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 <lk/debugfs.h>
17 #include "util/tool.h"
18 #include "util/stat.h"
21 #include <sys/prctl.h>
22 #include <sys/timerfd.h>
25 #include <semaphore.h>
29 #if defined(__i386__) || defined(__x86_64__)
35 #define INVALID_KEY (~0ULL)
40 struct kvm_event_stats {
46 struct list_head hash_entry;
51 struct kvm_event_stats total;
53 #define DEFAULT_VCPU_NUM 8
55 struct kvm_event_stats *vcpu;
58 typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
60 struct kvm_event_key {
68 struct kvm_events_ops {
69 bool (*is_begin_event)(struct perf_evsel *evsel,
70 struct perf_sample *sample,
71 struct event_key *key);
72 bool (*is_end_event)(struct perf_evsel *evsel,
73 struct perf_sample *sample, struct event_key *key);
74 void (*decode_key)(struct perf_kvm_stat *kvm, struct event_key *key,
79 struct exit_reasons_table {
80 unsigned long exit_code;
84 #define EVENTS_BITS 12
85 #define EVENTS_CACHE_SIZE (1UL << EVENTS_BITS)
87 struct perf_kvm_stat {
88 struct perf_tool tool;
89 struct perf_record_opts opts;
90 struct perf_evlist *evlist;
91 struct perf_session *session;
93 const char *file_name;
94 const char *report_event;
98 struct exit_reasons_table *exit_reasons;
99 int exit_reasons_size;
100 const char *exit_reasons_isa;
102 struct kvm_events_ops *events_ops;
104 struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
112 struct intlist *pid_list;
114 struct rb_root result;
117 unsigned int display_time;
122 static void exit_event_get_key(struct perf_evsel *evsel,
123 struct perf_sample *sample,
124 struct event_key *key)
127 key->key = perf_evsel__intval(evsel, sample, "exit_reason");
130 static bool kvm_exit_event(struct perf_evsel *evsel)
132 return !strcmp(evsel->name, "kvm:kvm_exit");
135 static bool exit_event_begin(struct perf_evsel *evsel,
136 struct perf_sample *sample, struct event_key *key)
138 if (kvm_exit_event(evsel)) {
139 exit_event_get_key(evsel, sample, key);
146 static bool kvm_entry_event(struct perf_evsel *evsel)
148 return !strcmp(evsel->name, "kvm:kvm_entry");
151 static bool exit_event_end(struct perf_evsel *evsel,
152 struct perf_sample *sample __maybe_unused,
153 struct event_key *key __maybe_unused)
155 return kvm_entry_event(evsel);
158 static struct exit_reasons_table vmx_exit_reasons[] = {
162 static struct exit_reasons_table svm_exit_reasons[] = {
166 static const char *get_exit_reason(struct perf_kvm_stat *kvm, u64 exit_code)
168 int i = kvm->exit_reasons_size;
169 struct exit_reasons_table *tbl = kvm->exit_reasons;
172 if (tbl->exit_code == exit_code)
177 pr_err("unknown kvm exit code:%lld on %s\n",
178 (unsigned long long)exit_code, kvm->exit_reasons_isa);
182 static void exit_event_decode_key(struct perf_kvm_stat *kvm,
183 struct event_key *key,
186 const char *exit_reason = get_exit_reason(kvm, key->key);
188 scnprintf(decode, 20, "%s", exit_reason);
191 static struct kvm_events_ops exit_events = {
192 .is_begin_event = exit_event_begin,
193 .is_end_event = exit_event_end,
194 .decode_key = exit_event_decode_key,
199 * For the mmio events, we treat:
200 * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
201 * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
203 static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
204 struct event_key *key)
206 key->key = perf_evsel__intval(evsel, sample, "gpa");
207 key->info = perf_evsel__intval(evsel, sample, "type");
210 #define KVM_TRACE_MMIO_READ_UNSATISFIED 0
211 #define KVM_TRACE_MMIO_READ 1
212 #define KVM_TRACE_MMIO_WRITE 2
214 static bool mmio_event_begin(struct perf_evsel *evsel,
215 struct perf_sample *sample, struct event_key *key)
217 /* MMIO read begin event in kernel. */
218 if (kvm_exit_event(evsel))
221 /* MMIO write begin event in kernel. */
222 if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
223 perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
224 mmio_event_get_key(evsel, sample, key);
231 static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
232 struct event_key *key)
234 /* MMIO write end event in kernel. */
235 if (kvm_entry_event(evsel))
238 /* MMIO read end event in kernel.*/
239 if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
240 perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
241 mmio_event_get_key(evsel, sample, key);
248 static void mmio_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
249 struct event_key *key,
252 scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
253 key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
256 static struct kvm_events_ops mmio_events = {
257 .is_begin_event = mmio_event_begin,
258 .is_end_event = mmio_event_end,
259 .decode_key = mmio_event_decode_key,
260 .name = "MMIO Access"
263 /* The time of emulation pio access is from kvm_pio to kvm_entry. */
264 static void ioport_event_get_key(struct perf_evsel *evsel,
265 struct perf_sample *sample,
266 struct event_key *key)
268 key->key = perf_evsel__intval(evsel, sample, "port");
269 key->info = perf_evsel__intval(evsel, sample, "rw");
272 static bool ioport_event_begin(struct perf_evsel *evsel,
273 struct perf_sample *sample,
274 struct event_key *key)
276 if (!strcmp(evsel->name, "kvm:kvm_pio")) {
277 ioport_event_get_key(evsel, sample, key);
284 static bool ioport_event_end(struct perf_evsel *evsel,
285 struct perf_sample *sample __maybe_unused,
286 struct event_key *key __maybe_unused)
288 return kvm_entry_event(evsel);
291 static void ioport_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
292 struct event_key *key,
295 scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
296 key->info ? "POUT" : "PIN");
299 static struct kvm_events_ops ioport_events = {
300 .is_begin_event = ioport_event_begin,
301 .is_end_event = ioport_event_end,
302 .decode_key = ioport_event_decode_key,
303 .name = "IO Port Access"
306 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
310 if (!strcmp(kvm->report_event, "vmexit"))
311 kvm->events_ops = &exit_events;
312 else if (!strcmp(kvm->report_event, "mmio"))
313 kvm->events_ops = &mmio_events;
314 else if (!strcmp(kvm->report_event, "ioport"))
315 kvm->events_ops = &ioport_events;
317 pr_err("Unknown report event:%s\n", kvm->report_event);
324 struct vcpu_event_record {
327 struct kvm_event *last_event;
331 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
335 for (i = 0; i < EVENTS_CACHE_SIZE; i++)
336 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
339 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
341 struct list_head *head;
342 struct kvm_event *event;
346 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
347 head = &kvm_events_cache[i];
348 list_for_each_entry(event, head, hash_entry) {
349 /* reset stats for event */
350 event->total.time = 0;
351 init_stats(&event->total.stats);
353 for (j = 0; j < event->max_vcpu; ++j) {
354 event->vcpu[j].time = 0;
355 init_stats(&event->vcpu[j].stats);
361 static int kvm_events_hash_fn(u64 key)
363 return key & (EVENTS_CACHE_SIZE - 1);
366 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
368 int old_max_vcpu = event->max_vcpu;
371 if (vcpu_id < event->max_vcpu)
374 while (event->max_vcpu <= vcpu_id)
375 event->max_vcpu += DEFAULT_VCPU_NUM;
378 event->vcpu = realloc(event->vcpu,
379 event->max_vcpu * sizeof(*event->vcpu));
382 pr_err("Not enough memory\n");
386 memset(event->vcpu + old_max_vcpu, 0,
387 (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
391 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
393 struct kvm_event *event;
395 event = zalloc(sizeof(*event));
397 pr_err("Not enough memory\n");
405 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
406 struct event_key *key)
408 struct kvm_event *event;
409 struct list_head *head;
411 BUG_ON(key->key == INVALID_KEY);
413 head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
414 list_for_each_entry(event, head, hash_entry) {
415 if (event->key.key == key->key && event->key.info == key->info)
419 event = kvm_alloc_init_event(key);
423 list_add(&event->hash_entry, head);
427 static bool handle_begin_event(struct perf_kvm_stat *kvm,
428 struct vcpu_event_record *vcpu_record,
429 struct event_key *key, u64 timestamp)
431 struct kvm_event *event = NULL;
433 if (key->key != INVALID_KEY)
434 event = find_create_kvm_event(kvm, key);
436 vcpu_record->last_event = event;
437 vcpu_record->start_time = timestamp;
442 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
444 kvm_stats->time += time_diff;
445 update_stats(&kvm_stats->stats, time_diff);
448 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
450 struct kvm_event_stats *kvm_stats = &event->total;
453 kvm_stats = &event->vcpu[vcpu_id];
455 return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
456 avg_stats(&kvm_stats->stats));
459 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
463 kvm_update_event_stats(&event->total, time_diff);
467 if (!kvm_event_expand(event, vcpu_id))
470 kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
474 static bool handle_end_event(struct perf_kvm_stat *kvm,
475 struct vcpu_event_record *vcpu_record,
476 struct event_key *key,
477 struct perf_sample *sample)
479 struct kvm_event *event;
480 u64 time_begin, time_diff;
483 if (kvm->trace_vcpu == -1)
486 vcpu = vcpu_record->vcpu_id;
488 event = vcpu_record->last_event;
489 time_begin = vcpu_record->start_time;
491 /* The begin event is not caught. */
496 * In some case, the 'begin event' only records the start timestamp,
497 * the actual event is recognized in the 'end event' (e.g. mmio-event).
500 /* Both begin and end events did not get the key. */
501 if (!event && key->key == INVALID_KEY)
505 event = find_create_kvm_event(kvm, key);
510 vcpu_record->last_event = NULL;
511 vcpu_record->start_time = 0;
513 /* seems to happen once in a while during live mode */
514 if (sample->time < time_begin) {
515 pr_debug("End time before begin time; skipping event.\n");
519 time_diff = sample->time - time_begin;
521 if (kvm->duration && time_diff > kvm->duration) {
524 kvm->events_ops->decode_key(kvm, &event->key, decode);
525 if (strcmp(decode, "HLT")) {
526 pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
527 sample->time, sample->pid, vcpu_record->vcpu_id,
528 decode, time_diff/1000);
532 return update_kvm_event(event, vcpu, time_diff);
536 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
537 struct perf_evsel *evsel,
538 struct perf_sample *sample)
540 /* Only kvm_entry records vcpu id. */
541 if (!thread->priv && kvm_entry_event(evsel)) {
542 struct vcpu_event_record *vcpu_record;
544 vcpu_record = zalloc(sizeof(*vcpu_record));
546 pr_err("%s: Not enough memory\n", __func__);
550 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
551 thread->priv = vcpu_record;
557 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
558 struct thread *thread,
559 struct perf_evsel *evsel,
560 struct perf_sample *sample)
562 struct vcpu_event_record *vcpu_record;
563 struct event_key key = {.key = INVALID_KEY};
565 vcpu_record = per_vcpu_record(thread, evsel, sample);
569 /* only process events for vcpus user cares about */
570 if ((kvm->trace_vcpu != -1) &&
571 (kvm->trace_vcpu != vcpu_record->vcpu_id))
574 if (kvm->events_ops->is_begin_event(evsel, sample, &key))
575 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
577 if (kvm->events_ops->is_end_event(evsel, sample, &key))
578 return handle_end_event(kvm, vcpu_record, &key, sample);
583 #define GET_EVENT_KEY(func, field) \
584 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
587 return event->total.field; \
589 if (vcpu >= event->max_vcpu) \
592 return event->vcpu[vcpu].field; \
595 #define COMPARE_EVENT_KEY(func, field) \
596 GET_EVENT_KEY(func, field) \
597 static int compare_kvm_event_ ## func(struct kvm_event *one, \
598 struct kvm_event *two, int vcpu)\
600 return get_event_ ##func(one, vcpu) > \
601 get_event_ ##func(two, vcpu); \
604 GET_EVENT_KEY(time, time);
605 COMPARE_EVENT_KEY(count, stats.n);
606 COMPARE_EVENT_KEY(mean, stats.mean);
607 GET_EVENT_KEY(max, stats.max);
608 GET_EVENT_KEY(min, stats.min);
610 #define DEF_SORT_NAME_KEY(name, compare_key) \
611 { #name, compare_kvm_event_ ## compare_key }
613 static struct kvm_event_key keys[] = {
614 DEF_SORT_NAME_KEY(sample, count),
615 DEF_SORT_NAME_KEY(time, mean),
619 static bool select_key(struct perf_kvm_stat *kvm)
623 for (i = 0; keys[i].name; i++) {
624 if (!strcmp(keys[i].name, kvm->sort_key)) {
625 kvm->compare = keys[i].key;
630 pr_err("Unknown compare key:%s\n", kvm->sort_key);
634 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
635 key_cmp_fun bigger, int vcpu)
637 struct rb_node **rb = &result->rb_node;
638 struct rb_node *parent = NULL;
642 p = container_of(*rb, struct kvm_event, rb);
645 if (bigger(event, p, vcpu))
646 rb = &(*rb)->rb_left;
648 rb = &(*rb)->rb_right;
651 rb_link_node(&event->rb, parent, rb);
652 rb_insert_color(&event->rb, result);
656 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
658 int vcpu = kvm->trace_vcpu;
660 kvm->total_count += get_event_count(event, vcpu);
661 kvm->total_time += get_event_time(event, vcpu);
664 static bool event_is_valid(struct kvm_event *event, int vcpu)
666 return !!get_event_count(event, vcpu);
669 static void sort_result(struct perf_kvm_stat *kvm)
672 int vcpu = kvm->trace_vcpu;
673 struct kvm_event *event;
675 for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
676 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
677 if (event_is_valid(event, vcpu)) {
678 update_total_count(kvm, event);
679 insert_to_result(&kvm->result, event,
686 /* returns left most element of result, and erase it */
687 static struct kvm_event *pop_from_result(struct rb_root *result)
689 struct rb_node *node = rb_first(result);
694 rb_erase(node, result);
695 return container_of(node, struct kvm_event, rb);
698 static void print_vcpu_info(struct perf_kvm_stat *kvm)
700 int vcpu = kvm->trace_vcpu;
702 pr_info("Analyze events for ");
705 if (kvm->opts.target.system_wide)
706 pr_info("all VMs, ");
707 else if (kvm->opts.target.pid)
708 pr_info("pid(s) %s, ", kvm->opts.target.pid);
710 pr_info("dazed and confused on what is monitored, ");
714 pr_info("all VCPUs:\n\n");
716 pr_info("VCPU %d:\n\n", vcpu);
719 static void show_timeofday(void)
725 gettimeofday(&tv, NULL);
726 if (localtime_r(&tv.tv_sec, <ime)) {
727 strftime(date, sizeof(date), "%H:%M:%S", <ime);
728 pr_info("%s.%06ld", date, tv.tv_usec);
730 pr_info("00:00:00.000000");
735 static void print_result(struct perf_kvm_stat *kvm)
738 struct kvm_event *event;
739 int vcpu = kvm->trace_vcpu;
747 print_vcpu_info(kvm);
748 pr_info("%20s ", kvm->events_ops->name);
749 pr_info("%10s ", "Samples");
750 pr_info("%9s ", "Samples%");
752 pr_info("%9s ", "Time%");
753 pr_info("%10s ", "Min Time");
754 pr_info("%10s ", "Max Time");
755 pr_info("%16s ", "Avg time");
758 while ((event = pop_from_result(&kvm->result))) {
759 u64 ecount, etime, max, min;
761 ecount = get_event_count(event, vcpu);
762 etime = get_event_time(event, vcpu);
763 max = get_event_max(event, vcpu);
764 min = get_event_min(event, vcpu);
766 kvm->events_ops->decode_key(kvm, &event->key, decode);
767 pr_info("%20s ", decode);
768 pr_info("%10llu ", (unsigned long long)ecount);
769 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
770 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
771 pr_info("%8" PRIu64 "us ", min / 1000);
772 pr_info("%8" PRIu64 "us ", max / 1000);
773 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
774 kvm_event_rel_stddev(vcpu, event));
778 pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
779 kvm->total_count, kvm->total_time / 1e3);
781 if (kvm->lost_events)
782 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
785 static int process_lost_event(struct perf_tool *tool,
786 union perf_event *event __maybe_unused,
787 struct perf_sample *sample __maybe_unused,
788 struct machine *machine __maybe_unused)
790 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
796 static bool skip_sample(struct perf_kvm_stat *kvm,
797 struct perf_sample *sample)
799 if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
805 static int process_sample_event(struct perf_tool *tool,
806 union perf_event *event,
807 struct perf_sample *sample,
808 struct perf_evsel *evsel,
809 struct machine *machine)
811 struct thread *thread;
812 struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
815 if (skip_sample(kvm, sample))
818 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
819 if (thread == NULL) {
820 pr_debug("problem processing %d event, skipping it.\n",
825 if (!handle_kvm_event(kvm, thread, evsel, sample))
831 static int cpu_isa_config(struct perf_kvm_stat *kvm)
833 char buf[64], *cpuid;
837 err = get_cpuid(buf, sizeof(buf));
839 pr_err("Failed to look up CPU type (Intel or AMD)\n");
844 cpuid = kvm->session->header.env.cpuid;
846 if (strstr(cpuid, "Intel"))
848 else if (strstr(cpuid, "AMD"))
851 pr_err("CPU %s is not supported.\n", cpuid);
856 kvm->exit_reasons = vmx_exit_reasons;
857 kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
858 kvm->exit_reasons_isa = "VMX";
864 static bool verify_vcpu(int vcpu)
866 if (vcpu != -1 && vcpu < 0) {
867 pr_err("Invalid vcpu:%d.\n", vcpu);
874 /* keeping the max events to a modest level to keep
875 * the processing of samples per mmap smooth.
877 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
879 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
882 union perf_event *event;
883 struct perf_sample sample;
887 *mmap_time = ULLONG_MAX;
888 while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
889 err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
891 pr_err("Failed to parse sample\n");
895 err = perf_session_queue_event(kvm->session, event, &sample, 0);
897 pr_err("Failed to enqueue sample: %d\n", err);
901 /* save time stamp of our first sample for this mmap */
903 *mmap_time = sample.time;
905 /* limit events per mmap handled all at once */
907 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
914 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
916 int i, err, throttled = 0;
918 u64 flush_time = ULLONG_MAX, mmap_time;
920 for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
921 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
925 /* flush time is going to be the minimum of all the individual
926 * mmap times. Essentially, we flush all the samples queued up
927 * from the last pass under our minimal start time -- that leaves
928 * a very small race for samples to come in with a lower timestamp.
929 * The ioctl to return the perf_clock timestamp should close the
932 if (mmap_time < flush_time)
933 flush_time = mmap_time;
936 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
940 /* flush queue after each round in which we processed events */
942 kvm->session->ordered_samples.next_flush = flush_time;
943 err = kvm->tool.finished_round(&kvm->tool, NULL, kvm->session);
945 if (kvm->lost_events)
946 pr_info("\nLost events: %" PRIu64 "\n\n",
955 static volatile int done;
957 static void sig_handler(int sig __maybe_unused)
962 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
964 struct itimerspec new_value;
967 kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
968 if (kvm->timerfd < 0) {
969 pr_err("timerfd_create failed\n");
973 new_value.it_value.tv_sec = kvm->display_time;
974 new_value.it_value.tv_nsec = 0;
975 new_value.it_interval.tv_sec = kvm->display_time;
976 new_value.it_interval.tv_nsec = 0;
978 if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
979 pr_err("timerfd_settime failed: %d\n", errno);
989 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
994 rc = read(kvm->timerfd, &c, sizeof(uint64_t));
999 pr_err("Failed to read timer fd: %d\n", errno);
1003 if (rc != sizeof(uint64_t)) {
1004 pr_err("Error reading timer fd - invalid size returned\n");
1009 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
1011 /* update display */
1016 clear_events_cache_stats(kvm->kvm_events_cache);
1017 kvm->total_count = 0;
1018 kvm->total_time = 0;
1019 kvm->lost_events = 0;
1024 static int fd_set_nonblock(int fd)
1028 arg = fcntl(fd, F_GETFL);
1030 pr_err("Failed to get current flags for fd %d\n", fd);
1034 if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
1035 pr_err("Failed to set non-block option on fd %d\n", fd);
1043 int perf_kvm__handle_stdin(struct termios *tc_now, struct termios *tc_save)
1047 tcsetattr(0, TCSANOW, tc_now);
1049 tcsetattr(0, TCSAFLUSH, tc_save);
1057 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
1059 struct pollfd *pollfds = NULL;
1060 int nr_fds, nr_stdin, ret, err = -EINVAL;
1061 struct termios tc, save;
1063 /* live flag must be set first */
1066 ret = cpu_isa_config(kvm);
1070 if (!verify_vcpu(kvm->trace_vcpu) ||
1072 !register_kvm_events_ops(kvm)) {
1076 init_kvm_event_record(kvm);
1078 tcgetattr(0, &save);
1080 tc.c_lflag &= ~(ICANON | ECHO);
1084 signal(SIGINT, sig_handler);
1085 signal(SIGTERM, sig_handler);
1087 /* copy pollfds -- need to add timerfd and stdin */
1088 nr_fds = kvm->evlist->nr_fds;
1089 pollfds = zalloc(sizeof(struct pollfd) * (nr_fds + 2));
1094 memcpy(pollfds, kvm->evlist->pollfd,
1095 sizeof(struct pollfd) * kvm->evlist->nr_fds);
1098 if (perf_kvm__timerfd_create(kvm) < 0) {
1103 pollfds[nr_fds].fd = kvm->timerfd;
1104 pollfds[nr_fds].events = POLLIN;
1107 pollfds[nr_fds].fd = fileno(stdin);
1108 pollfds[nr_fds].events = POLLIN;
1111 if (fd_set_nonblock(fileno(stdin)) != 0)
1114 /* everything is good - enable the events and process */
1115 perf_evlist__enable(kvm->evlist);
1120 rc = perf_kvm__mmap_read(kvm);
1124 err = perf_kvm__handle_timerfd(kvm);
1128 if (pollfds[nr_stdin].revents & POLLIN)
1129 done = perf_kvm__handle_stdin(&tc, &save);
1132 err = poll(pollfds, nr_fds, 100);
1135 perf_evlist__disable(kvm->evlist);
1143 if (kvm->timerfd >= 0)
1144 close(kvm->timerfd);
1152 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1155 struct perf_evsel *pos;
1156 struct perf_evlist *evlist = kvm->evlist;
1158 perf_evlist__config(evlist, &kvm->opts);
1161 * Note: exclude_{guest,host} do not apply here.
1162 * This command processes KVM tracepoints from host only
1164 list_for_each_entry(pos, &evlist->entries, node) {
1165 struct perf_event_attr *attr = &pos->attr;
1167 /* make sure these *are* set */
1168 perf_evsel__set_sample_bit(pos, TID);
1169 perf_evsel__set_sample_bit(pos, TIME);
1170 perf_evsel__set_sample_bit(pos, CPU);
1171 perf_evsel__set_sample_bit(pos, RAW);
1172 /* make sure these are *not*; want as small a sample as possible */
1173 perf_evsel__reset_sample_bit(pos, PERIOD);
1174 perf_evsel__reset_sample_bit(pos, IP);
1175 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1176 perf_evsel__reset_sample_bit(pos, ADDR);
1177 perf_evsel__reset_sample_bit(pos, READ);
1182 attr->sample_period = 1;
1184 attr->watermark = 0;
1185 attr->wakeup_events = 1000;
1187 /* will enable all once we are ready */
1191 err = perf_evlist__open(evlist);
1193 printf("Couldn't create the events: %s\n", strerror(errno));
1197 if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1198 ui__error("Failed to mmap the events: %s\n", strerror(errno));
1199 perf_evlist__close(evlist);
1209 static int read_events(struct perf_kvm_stat *kvm)
1213 struct perf_tool eops = {
1214 .sample = process_sample_event,
1215 .comm = perf_event__process_comm,
1216 .ordered_samples = true,
1220 kvm->session = perf_session__new(kvm->file_name, O_RDONLY, 0, false,
1222 if (!kvm->session) {
1223 pr_err("Initializing perf session failed\n");
1227 if (!perf_session__has_traces(kvm->session, "kvm record"))
1231 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1232 * traced in the old kernel.
1234 ret = cpu_isa_config(kvm);
1238 return perf_session__process_events(kvm->session, &kvm->tool);
1241 static int parse_target_str(struct perf_kvm_stat *kvm)
1244 kvm->pid_list = intlist__new(kvm->pid_str);
1245 if (kvm->pid_list == NULL) {
1246 pr_err("Error parsing process id string\n");
1254 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1257 int vcpu = kvm->trace_vcpu;
1259 if (parse_target_str(kvm) != 0)
1262 if (!verify_vcpu(vcpu))
1265 if (!select_key(kvm))
1268 if (!register_kvm_events_ops(kvm))
1271 init_kvm_event_record(kvm);
1274 ret = read_events(kvm);
1285 static const char * const kvm_events_tp[] = {
1292 #define STRDUP_FAIL_EXIT(s) \
1301 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1303 unsigned int rec_argc, i, j;
1304 const char **rec_argv;
1305 const char * const record_args[] = {
1312 rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1313 2 * ARRAY_SIZE(kvm_events_tp);
1314 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1316 if (rec_argv == NULL)
1319 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1320 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1322 for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1323 rec_argv[i++] = "-e";
1324 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1327 rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1328 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1330 for (j = 1; j < (unsigned int)argc; j++, i++)
1331 rec_argv[i] = argv[j];
1333 return cmd_record(i, rec_argv, NULL);
1337 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1339 const struct option kvm_events_report_options[] = {
1340 OPT_STRING(0, "event", &kvm->report_event, "report event",
1341 "event for reporting: vmexit, mmio, ioport"),
1342 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1343 "vcpu id to report"),
1344 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1345 "key for sorting: sample(sort by samples number)"
1346 " time (sort by avg time)"),
1347 OPT_STRING('p', "pid", &kvm->pid_str, "pid",
1348 "analyze events only for given process id(s)"),
1352 const char * const kvm_events_report_usage[] = {
1353 "perf kvm stat report [<options>]",
1360 argc = parse_options(argc, argv,
1361 kvm_events_report_options,
1362 kvm_events_report_usage, 0);
1364 usage_with_options(kvm_events_report_usage,
1365 kvm_events_report_options);
1368 return kvm_events_report_vcpu(kvm);
1371 static struct perf_evlist *kvm_live_event_list(void)
1373 struct perf_evlist *evlist;
1374 char *tp, *name, *sys;
1378 evlist = perf_evlist__new();
1382 for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1384 tp = strdup(kvm_events_tp[j]);
1388 /* split tracepoint into subsystem and name */
1390 name = strchr(tp, ':');
1392 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1400 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1401 pr_err("Failed to add %s tracepoint to the list\n", kvm_events_tp[j]);
1413 perf_evlist__delete(evlist);
1420 static int kvm_events_live(struct perf_kvm_stat *kvm,
1421 int argc, const char **argv)
1423 char errbuf[BUFSIZ];
1426 const struct option live_options[] = {
1427 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1428 "record events on existing process id"),
1429 OPT_UINTEGER('m', "mmap-pages", &kvm->opts.mmap_pages,
1430 "number of mmap data pages"),
1431 OPT_INCR('v', "verbose", &verbose,
1432 "be more verbose (show counter open errors, etc)"),
1433 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1434 "system-wide collection from all CPUs"),
1435 OPT_UINTEGER('d', "display", &kvm->display_time,
1436 "time in seconds between display updates"),
1437 OPT_STRING(0, "event", &kvm->report_event, "report event",
1438 "event for reporting: vmexit, mmio, ioport"),
1439 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1440 "vcpu id to report"),
1441 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1442 "key for sorting: sample(sort by samples number)"
1443 " time (sort by avg time)"),
1444 OPT_U64(0, "duration", &kvm->duration,
1445 "show events other than HALT that take longer than duration usecs"),
1448 const char * const live_usage[] = {
1449 "perf kvm stat live [<options>]",
1454 /* event handling */
1455 kvm->tool.sample = process_sample_event;
1456 kvm->tool.comm = perf_event__process_comm;
1457 kvm->tool.exit = perf_event__process_exit;
1458 kvm->tool.fork = perf_event__process_fork;
1459 kvm->tool.lost = process_lost_event;
1460 kvm->tool.ordered_samples = true;
1461 perf_tool__fill_defaults(&kvm->tool);
1464 kvm->display_time = 1;
1465 kvm->opts.user_interval = 1;
1466 kvm->opts.mmap_pages = 512;
1467 kvm->opts.target.uses_mmap = false;
1468 kvm->opts.target.uid_str = NULL;
1469 kvm->opts.target.uid = UINT_MAX;
1472 disable_buildid_cache();
1475 setup_browser(false);
1478 argc = parse_options(argc, argv, live_options,
1481 usage_with_options(live_usage, live_options);
1484 kvm->duration *= NSEC_PER_USEC; /* convert usec to nsec */
1487 * target related setups
1489 err = perf_target__validate(&kvm->opts.target);
1491 perf_target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1492 ui__warning("%s", errbuf);
1495 if (perf_target__none(&kvm->opts.target))
1496 kvm->opts.target.system_wide = true;
1500 * generate the event list
1502 kvm->evlist = kvm_live_event_list();
1503 if (kvm->evlist == NULL) {
1508 symbol_conf.nr_events = kvm->evlist->nr_entries;
1510 if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1511 usage_with_options(live_usage, live_options);
1516 kvm->session = perf_session__new(NULL, O_WRONLY, false, false, &kvm->tool);
1517 if (kvm->session == NULL) {
1521 kvm->session->evlist = kvm->evlist;
1522 perf_session__set_id_hdr_size(kvm->session);
1525 if (perf_target__has_task(&kvm->opts.target))
1526 perf_event__synthesize_thread_map(&kvm->tool,
1527 kvm->evlist->threads,
1528 perf_event__process,
1529 &kvm->session->machines.host);
1531 perf_event__synthesize_threads(&kvm->tool, perf_event__process,
1532 &kvm->session->machines.host);
1535 err = kvm_live_open_events(kvm);
1539 err = kvm_events_live_report(kvm);
1545 perf_session__delete(kvm->session);
1546 kvm->session = NULL;
1548 perf_evlist__delete_maps(kvm->evlist);
1549 perf_evlist__delete(kvm->evlist);
1555 static void print_kvm_stat_usage(void)
1557 printf("Usage: perf kvm stat <command>\n\n");
1559 printf("# Available commands:\n");
1560 printf("\trecord: record kvm events\n");
1561 printf("\treport: report statistical data of kvm events\n");
1562 printf("\tlive: live reporting of statistical data of kvm events\n");
1564 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1567 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1569 struct perf_kvm_stat kvm = {
1570 .file_name = file_name,
1573 .report_event = "vmexit",
1574 .sort_key = "sample",
1576 .exit_reasons = svm_exit_reasons,
1577 .exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
1578 .exit_reasons_isa = "SVM",
1582 print_kvm_stat_usage();
1586 if (!strncmp(argv[1], "rec", 3))
1587 return kvm_events_record(&kvm, argc - 1, argv + 1);
1589 if (!strncmp(argv[1], "rep", 3))
1590 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1592 if (!strncmp(argv[1], "live", 4))
1593 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1596 return cmd_stat(argc, argv, NULL);
1600 static int __cmd_record(const char *file_name, int argc, const char **argv)
1602 int rec_argc, i = 0, j;
1603 const char **rec_argv;
1605 rec_argc = argc + 2;
1606 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1607 rec_argv[i++] = strdup("record");
1608 rec_argv[i++] = strdup("-o");
1609 rec_argv[i++] = strdup(file_name);
1610 for (j = 1; j < argc; j++, i++)
1611 rec_argv[i] = argv[j];
1613 BUG_ON(i != rec_argc);
1615 return cmd_record(i, rec_argv, NULL);
1618 static int __cmd_report(const char *file_name, int argc, const char **argv)
1620 int rec_argc, i = 0, j;
1621 const char **rec_argv;
1623 rec_argc = argc + 2;
1624 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1625 rec_argv[i++] = strdup("report");
1626 rec_argv[i++] = strdup("-i");
1627 rec_argv[i++] = strdup(file_name);
1628 for (j = 1; j < argc; j++, i++)
1629 rec_argv[i] = argv[j];
1631 BUG_ON(i != rec_argc);
1633 return cmd_report(i, rec_argv, NULL);
1637 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1639 int rec_argc, i = 0, j;
1640 const char **rec_argv;
1642 rec_argc = argc + 2;
1643 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1644 rec_argv[i++] = strdup("buildid-list");
1645 rec_argv[i++] = strdup("-i");
1646 rec_argv[i++] = strdup(file_name);
1647 for (j = 1; j < argc; j++, i++)
1648 rec_argv[i] = argv[j];
1650 BUG_ON(i != rec_argc);
1652 return cmd_buildid_list(i, rec_argv, NULL);
1655 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1657 const char *file_name = NULL;
1658 const struct option kvm_options[] = {
1659 OPT_STRING('i', "input", &file_name, "file",
1661 OPT_STRING('o', "output", &file_name, "file",
1662 "Output file name"),
1663 OPT_BOOLEAN(0, "guest", &perf_guest,
1664 "Collect guest os data"),
1665 OPT_BOOLEAN(0, "host", &perf_host,
1666 "Collect host os data"),
1667 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1668 "guest mount directory under which every guest os"
1669 " instance has a subdir"),
1670 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1671 "file", "file saving guest os vmlinux"),
1672 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1673 "file", "file saving guest os /proc/kallsyms"),
1674 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1675 "file", "file saving guest os /proc/modules"),
1680 const char * const kvm_usage[] = {
1681 "perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
1688 argc = parse_options(argc, argv, kvm_options, kvm_usage,
1689 PARSE_OPT_STOP_AT_NON_OPTION);
1691 usage_with_options(kvm_usage, kvm_options);
1697 if (perf_host && !perf_guest)
1698 file_name = strdup("perf.data.host");
1699 else if (!perf_host && perf_guest)
1700 file_name = strdup("perf.data.guest");
1702 file_name = strdup("perf.data.kvm");
1705 pr_err("Failed to allocate memory for filename\n");
1710 if (!strncmp(argv[0], "rec", 3))
1711 return __cmd_record(file_name, argc, argv);
1712 else if (!strncmp(argv[0], "rep", 3))
1713 return __cmd_report(file_name, argc, argv);
1714 else if (!strncmp(argv[0], "diff", 4))
1715 return cmd_diff(argc, argv, NULL);
1716 else if (!strncmp(argv[0], "top", 3))
1717 return cmd_top(argc, argv, NULL);
1718 else if (!strncmp(argv[0], "buildid-list", 12))
1719 return __cmd_buildid_list(file_name, argc, argv);
1720 #if defined(__i386__) || defined(__x86_64__)
1721 else if (!strncmp(argv[0], "stat", 4))
1722 return kvm_cmd_stat(file_name, argc, argv);
1725 usage_with_options(kvm_usage, kvm_options);