event(key);
429 list_add(&event->hash_entry, head);
433 static bool handle_begin_event(struct perf_kvm_stat *kvm,
434 struct vcpu_event_record *vcpu_record,
435 struct event_key *key, u64 timestamp)
437 struct kvm_event *event = NULL;
439 if (key->key != INVALID_KEY)
440 event = find_create_kvm_event(kvm, key);
442 vcpu_record->last_event = event;
443 vcpu_record->start_time = timestamp;
448 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
450 kvm_stats->time += time_diff;
451 update_stats(&kvm_stats->stats, time_diff);
454 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
456 struct kvm_event_stats *kvm_stats = &event->total;
459 kvm_stats = &event->vcpu[vcpu_id];
461 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
462 avg_stats(&kvm_stats->stats));
465 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
469 kvm_update_event_stats(&event->total, time_diff);
473 if (!kvm_event_expand(event, vcpu_id))
476 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
480 static bool handle_end_event(struct perf_kvm_stat *kvm,
481 struct vcpu_event_record *vcpu_record,
482 struct event_key *key,
483 struct perf_sample *sample)
485 struct kvm_event *event;
486 u64 time_begin, time_diff;
489 if (kvm->trace_vcpu == -1)
492 vcpu = vcpu_record->vcpu_id;
494 event = vcpu_record->last_event;
495 time_begin = vcpu_record->start_time;
497 /* The begin event is not caught. */
502 * In some case, the 'begin event' only records the start timestamp,
503 * the actual event is recognized in the 'end event' (e.g. mmio-event).
506 /* Both begin and end events did not get the key. */
507 if (!event && key->key == INVALID_KEY)
511 event = find_create_kvm_event(kvm, key);
516 vcpu_record->last_event = NULL;
517 vcpu_record->start_time = 0;
519 /* seems to happen once in a while during live mode */
520 if (sample->time < time_begin) {
521 pr_debug("End time before begin time; skipping event.\n");
525 time_diff = sample->time - time_begin;
527 if (kvm->duration && time_diff > kvm->duration) {
530 kvm->events_ops->decode_key(kvm, &event->key, decode);
531 if (strcmp(decode, "HLT")) {
532 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
533 sample->time, sample->pid, vcpu_record->vcpu_id,
534 decode, time_diff/1000);
538 return update_kvm_event(event, vcpu, time_diff);
542 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
543 struct perf_evsel *evsel,
544 struct perf_sample *sample)
546 /* Only kvm_entry records vcpu id. */
547 if (!thread->priv && kvm_entry_event(evsel)) {
548 struct vcpu_event_record *vcpu_record;
550 vcpu_record = zalloc(sizeof(*vcpu_record));
552 pr_err("%s: Not enough memory\n", __func__);
556 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
557 thread->priv = vcpu_record;
563 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
564 struct thread *thread,
565 struct perf_evsel *evsel,
566 struct perf_sample *sample)
568 struct vcpu_event_record *vcpu_record;
569 struct event_key key = {.key = INVALID_KEY};
571 vcpu_record = per_vcpu_record(thread, evsel, sample);
575 /* only process events for vcpus user cares about */
576 if ((kvm->trace_vcpu != -1) &&
577 (kvm->trace_vcpu != vcpu_record->vcpu_id))
580 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
581 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
583 if (kvm->events_ops->is_end_event(evsel, sample, &key))
584 return handle_end_event(kvm, vcpu_record, &key, sample);
589 #define GET_EVENT_KEY(func, field) \
590 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
593 return event->total.field; \
595 if (vcpu >= event->max_vcpu) \
598 return event->vcpu[vcpu].field; \
601 #define COMPARE_EVENT_KEY(func, field) \
602 GET_EVENT_KEY(func, field) \
603 static int compare_kvm_event_ ## func(struct kvm_event *one, \
604 struct kvm_event *two, int vcpu)\
606 return get_event_ ##func(one, vcpu) > \
607 get_event_ ##func(two, vcpu); \
610 GET_EVENT_KEY(time, time);
611 COMPARE_EVENT_KEY(count, stats.n);
612 COMPARE_EVENT_KEY(mean, stats.mean);
613 GET_EVENT_KEY(max, stats.max);
614 GET_EVENT_KEY(min, stats.min);
616 #define DEF_SORT_NAME_KEY(name, compare_key) \
617 { #name, compare_kvm_event_ ## compare_key }
619 static struct kvm_event_key keys[] = {
620 DEF_SORT_NAME_KEY(sample, count),
621 DEF_SORT_NAME_KEY(time, mean),
625 static bool select_key(struct perf_kvm_stat *kvm)
629 for (i = 0; keys[i].name; i++) {
630 if (!strcmp(keys[i].name, kvm->sort_key)) {
631 kvm->compare = keys[i].key;
636 pr_err("Unknown compare key:%s\n", kvm->sort_key);
640 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
641 key_cmp_fun bigger, int vcpu)
643 struct rb_node **rb = &result->rb_node;
644 struct rb_node *parent = NULL;
648 p = container_of(*rb, struct kvm_event, rb);
651 if (bigger(event, p, vcpu))
652 rb = &(*rb)->rb_left;
654 rb = &(*rb)->rb_right;
657 rb_link_node(&event->rb, parent, rb);
658 rb_insert_color(&event->rb, result);
662 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
664 int vcpu = kvm->trace_vcpu;
666 kvm->total_count += get_event_count(event, vcpu);
667 kvm->total_time += get_event_time(event, vcpu);
670 static bool event_is_valid(struct kvm_event *event, int vcpu)
672 return !!get_event_count(event, vcpu);
675 static void sort_result(struct perf_kvm_stat *kvm)
678 int vcpu = kvm->trace_vcpu;
679 struct kvm_event *event;
681 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
682 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
683 if (event_is_valid(event, vcpu)) {
684 update_total_count(kvm, event);
685 insert_to_result(&kvm->result, event,
692 /* returns left most element of result, and erase it */
693 static struct kvm_event *pop_from_result(struct rb_root *result)
695 struct rb_node *node = rb_first(result);
700 rb_erase(node, result);
701 return container_of(node, struct kvm_event, rb);
704 static void print_vcpu_info(struct perf_kvm_stat *kvm)
706 int vcpu = kvm->trace_vcpu;
708 pr_info("Analyze events for ");
711 if (kvm->opts.target.system_wide)
712 pr_info("all VMs, ");
713 else if (kvm->opts.target.pid)
714 pr_info("pid(s) %s, ", kvm->opts.target.pid);
716 pr_info("dazed and confused on what is monitored, ");
720 pr_info("all VCPUs:\n\n");
722 pr_info("VCPU %d:\n\n", vcpu);
725 static void show_timeofday(void)
731 gettimeofday(&tv, NULL);
732 if (localtime_r(&tv.tv_sec, <ime)) {
733 strftime(date, sizeof(date), "%H:%M:%S", <ime);
734 pr_info("%s.%06ld", date, tv.tv_usec);
736 pr_info("00:00:00.000000");
741 static void print_result(struct perf_kvm_stat *kvm)
744 struct kvm_event *event;
745 int vcpu = kvm->trace_vcpu;
753 print_vcpu_info(kvm);
754 pr_info("%20s ", kvm->events_ops->name);
755 pr_info("%10s ", "Samples");
756 pr_info("%9s ", "Samples%");
758 pr_info("%9s ", "Time%");
759 pr_info("%10s ", "Min Time");
760 pr_info("%10s ", "Max Time");
761 pr_info("%16s ", "Avg time");
764 while ((event = pop_from_result(&kvm->result))) {
765 u64 ecount, etime, max, min;
767 ecount = get_event_count(event, vcpu);
768 etime = get_event_time(event, vcpu);
769 max = get_event_max(event, vcpu);
770 min = get_event_min(event, vcpu);
772 kvm->events_ops->decode_key(kvm, &event->key, decode);
773 pr_info("%20s ", decode);
774 pr_info("%10llu ", (unsigned long long)ecount);
775 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
776 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
777 pr_info("%8" PRIu64 "us ", min / 1000);
778 pr_info("%8" PRIu64 "us ", max / 1000);
779 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
780 kvm_event_rel_stddev(vcpu, event));
784 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
785 kvm->total_count, kvm->total_time / 1e3);
787 if (kvm->lost_events)
788 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
791 #ifdef HAVE_TIMERFD_SUPPORT
792 static int process_lost_event(struct perf_tool *tool,
793 union perf_event *event __maybe_unused,
794 struct perf_sample *sample __maybe_unused,
795 struct machine *machine __maybe_unused)
797 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
804 static bool skip_sample(struct perf_kvm_stat *kvm,
805 struct perf_sample *sample)
807 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
813 static int process_sample_event(struct perf_tool *tool,
814 union perf_event *event,
815 struct perf_sample *sample,
816 struct perf_evsel *evsel,
817 struct machine *machine)
819 struct thread *thread;
820 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
823 if (skip_sample(kvm, sample))
826 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
827 if (thread == NULL) {
828 pr_debug("problem processing %d event, skipping it.\n",
833 if (!handle_kvm_event(kvm, thread, evsel, sample))
839 static int cpu_isa_config(struct perf_kvm_stat *kvm)
841 char buf[64], *cpuid;
845 err = get_cpuid(buf, sizeof(buf));
847 pr_err("Failed to look up CPU type (Intel or AMD)\n");
852 cpuid = kvm->session->header.env.cpuid;
854 if (strstr(cpuid, "Intel"))
856 else if (strstr(cpuid, "AMD"))
859 pr_err("CPU %s is not supported.\n", cpuid);
864 kvm->exit_reasons = vmx_exit_reasons;
865 kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
866 kvm->exit_reasons_isa = "VMX";
872 static bool verify_vcpu(int vcpu)
874 if (vcpu != -1 && vcpu < 0) {
875 pr_err("Invalid vcpu:%d.\n", vcpu);
882 #ifdef HAVE_TIMERFD_SUPPORT
883 /* keeping the max events to a modest level to keep
884 * the processing of samples per mmap smooth.
886 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
888 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
891 union perf_event *event;
892 struct perf_sample sample;
896 *mmap_time = ULLONG_MAX;
897 while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
898 err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
900 perf_evlist__mmap_consume(kvm->evlist, idx);
901 pr_err("Failed to parse sample\n");
905 err = perf_session_queue_event(kvm->session, event, &sample, 0);
907 * FIXME: Here we can't consume the event, as perf_session_queue_event will
908 * point to it, and it'll get possibly overwritten by the kernel.
910 perf_evlist__mmap_consume(kvm->evlist, idx);
913 pr_err("Failed to enqueue sample: %d\n", err);
917 /* save time stamp of our first sample for this mmap */
919 *mmap_time = sample.time;
921 /* limit events per mmap handled all at once */
923 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
930 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
932 int i, err, throttled = 0;
934 u64 flush_time = ULLONG_MAX, mmap_time;
936 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
937 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
941 /* flush time is going to be the minimum of all the individual
942 * mmap times. Essentially, we flush all the samples queued up
943 * from the last pass under our minimal start time -- that leaves
944 * a very small race for samples to come in with a lower timestamp.
945 * The ioctl to return the perf_clock timestamp should close the
948 if (mmap_time < flush_time)
949 flush_time = mmap_time;
952 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
956 /* flush queue after each round in which we processed events */
958 kvm->session->ordered_samples.next_flush = flush_time;
959 err = kvm->tool.finished_round(&kvm->tool, NULL, kvm->session);
961 if (kvm->lost_events)
962 pr_info("\nLost events: %" PRIu64 "\n\n",
971 static volatile int done;
973 static void sig_handler(int sig __maybe_unused)
978 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
980 struct itimerspec new_value;
983 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
984 if (kvm->timerfd < 0) {
985 pr_err("timerfd_create failed\n");
989 new_value.it_value.tv_sec = kvm->display_time;
990 new_value.it_value.tv_nsec = 0;
991 new_value.it_interval.tv_sec = kvm->display_time;
992 new_value.it_interval.tv_nsec = 0;
994 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
995 pr_err("timerfd_settime failed: %d\n", errno);
1005 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
1010 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
1012 if (errno == EAGAIN)
1015 pr_err("Failed to read timer fd: %d\n", errno);
1019 if (rc != sizeof(uint64_t)) {
1020 pr_err("Error reading timer fd - invalid size returned\n");
1025 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
1027 /* update display */
1032 clear_events_cache_stats(kvm->kvm_events_cache);
1033 kvm->total_count = 0;
1034 kvm->total_time = 0;
1035 kvm->lost_events = 0;
1040 static int fd_set_nonblock(int fd)
1044 arg = fcntl(fd, F_GETFL);
1046 pr_err("Failed to get current flags for fd %d\n", fd);
1050 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
1051 pr_err("Failed to set non-block option on fd %d\n", fd);
1059 int perf_kvm__handle_stdin(struct termios *tc_now, struct termios *tc_save)
1063 tcsetattr(0, TCSANOW, tc_now);
1065 tcsetattr(0, TCSAFLUSH, tc_save);
1073 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
1075 struct pollfd *pollfds = NULL;
1076 int nr_fds, nr_stdin, ret, err = -EINVAL;
1077 struct termios tc, save;
1079 /* live flag must be set first */
1082 ret = cpu_isa_config(kvm);
1086 if (!verify_vcpu(kvm->trace_vcpu) ||
1088 !register_kvm_events_ops(kvm)) {
1092 init_kvm_event_record(kvm);
1094 tcgetattr(0, &save);
1096 tc.c_lflag &= ~(ICANON | ECHO);
1100 signal(SIGINT, sig_handler);
1101 signal(SIGTERM, sig_handler);
1103 /* copy pollfds -- need to add timerfd and stdin */
1104 nr_fds = kvm->evlist->nr_fds;
1105 pollfds = zalloc(sizeof(struct pollfd) * (nr_fds + 2));
1110 memcpy(pollfds, kvm->evlist->pollfd,
1111 sizeof(struct pollfd) * kvm->evlist->nr_fds);
1114 if (perf_kvm__timerfd_create(kvm) < 0) {
1119 pollfds[nr_fds].fd = kvm->timerfd;
1120 pollfds[nr_fds].events = POLLIN;
1123 pollfds[nr_fds].fd = fileno(stdin);
1124 pollfds[nr_fds].events = POLLIN;
1127 if (fd_set_nonblock(fileno(stdin)) != 0)
1130 /* everything is good - enable the events and process */
1131 perf_evlist__enable(kvm->evlist);
1136 rc = perf_kvm__mmap_read(kvm);
1140 err = perf_kvm__handle_timerfd(kvm);
1144 if (pollfds[nr_stdin].revents & POLLIN)
1145 done = perf_kvm__handle_stdin(&tc, &save);
1148 err = poll(pollfds, nr_fds, 100);
1151 perf_evlist__disable(kvm->evlist);
1159 if (kvm->timerfd >= 0)
1160 close(kvm->timerfd);
1166 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1169 struct perf_evsel *pos;
1170 struct perf_evlist *evlist = kvm->evlist;
1172 perf_evlist__config(evlist, &kvm->opts);
1175 * Note: exclude_{guest,host} do not apply here.
1176 * This command processes KVM tracepoints from host only
1178 evlist__for_each(evlist, pos) {
1179 struct perf_event_attr *attr = &pos->attr;
1181 /* make sure these *are* set */
1182 perf_evsel__set_sample_bit(pos, TID);
1183 perf_evsel__set_sample_bit(pos, TIME);
1184 perf_evsel__set_sample_bit(pos, CPU);
1185 perf_evsel__set_sample_bit(pos, RAW);
1186 /* make sure these are *not*; want as small a sample as possible */
1187 perf_evsel__reset_sample_bit(pos, PERIOD);
1188 perf_evsel__reset_sample_bit(pos, IP);
1189 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1190 perf_evsel__reset_sample_bit(pos, ADDR);
1191 perf_evsel__reset_sample_bit(pos, READ);
1196 attr->sample_period = 1;
1198 attr->watermark = 0;
1199 attr->wakeup_events = 1000;
1201 /* will enable all once we are ready */
1205 err = perf_evlist__open(evlist);
1207 printf("Couldn't create the events: %s\n", strerror(errno));
1211 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1212 ui__error("Failed to mmap the events: %s\n", strerror(errno));
1213 perf_evlist__close(evlist);
1224 static int read_events(struct perf_kvm_stat *kvm)
1228 struct perf_tool eops = {
1229 .sample = process_sample_event,
1230 .comm = perf_event__process_comm,
1231 .ordered_samples = true,
1233 struct perf_data_file file = {
1234 .path = kvm->file_name,
1235 .mode = PERF_DATA_MODE_READ,
1239 kvm->session = perf_session__new(&file, false, &kvm->tool);
1240 if (!kvm->session) {
1241 pr_err("Initializing perf session failed\n");
1245 if (!perf_session__has_traces(kvm->session, "kvm record"))
1249 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1250 * traced in the old kernel.
1252 ret = cpu_isa_config(kvm);
1256 return perf_session__process_events(kvm->session, &kvm->tool);
1259 static int parse_target_str(struct perf_kvm_stat *kvm)
1262 kvm->pid_list = intlist__new(kvm->pid_str);
1263 if (kvm->pid_list == NULL) {
1264 pr_err("Error parsing process id string\n");
1272 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1275 int vcpu = kvm->trace_vcpu;
1277 if (parse_target_str(kvm) != 0)
1280 if (!verify_vcpu(vcpu))
1283 if (!select_key(kvm))
1286 if (!register_kvm_events_ops(kvm))
1289 init_kvm_event_record(kvm);
1292 ret = read_events(kvm);
1303 static const char * const kvm_events_tp[] = {
1310 #define STRDUP_FAIL_EXIT(s) \
1319 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1321 unsigned int rec_argc, i, j;
1322 const char **rec_argv;
1323 const char * const record_args[] = {
1330 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1331 2 * ARRAY_SIZE(kvm_events_tp);
1332 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1334 if (rec_argv == NULL)
1337 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1338 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1340 for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1341 rec_argv[i++] = "-e";
1342 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1345 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1346 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1348 for (j = 1; j < (unsigned int)argc; j++, i++)
1349 rec_argv[i] = argv[j];
1351 return cmd_record(i, rec_argv, NULL);
1355 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1357 const struct option kvm_events_report_options[] = {
1358 OPT_STRING(0, "event", &kvm->report_event, "report event",
1359 "event for reporting: vmexit, mmio, ioport"),
1360 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1361 "vcpu id to report"),
1362 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1363 "key for sorting: sample(sort by samples number)"
1364 " time (sort by avg time)"),
1365 OPT_STRING('p', "pid", &kvm->pid_str, "pid",
1366 "analyze events only for given process id(s)"),
1370 const char * const kvm_events_report_usage[] = {
1371 "perf kvm stat report [<options>]",
1378 argc = parse_options(argc, argv,
1379 kvm_events_report_options,
1380 kvm_events_report_usage, 0);
1382 usage_with_options(kvm_events_report_usage,
1383 kvm_events_report_options);
1386 return kvm_events_report_vcpu(kvm);
1389 #ifdef HAVE_TIMERFD_SUPPORT
1390 static struct perf_evlist *kvm_live_event_list(void)
1392 struct perf_evlist *evlist;
1393 char *tp, *name, *sys;
1397 evlist = perf_evlist__new();
1401 for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1403 tp = strdup(kvm_events_tp[j]);
1407 /* split tracepoint into subsystem and name */
1409 name = strchr(tp, ':');
1411 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1419 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1420 pr_err("Failed to add %s tracepoint to the list\n", kvm_events_tp[j]);
1432 perf_evlist__delete(evlist);
1439 static int kvm_events_live(struct perf_kvm_stat *kvm,
1440 int argc, const char **argv)
1442 char errbuf[BUFSIZ];
1445 const struct option live_options[] = {
1446 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1447 "record events on existing process id"),
1448 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1449 "number of mmap data pages",
1450 perf_evlist__parse_mmap_pages),
1451 OPT_INCR('v', "verbose", &verbose,
1452 "be more verbose (show counter open errors, etc)"),
1453 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1454 "system-wide collection from all CPUs"),
1455 OPT_UINTEGER('d', "display", &kvm->display_time,
1456 "time in seconds between display updates"),
1457 OPT_STRING(0, "event", &kvm->report_event, "report event",
1458 "event for reporting: vmexit, mmio, ioport"),
1459 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1460 "vcpu id to report"),
1461 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1462 "key for sorting: sample(sort by samples number)"
1463 " time (sort by avg time)"),
1464 OPT_U64(0, "duration", &kvm->duration,
1465 "show events other than HALT that take longer than duration usecs"),
1468 const char * const live_usage[] = {
1469 "perf kvm stat live [<options>]",
1472 struct perf_data_file file = {
1473 .mode = PERF_DATA_MODE_WRITE,
1477 /* event handling */
1478 kvm->tool.sample = process_sample_event;
1479 kvm->tool.comm = perf_event__process_comm;
1480 kvm->tool.exit = perf_event__process_exit;
1481 kvm->tool.fork = perf_event__process_fork;
1482 kvm->tool.lost = process_lost_event;
1483 kvm->tool.ordered_samples = true;
1484 perf_tool__fill_defaults(&kvm->tool);
1487 kvm->display_time = 1;
1488 kvm->opts.user_interval = 1;
1489 kvm->opts.mmap_pages = 512;
1490 kvm->opts.target.uses_mmap = false;
1491 kvm->opts.target.uid_str = NULL;
1492 kvm->opts.target.uid = UINT_MAX;
1495 disable_buildid_cache();
1498 setup_browser(false);
1501 argc = parse_options(argc, argv, live_options,
1504 usage_with_options(live_usage, live_options);
1507 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1510 * target related setups
1512 err = target__validate(&kvm->opts.target);
1514 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1515 ui__warning("%s", errbuf);
1518 if (target__none(&kvm->opts.target))
1519 kvm->opts.target.system_wide = true;
1523 * generate the event list
1525 kvm->evlist = kvm_live_event_list();
1526 if (kvm->evlist == NULL) {
1531 symbol_conf.nr_events = kvm->evlist->nr_entries;
1533 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1534 usage_with_options(live_usage, live_options);
1539 kvm->session = perf_session__new(&file, false, &kvm->tool);
1540 if (kvm->session == NULL) {
1544 kvm->session->evlist = kvm->evlist;
1545 perf_session__set_id_hdr_size(kvm->session);
1546 machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1547 kvm->evlist->threads, false);
1548 err = kvm_live_open_events(kvm);
1552 err = kvm_events_live_report(kvm);
1558 perf_session__delete(kvm->session);
1559 kvm->session = NULL;
1561 perf_evlist__delete(kvm->evlist);
1567 static void print_kvm_stat_usage(void)
1569 printf("Usage: perf kvm stat <command>\n\n");
1571 printf("# Available commands:\n");
1572 printf("\trecord: record kvm events\n");
1573 printf("\treport: report statistical data of kvm events\n");
1574 printf("\tlive: live reporting of statistical data of kvm events\n");
1576 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1579 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1581 struct perf_kvm_stat kvm = {
1582 .file_name = file_name,
1585 .report_event = "vmexit",
1586 .sort_key = "sample",
1588 .exit_reasons = svm_exit_reasons,
1589 .exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
1590 .exit_reasons_isa = "SVM",
1594 print_kvm_stat_usage();
1598 if (!strncmp(argv[1], "rec", 3))
1599 return kvm_events_record(&kvm, argc - 1, argv + 1);
1601 if (!strncmp(argv[1], "rep", 3))
1602 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1604 #ifdef HAVE_TIMERFD_SUPPORT
1605 if (!strncmp(argv[1], "live", 4))
1606 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1610 return cmd_stat(argc, argv, NULL);
1614 static int __cmd_record(const char *file_name, int argc, const char **argv)
1616 int rec_argc, i = 0, j;
1617 const char **rec_argv;
1619 rec_argc = argc + 2;
1620 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1621 rec_argv[i++] = strdup("record");
1622 rec_argv[i++] = strdup("-o");
1623 rec_argv[i++] = strdup(file_name);
1624 for (j = 1; j < argc; j++, i++)
1625 rec_argv[i] = argv[j];
1627 BUG_ON(i != rec_argc);
1629 return cmd_record(i, rec_argv, NULL);
1632 static int __cmd_report(const char *file_name, int argc, const char **argv)
1634 int rec_argc, i = 0, j;
1635 const char **rec_argv;
1637 rec_argc = argc + 2;
1638 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1639 rec_argv[i++] = strdup("report");
1640 rec_argv[i++] = strdup("-i");
1641 rec_argv[i++] = strdup(file_name);
1642 for (j = 1; j < argc; j++, i++)
1643 rec_argv[i] = argv[j];
1645 BUG_ON(i != rec_argc);
1647 return cmd_report(i, rec_argv, NULL);
1651 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1653 int rec_argc, i = 0, j;
1654 const char **rec_argv;
1656 rec_argc = argc + 2;
1657 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1658 rec_argv[i++] = strdup("buildid-list");
1659 rec_argv[i++] = strdup("-i");
1660 rec_argv[i++] = strdup(file_name);
1661 for (j = 1; j < argc; j++, i++)
1662 rec_argv[i] = argv[j];
1664 BUG_ON(i != rec_argc);
1666 return cmd_buildid_list(i, rec_argv, NULL);
1669 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1671 const char *file_name = NULL;
1672 const struct option kvm_options[] = {
1673 OPT_STRING('i', "input", &file_name, "file",
1675 OPT_STRING('o', "output", &file_name, "file",
1676 "Output file name"),
1677 OPT_BOOLEAN(0, "guest", &perf_guest,
1678 "Collect guest os data"),
1679 OPT_BOOLEAN(0, "host", &perf_host,
1680 "Collect host os data"),
1681 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1682 "guest mount directory under which every guest os"
1683 " instance has a subdir"),
1684 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1685 "file", "file saving guest os vmlinux"),
1686 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1687 "file", "file saving guest os /proc/kallsyms"),
1688 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1689 "file", "file saving guest os /proc/modules"),
1690 OPT_INCR('v', "verbose", &verbose,
1691 "be more verbose (show counter open errors, etc)"),
1695 const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1696 "buildid-list", "stat", NULL };
1697 const char *kvm_usage[] = { NULL, NULL };
1702 argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1703 PARSE_OPT_STOP_AT_NON_OPTION);
1705 usage_with_options(kvm_usage, kvm_options);
1711 file_name = get_filename_for_perf_kvm();
1714 pr_err("Failed to allocate memory for filename\n");
1719 if (!strncmp(argv[0], "rec", 3))
1720 return __cmd_record(file_name, argc, argv);
1721 else if (!strncmp(argv[0], "rep", 3))
1722 return __cmd_report(file_name, argc, argv);
1723 else if (!strncmp(argv[0], "diff", 4))
1724 return cmd_diff(argc, argv, NULL);
1725 else if (!strncmp(argv[0], "top", 3))
1726 return cmd_top(argc, argv, NULL);
1727 else if (!strncmp(argv[0], "buildid-list", 12))
1728 return __cmd_buildid_list(file_name, argc, argv);
1729 #if defined(__i386__) || defined(__x86_64__)
1730 else if (!strncmp(argv[0], "stat", 4))
1731 return kvm_cmd_stat(file_name, argc, argv);
1734 usage_with_options(kvm_usage, kvm_options);