1 // SPDX-License-Identifier: GPL-2.0
5 #include "util/build-id.h"
6 #include "util/evsel.h"
7 #include "util/evlist.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/intlist.h"
15 #include <subcmd/pager.h>
16 #include <subcmd/parse-options.h>
17 #include "util/trace-event.h"
18 #include "util/debug.h"
19 #include "util/tool.h"
20 #include "util/stat.h"
21 #include "util/synthetic-events.h"
23 #include "util/data.h"
24 #include "util/ordered-events.h"
25 #include "util/kvm-stat.h"
28 #include <sys/prctl.h>
29 #ifdef HAVE_TIMERFD_SUPPORT
30 #include <sys/timerfd.h>
33 #include <sys/types.h>
37 #include <linux/err.h>
38 #include <linux/kernel.h>
39 #include <linux/string.h>
40 #include <linux/time64.h>
41 #include <linux/zalloc.h>
46 #include <semaphore.h>
49 #include <perf/mmap.h>
51 static const char *get_filename_for_perf_kvm(void)
55 if (perf_host
&& !perf_guest
)
56 filename
= strdup("perf.data.host");
57 else if (!perf_host
&& perf_guest
)
58 filename
= strdup("perf.data.guest");
60 filename
= strdup("perf.data.kvm");
65 #ifdef HAVE_KVM_STAT_SUPPORT
67 void exit_event_get_key(struct evsel
*evsel
,
68 struct perf_sample
*sample
,
69 struct event_key
*key
)
72 key
->key
= evsel__intval(evsel
, sample
, kvm_exit_reason
);
75 bool kvm_exit_event(struct evsel
*evsel
)
77 return !strcmp(evsel
->name
, kvm_exit_trace
);
80 bool exit_event_begin(struct evsel
*evsel
,
81 struct perf_sample
*sample
, struct event_key
*key
)
83 if (kvm_exit_event(evsel
)) {
84 exit_event_get_key(evsel
, sample
, key
);
91 bool kvm_entry_event(struct evsel
*evsel
)
93 return !strcmp(evsel
->name
, kvm_entry_trace
);
96 bool exit_event_end(struct evsel
*evsel
,
97 struct perf_sample
*sample __maybe_unused
,
98 struct event_key
*key __maybe_unused
)
100 return kvm_entry_event(evsel
);
103 static const char *get_exit_reason(struct perf_kvm_stat
*kvm
,
104 struct exit_reasons_table
*tbl
,
107 while (tbl
->reason
!= NULL
) {
108 if (tbl
->exit_code
== exit_code
)
113 pr_err("unknown kvm exit code:%lld on %s\n",
114 (unsigned long long)exit_code
, kvm
->exit_reasons_isa
);
118 void exit_event_decode_key(struct perf_kvm_stat
*kvm
,
119 struct event_key
*key
,
122 const char *exit_reason
= get_exit_reason(kvm
, key
->exit_reasons
,
125 scnprintf(decode
, decode_str_len
, "%s", exit_reason
);
128 static bool register_kvm_events_ops(struct perf_kvm_stat
*kvm
)
130 struct kvm_reg_events_ops
*events_ops
= kvm_reg_events_ops
;
132 for (events_ops
= kvm_reg_events_ops
; events_ops
->name
; events_ops
++) {
133 if (!strcmp(events_ops
->name
, kvm
->report_event
)) {
134 kvm
->events_ops
= events_ops
->ops
;
142 struct vcpu_event_record
{
145 struct kvm_event
*last_event
;
149 static void init_kvm_event_record(struct perf_kvm_stat
*kvm
)
153 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++)
154 INIT_LIST_HEAD(&kvm
->kvm_events_cache
[i
]);
157 #ifdef HAVE_TIMERFD_SUPPORT
158 static void clear_events_cache_stats(struct list_head
*kvm_events_cache
)
160 struct list_head
*head
;
161 struct kvm_event
*event
;
165 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++) {
166 head
= &kvm_events_cache
[i
];
167 list_for_each_entry(event
, head
, hash_entry
) {
168 /* reset stats for event */
169 event
->total
.time
= 0;
170 init_stats(&event
->total
.stats
);
172 for (j
= 0; j
< event
->max_vcpu
; ++j
) {
173 event
->vcpu
[j
].time
= 0;
174 init_stats(&event
->vcpu
[j
].stats
);
181 static int kvm_events_hash_fn(u64 key
)
183 return key
& (EVENTS_CACHE_SIZE
- 1);
186 static bool kvm_event_expand(struct kvm_event
*event
, int vcpu_id
)
188 int old_max_vcpu
= event
->max_vcpu
;
191 if (vcpu_id
< event
->max_vcpu
)
194 while (event
->max_vcpu
<= vcpu_id
)
195 event
->max_vcpu
+= DEFAULT_VCPU_NUM
;
198 event
->vcpu
= realloc(event
->vcpu
,
199 event
->max_vcpu
* sizeof(*event
->vcpu
));
202 pr_err("Not enough memory\n");
206 memset(event
->vcpu
+ old_max_vcpu
, 0,
207 (event
->max_vcpu
- old_max_vcpu
) * sizeof(*event
->vcpu
));
211 static struct kvm_event
*kvm_alloc_init_event(struct event_key
*key
)
213 struct kvm_event
*event
;
215 event
= zalloc(sizeof(*event
));
217 pr_err("Not enough memory\n");
222 init_stats(&event
->total
.stats
);
226 static struct kvm_event
*find_create_kvm_event(struct perf_kvm_stat
*kvm
,
227 struct event_key
*key
)
229 struct kvm_event
*event
;
230 struct list_head
*head
;
232 BUG_ON(key
->key
== INVALID_KEY
);
234 head
= &kvm
->kvm_events_cache
[kvm_events_hash_fn(key
->key
)];
235 list_for_each_entry(event
, head
, hash_entry
) {
236 if (event
->key
.key
== key
->key
&& event
->key
.info
== key
->info
)
240 event
= kvm_alloc_init_event(key
);
244 list_add(&event
->hash_entry
, head
);
248 static bool handle_begin_event(struct perf_kvm_stat
*kvm
,
249 struct vcpu_event_record
*vcpu_record
,
250 struct event_key
*key
, u64 timestamp
)
252 struct kvm_event
*event
= NULL
;
254 if (key
->key
!= INVALID_KEY
)
255 event
= find_create_kvm_event(kvm
, key
);
257 vcpu_record
->last_event
= event
;
258 vcpu_record
->start_time
= timestamp
;
263 kvm_update_event_stats(struct kvm_event_stats
*kvm_stats
, u64 time_diff
)
265 kvm_stats
->time
+= time_diff
;
266 update_stats(&kvm_stats
->stats
, time_diff
);
269 static double kvm_event_rel_stddev(int vcpu_id
, struct kvm_event
*event
)
271 struct kvm_event_stats
*kvm_stats
= &event
->total
;
274 kvm_stats
= &event
->vcpu
[vcpu_id
];
276 return rel_stddev_stats(stddev_stats(&kvm_stats
->stats
),
277 avg_stats(&kvm_stats
->stats
));
280 static bool update_kvm_event(struct kvm_event
*event
, int vcpu_id
,
284 kvm_update_event_stats(&event
->total
, time_diff
);
288 if (!kvm_event_expand(event
, vcpu_id
))
291 kvm_update_event_stats(&event
->vcpu
[vcpu_id
], time_diff
);
295 static bool is_child_event(struct perf_kvm_stat
*kvm
,
297 struct perf_sample
*sample
,
298 struct event_key
*key
)
300 struct child_event_ops
*child_ops
;
302 child_ops
= kvm
->events_ops
->child_ops
;
307 for (; child_ops
->name
; child_ops
++) {
308 if (!strcmp(evsel
->name
, child_ops
->name
)) {
309 child_ops
->get_key(evsel
, sample
, key
);
317 static bool handle_child_event(struct perf_kvm_stat
*kvm
,
318 struct vcpu_event_record
*vcpu_record
,
319 struct event_key
*key
,
320 struct perf_sample
*sample __maybe_unused
)
322 struct kvm_event
*event
= NULL
;
324 if (key
->key
!= INVALID_KEY
)
325 event
= find_create_kvm_event(kvm
, key
);
327 vcpu_record
->last_event
= event
;
332 static bool skip_event(const char *event
)
334 const char * const *skip_events
;
336 for (skip_events
= kvm_skip_events
; *skip_events
; skip_events
++)
337 if (!strcmp(event
, *skip_events
))
343 static bool handle_end_event(struct perf_kvm_stat
*kvm
,
344 struct vcpu_event_record
*vcpu_record
,
345 struct event_key
*key
,
346 struct perf_sample
*sample
)
348 struct kvm_event
*event
;
349 u64 time_begin
, time_diff
;
352 if (kvm
->trace_vcpu
== -1)
355 vcpu
= vcpu_record
->vcpu_id
;
357 event
= vcpu_record
->last_event
;
358 time_begin
= vcpu_record
->start_time
;
360 /* The begin event is not caught. */
365 * In some case, the 'begin event' only records the start timestamp,
366 * the actual event is recognized in the 'end event' (e.g. mmio-event).
369 /* Both begin and end events did not get the key. */
370 if (!event
&& key
->key
== INVALID_KEY
)
374 event
= find_create_kvm_event(kvm
, key
);
379 vcpu_record
->last_event
= NULL
;
380 vcpu_record
->start_time
= 0;
382 /* seems to happen once in a while during live mode */
383 if (sample
->time
< time_begin
) {
384 pr_debug("End time before begin time; skipping event.\n");
388 time_diff
= sample
->time
- time_begin
;
390 if (kvm
->duration
&& time_diff
> kvm
->duration
) {
391 char decode
[decode_str_len
];
393 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
394 if (!skip_event(decode
)) {
395 pr_info("%" PRIu64
" VM %d, vcpu %d: %s event took %" PRIu64
"usec\n",
396 sample
->time
, sample
->pid
, vcpu_record
->vcpu_id
,
397 decode
, time_diff
/ NSEC_PER_USEC
);
401 return update_kvm_event(event
, vcpu
, time_diff
);
405 struct vcpu_event_record
*per_vcpu_record(struct thread
*thread
,
407 struct perf_sample
*sample
)
409 /* Only kvm_entry records vcpu id. */
410 if (!thread__priv(thread
) && kvm_entry_event(evsel
)) {
411 struct vcpu_event_record
*vcpu_record
;
413 vcpu_record
= zalloc(sizeof(*vcpu_record
));
415 pr_err("%s: Not enough memory\n", __func__
);
419 vcpu_record
->vcpu_id
= evsel__intval(evsel
, sample
, vcpu_id_str
);
420 thread__set_priv(thread
, vcpu_record
);
423 return thread__priv(thread
);
426 static bool handle_kvm_event(struct perf_kvm_stat
*kvm
,
427 struct thread
*thread
,
429 struct perf_sample
*sample
)
431 struct vcpu_event_record
*vcpu_record
;
432 struct event_key key
= { .key
= INVALID_KEY
,
433 .exit_reasons
= kvm
->exit_reasons
};
435 vcpu_record
= per_vcpu_record(thread
, evsel
, sample
);
439 /* only process events for vcpus user cares about */
440 if ((kvm
->trace_vcpu
!= -1) &&
441 (kvm
->trace_vcpu
!= vcpu_record
->vcpu_id
))
444 if (kvm
->events_ops
->is_begin_event(evsel
, sample
, &key
))
445 return handle_begin_event(kvm
, vcpu_record
, &key
, sample
->time
);
447 if (is_child_event(kvm
, evsel
, sample
, &key
))
448 return handle_child_event(kvm
, vcpu_record
, &key
, sample
);
450 if (kvm
->events_ops
->is_end_event(evsel
, sample
, &key
))
451 return handle_end_event(kvm
, vcpu_record
, &key
, sample
);
456 #define GET_EVENT_KEY(func, field) \
457 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
460 return event->total.field; \
462 if (vcpu >= event->max_vcpu) \
465 return event->vcpu[vcpu].field; \
468 #define COMPARE_EVENT_KEY(func, field) \
469 GET_EVENT_KEY(func, field) \
470 static int compare_kvm_event_ ## func(struct kvm_event *one, \
471 struct kvm_event *two, int vcpu)\
473 return get_event_ ##func(one, vcpu) > \
474 get_event_ ##func(two, vcpu); \
477 GET_EVENT_KEY(time
, time
);
478 COMPARE_EVENT_KEY(count
, stats
.n
);
479 COMPARE_EVENT_KEY(mean
, stats
.mean
);
480 GET_EVENT_KEY(max
, stats
.max
);
481 GET_EVENT_KEY(min
, stats
.min
);
483 #define DEF_SORT_NAME_KEY(name, compare_key) \
484 { #name, compare_kvm_event_ ## compare_key }
486 static struct kvm_event_key keys
[] = {
487 DEF_SORT_NAME_KEY(sample
, count
),
488 DEF_SORT_NAME_KEY(time
, mean
),
492 static bool select_key(struct perf_kvm_stat
*kvm
)
496 for (i
= 0; keys
[i
].name
; i
++) {
497 if (!strcmp(keys
[i
].name
, kvm
->sort_key
)) {
498 kvm
->compare
= keys
[i
].key
;
503 pr_err("Unknown compare key:%s\n", kvm
->sort_key
);
507 static void insert_to_result(struct rb_root
*result
, struct kvm_event
*event
,
508 key_cmp_fun bigger
, int vcpu
)
510 struct rb_node
**rb
= &result
->rb_node
;
511 struct rb_node
*parent
= NULL
;
515 p
= container_of(*rb
, struct kvm_event
, rb
);
518 if (bigger(event
, p
, vcpu
))
519 rb
= &(*rb
)->rb_left
;
521 rb
= &(*rb
)->rb_right
;
524 rb_link_node(&event
->rb
, parent
, rb
);
525 rb_insert_color(&event
->rb
, result
);
529 update_total_count(struct perf_kvm_stat
*kvm
, struct kvm_event
*event
)
531 int vcpu
= kvm
->trace_vcpu
;
533 kvm
->total_count
+= get_event_count(event
, vcpu
);
534 kvm
->total_time
+= get_event_time(event
, vcpu
);
537 static bool event_is_valid(struct kvm_event
*event
, int vcpu
)
539 return !!get_event_count(event
, vcpu
);
542 static void sort_result(struct perf_kvm_stat
*kvm
)
545 int vcpu
= kvm
->trace_vcpu
;
546 struct kvm_event
*event
;
548 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++) {
549 list_for_each_entry(event
, &kvm
->kvm_events_cache
[i
], hash_entry
) {
550 if (event_is_valid(event
, vcpu
)) {
551 update_total_count(kvm
, event
);
552 insert_to_result(&kvm
->result
, event
,
559 /* returns left most element of result, and erase it */
560 static struct kvm_event
*pop_from_result(struct rb_root
*result
)
562 struct rb_node
*node
= rb_first(result
);
567 rb_erase(node
, result
);
568 return container_of(node
, struct kvm_event
, rb
);
571 static void print_vcpu_info(struct perf_kvm_stat
*kvm
)
573 int vcpu
= kvm
->trace_vcpu
;
575 pr_info("Analyze events for ");
577 if (kvm
->opts
.target
.system_wide
)
578 pr_info("all VMs, ");
579 else if (kvm
->opts
.target
.pid
)
580 pr_info("pid(s) %s, ", kvm
->opts
.target
.pid
);
582 pr_info("dazed and confused on what is monitored, ");
585 pr_info("all VCPUs:\n\n");
587 pr_info("VCPU %d:\n\n", vcpu
);
590 static void show_timeofday(void)
596 gettimeofday(&tv
, NULL
);
597 if (localtime_r(&tv
.tv_sec
, <ime
)) {
598 strftime(date
, sizeof(date
), "%H:%M:%S", <ime
);
599 pr_info("%s.%06ld", date
, tv
.tv_usec
);
601 pr_info("00:00:00.000000");
606 static void print_result(struct perf_kvm_stat
*kvm
)
608 char decode
[decode_str_len
];
609 struct kvm_event
*event
;
610 int vcpu
= kvm
->trace_vcpu
;
618 print_vcpu_info(kvm
);
619 pr_info("%*s ", decode_str_len
, kvm
->events_ops
->name
);
620 pr_info("%10s ", "Samples");
621 pr_info("%9s ", "Samples%");
623 pr_info("%9s ", "Time%");
624 pr_info("%11s ", "Min Time");
625 pr_info("%11s ", "Max Time");
626 pr_info("%16s ", "Avg time");
629 while ((event
= pop_from_result(&kvm
->result
))) {
630 u64 ecount
, etime
, max
, min
;
632 ecount
= get_event_count(event
, vcpu
);
633 etime
= get_event_time(event
, vcpu
);
634 max
= get_event_max(event
, vcpu
);
635 min
= get_event_min(event
, vcpu
);
637 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
638 pr_info("%*s ", decode_str_len
, decode
);
639 pr_info("%10llu ", (unsigned long long)ecount
);
640 pr_info("%8.2f%% ", (double)ecount
/ kvm
->total_count
* 100);
641 pr_info("%8.2f%% ", (double)etime
/ kvm
->total_time
* 100);
642 pr_info("%9.2fus ", (double)min
/ NSEC_PER_USEC
);
643 pr_info("%9.2fus ", (double)max
/ NSEC_PER_USEC
);
644 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime
/ ecount
/ NSEC_PER_USEC
,
645 kvm_event_rel_stddev(vcpu
, event
));
649 pr_info("\nTotal Samples:%" PRIu64
", Total events handled time:%.2fus.\n\n",
650 kvm
->total_count
, kvm
->total_time
/ (double)NSEC_PER_USEC
);
652 if (kvm
->lost_events
)
653 pr_info("\nLost events: %" PRIu64
"\n\n", kvm
->lost_events
);
656 #ifdef HAVE_TIMERFD_SUPPORT
657 static int process_lost_event(struct perf_tool
*tool
,
658 union perf_event
*event __maybe_unused
,
659 struct perf_sample
*sample __maybe_unused
,
660 struct machine
*machine __maybe_unused
)
662 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
, tool
);
669 static bool skip_sample(struct perf_kvm_stat
*kvm
,
670 struct perf_sample
*sample
)
672 if (kvm
->pid_list
&& intlist__find(kvm
->pid_list
, sample
->pid
) == NULL
)
678 static int process_sample_event(struct perf_tool
*tool
,
679 union perf_event
*event
,
680 struct perf_sample
*sample
,
682 struct machine
*machine
)
685 struct thread
*thread
;
686 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
,
689 if (skip_sample(kvm
, sample
))
692 thread
= machine__findnew_thread(machine
, sample
->pid
, sample
->tid
);
693 if (thread
== NULL
) {
694 pr_debug("problem processing %d event, skipping it.\n",
699 if (!handle_kvm_event(kvm
, thread
, evsel
, sample
))
706 static int cpu_isa_config(struct perf_kvm_stat
*kvm
)
708 char buf
[128], *cpuid
;
712 err
= get_cpuid(buf
, sizeof(buf
));
714 pr_err("Failed to look up CPU type: %s\n",
715 str_error_r(err
, buf
, sizeof(buf
)));
720 cpuid
= kvm
->session
->header
.env
.cpuid
;
723 pr_err("Failed to look up CPU type\n");
727 err
= cpu_isa_init(kvm
, cpuid
);
729 pr_err("CPU %s is not supported.\n", cpuid
);
734 static bool verify_vcpu(int vcpu
)
736 if (vcpu
!= -1 && vcpu
< 0) {
737 pr_err("Invalid vcpu:%d.\n", vcpu
);
744 #ifdef HAVE_TIMERFD_SUPPORT
745 /* keeping the max events to a modest level to keep
746 * the processing of samples per mmap smooth.
748 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
750 static s64
perf_kvm__mmap_read_idx(struct perf_kvm_stat
*kvm
, int idx
,
753 struct evlist
*evlist
= kvm
->evlist
;
754 union perf_event
*event
;
760 *mmap_time
= ULLONG_MAX
;
761 md
= &evlist
->mmap
[idx
];
762 err
= perf_mmap__read_init(&md
->core
);
764 return (err
== -EAGAIN
) ? 0 : -1;
766 while ((event
= perf_mmap__read_event(&md
->core
)) != NULL
) {
767 err
= perf_evlist__parse_sample_timestamp(evlist
, event
, ×tamp
);
769 perf_mmap__consume(&md
->core
);
770 pr_err("Failed to parse sample\n");
774 err
= perf_session__queue_event(kvm
->session
, event
, timestamp
, 0);
776 * FIXME: Here we can't consume the event, as perf_session__queue_event will
777 * point to it, and it'll get possibly overwritten by the kernel.
779 perf_mmap__consume(&md
->core
);
782 pr_err("Failed to enqueue sample: %d\n", err
);
786 /* save time stamp of our first sample for this mmap */
788 *mmap_time
= timestamp
;
790 /* limit events per mmap handled all at once */
792 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
796 perf_mmap__read_done(&md
->core
);
800 static int perf_kvm__mmap_read(struct perf_kvm_stat
*kvm
)
802 int i
, err
, throttled
= 0;
804 u64 flush_time
= ULLONG_MAX
, mmap_time
;
806 for (i
= 0; i
< kvm
->evlist
->core
.nr_mmaps
; i
++) {
807 n
= perf_kvm__mmap_read_idx(kvm
, i
, &mmap_time
);
811 /* flush time is going to be the minimum of all the individual
812 * mmap times. Essentially, we flush all the samples queued up
813 * from the last pass under our minimal start time -- that leaves
814 * a very small race for samples to come in with a lower timestamp.
815 * The ioctl to return the perf_clock timestamp should close the
818 if (mmap_time
< flush_time
)
819 flush_time
= mmap_time
;
822 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
826 /* flush queue after each round in which we processed events */
828 struct ordered_events
*oe
= &kvm
->session
->ordered_events
;
830 oe
->next_flush
= flush_time
;
831 err
= ordered_events__flush(oe
, OE_FLUSH__ROUND
);
833 if (kvm
->lost_events
)
834 pr_info("\nLost events: %" PRIu64
"\n\n",
843 static volatile int done
;
845 static void sig_handler(int sig __maybe_unused
)
850 static int perf_kvm__timerfd_create(struct perf_kvm_stat
*kvm
)
852 struct itimerspec new_value
;
855 kvm
->timerfd
= timerfd_create(CLOCK_MONOTONIC
, TFD_NONBLOCK
);
856 if (kvm
->timerfd
< 0) {
857 pr_err("timerfd_create failed\n");
861 new_value
.it_value
.tv_sec
= kvm
->display_time
;
862 new_value
.it_value
.tv_nsec
= 0;
863 new_value
.it_interval
.tv_sec
= kvm
->display_time
;
864 new_value
.it_interval
.tv_nsec
= 0;
866 if (timerfd_settime(kvm
->timerfd
, 0, &new_value
, NULL
) != 0) {
867 pr_err("timerfd_settime failed: %d\n", errno
);
877 static int perf_kvm__handle_timerfd(struct perf_kvm_stat
*kvm
)
882 rc
= read(kvm
->timerfd
, &c
, sizeof(uint64_t));
887 pr_err("Failed to read timer fd: %d\n", errno
);
891 if (rc
!= sizeof(uint64_t)) {
892 pr_err("Error reading timer fd - invalid size returned\n");
897 pr_debug("Missed timer beats: %" PRIu64
"\n", c
-1);
904 clear_events_cache_stats(kvm
->kvm_events_cache
);
905 kvm
->total_count
= 0;
907 kvm
->lost_events
= 0;
912 static int fd_set_nonblock(int fd
)
916 arg
= fcntl(fd
, F_GETFL
);
918 pr_err("Failed to get current flags for fd %d\n", fd
);
922 if (fcntl(fd
, F_SETFL
, arg
| O_NONBLOCK
) < 0) {
923 pr_err("Failed to set non-block option on fd %d\n", fd
);
930 static int perf_kvm__handle_stdin(void)
941 static int kvm_events_live_report(struct perf_kvm_stat
*kvm
)
943 int nr_stdin
, ret
, err
= -EINVAL
;
946 /* live flag must be set first */
949 ret
= cpu_isa_config(kvm
);
953 if (!verify_vcpu(kvm
->trace_vcpu
) ||
955 !register_kvm_events_ops(kvm
)) {
959 set_term_quiet_input(&save
);
960 init_kvm_event_record(kvm
);
962 signal(SIGINT
, sig_handler
);
963 signal(SIGTERM
, sig_handler
);
966 if (perf_kvm__timerfd_create(kvm
) < 0) {
971 if (evlist__add_pollfd(kvm
->evlist
, kvm
->timerfd
) < 0)
974 nr_stdin
= evlist__add_pollfd(kvm
->evlist
, fileno(stdin
));
978 if (fd_set_nonblock(fileno(stdin
)) != 0)
981 /* everything is good - enable the events and process */
982 evlist__enable(kvm
->evlist
);
985 struct fdarray
*fda
= &kvm
->evlist
->core
.pollfd
;
988 rc
= perf_kvm__mmap_read(kvm
);
992 err
= perf_kvm__handle_timerfd(kvm
);
996 if (fda
->entries
[nr_stdin
].revents
& POLLIN
)
997 done
= perf_kvm__handle_stdin();
1000 err
= evlist__poll(kvm
->evlist
, 100);
1003 evlist__disable(kvm
->evlist
);
1011 if (kvm
->timerfd
>= 0)
1012 close(kvm
->timerfd
);
1014 tcsetattr(0, TCSAFLUSH
, &save
);
1018 static int kvm_live_open_events(struct perf_kvm_stat
*kvm
)
1022 struct evlist
*evlist
= kvm
->evlist
;
1023 char sbuf
[STRERR_BUFSIZE
];
1025 perf_evlist__config(evlist
, &kvm
->opts
, NULL
);
1028 * Note: exclude_{guest,host} do not apply here.
1029 * This command processes KVM tracepoints from host only
1031 evlist__for_each_entry(evlist
, pos
) {
1032 struct perf_event_attr
*attr
= &pos
->core
.attr
;
1034 /* make sure these *are* set */
1035 evsel__set_sample_bit(pos
, TID
);
1036 evsel__set_sample_bit(pos
, TIME
);
1037 evsel__set_sample_bit(pos
, CPU
);
1038 evsel__set_sample_bit(pos
, RAW
);
1039 /* make sure these are *not*; want as small a sample as possible */
1040 evsel__reset_sample_bit(pos
, PERIOD
);
1041 evsel__reset_sample_bit(pos
, IP
);
1042 evsel__reset_sample_bit(pos
, CALLCHAIN
);
1043 evsel__reset_sample_bit(pos
, ADDR
);
1044 evsel__reset_sample_bit(pos
, READ
);
1049 attr
->sample_period
= 1;
1051 attr
->watermark
= 0;
1052 attr
->wakeup_events
= 1000;
1054 /* will enable all once we are ready */
1058 err
= evlist__open(evlist
);
1060 printf("Couldn't create the events: %s\n",
1061 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
1065 if (evlist__mmap(evlist
, kvm
->opts
.mmap_pages
) < 0) {
1066 ui__error("Failed to mmap the events: %s\n",
1067 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
1068 evlist__close(evlist
);
1079 static int read_events(struct perf_kvm_stat
*kvm
)
1083 struct perf_tool eops
= {
1084 .sample
= process_sample_event
,
1085 .comm
= perf_event__process_comm
,
1086 .namespaces
= perf_event__process_namespaces
,
1087 .ordered_events
= true,
1089 struct perf_data file
= {
1090 .path
= kvm
->file_name
,
1091 .mode
= PERF_DATA_MODE_READ
,
1092 .force
= kvm
->force
,
1096 kvm
->session
= perf_session__new(&file
, false, &kvm
->tool
);
1097 if (IS_ERR(kvm
->session
)) {
1098 pr_err("Initializing perf session failed\n");
1099 return PTR_ERR(kvm
->session
);
1102 symbol__init(&kvm
->session
->header
.env
);
1104 if (!perf_session__has_traces(kvm
->session
, "kvm record")) {
1110 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1111 * traced in the old kernel.
1113 ret
= cpu_isa_config(kvm
);
1117 ret
= perf_session__process_events(kvm
->session
);
1120 perf_session__delete(kvm
->session
);
1124 static int parse_target_str(struct perf_kvm_stat
*kvm
)
1126 if (kvm
->opts
.target
.pid
) {
1127 kvm
->pid_list
= intlist__new(kvm
->opts
.target
.pid
);
1128 if (kvm
->pid_list
== NULL
) {
1129 pr_err("Error parsing process id string\n");
1137 static int kvm_events_report_vcpu(struct perf_kvm_stat
*kvm
)
1140 int vcpu
= kvm
->trace_vcpu
;
1142 if (parse_target_str(kvm
) != 0)
1145 if (!verify_vcpu(vcpu
))
1148 if (!select_key(kvm
))
1151 if (!register_kvm_events_ops(kvm
))
1154 init_kvm_event_record(kvm
);
1157 ret
= read_events(kvm
);
1168 #define STRDUP_FAIL_EXIT(s) \
1176 int __weak
setup_kvm_events_tp(struct perf_kvm_stat
*kvm __maybe_unused
)
1182 kvm_events_record(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1184 unsigned int rec_argc
, i
, j
, events_tp_size
;
1185 const char **rec_argv
;
1186 const char * const record_args
[] = {
1192 const char * const kvm_stat_record_usage
[] = {
1193 "perf kvm stat record [<options>]",
1196 const char * const *events_tp
;
1200 ret
= setup_kvm_events_tp(kvm
);
1202 pr_err("Unable to setup the kvm tracepoints\n");
1206 for (events_tp
= kvm_events_tp
; *events_tp
; events_tp
++)
1209 rec_argc
= ARRAY_SIZE(record_args
) + argc
+ 2 +
1211 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1213 if (rec_argv
== NULL
)
1216 for (i
= 0; i
< ARRAY_SIZE(record_args
); i
++)
1217 rec_argv
[i
] = STRDUP_FAIL_EXIT(record_args
[i
]);
1219 for (j
= 0; j
< events_tp_size
; j
++) {
1220 rec_argv
[i
++] = "-e";
1221 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm_events_tp
[j
]);
1224 rec_argv
[i
++] = STRDUP_FAIL_EXIT("-o");
1225 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm
->file_name
);
1227 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
1228 rec_argv
[i
] = argv
[j
];
1230 set_option_flag(record_options
, 'e', "event", PARSE_OPT_HIDDEN
);
1231 set_option_flag(record_options
, 0, "filter", PARSE_OPT_HIDDEN
);
1232 set_option_flag(record_options
, 'R', "raw-samples", PARSE_OPT_HIDDEN
);
1234 set_option_flag(record_options
, 'F', "freq", PARSE_OPT_DISABLED
);
1235 set_option_flag(record_options
, 0, "group", PARSE_OPT_DISABLED
);
1236 set_option_flag(record_options
, 'g', NULL
, PARSE_OPT_DISABLED
);
1237 set_option_flag(record_options
, 0, "call-graph", PARSE_OPT_DISABLED
);
1238 set_option_flag(record_options
, 'd', "data", PARSE_OPT_DISABLED
);
1239 set_option_flag(record_options
, 'T', "timestamp", PARSE_OPT_DISABLED
);
1240 set_option_flag(record_options
, 'P', "period", PARSE_OPT_DISABLED
);
1241 set_option_flag(record_options
, 'n', "no-samples", PARSE_OPT_DISABLED
);
1242 set_option_flag(record_options
, 'N', "no-buildid-cache", PARSE_OPT_DISABLED
);
1243 set_option_flag(record_options
, 'B', "no-buildid", PARSE_OPT_DISABLED
);
1244 set_option_flag(record_options
, 'G', "cgroup", PARSE_OPT_DISABLED
);
1245 set_option_flag(record_options
, 'b', "branch-any", PARSE_OPT_DISABLED
);
1246 set_option_flag(record_options
, 'j', "branch-filter", PARSE_OPT_DISABLED
);
1247 set_option_flag(record_options
, 'W', "weight", PARSE_OPT_DISABLED
);
1248 set_option_flag(record_options
, 0, "transaction", PARSE_OPT_DISABLED
);
1250 record_usage
= kvm_stat_record_usage
;
1251 return cmd_record(i
, rec_argv
);
1255 kvm_events_report(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1257 const struct option kvm_events_report_options
[] = {
1258 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1259 "event for reporting: vmexit, "
1260 "mmio (x86 only), ioport (x86 only)"),
1261 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1262 "vcpu id to report"),
1263 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1264 "key for sorting: sample(sort by samples number)"
1265 " time (sort by avg time)"),
1266 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1267 "analyze events only for given process id(s)"),
1268 OPT_BOOLEAN('f', "force", &kvm
->force
, "don't complain, do it"),
1272 const char * const kvm_events_report_usage
[] = {
1273 "perf kvm stat report [<options>]",
1278 argc
= parse_options(argc
, argv
,
1279 kvm_events_report_options
,
1280 kvm_events_report_usage
, 0);
1282 usage_with_options(kvm_events_report_usage
,
1283 kvm_events_report_options
);
1286 if (!kvm
->opts
.target
.pid
)
1287 kvm
->opts
.target
.system_wide
= true;
1289 return kvm_events_report_vcpu(kvm
);
1292 #ifdef HAVE_TIMERFD_SUPPORT
1293 static struct evlist
*kvm_live_event_list(void)
1295 struct evlist
*evlist
;
1296 char *tp
, *name
, *sys
;
1298 const char * const *events_tp
;
1300 evlist
= evlist__new();
1304 for (events_tp
= kvm_events_tp
; *events_tp
; events_tp
++) {
1306 tp
= strdup(*events_tp
);
1310 /* split tracepoint into subsystem and name */
1312 name
= strchr(tp
, ':');
1314 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1322 if (evlist__add_newtp(evlist
, sys
, name
, NULL
)) {
1323 pr_err("Failed to add %s tracepoint to the list\n", *events_tp
);
1335 evlist__delete(evlist
);
1342 static int kvm_events_live(struct perf_kvm_stat
*kvm
,
1343 int argc
, const char **argv
)
1345 char errbuf
[BUFSIZ
];
1348 const struct option live_options
[] = {
1349 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1350 "record events on existing process id"),
1351 OPT_CALLBACK('m', "mmap-pages", &kvm
->opts
.mmap_pages
, "pages",
1352 "number of mmap data pages",
1353 perf_evlist__parse_mmap_pages
),
1354 OPT_INCR('v', "verbose", &verbose
,
1355 "be more verbose (show counter open errors, etc)"),
1356 OPT_BOOLEAN('a', "all-cpus", &kvm
->opts
.target
.system_wide
,
1357 "system-wide collection from all CPUs"),
1358 OPT_UINTEGER('d', "display", &kvm
->display_time
,
1359 "time in seconds between display updates"),
1360 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1361 "event for reporting: "
1362 "vmexit, mmio (x86 only), ioport (x86 only)"),
1363 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1364 "vcpu id to report"),
1365 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1366 "key for sorting: sample(sort by samples number)"
1367 " time (sort by avg time)"),
1368 OPT_U64(0, "duration", &kvm
->duration
,
1369 "show events other than"
1370 " HLT (x86 only) or Wait state (s390 only)"
1371 " that take longer than duration usecs"),
1372 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout
,
1373 "per thread proc mmap processing timeout in ms"),
1376 const char * const live_usage
[] = {
1377 "perf kvm stat live [<options>]",
1380 struct perf_data data
= {
1381 .mode
= PERF_DATA_MODE_WRITE
,
1385 /* event handling */
1386 kvm
->tool
.sample
= process_sample_event
;
1387 kvm
->tool
.comm
= perf_event__process_comm
;
1388 kvm
->tool
.exit
= perf_event__process_exit
;
1389 kvm
->tool
.fork
= perf_event__process_fork
;
1390 kvm
->tool
.lost
= process_lost_event
;
1391 kvm
->tool
.namespaces
= perf_event__process_namespaces
;
1392 kvm
->tool
.ordered_events
= true;
1393 perf_tool__fill_defaults(&kvm
->tool
);
1396 kvm
->display_time
= 1;
1397 kvm
->opts
.user_interval
= 1;
1398 kvm
->opts
.mmap_pages
= 512;
1399 kvm
->opts
.target
.uses_mmap
= false;
1400 kvm
->opts
.target
.uid_str
= NULL
;
1401 kvm
->opts
.target
.uid
= UINT_MAX
;
1404 disable_buildid_cache();
1409 argc
= parse_options(argc
, argv
, live_options
,
1412 usage_with_options(live_usage
, live_options
);
1415 kvm
->duration
*= NSEC_PER_USEC
; /* convert usec to nsec */
1418 * target related setups
1420 err
= target__validate(&kvm
->opts
.target
);
1422 target__strerror(&kvm
->opts
.target
, err
, errbuf
, BUFSIZ
);
1423 ui__warning("%s", errbuf
);
1426 if (target__none(&kvm
->opts
.target
))
1427 kvm
->opts
.target
.system_wide
= true;
1431 * generate the event list
1433 err
= setup_kvm_events_tp(kvm
);
1435 pr_err("Unable to setup the kvm tracepoints\n");
1439 kvm
->evlist
= kvm_live_event_list();
1440 if (kvm
->evlist
== NULL
) {
1445 if (perf_evlist__create_maps(kvm
->evlist
, &kvm
->opts
.target
) < 0)
1446 usage_with_options(live_usage
, live_options
);
1451 kvm
->session
= perf_session__new(&data
, false, &kvm
->tool
);
1452 if (IS_ERR(kvm
->session
)) {
1453 err
= PTR_ERR(kvm
->session
);
1456 kvm
->session
->evlist
= kvm
->evlist
;
1457 perf_session__set_id_hdr_size(kvm
->session
);
1458 ordered_events__set_copy_on_queue(&kvm
->session
->ordered_events
, true);
1459 machine__synthesize_threads(&kvm
->session
->machines
.host
, &kvm
->opts
.target
,
1460 kvm
->evlist
->core
.threads
, false, 1);
1461 err
= kvm_live_open_events(kvm
);
1465 err
= kvm_events_live_report(kvm
);
1468 perf_session__delete(kvm
->session
);
1469 kvm
->session
= NULL
;
1470 evlist__delete(kvm
->evlist
);
1476 static void print_kvm_stat_usage(void)
1478 printf("Usage: perf kvm stat <command>\n\n");
1480 printf("# Available commands:\n");
1481 printf("\trecord: record kvm events\n");
1482 printf("\treport: report statistical data of kvm events\n");
1483 printf("\tlive: live reporting of statistical data of kvm events\n");
1485 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1488 static int kvm_cmd_stat(const char *file_name
, int argc
, const char **argv
)
1490 struct perf_kvm_stat kvm
= {
1491 .file_name
= file_name
,
1494 .report_event
= "vmexit",
1495 .sort_key
= "sample",
1500 print_kvm_stat_usage();
1504 if (!strncmp(argv
[1], "rec", 3))
1505 return kvm_events_record(&kvm
, argc
- 1, argv
+ 1);
1507 if (!strncmp(argv
[1], "rep", 3))
1508 return kvm_events_report(&kvm
, argc
- 1 , argv
+ 1);
1510 #ifdef HAVE_TIMERFD_SUPPORT
1511 if (!strncmp(argv
[1], "live", 4))
1512 return kvm_events_live(&kvm
, argc
- 1 , argv
+ 1);
1516 return cmd_stat(argc
, argv
);
1518 #endif /* HAVE_KVM_STAT_SUPPORT */
1520 int __weak
kvm_add_default_arch_event(int *argc __maybe_unused
,
1521 const char **argv __maybe_unused
)
1526 static int __cmd_record(const char *file_name
, int argc
, const char **argv
)
1528 int rec_argc
, i
= 0, j
, ret
;
1529 const char **rec_argv
;
1531 ret
= kvm_add_default_arch_event(&argc
, argv
);
1535 rec_argc
= argc
+ 2;
1536 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1537 rec_argv
[i
++] = strdup("record");
1538 rec_argv
[i
++] = strdup("-o");
1539 rec_argv
[i
++] = strdup(file_name
);
1540 for (j
= 1; j
< argc
; j
++, i
++)
1541 rec_argv
[i
] = argv
[j
];
1543 BUG_ON(i
!= rec_argc
);
1545 return cmd_record(i
, rec_argv
);
1548 static int __cmd_report(const char *file_name
, int argc
, const char **argv
)
1550 int rec_argc
, i
= 0, j
;
1551 const char **rec_argv
;
1553 rec_argc
= argc
+ 2;
1554 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1555 rec_argv
[i
++] = strdup("report");
1556 rec_argv
[i
++] = strdup("-i");
1557 rec_argv
[i
++] = strdup(file_name
);
1558 for (j
= 1; j
< argc
; j
++, i
++)
1559 rec_argv
[i
] = argv
[j
];
1561 BUG_ON(i
!= rec_argc
);
1563 return cmd_report(i
, rec_argv
);
1567 __cmd_buildid_list(const char *file_name
, int argc
, const char **argv
)
1569 int rec_argc
, i
= 0, j
;
1570 const char **rec_argv
;
1572 rec_argc
= argc
+ 2;
1573 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1574 rec_argv
[i
++] = strdup("buildid-list");
1575 rec_argv
[i
++] = strdup("-i");
1576 rec_argv
[i
++] = strdup(file_name
);
1577 for (j
= 1; j
< argc
; j
++, i
++)
1578 rec_argv
[i
] = argv
[j
];
1580 BUG_ON(i
!= rec_argc
);
1582 return cmd_buildid_list(i
, rec_argv
);
1585 int cmd_kvm(int argc
, const char **argv
)
1587 const char *file_name
= NULL
;
1588 const struct option kvm_options
[] = {
1589 OPT_STRING('i', "input", &file_name
, "file",
1591 OPT_STRING('o', "output", &file_name
, "file",
1592 "Output file name"),
1593 OPT_BOOLEAN(0, "guest", &perf_guest
,
1594 "Collect guest os data"),
1595 OPT_BOOLEAN(0, "host", &perf_host
,
1596 "Collect host os data"),
1597 OPT_STRING(0, "guestmount", &symbol_conf
.guestmount
, "directory",
1598 "guest mount directory under which every guest os"
1599 " instance has a subdir"),
1600 OPT_STRING(0, "guestvmlinux", &symbol_conf
.default_guest_vmlinux_name
,
1601 "file", "file saving guest os vmlinux"),
1602 OPT_STRING(0, "guestkallsyms", &symbol_conf
.default_guest_kallsyms
,
1603 "file", "file saving guest os /proc/kallsyms"),
1604 OPT_STRING(0, "guestmodules", &symbol_conf
.default_guest_modules
,
1605 "file", "file saving guest os /proc/modules"),
1606 OPT_INCR('v', "verbose", &verbose
,
1607 "be more verbose (show counter open errors, etc)"),
1611 const char *const kvm_subcommands
[] = { "top", "record", "report", "diff",
1612 "buildid-list", "stat", NULL
};
1613 const char *kvm_usage
[] = { NULL
, NULL
};
1618 argc
= parse_options_subcommand(argc
, argv
, kvm_options
, kvm_subcommands
, kvm_usage
,
1619 PARSE_OPT_STOP_AT_NON_OPTION
);
1621 usage_with_options(kvm_usage
, kvm_options
);
1627 file_name
= get_filename_for_perf_kvm();
1630 pr_err("Failed to allocate memory for filename\n");
1635 if (!strncmp(argv
[0], "rec", 3))
1636 return __cmd_record(file_name
, argc
, argv
);
1637 else if (!strncmp(argv
[0], "rep", 3))
1638 return __cmd_report(file_name
, argc
, argv
);
1639 else if (!strncmp(argv
[0], "diff", 4))
1640 return cmd_diff(argc
, argv
);
1641 else if (!strncmp(argv
[0], "top", 3))
1642 return cmd_top(argc
, argv
);
1643 else if (!strncmp(argv
[0], "buildid-list", 12))
1644 return __cmd_buildid_list(file_name
, argc
, argv
);
1645 #ifdef HAVE_KVM_STAT_SUPPORT
1646 else if (!strncmp(argv
[0], "stat", 4))
1647 return kvm_cmd_stat(file_name
, argc
, argv
);
1650 usage_with_options(kvm_usage
, kvm_options
);