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");
410 static struct kvm_event
*find_create_kvm_event(struct perf_kvm_stat
*kvm
,
411 struct event_key
*key
)
413 struct kvm_event
*event
;
414 struct list_head
*head
;
416 BUG_ON(key
->key
== INVALID_KEY
);
418 head
= &kvm
->kvm_events_cache
[kvm_events_hash_fn(key
->key
)];
419 list_for_each_entry(event
, head
, hash_entry
) {
420 if (event
->key
.key
== key
->key
&& event
->key
.info
== key
->info
)
424 event
= kvm_alloc_init_event(key
);
428 list_add(&event
->hash_entry
, head
);
432 static bool handle_begin_event(struct perf_kvm_stat
*kvm
,
433 struct vcpu_event_record
*vcpu_record
,
434 struct event_key
*key
, u64 timestamp
)
436 struct kvm_event
*event
= NULL
;
438 if (key
->key
!= INVALID_KEY
)
439 event
= find_create_kvm_event(kvm
, key
);
441 vcpu_record
->last_event
= event
;
442 vcpu_record
->start_time
= timestamp
;
447 kvm_update_event_stats(struct kvm_event_stats
*kvm_stats
, u64 time_diff
)
449 kvm_stats
->time
+= time_diff
;
450 update_stats(&kvm_stats
->stats
, time_diff
);
453 static double kvm_event_rel_stddev(int vcpu_id
, struct kvm_event
*event
)
455 struct kvm_event_stats
*kvm_stats
= &event
->total
;
458 kvm_stats
= &event
->vcpu
[vcpu_id
];
460 return rel_stddev_stats(stddev_stats(&kvm_stats
->stats
),
461 avg_stats(&kvm_stats
->stats
));
464 static bool update_kvm_event(struct kvm_event
*event
, int vcpu_id
,
468 kvm_update_event_stats(&event
->total
, time_diff
);
472 if (!kvm_event_expand(event
, vcpu_id
))
475 kvm_update_event_stats(&event
->vcpu
[vcpu_id
], time_diff
);
479 static bool handle_end_event(struct perf_kvm_stat
*kvm
,
480 struct vcpu_event_record
*vcpu_record
,
481 struct event_key
*key
,
482 struct perf_sample
*sample
)
484 struct kvm_event
*event
;
485 u64 time_begin
, time_diff
;
488 if (kvm
->trace_vcpu
== -1)
491 vcpu
= vcpu_record
->vcpu_id
;
493 event
= vcpu_record
->last_event
;
494 time_begin
= vcpu_record
->start_time
;
496 /* The begin event is not caught. */
501 * In some case, the 'begin event' only records the start timestamp,
502 * the actual event is recognized in the 'end event' (e.g. mmio-event).
505 /* Both begin and end events did not get the key. */
506 if (!event
&& key
->key
== INVALID_KEY
)
510 event
= find_create_kvm_event(kvm
, key
);
515 vcpu_record
->last_event
= NULL
;
516 vcpu_record
->start_time
= 0;
518 /* seems to happen once in a while during live mode */
519 if (sample
->time
< time_begin
) {
520 pr_debug("End time before begin time; skipping event.\n");
524 time_diff
= sample
->time
- time_begin
;
526 if (kvm
->duration
&& time_diff
> kvm
->duration
) {
529 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
530 if (strcmp(decode
, "HLT")) {
531 pr_info("%" PRIu64
" VM %d, vcpu %d: %s event took %" PRIu64
"usec\n",
532 sample
->time
, sample
->pid
, vcpu_record
->vcpu_id
,
533 decode
, time_diff
/1000);
537 return update_kvm_event(event
, vcpu
, time_diff
);
541 struct vcpu_event_record
*per_vcpu_record(struct thread
*thread
,
542 struct perf_evsel
*evsel
,
543 struct perf_sample
*sample
)
545 /* Only kvm_entry records vcpu id. */
546 if (!thread
->priv
&& kvm_entry_event(evsel
)) {
547 struct vcpu_event_record
*vcpu_record
;
549 vcpu_record
= zalloc(sizeof(*vcpu_record
));
551 pr_err("%s: Not enough memory\n", __func__
);
555 vcpu_record
->vcpu_id
= perf_evsel__intval(evsel
, sample
, "vcpu_id");
556 thread
->priv
= vcpu_record
;
562 static bool handle_kvm_event(struct perf_kvm_stat
*kvm
,
563 struct thread
*thread
,
564 struct perf_evsel
*evsel
,
565 struct perf_sample
*sample
)
567 struct vcpu_event_record
*vcpu_record
;
568 struct event_key key
= {.key
= INVALID_KEY
};
570 vcpu_record
= per_vcpu_record(thread
, evsel
, sample
);
574 /* only process events for vcpus user cares about */
575 if ((kvm
->trace_vcpu
!= -1) &&
576 (kvm
->trace_vcpu
!= vcpu_record
->vcpu_id
))
579 if (kvm
->events_ops
->is_begin_event(evsel
, sample
, &key
))
580 return handle_begin_event(kvm
, vcpu_record
, &key
, sample
->time
);
582 if (kvm
->events_ops
->is_end_event(evsel
, sample
, &key
))
583 return handle_end_event(kvm
, vcpu_record
, &key
, sample
);
588 #define GET_EVENT_KEY(func, field) \
589 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
592 return event->total.field; \
594 if (vcpu >= event->max_vcpu) \
597 return event->vcpu[vcpu].field; \
600 #define COMPARE_EVENT_KEY(func, field) \
601 GET_EVENT_KEY(func, field) \
602 static int compare_kvm_event_ ## func(struct kvm_event *one, \
603 struct kvm_event *two, int vcpu)\
605 return get_event_ ##func(one, vcpu) > \
606 get_event_ ##func(two, vcpu); \
609 GET_EVENT_KEY(time
, time
);
610 COMPARE_EVENT_KEY(count
, stats
.n
);
611 COMPARE_EVENT_KEY(mean
, stats
.mean
);
612 GET_EVENT_KEY(max
, stats
.max
);
613 GET_EVENT_KEY(min
, stats
.min
);
615 #define DEF_SORT_NAME_KEY(name, compare_key) \
616 { #name, compare_kvm_event_ ## compare_key }
618 static struct kvm_event_key keys
[] = {
619 DEF_SORT_NAME_KEY(sample
, count
),
620 DEF_SORT_NAME_KEY(time
, mean
),
624 static bool select_key(struct perf_kvm_stat
*kvm
)
628 for (i
= 0; keys
[i
].name
; i
++) {
629 if (!strcmp(keys
[i
].name
, kvm
->sort_key
)) {
630 kvm
->compare
= keys
[i
].key
;
635 pr_err("Unknown compare key:%s\n", kvm
->sort_key
);
639 static void insert_to_result(struct rb_root
*result
, struct kvm_event
*event
,
640 key_cmp_fun bigger
, int vcpu
)
642 struct rb_node
**rb
= &result
->rb_node
;
643 struct rb_node
*parent
= NULL
;
647 p
= container_of(*rb
, struct kvm_event
, rb
);
650 if (bigger(event
, p
, vcpu
))
651 rb
= &(*rb
)->rb_left
;
653 rb
= &(*rb
)->rb_right
;
656 rb_link_node(&event
->rb
, parent
, rb
);
657 rb_insert_color(&event
->rb
, result
);
661 update_total_count(struct perf_kvm_stat
*kvm
, struct kvm_event
*event
)
663 int vcpu
= kvm
->trace_vcpu
;
665 kvm
->total_count
+= get_event_count(event
, vcpu
);
666 kvm
->total_time
+= get_event_time(event
, vcpu
);
669 static bool event_is_valid(struct kvm_event
*event
, int vcpu
)
671 return !!get_event_count(event
, vcpu
);
674 static void sort_result(struct perf_kvm_stat
*kvm
)
677 int vcpu
= kvm
->trace_vcpu
;
678 struct kvm_event
*event
;
680 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++) {
681 list_for_each_entry(event
, &kvm
->kvm_events_cache
[i
], hash_entry
) {
682 if (event_is_valid(event
, vcpu
)) {
683 update_total_count(kvm
, event
);
684 insert_to_result(&kvm
->result
, event
,
691 /* returns left most element of result, and erase it */
692 static struct kvm_event
*pop_from_result(struct rb_root
*result
)
694 struct rb_node
*node
= rb_first(result
);
699 rb_erase(node
, result
);
700 return container_of(node
, struct kvm_event
, rb
);
703 static void print_vcpu_info(struct perf_kvm_stat
*kvm
)
705 int vcpu
= kvm
->trace_vcpu
;
707 pr_info("Analyze events for ");
710 if (kvm
->opts
.target
.system_wide
)
711 pr_info("all VMs, ");
712 else if (kvm
->opts
.target
.pid
)
713 pr_info("pid(s) %s, ", kvm
->opts
.target
.pid
);
715 pr_info("dazed and confused on what is monitored, ");
719 pr_info("all VCPUs:\n\n");
721 pr_info("VCPU %d:\n\n", vcpu
);
724 static void show_timeofday(void)
730 gettimeofday(&tv
, NULL
);
731 if (localtime_r(&tv
.tv_sec
, <ime
)) {
732 strftime(date
, sizeof(date
), "%H:%M:%S", <ime
);
733 pr_info("%s.%06ld", date
, tv
.tv_usec
);
735 pr_info("00:00:00.000000");
740 static void print_result(struct perf_kvm_stat
*kvm
)
743 struct kvm_event
*event
;
744 int vcpu
= kvm
->trace_vcpu
;
752 print_vcpu_info(kvm
);
753 pr_info("%20s ", kvm
->events_ops
->name
);
754 pr_info("%10s ", "Samples");
755 pr_info("%9s ", "Samples%");
757 pr_info("%9s ", "Time%");
758 pr_info("%10s ", "Min Time");
759 pr_info("%10s ", "Max Time");
760 pr_info("%16s ", "Avg time");
763 while ((event
= pop_from_result(&kvm
->result
))) {
764 u64 ecount
, etime
, max
, min
;
766 ecount
= get_event_count(event
, vcpu
);
767 etime
= get_event_time(event
, vcpu
);
768 max
= get_event_max(event
, vcpu
);
769 min
= get_event_min(event
, vcpu
);
771 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
772 pr_info("%20s ", decode
);
773 pr_info("%10llu ", (unsigned long long)ecount
);
774 pr_info("%8.2f%% ", (double)ecount
/ kvm
->total_count
* 100);
775 pr_info("%8.2f%% ", (double)etime
/ kvm
->total_time
* 100);
776 pr_info("%8" PRIu64
"us ", min
/ 1000);
777 pr_info("%8" PRIu64
"us ", max
/ 1000);
778 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime
/ ecount
/1e3
,
779 kvm_event_rel_stddev(vcpu
, event
));
783 pr_info("\nTotal Samples:%" PRIu64
", Total events handled time:%.2fus.\n\n",
784 kvm
->total_count
, kvm
->total_time
/ 1e3
);
786 if (kvm
->lost_events
)
787 pr_info("\nLost events: %" PRIu64
"\n\n", kvm
->lost_events
);
790 #ifdef HAVE_TIMERFD_SUPPORT
791 static int process_lost_event(struct perf_tool
*tool
,
792 union perf_event
*event __maybe_unused
,
793 struct perf_sample
*sample __maybe_unused
,
794 struct machine
*machine __maybe_unused
)
796 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
, tool
);
803 static bool skip_sample(struct perf_kvm_stat
*kvm
,
804 struct perf_sample
*sample
)
806 if (kvm
->pid_list
&& intlist__find(kvm
->pid_list
, sample
->pid
) == NULL
)
812 static int process_sample_event(struct perf_tool
*tool
,
813 union perf_event
*event
,
814 struct perf_sample
*sample
,
815 struct perf_evsel
*evsel
,
816 struct machine
*machine
)
818 struct thread
*thread
;
819 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
,
822 if (skip_sample(kvm
, sample
))
825 thread
= machine__findnew_thread(machine
, sample
->pid
, sample
->tid
);
826 if (thread
== NULL
) {
827 pr_debug("problem processing %d event, skipping it.\n",
832 if (!handle_kvm_event(kvm
, thread
, evsel
, sample
))
838 static int cpu_isa_config(struct perf_kvm_stat
*kvm
)
840 char buf
[64], *cpuid
;
844 err
= get_cpuid(buf
, sizeof(buf
));
846 pr_err("Failed to look up CPU type (Intel or AMD)\n");
851 cpuid
= kvm
->session
->header
.env
.cpuid
;
853 if (strstr(cpuid
, "Intel"))
855 else if (strstr(cpuid
, "AMD"))
858 pr_err("CPU %s is not supported.\n", cpuid
);
863 kvm
->exit_reasons
= vmx_exit_reasons
;
864 kvm
->exit_reasons_size
= ARRAY_SIZE(vmx_exit_reasons
);
865 kvm
->exit_reasons_isa
= "VMX";
871 static bool verify_vcpu(int vcpu
)
873 if (vcpu
!= -1 && vcpu
< 0) {
874 pr_err("Invalid vcpu:%d.\n", vcpu
);
881 #ifdef HAVE_TIMERFD_SUPPORT
882 /* keeping the max events to a modest level to keep
883 * the processing of samples per mmap smooth.
885 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
887 static s64
perf_kvm__mmap_read_idx(struct perf_kvm_stat
*kvm
, int idx
,
890 union perf_event
*event
;
891 struct perf_sample sample
;
895 *mmap_time
= ULLONG_MAX
;
896 while ((event
= perf_evlist__mmap_read(kvm
->evlist
, idx
)) != NULL
) {
897 err
= perf_evlist__parse_sample(kvm
->evlist
, event
, &sample
);
899 perf_evlist__mmap_consume(kvm
->evlist
, idx
);
900 pr_err("Failed to parse sample\n");
904 err
= perf_session_queue_event(kvm
->session
, event
, &sample
, 0);
906 * FIXME: Here we can't consume the event, as perf_session_queue_event will
907 * point to it, and it'll get possibly overwritten by the kernel.
909 perf_evlist__mmap_consume(kvm
->evlist
, idx
);
912 pr_err("Failed to enqueue sample: %d\n", err
);
916 /* save time stamp of our first sample for this mmap */
918 *mmap_time
= sample
.time
;
920 /* limit events per mmap handled all at once */
922 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
929 static int perf_kvm__mmap_read(struct perf_kvm_stat
*kvm
)
931 int i
, err
, throttled
= 0;
933 u64 flush_time
= ULLONG_MAX
, mmap_time
;
935 for (i
= 0; i
< kvm
->evlist
->nr_mmaps
; i
++) {
936 n
= perf_kvm__mmap_read_idx(kvm
, i
, &mmap_time
);
940 /* flush time is going to be the minimum of all the individual
941 * mmap times. Essentially, we flush all the samples queued up
942 * from the last pass under our minimal start time -- that leaves
943 * a very small race for samples to come in with a lower timestamp.
944 * The ioctl to return the perf_clock timestamp should close the
947 if (mmap_time
< flush_time
)
948 flush_time
= mmap_time
;
951 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
955 /* flush queue after each round in which we processed events */
957 kvm
->session
->ordered_samples
.next_flush
= flush_time
;
958 err
= kvm
->tool
.finished_round(&kvm
->tool
, NULL
, kvm
->session
);
960 if (kvm
->lost_events
)
961 pr_info("\nLost events: %" PRIu64
"\n\n",
970 static volatile int done
;
972 static void sig_handler(int sig __maybe_unused
)
977 static int perf_kvm__timerfd_create(struct perf_kvm_stat
*kvm
)
979 struct itimerspec new_value
;
982 kvm
->timerfd
= timerfd_create(CLOCK_MONOTONIC
, TFD_NONBLOCK
);
983 if (kvm
->timerfd
< 0) {
984 pr_err("timerfd_create failed\n");
988 new_value
.it_value
.tv_sec
= kvm
->display_time
;
989 new_value
.it_value
.tv_nsec
= 0;
990 new_value
.it_interval
.tv_sec
= kvm
->display_time
;
991 new_value
.it_interval
.tv_nsec
= 0;
993 if (timerfd_settime(kvm
->timerfd
, 0, &new_value
, NULL
) != 0) {
994 pr_err("timerfd_settime failed: %d\n", errno
);
1004 static int perf_kvm__handle_timerfd(struct perf_kvm_stat
*kvm
)
1009 rc
= read(kvm
->timerfd
, &c
, sizeof(uint64_t));
1011 if (errno
== EAGAIN
)
1014 pr_err("Failed to read timer fd: %d\n", errno
);
1018 if (rc
!= sizeof(uint64_t)) {
1019 pr_err("Error reading timer fd - invalid size returned\n");
1024 pr_debug("Missed timer beats: %" PRIu64
"\n", c
-1);
1026 /* update display */
1031 clear_events_cache_stats(kvm
->kvm_events_cache
);
1032 kvm
->total_count
= 0;
1033 kvm
->total_time
= 0;
1034 kvm
->lost_events
= 0;
1039 static int fd_set_nonblock(int fd
)
1043 arg
= fcntl(fd
, F_GETFL
);
1045 pr_err("Failed to get current flags for fd %d\n", fd
);
1049 if (fcntl(fd
, F_SETFL
, arg
| O_NONBLOCK
) < 0) {
1050 pr_err("Failed to set non-block option on fd %d\n", fd
);
1058 int perf_kvm__handle_stdin(struct termios
*tc_now
, struct termios
*tc_save
)
1062 tcsetattr(0, TCSANOW
, tc_now
);
1064 tcsetattr(0, TCSAFLUSH
, tc_save
);
1072 static int kvm_events_live_report(struct perf_kvm_stat
*kvm
)
1074 struct pollfd
*pollfds
= NULL
;
1075 int nr_fds
, nr_stdin
, ret
, err
= -EINVAL
;
1076 struct termios tc
, save
;
1078 /* live flag must be set first */
1081 ret
= cpu_isa_config(kvm
);
1085 if (!verify_vcpu(kvm
->trace_vcpu
) ||
1087 !register_kvm_events_ops(kvm
)) {
1091 init_kvm_event_record(kvm
);
1093 tcgetattr(0, &save
);
1095 tc
.c_lflag
&= ~(ICANON
| ECHO
);
1099 signal(SIGINT
, sig_handler
);
1100 signal(SIGTERM
, sig_handler
);
1102 /* copy pollfds -- need to add timerfd and stdin */
1103 nr_fds
= kvm
->evlist
->nr_fds
;
1104 pollfds
= zalloc(sizeof(struct pollfd
) * (nr_fds
+ 2));
1109 memcpy(pollfds
, kvm
->evlist
->pollfd
,
1110 sizeof(struct pollfd
) * kvm
->evlist
->nr_fds
);
1113 if (perf_kvm__timerfd_create(kvm
) < 0) {
1118 pollfds
[nr_fds
].fd
= kvm
->timerfd
;
1119 pollfds
[nr_fds
].events
= POLLIN
;
1122 pollfds
[nr_fds
].fd
= fileno(stdin
);
1123 pollfds
[nr_fds
].events
= POLLIN
;
1126 if (fd_set_nonblock(fileno(stdin
)) != 0)
1129 /* everything is good - enable the events and process */
1130 perf_evlist__enable(kvm
->evlist
);
1135 rc
= perf_kvm__mmap_read(kvm
);
1139 err
= perf_kvm__handle_timerfd(kvm
);
1143 if (pollfds
[nr_stdin
].revents
& POLLIN
)
1144 done
= perf_kvm__handle_stdin(&tc
, &save
);
1147 err
= poll(pollfds
, nr_fds
, 100);
1150 perf_evlist__disable(kvm
->evlist
);
1158 if (kvm
->timerfd
>= 0)
1159 close(kvm
->timerfd
);
1165 static int kvm_live_open_events(struct perf_kvm_stat
*kvm
)
1168 struct perf_evsel
*pos
;
1169 struct perf_evlist
*evlist
= kvm
->evlist
;
1171 perf_evlist__config(evlist
, &kvm
->opts
);
1174 * Note: exclude_{guest,host} do not apply here.
1175 * This command processes KVM tracepoints from host only
1177 evlist__for_each(evlist
, pos
) {
1178 struct perf_event_attr
*attr
= &pos
->attr
;
1180 /* make sure these *are* set */
1181 perf_evsel__set_sample_bit(pos
, TID
);
1182 perf_evsel__set_sample_bit(pos
, TIME
);
1183 perf_evsel__set_sample_bit(pos
, CPU
);
1184 perf_evsel__set_sample_bit(pos
, RAW
);
1185 /* make sure these are *not*; want as small a sample as possible */
1186 perf_evsel__reset_sample_bit(pos
, PERIOD
);
1187 perf_evsel__reset_sample_bit(pos
, IP
);
1188 perf_evsel__reset_sample_bit(pos
, CALLCHAIN
);
1189 perf_evsel__reset_sample_bit(pos
, ADDR
);
1190 perf_evsel__reset_sample_bit(pos
, READ
);
1195 attr
->sample_period
= 1;
1197 attr
->watermark
= 0;
1198 attr
->wakeup_events
= 1000;
1200 /* will enable all once we are ready */
1204 err
= perf_evlist__open(evlist
);
1206 printf("Couldn't create the events: %s\n", strerror(errno
));
1210 if (perf_evlist__mmap(evlist
, kvm
->opts
.mmap_pages
, false) < 0) {
1211 ui__error("Failed to mmap the events: %s\n", strerror(errno
));
1212 perf_evlist__close(evlist
);
1223 static int read_events(struct perf_kvm_stat
*kvm
)
1227 struct perf_tool eops
= {
1228 .sample
= process_sample_event
,
1229 .comm
= perf_event__process_comm
,
1230 .ordered_samples
= true,
1232 struct perf_data_file file
= {
1233 .path
= kvm
->file_name
,
1234 .mode
= PERF_DATA_MODE_READ
,
1238 kvm
->session
= perf_session__new(&file
, false, &kvm
->tool
);
1239 if (!kvm
->session
) {
1240 pr_err("Initializing perf session failed\n");
1244 if (!perf_session__has_traces(kvm
->session
, "kvm record"))
1248 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1249 * traced in the old kernel.
1251 ret
= cpu_isa_config(kvm
);
1255 return perf_session__process_events(kvm
->session
, &kvm
->tool
);
1258 static int parse_target_str(struct perf_kvm_stat
*kvm
)
1261 kvm
->pid_list
= intlist__new(kvm
->pid_str
);
1262 if (kvm
->pid_list
== NULL
) {
1263 pr_err("Error parsing process id string\n");
1271 static int kvm_events_report_vcpu(struct perf_kvm_stat
*kvm
)
1274 int vcpu
= kvm
->trace_vcpu
;
1276 if (parse_target_str(kvm
) != 0)
1279 if (!verify_vcpu(vcpu
))
1282 if (!select_key(kvm
))
1285 if (!register_kvm_events_ops(kvm
))
1288 init_kvm_event_record(kvm
);
1291 ret
= read_events(kvm
);
1302 static const char * const kvm_events_tp
[] = {
1309 #define STRDUP_FAIL_EXIT(s) \
1318 kvm_events_record(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1320 unsigned int rec_argc
, i
, j
;
1321 const char **rec_argv
;
1322 const char * const record_args
[] = {
1329 rec_argc
= ARRAY_SIZE(record_args
) + argc
+ 2 +
1330 2 * ARRAY_SIZE(kvm_events_tp
);
1331 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1333 if (rec_argv
== NULL
)
1336 for (i
= 0; i
< ARRAY_SIZE(record_args
); i
++)
1337 rec_argv
[i
] = STRDUP_FAIL_EXIT(record_args
[i
]);
1339 for (j
= 0; j
< ARRAY_SIZE(kvm_events_tp
); j
++) {
1340 rec_argv
[i
++] = "-e";
1341 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm_events_tp
[j
]);
1344 rec_argv
[i
++] = STRDUP_FAIL_EXIT("-o");
1345 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm
->file_name
);
1347 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
1348 rec_argv
[i
] = argv
[j
];
1350 return cmd_record(i
, rec_argv
, NULL
);
1354 kvm_events_report(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1356 const struct option kvm_events_report_options
[] = {
1357 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1358 "event for reporting: vmexit, mmio, ioport"),
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_STRING('p', "pid", &kvm
->pid_str
, "pid",
1365 "analyze events only for given process id(s)"),
1369 const char * const kvm_events_report_usage
[] = {
1370 "perf kvm stat report [<options>]",
1377 argc
= parse_options(argc
, argv
,
1378 kvm_events_report_options
,
1379 kvm_events_report_usage
, 0);
1381 usage_with_options(kvm_events_report_usage
,
1382 kvm_events_report_options
);
1385 return kvm_events_report_vcpu(kvm
);
1388 #ifdef HAVE_TIMERFD_SUPPORT
1389 static struct perf_evlist
*kvm_live_event_list(void)
1391 struct perf_evlist
*evlist
;
1392 char *tp
, *name
, *sys
;
1396 evlist
= perf_evlist__new();
1400 for (j
= 0; j
< ARRAY_SIZE(kvm_events_tp
); j
++) {
1402 tp
= strdup(kvm_events_tp
[j
]);
1406 /* split tracepoint into subsystem and name */
1408 name
= strchr(tp
, ':');
1410 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1418 if (perf_evlist__add_newtp(evlist
, sys
, name
, NULL
)) {
1419 pr_err("Failed to add %s tracepoint to the list\n", kvm_events_tp
[j
]);
1431 perf_evlist__delete(evlist
);
1438 static int kvm_events_live(struct perf_kvm_stat
*kvm
,
1439 int argc
, const char **argv
)
1441 char errbuf
[BUFSIZ
];
1444 const struct option live_options
[] = {
1445 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1446 "record events on existing process id"),
1447 OPT_CALLBACK('m', "mmap-pages", &kvm
->opts
.mmap_pages
, "pages",
1448 "number of mmap data pages",
1449 perf_evlist__parse_mmap_pages
),
1450 OPT_INCR('v', "verbose", &verbose
,
1451 "be more verbose (show counter open errors, etc)"),
1452 OPT_BOOLEAN('a', "all-cpus", &kvm
->opts
.target
.system_wide
,
1453 "system-wide collection from all CPUs"),
1454 OPT_UINTEGER('d', "display", &kvm
->display_time
,
1455 "time in seconds between display updates"),
1456 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1457 "event for reporting: vmexit, mmio, ioport"),
1458 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1459 "vcpu id to report"),
1460 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1461 "key for sorting: sample(sort by samples number)"
1462 " time (sort by avg time)"),
1463 OPT_U64(0, "duration", &kvm
->duration
,
1464 "show events other than HALT that take longer than duration usecs"),
1467 const char * const live_usage
[] = {
1468 "perf kvm stat live [<options>]",
1471 struct perf_data_file file
= {
1472 .mode
= PERF_DATA_MODE_WRITE
,
1476 /* event handling */
1477 kvm
->tool
.sample
= process_sample_event
;
1478 kvm
->tool
.comm
= perf_event__process_comm
;
1479 kvm
->tool
.exit
= perf_event__process_exit
;
1480 kvm
->tool
.fork
= perf_event__process_fork
;
1481 kvm
->tool
.lost
= process_lost_event
;
1482 kvm
->tool
.ordered_samples
= true;
1483 perf_tool__fill_defaults(&kvm
->tool
);
1486 kvm
->display_time
= 1;
1487 kvm
->opts
.user_interval
= 1;
1488 kvm
->opts
.mmap_pages
= 512;
1489 kvm
->opts
.target
.uses_mmap
= false;
1490 kvm
->opts
.target
.uid_str
= NULL
;
1491 kvm
->opts
.target
.uid
= UINT_MAX
;
1494 disable_buildid_cache();
1497 setup_browser(false);
1500 argc
= parse_options(argc
, argv
, live_options
,
1503 usage_with_options(live_usage
, live_options
);
1506 kvm
->duration
*= NSEC_PER_USEC
; /* convert usec to nsec */
1509 * target related setups
1511 err
= target__validate(&kvm
->opts
.target
);
1513 target__strerror(&kvm
->opts
.target
, err
, errbuf
, BUFSIZ
);
1514 ui__warning("%s", errbuf
);
1517 if (target__none(&kvm
->opts
.target
))
1518 kvm
->opts
.target
.system_wide
= true;
1522 * generate the event list
1524 kvm
->evlist
= kvm_live_event_list();
1525 if (kvm
->evlist
== NULL
) {
1530 symbol_conf
.nr_events
= kvm
->evlist
->nr_entries
;
1532 if (perf_evlist__create_maps(kvm
->evlist
, &kvm
->opts
.target
) < 0)
1533 usage_with_options(live_usage
, live_options
);
1538 kvm
->session
= perf_session__new(&file
, false, &kvm
->tool
);
1539 if (kvm
->session
== NULL
) {
1543 kvm
->session
->evlist
= kvm
->evlist
;
1544 perf_session__set_id_hdr_size(kvm
->session
);
1545 machine__synthesize_threads(&kvm
->session
->machines
.host
, &kvm
->opts
.target
,
1546 kvm
->evlist
->threads
, false);
1547 err
= kvm_live_open_events(kvm
);
1551 err
= kvm_events_live_report(kvm
);
1557 perf_session__delete(kvm
->session
);
1558 kvm
->session
= NULL
;
1560 perf_evlist__delete(kvm
->evlist
);
1566 static void print_kvm_stat_usage(void)
1568 printf("Usage: perf kvm stat <command>\n\n");
1570 printf("# Available commands:\n");
1571 printf("\trecord: record kvm events\n");
1572 printf("\treport: report statistical data of kvm events\n");
1573 printf("\tlive: live reporting of statistical data of kvm events\n");
1575 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1578 static int kvm_cmd_stat(const char *file_name
, int argc
, const char **argv
)
1580 struct perf_kvm_stat kvm
= {
1581 .file_name
= file_name
,
1584 .report_event
= "vmexit",
1585 .sort_key
= "sample",
1587 .exit_reasons
= svm_exit_reasons
,
1588 .exit_reasons_size
= ARRAY_SIZE(svm_exit_reasons
),
1589 .exit_reasons_isa
= "SVM",
1593 print_kvm_stat_usage();
1597 if (!strncmp(argv
[1], "rec", 3))
1598 return kvm_events_record(&kvm
, argc
- 1, argv
+ 1);
1600 if (!strncmp(argv
[1], "rep", 3))
1601 return kvm_events_report(&kvm
, argc
- 1 , argv
+ 1);
1603 #ifdef HAVE_TIMERFD_SUPPORT
1604 if (!strncmp(argv
[1], "live", 4))
1605 return kvm_events_live(&kvm
, argc
- 1 , argv
+ 1);
1609 return cmd_stat(argc
, argv
, NULL
);
1613 static int __cmd_record(const char *file_name
, int argc
, const char **argv
)
1615 int rec_argc
, i
= 0, j
;
1616 const char **rec_argv
;
1618 rec_argc
= argc
+ 2;
1619 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1620 rec_argv
[i
++] = strdup("record");
1621 rec_argv
[i
++] = strdup("-o");
1622 rec_argv
[i
++] = strdup(file_name
);
1623 for (j
= 1; j
< argc
; j
++, i
++)
1624 rec_argv
[i
] = argv
[j
];
1626 BUG_ON(i
!= rec_argc
);
1628 return cmd_record(i
, rec_argv
, NULL
);
1631 static int __cmd_report(const char *file_name
, int argc
, const char **argv
)
1633 int rec_argc
, i
= 0, j
;
1634 const char **rec_argv
;
1636 rec_argc
= argc
+ 2;
1637 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1638 rec_argv
[i
++] = strdup("report");
1639 rec_argv
[i
++] = strdup("-i");
1640 rec_argv
[i
++] = strdup(file_name
);
1641 for (j
= 1; j
< argc
; j
++, i
++)
1642 rec_argv
[i
] = argv
[j
];
1644 BUG_ON(i
!= rec_argc
);
1646 return cmd_report(i
, rec_argv
, NULL
);
1650 __cmd_buildid_list(const char *file_name
, int argc
, const char **argv
)
1652 int rec_argc
, i
= 0, j
;
1653 const char **rec_argv
;
1655 rec_argc
= argc
+ 2;
1656 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1657 rec_argv
[i
++] = strdup("buildid-list");
1658 rec_argv
[i
++] = strdup("-i");
1659 rec_argv
[i
++] = strdup(file_name
);
1660 for (j
= 1; j
< argc
; j
++, i
++)
1661 rec_argv
[i
] = argv
[j
];
1663 BUG_ON(i
!= rec_argc
);
1665 return cmd_buildid_list(i
, rec_argv
, NULL
);
1668 int cmd_kvm(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1670 const char *file_name
= NULL
;
1671 const struct option kvm_options
[] = {
1672 OPT_STRING('i', "input", &file_name
, "file",
1674 OPT_STRING('o', "output", &file_name
, "file",
1675 "Output file name"),
1676 OPT_BOOLEAN(0, "guest", &perf_guest
,
1677 "Collect guest os data"),
1678 OPT_BOOLEAN(0, "host", &perf_host
,
1679 "Collect host os data"),
1680 OPT_STRING(0, "guestmount", &symbol_conf
.guestmount
, "directory",
1681 "guest mount directory under which every guest os"
1682 " instance has a subdir"),
1683 OPT_STRING(0, "guestvmlinux", &symbol_conf
.default_guest_vmlinux_name
,
1684 "file", "file saving guest os vmlinux"),
1685 OPT_STRING(0, "guestkallsyms", &symbol_conf
.default_guest_kallsyms
,
1686 "file", "file saving guest os /proc/kallsyms"),
1687 OPT_STRING(0, "guestmodules", &symbol_conf
.default_guest_modules
,
1688 "file", "file saving guest os /proc/modules"),
1689 OPT_INCR('v', "verbose", &verbose
,
1690 "be more verbose (show counter open errors, etc)"),
1695 const char * const kvm_usage
[] = {
1696 "perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
1703 argc
= parse_options(argc
, argv
, kvm_options
, kvm_usage
,
1704 PARSE_OPT_STOP_AT_NON_OPTION
);
1706 usage_with_options(kvm_usage
, kvm_options
);
1712 file_name
= get_filename_for_perf_kvm();
1715 pr_err("Failed to allocate memory for filename\n");
1720 if (!strncmp(argv
[0], "rec", 3))
1721 return __cmd_record(file_name
, argc
, argv
);
1722 else if (!strncmp(argv
[0], "rep", 3))
1723 return __cmd_report(file_name
, argc
, argv
);
1724 else if (!strncmp(argv
[0], "diff", 4))
1725 return cmd_diff(argc
, argv
, NULL
);
1726 else if (!strncmp(argv
[0], "top", 3))
1727 return cmd_top(argc
, argv
, NULL
);
1728 else if (!strncmp(argv
[0], "buildid-list", 12))
1729 return __cmd_buildid_list(file_name
, argc
, argv
);
1730 #if defined(__i386__) || defined(__x86_64__)
1731 else if (!strncmp(argv
[0], "stat", 4))
1732 return kvm_cmd_stat(file_name
, argc
, argv
);
1735 usage_with_options(kvm_usage
, kvm_options
);