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
= perf_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
= perf_evsel__intval(evsel
, sample
,
421 thread__set_priv(thread
, vcpu_record
);
424 return thread__priv(thread
);
427 static bool handle_kvm_event(struct perf_kvm_stat
*kvm
,
428 struct thread
*thread
,
430 struct perf_sample
*sample
)
432 struct vcpu_event_record
*vcpu_record
;
433 struct event_key key
= { .key
= INVALID_KEY
,
434 .exit_reasons
= kvm
->exit_reasons
};
436 vcpu_record
= per_vcpu_record(thread
, evsel
, sample
);
440 /* only process events for vcpus user cares about */
441 if ((kvm
->trace_vcpu
!= -1) &&
442 (kvm
->trace_vcpu
!= vcpu_record
->vcpu_id
))
445 if (kvm
->events_ops
->is_begin_event(evsel
, sample
, &key
))
446 return handle_begin_event(kvm
, vcpu_record
, &key
, sample
->time
);
448 if (is_child_event(kvm
, evsel
, sample
, &key
))
449 return handle_child_event(kvm
, vcpu_record
, &key
, sample
);
451 if (kvm
->events_ops
->is_end_event(evsel
, sample
, &key
))
452 return handle_end_event(kvm
, vcpu_record
, &key
, sample
);
457 #define GET_EVENT_KEY(func, field) \
458 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
461 return event->total.field; \
463 if (vcpu >= event->max_vcpu) \
466 return event->vcpu[vcpu].field; \
469 #define COMPARE_EVENT_KEY(func, field) \
470 GET_EVENT_KEY(func, field) \
471 static int compare_kvm_event_ ## func(struct kvm_event *one, \
472 struct kvm_event *two, int vcpu)\
474 return get_event_ ##func(one, vcpu) > \
475 get_event_ ##func(two, vcpu); \
478 GET_EVENT_KEY(time
, time
);
479 COMPARE_EVENT_KEY(count
, stats
.n
);
480 COMPARE_EVENT_KEY(mean
, stats
.mean
);
481 GET_EVENT_KEY(max
, stats
.max
);
482 GET_EVENT_KEY(min
, stats
.min
);
484 #define DEF_SORT_NAME_KEY(name, compare_key) \
485 { #name, compare_kvm_event_ ## compare_key }
487 static struct kvm_event_key keys
[] = {
488 DEF_SORT_NAME_KEY(sample
, count
),
489 DEF_SORT_NAME_KEY(time
, mean
),
493 static bool select_key(struct perf_kvm_stat
*kvm
)
497 for (i
= 0; keys
[i
].name
; i
++) {
498 if (!strcmp(keys
[i
].name
, kvm
->sort_key
)) {
499 kvm
->compare
= keys
[i
].key
;
504 pr_err("Unknown compare key:%s\n", kvm
->sort_key
);
508 static void insert_to_result(struct rb_root
*result
, struct kvm_event
*event
,
509 key_cmp_fun bigger
, int vcpu
)
511 struct rb_node
**rb
= &result
->rb_node
;
512 struct rb_node
*parent
= NULL
;
516 p
= container_of(*rb
, struct kvm_event
, rb
);
519 if (bigger(event
, p
, vcpu
))
520 rb
= &(*rb
)->rb_left
;
522 rb
= &(*rb
)->rb_right
;
525 rb_link_node(&event
->rb
, parent
, rb
);
526 rb_insert_color(&event
->rb
, result
);
530 update_total_count(struct perf_kvm_stat
*kvm
, struct kvm_event
*event
)
532 int vcpu
= kvm
->trace_vcpu
;
534 kvm
->total_count
+= get_event_count(event
, vcpu
);
535 kvm
->total_time
+= get_event_time(event
, vcpu
);
538 static bool event_is_valid(struct kvm_event
*event
, int vcpu
)
540 return !!get_event_count(event
, vcpu
);
543 static void sort_result(struct perf_kvm_stat
*kvm
)
546 int vcpu
= kvm
->trace_vcpu
;
547 struct kvm_event
*event
;
549 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++) {
550 list_for_each_entry(event
, &kvm
->kvm_events_cache
[i
], hash_entry
) {
551 if (event_is_valid(event
, vcpu
)) {
552 update_total_count(kvm
, event
);
553 insert_to_result(&kvm
->result
, event
,
560 /* returns left most element of result, and erase it */
561 static struct kvm_event
*pop_from_result(struct rb_root
*result
)
563 struct rb_node
*node
= rb_first(result
);
568 rb_erase(node
, result
);
569 return container_of(node
, struct kvm_event
, rb
);
572 static void print_vcpu_info(struct perf_kvm_stat
*kvm
)
574 int vcpu
= kvm
->trace_vcpu
;
576 pr_info("Analyze events for ");
578 if (kvm
->opts
.target
.system_wide
)
579 pr_info("all VMs, ");
580 else if (kvm
->opts
.target
.pid
)
581 pr_info("pid(s) %s, ", kvm
->opts
.target
.pid
);
583 pr_info("dazed and confused on what is monitored, ");
586 pr_info("all VCPUs:\n\n");
588 pr_info("VCPU %d:\n\n", vcpu
);
591 static void show_timeofday(void)
597 gettimeofday(&tv
, NULL
);
598 if (localtime_r(&tv
.tv_sec
, <ime
)) {
599 strftime(date
, sizeof(date
), "%H:%M:%S", <ime
);
600 pr_info("%s.%06ld", date
, tv
.tv_usec
);
602 pr_info("00:00:00.000000");
607 static void print_result(struct perf_kvm_stat
*kvm
)
609 char decode
[decode_str_len
];
610 struct kvm_event
*event
;
611 int vcpu
= kvm
->trace_vcpu
;
619 print_vcpu_info(kvm
);
620 pr_info("%*s ", decode_str_len
, kvm
->events_ops
->name
);
621 pr_info("%10s ", "Samples");
622 pr_info("%9s ", "Samples%");
624 pr_info("%9s ", "Time%");
625 pr_info("%11s ", "Min Time");
626 pr_info("%11s ", "Max Time");
627 pr_info("%16s ", "Avg time");
630 while ((event
= pop_from_result(&kvm
->result
))) {
631 u64 ecount
, etime
, max
, min
;
633 ecount
= get_event_count(event
, vcpu
);
634 etime
= get_event_time(event
, vcpu
);
635 max
= get_event_max(event
, vcpu
);
636 min
= get_event_min(event
, vcpu
);
638 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
639 pr_info("%*s ", decode_str_len
, decode
);
640 pr_info("%10llu ", (unsigned long long)ecount
);
641 pr_info("%8.2f%% ", (double)ecount
/ kvm
->total_count
* 100);
642 pr_info("%8.2f%% ", (double)etime
/ kvm
->total_time
* 100);
643 pr_info("%9.2fus ", (double)min
/ NSEC_PER_USEC
);
644 pr_info("%9.2fus ", (double)max
/ NSEC_PER_USEC
);
645 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime
/ ecount
/ NSEC_PER_USEC
,
646 kvm_event_rel_stddev(vcpu
, event
));
650 pr_info("\nTotal Samples:%" PRIu64
", Total events handled time:%.2fus.\n\n",
651 kvm
->total_count
, kvm
->total_time
/ (double)NSEC_PER_USEC
);
653 if (kvm
->lost_events
)
654 pr_info("\nLost events: %" PRIu64
"\n\n", kvm
->lost_events
);
657 #ifdef HAVE_TIMERFD_SUPPORT
658 static int process_lost_event(struct perf_tool
*tool
,
659 union perf_event
*event __maybe_unused
,
660 struct perf_sample
*sample __maybe_unused
,
661 struct machine
*machine __maybe_unused
)
663 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
, tool
);
670 static bool skip_sample(struct perf_kvm_stat
*kvm
,
671 struct perf_sample
*sample
)
673 if (kvm
->pid_list
&& intlist__find(kvm
->pid_list
, sample
->pid
) == NULL
)
679 static int process_sample_event(struct perf_tool
*tool
,
680 union perf_event
*event
,
681 struct perf_sample
*sample
,
683 struct machine
*machine
)
686 struct thread
*thread
;
687 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
,
690 if (skip_sample(kvm
, sample
))
693 thread
= machine__findnew_thread(machine
, sample
->pid
, sample
->tid
);
694 if (thread
== NULL
) {
695 pr_debug("problem processing %d event, skipping it.\n",
700 if (!handle_kvm_event(kvm
, thread
, evsel
, sample
))
707 static int cpu_isa_config(struct perf_kvm_stat
*kvm
)
709 char buf
[128], *cpuid
;
713 err
= get_cpuid(buf
, sizeof(buf
));
715 pr_err("Failed to look up CPU type: %s\n",
716 str_error_r(err
, buf
, sizeof(buf
)));
721 cpuid
= kvm
->session
->header
.env
.cpuid
;
724 pr_err("Failed to look up CPU type\n");
728 err
= cpu_isa_init(kvm
, cpuid
);
730 pr_err("CPU %s is not supported.\n", cpuid
);
735 static bool verify_vcpu(int vcpu
)
737 if (vcpu
!= -1 && vcpu
< 0) {
738 pr_err("Invalid vcpu:%d.\n", vcpu
);
745 #ifdef HAVE_TIMERFD_SUPPORT
746 /* keeping the max events to a modest level to keep
747 * the processing of samples per mmap smooth.
749 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
751 static s64
perf_kvm__mmap_read_idx(struct perf_kvm_stat
*kvm
, int idx
,
754 struct evlist
*evlist
= kvm
->evlist
;
755 union perf_event
*event
;
761 *mmap_time
= ULLONG_MAX
;
762 md
= &evlist
->mmap
[idx
];
763 err
= perf_mmap__read_init(&md
->core
);
765 return (err
== -EAGAIN
) ? 0 : -1;
767 while ((event
= perf_mmap__read_event(&md
->core
)) != NULL
) {
768 err
= perf_evlist__parse_sample_timestamp(evlist
, event
, ×tamp
);
770 perf_mmap__consume(&md
->core
);
771 pr_err("Failed to parse sample\n");
775 err
= perf_session__queue_event(kvm
->session
, event
, timestamp
, 0);
777 * FIXME: Here we can't consume the event, as perf_session__queue_event will
778 * point to it, and it'll get possibly overwritten by the kernel.
780 perf_mmap__consume(&md
->core
);
783 pr_err("Failed to enqueue sample: %d\n", err
);
787 /* save time stamp of our first sample for this mmap */
789 *mmap_time
= timestamp
;
791 /* limit events per mmap handled all at once */
793 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
797 perf_mmap__read_done(&md
->core
);
801 static int perf_kvm__mmap_read(struct perf_kvm_stat
*kvm
)
803 int i
, err
, throttled
= 0;
805 u64 flush_time
= ULLONG_MAX
, mmap_time
;
807 for (i
= 0; i
< kvm
->evlist
->core
.nr_mmaps
; i
++) {
808 n
= perf_kvm__mmap_read_idx(kvm
, i
, &mmap_time
);
812 /* flush time is going to be the minimum of all the individual
813 * mmap times. Essentially, we flush all the samples queued up
814 * from the last pass under our minimal start time -- that leaves
815 * a very small race for samples to come in with a lower timestamp.
816 * The ioctl to return the perf_clock timestamp should close the
819 if (mmap_time
< flush_time
)
820 flush_time
= mmap_time
;
823 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
827 /* flush queue after each round in which we processed events */
829 struct ordered_events
*oe
= &kvm
->session
->ordered_events
;
831 oe
->next_flush
= flush_time
;
832 err
= ordered_events__flush(oe
, OE_FLUSH__ROUND
);
834 if (kvm
->lost_events
)
835 pr_info("\nLost events: %" PRIu64
"\n\n",
844 static volatile int done
;
846 static void sig_handler(int sig __maybe_unused
)
851 static int perf_kvm__timerfd_create(struct perf_kvm_stat
*kvm
)
853 struct itimerspec new_value
;
856 kvm
->timerfd
= timerfd_create(CLOCK_MONOTONIC
, TFD_NONBLOCK
);
857 if (kvm
->timerfd
< 0) {
858 pr_err("timerfd_create failed\n");
862 new_value
.it_value
.tv_sec
= kvm
->display_time
;
863 new_value
.it_value
.tv_nsec
= 0;
864 new_value
.it_interval
.tv_sec
= kvm
->display_time
;
865 new_value
.it_interval
.tv_nsec
= 0;
867 if (timerfd_settime(kvm
->timerfd
, 0, &new_value
, NULL
) != 0) {
868 pr_err("timerfd_settime failed: %d\n", errno
);
878 static int perf_kvm__handle_timerfd(struct perf_kvm_stat
*kvm
)
883 rc
= read(kvm
->timerfd
, &c
, sizeof(uint64_t));
888 pr_err("Failed to read timer fd: %d\n", errno
);
892 if (rc
!= sizeof(uint64_t)) {
893 pr_err("Error reading timer fd - invalid size returned\n");
898 pr_debug("Missed timer beats: %" PRIu64
"\n", c
-1);
905 clear_events_cache_stats(kvm
->kvm_events_cache
);
906 kvm
->total_count
= 0;
908 kvm
->lost_events
= 0;
913 static int fd_set_nonblock(int fd
)
917 arg
= fcntl(fd
, F_GETFL
);
919 pr_err("Failed to get current flags for fd %d\n", fd
);
923 if (fcntl(fd
, F_SETFL
, arg
| O_NONBLOCK
) < 0) {
924 pr_err("Failed to set non-block option on fd %d\n", fd
);
931 static int perf_kvm__handle_stdin(void)
942 static int kvm_events_live_report(struct perf_kvm_stat
*kvm
)
944 int nr_stdin
, ret
, err
= -EINVAL
;
947 /* live flag must be set first */
950 ret
= cpu_isa_config(kvm
);
954 if (!verify_vcpu(kvm
->trace_vcpu
) ||
956 !register_kvm_events_ops(kvm
)) {
960 set_term_quiet_input(&save
);
961 init_kvm_event_record(kvm
);
963 signal(SIGINT
, sig_handler
);
964 signal(SIGTERM
, sig_handler
);
967 if (perf_kvm__timerfd_create(kvm
) < 0) {
972 if (evlist__add_pollfd(kvm
->evlist
, kvm
->timerfd
) < 0)
975 nr_stdin
= evlist__add_pollfd(kvm
->evlist
, fileno(stdin
));
979 if (fd_set_nonblock(fileno(stdin
)) != 0)
982 /* everything is good - enable the events and process */
983 evlist__enable(kvm
->evlist
);
986 struct fdarray
*fda
= &kvm
->evlist
->core
.pollfd
;
989 rc
= perf_kvm__mmap_read(kvm
);
993 err
= perf_kvm__handle_timerfd(kvm
);
997 if (fda
->entries
[nr_stdin
].revents
& POLLIN
)
998 done
= perf_kvm__handle_stdin();
1001 err
= evlist__poll(kvm
->evlist
, 100);
1004 evlist__disable(kvm
->evlist
);
1012 if (kvm
->timerfd
>= 0)
1013 close(kvm
->timerfd
);
1015 tcsetattr(0, TCSAFLUSH
, &save
);
1019 static int kvm_live_open_events(struct perf_kvm_stat
*kvm
)
1023 struct evlist
*evlist
= kvm
->evlist
;
1024 char sbuf
[STRERR_BUFSIZE
];
1026 perf_evlist__config(evlist
, &kvm
->opts
, NULL
);
1029 * Note: exclude_{guest,host} do not apply here.
1030 * This command processes KVM tracepoints from host only
1032 evlist__for_each_entry(evlist
, pos
) {
1033 struct perf_event_attr
*attr
= &pos
->core
.attr
;
1035 /* make sure these *are* set */
1036 perf_evsel__set_sample_bit(pos
, TID
);
1037 perf_evsel__set_sample_bit(pos
, TIME
);
1038 perf_evsel__set_sample_bit(pos
, CPU
);
1039 perf_evsel__set_sample_bit(pos
, RAW
);
1040 /* make sure these are *not*; want as small a sample as possible */
1041 perf_evsel__reset_sample_bit(pos
, PERIOD
);
1042 perf_evsel__reset_sample_bit(pos
, IP
);
1043 perf_evsel__reset_sample_bit(pos
, CALLCHAIN
);
1044 perf_evsel__reset_sample_bit(pos
, ADDR
);
1045 perf_evsel__reset_sample_bit(pos
, READ
);
1050 attr
->sample_period
= 1;
1052 attr
->watermark
= 0;
1053 attr
->wakeup_events
= 1000;
1055 /* will enable all once we are ready */
1059 err
= evlist__open(evlist
);
1061 printf("Couldn't create the events: %s\n",
1062 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
1066 if (evlist__mmap(evlist
, kvm
->opts
.mmap_pages
) < 0) {
1067 ui__error("Failed to mmap the events: %s\n",
1068 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
1069 evlist__close(evlist
);
1080 static int read_events(struct perf_kvm_stat
*kvm
)
1084 struct perf_tool eops
= {
1085 .sample
= process_sample_event
,
1086 .comm
= perf_event__process_comm
,
1087 .namespaces
= perf_event__process_namespaces
,
1088 .ordered_events
= true,
1090 struct perf_data file
= {
1091 .path
= kvm
->file_name
,
1092 .mode
= PERF_DATA_MODE_READ
,
1093 .force
= kvm
->force
,
1097 kvm
->session
= perf_session__new(&file
, false, &kvm
->tool
);
1098 if (IS_ERR(kvm
->session
)) {
1099 pr_err("Initializing perf session failed\n");
1100 return PTR_ERR(kvm
->session
);
1103 symbol__init(&kvm
->session
->header
.env
);
1105 if (!perf_session__has_traces(kvm
->session
, "kvm record")) {
1111 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1112 * traced in the old kernel.
1114 ret
= cpu_isa_config(kvm
);
1118 ret
= perf_session__process_events(kvm
->session
);
1121 perf_session__delete(kvm
->session
);
1125 static int parse_target_str(struct perf_kvm_stat
*kvm
)
1127 if (kvm
->opts
.target
.pid
) {
1128 kvm
->pid_list
= intlist__new(kvm
->opts
.target
.pid
);
1129 if (kvm
->pid_list
== NULL
) {
1130 pr_err("Error parsing process id string\n");
1138 static int kvm_events_report_vcpu(struct perf_kvm_stat
*kvm
)
1141 int vcpu
= kvm
->trace_vcpu
;
1143 if (parse_target_str(kvm
) != 0)
1146 if (!verify_vcpu(vcpu
))
1149 if (!select_key(kvm
))
1152 if (!register_kvm_events_ops(kvm
))
1155 init_kvm_event_record(kvm
);
1158 ret
= read_events(kvm
);
1169 #define STRDUP_FAIL_EXIT(s) \
1177 int __weak
setup_kvm_events_tp(struct perf_kvm_stat
*kvm __maybe_unused
)
1183 kvm_events_record(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1185 unsigned int rec_argc
, i
, j
, events_tp_size
;
1186 const char **rec_argv
;
1187 const char * const record_args
[] = {
1193 const char * const kvm_stat_record_usage
[] = {
1194 "perf kvm stat record [<options>]",
1197 const char * const *events_tp
;
1201 ret
= setup_kvm_events_tp(kvm
);
1203 pr_err("Unable to setup the kvm tracepoints\n");
1207 for (events_tp
= kvm_events_tp
; *events_tp
; events_tp
++)
1210 rec_argc
= ARRAY_SIZE(record_args
) + argc
+ 2 +
1212 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1214 if (rec_argv
== NULL
)
1217 for (i
= 0; i
< ARRAY_SIZE(record_args
); i
++)
1218 rec_argv
[i
] = STRDUP_FAIL_EXIT(record_args
[i
]);
1220 for (j
= 0; j
< events_tp_size
; j
++) {
1221 rec_argv
[i
++] = "-e";
1222 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm_events_tp
[j
]);
1225 rec_argv
[i
++] = STRDUP_FAIL_EXIT("-o");
1226 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm
->file_name
);
1228 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
1229 rec_argv
[i
] = argv
[j
];
1231 set_option_flag(record_options
, 'e', "event", PARSE_OPT_HIDDEN
);
1232 set_option_flag(record_options
, 0, "filter", PARSE_OPT_HIDDEN
);
1233 set_option_flag(record_options
, 'R', "raw-samples", PARSE_OPT_HIDDEN
);
1235 set_option_flag(record_options
, 'F', "freq", PARSE_OPT_DISABLED
);
1236 set_option_flag(record_options
, 0, "group", PARSE_OPT_DISABLED
);
1237 set_option_flag(record_options
, 'g', NULL
, PARSE_OPT_DISABLED
);
1238 set_option_flag(record_options
, 0, "call-graph", PARSE_OPT_DISABLED
);
1239 set_option_flag(record_options
, 'd', "data", PARSE_OPT_DISABLED
);
1240 set_option_flag(record_options
, 'T', "timestamp", PARSE_OPT_DISABLED
);
1241 set_option_flag(record_options
, 'P', "period", PARSE_OPT_DISABLED
);
1242 set_option_flag(record_options
, 'n', "no-samples", PARSE_OPT_DISABLED
);
1243 set_option_flag(record_options
, 'N', "no-buildid-cache", PARSE_OPT_DISABLED
);
1244 set_option_flag(record_options
, 'B', "no-buildid", PARSE_OPT_DISABLED
);
1245 set_option_flag(record_options
, 'G', "cgroup", PARSE_OPT_DISABLED
);
1246 set_option_flag(record_options
, 'b', "branch-any", PARSE_OPT_DISABLED
);
1247 set_option_flag(record_options
, 'j', "branch-filter", PARSE_OPT_DISABLED
);
1248 set_option_flag(record_options
, 'W', "weight", PARSE_OPT_DISABLED
);
1249 set_option_flag(record_options
, 0, "transaction", PARSE_OPT_DISABLED
);
1251 record_usage
= kvm_stat_record_usage
;
1252 return cmd_record(i
, rec_argv
);
1256 kvm_events_report(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1258 const struct option kvm_events_report_options
[] = {
1259 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1260 "event for reporting: vmexit, "
1261 "mmio (x86 only), ioport (x86 only)"),
1262 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1263 "vcpu id to report"),
1264 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1265 "key for sorting: sample(sort by samples number)"
1266 " time (sort by avg time)"),
1267 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1268 "analyze events only for given process id(s)"),
1269 OPT_BOOLEAN('f', "force", &kvm
->force
, "don't complain, do it"),
1273 const char * const kvm_events_report_usage
[] = {
1274 "perf kvm stat report [<options>]",
1279 argc
= parse_options(argc
, argv
,
1280 kvm_events_report_options
,
1281 kvm_events_report_usage
, 0);
1283 usage_with_options(kvm_events_report_usage
,
1284 kvm_events_report_options
);
1287 if (!kvm
->opts
.target
.pid
)
1288 kvm
->opts
.target
.system_wide
= true;
1290 return kvm_events_report_vcpu(kvm
);
1293 #ifdef HAVE_TIMERFD_SUPPORT
1294 static struct evlist
*kvm_live_event_list(void)
1296 struct evlist
*evlist
;
1297 char *tp
, *name
, *sys
;
1299 const char * const *events_tp
;
1301 evlist
= evlist__new();
1305 for (events_tp
= kvm_events_tp
; *events_tp
; events_tp
++) {
1307 tp
= strdup(*events_tp
);
1311 /* split tracepoint into subsystem and name */
1313 name
= strchr(tp
, ':');
1315 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1323 if (perf_evlist__add_newtp(evlist
, sys
, name
, NULL
)) {
1324 pr_err("Failed to add %s tracepoint to the list\n", *events_tp
);
1336 evlist__delete(evlist
);
1343 static int kvm_events_live(struct perf_kvm_stat
*kvm
,
1344 int argc
, const char **argv
)
1346 char errbuf
[BUFSIZ
];
1349 const struct option live_options
[] = {
1350 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1351 "record events on existing process id"),
1352 OPT_CALLBACK('m', "mmap-pages", &kvm
->opts
.mmap_pages
, "pages",
1353 "number of mmap data pages",
1354 perf_evlist__parse_mmap_pages
),
1355 OPT_INCR('v', "verbose", &verbose
,
1356 "be more verbose (show counter open errors, etc)"),
1357 OPT_BOOLEAN('a', "all-cpus", &kvm
->opts
.target
.system_wide
,
1358 "system-wide collection from all CPUs"),
1359 OPT_UINTEGER('d', "display", &kvm
->display_time
,
1360 "time in seconds between display updates"),
1361 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1362 "event for reporting: "
1363 "vmexit, mmio (x86 only), ioport (x86 only)"),
1364 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1365 "vcpu id to report"),
1366 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1367 "key for sorting: sample(sort by samples number)"
1368 " time (sort by avg time)"),
1369 OPT_U64(0, "duration", &kvm
->duration
,
1370 "show events other than"
1371 " HLT (x86 only) or Wait state (s390 only)"
1372 " that take longer than duration usecs"),
1373 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout
,
1374 "per thread proc mmap processing timeout in ms"),
1377 const char * const live_usage
[] = {
1378 "perf kvm stat live [<options>]",
1381 struct perf_data data
= {
1382 .mode
= PERF_DATA_MODE_WRITE
,
1386 /* event handling */
1387 kvm
->tool
.sample
= process_sample_event
;
1388 kvm
->tool
.comm
= perf_event__process_comm
;
1389 kvm
->tool
.exit
= perf_event__process_exit
;
1390 kvm
->tool
.fork
= perf_event__process_fork
;
1391 kvm
->tool
.lost
= process_lost_event
;
1392 kvm
->tool
.namespaces
= perf_event__process_namespaces
;
1393 kvm
->tool
.ordered_events
= true;
1394 perf_tool__fill_defaults(&kvm
->tool
);
1397 kvm
->display_time
= 1;
1398 kvm
->opts
.user_interval
= 1;
1399 kvm
->opts
.mmap_pages
= 512;
1400 kvm
->opts
.target
.uses_mmap
= false;
1401 kvm
->opts
.target
.uid_str
= NULL
;
1402 kvm
->opts
.target
.uid
= UINT_MAX
;
1405 disable_buildid_cache();
1410 argc
= parse_options(argc
, argv
, live_options
,
1413 usage_with_options(live_usage
, live_options
);
1416 kvm
->duration
*= NSEC_PER_USEC
; /* convert usec to nsec */
1419 * target related setups
1421 err
= target__validate(&kvm
->opts
.target
);
1423 target__strerror(&kvm
->opts
.target
, err
, errbuf
, BUFSIZ
);
1424 ui__warning("%s", errbuf
);
1427 if (target__none(&kvm
->opts
.target
))
1428 kvm
->opts
.target
.system_wide
= true;
1432 * generate the event list
1434 err
= setup_kvm_events_tp(kvm
);
1436 pr_err("Unable to setup the kvm tracepoints\n");
1440 kvm
->evlist
= kvm_live_event_list();
1441 if (kvm
->evlist
== NULL
) {
1446 if (perf_evlist__create_maps(kvm
->evlist
, &kvm
->opts
.target
) < 0)
1447 usage_with_options(live_usage
, live_options
);
1452 kvm
->session
= perf_session__new(&data
, false, &kvm
->tool
);
1453 if (IS_ERR(kvm
->session
)) {
1454 err
= PTR_ERR(kvm
->session
);
1457 kvm
->session
->evlist
= kvm
->evlist
;
1458 perf_session__set_id_hdr_size(kvm
->session
);
1459 ordered_events__set_copy_on_queue(&kvm
->session
->ordered_events
, true);
1460 machine__synthesize_threads(&kvm
->session
->machines
.host
, &kvm
->opts
.target
,
1461 kvm
->evlist
->core
.threads
, false, 1);
1462 err
= kvm_live_open_events(kvm
);
1466 err
= kvm_events_live_report(kvm
);
1469 perf_session__delete(kvm
->session
);
1470 kvm
->session
= NULL
;
1471 evlist__delete(kvm
->evlist
);
1477 static void print_kvm_stat_usage(void)
1479 printf("Usage: perf kvm stat <command>\n\n");
1481 printf("# Available commands:\n");
1482 printf("\trecord: record kvm events\n");
1483 printf("\treport: report statistical data of kvm events\n");
1484 printf("\tlive: live reporting of statistical data of kvm events\n");
1486 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1489 static int kvm_cmd_stat(const char *file_name
, int argc
, const char **argv
)
1491 struct perf_kvm_stat kvm
= {
1492 .file_name
= file_name
,
1495 .report_event
= "vmexit",
1496 .sort_key
= "sample",
1501 print_kvm_stat_usage();
1505 if (!strncmp(argv
[1], "rec", 3))
1506 return kvm_events_record(&kvm
, argc
- 1, argv
+ 1);
1508 if (!strncmp(argv
[1], "rep", 3))
1509 return kvm_events_report(&kvm
, argc
- 1 , argv
+ 1);
1511 #ifdef HAVE_TIMERFD_SUPPORT
1512 if (!strncmp(argv
[1], "live", 4))
1513 return kvm_events_live(&kvm
, argc
- 1 , argv
+ 1);
1517 return cmd_stat(argc
, argv
);
1519 #endif /* HAVE_KVM_STAT_SUPPORT */
1521 int __weak
kvm_add_default_arch_event(int *argc __maybe_unused
,
1522 const char **argv __maybe_unused
)
1527 static int __cmd_record(const char *file_name
, int argc
, const char **argv
)
1529 int rec_argc
, i
= 0, j
, ret
;
1530 const char **rec_argv
;
1532 ret
= kvm_add_default_arch_event(&argc
, argv
);
1536 rec_argc
= argc
+ 2;
1537 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1538 rec_argv
[i
++] = strdup("record");
1539 rec_argv
[i
++] = strdup("-o");
1540 rec_argv
[i
++] = strdup(file_name
);
1541 for (j
= 1; j
< argc
; j
++, i
++)
1542 rec_argv
[i
] = argv
[j
];
1544 BUG_ON(i
!= rec_argc
);
1546 return cmd_record(i
, rec_argv
);
1549 static int __cmd_report(const char *file_name
, int argc
, const char **argv
)
1551 int rec_argc
, i
= 0, j
;
1552 const char **rec_argv
;
1554 rec_argc
= argc
+ 2;
1555 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1556 rec_argv
[i
++] = strdup("report");
1557 rec_argv
[i
++] = strdup("-i");
1558 rec_argv
[i
++] = strdup(file_name
);
1559 for (j
= 1; j
< argc
; j
++, i
++)
1560 rec_argv
[i
] = argv
[j
];
1562 BUG_ON(i
!= rec_argc
);
1564 return cmd_report(i
, rec_argv
);
1568 __cmd_buildid_list(const char *file_name
, int argc
, const char **argv
)
1570 int rec_argc
, i
= 0, j
;
1571 const char **rec_argv
;
1573 rec_argc
= argc
+ 2;
1574 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1575 rec_argv
[i
++] = strdup("buildid-list");
1576 rec_argv
[i
++] = strdup("-i");
1577 rec_argv
[i
++] = strdup(file_name
);
1578 for (j
= 1; j
< argc
; j
++, i
++)
1579 rec_argv
[i
] = argv
[j
];
1581 BUG_ON(i
!= rec_argc
);
1583 return cmd_buildid_list(i
, rec_argv
);
1586 int cmd_kvm(int argc
, const char **argv
)
1588 const char *file_name
= NULL
;
1589 const struct option kvm_options
[] = {
1590 OPT_STRING('i', "input", &file_name
, "file",
1592 OPT_STRING('o', "output", &file_name
, "file",
1593 "Output file name"),
1594 OPT_BOOLEAN(0, "guest", &perf_guest
,
1595 "Collect guest os data"),
1596 OPT_BOOLEAN(0, "host", &perf_host
,
1597 "Collect host os data"),
1598 OPT_STRING(0, "guestmount", &symbol_conf
.guestmount
, "directory",
1599 "guest mount directory under which every guest os"
1600 " instance has a subdir"),
1601 OPT_STRING(0, "guestvmlinux", &symbol_conf
.default_guest_vmlinux_name
,
1602 "file", "file saving guest os vmlinux"),
1603 OPT_STRING(0, "guestkallsyms", &symbol_conf
.default_guest_kallsyms
,
1604 "file", "file saving guest os /proc/kallsyms"),
1605 OPT_STRING(0, "guestmodules", &symbol_conf
.default_guest_modules
,
1606 "file", "file saving guest os /proc/modules"),
1607 OPT_INCR('v', "verbose", &verbose
,
1608 "be more verbose (show counter open errors, etc)"),
1612 const char *const kvm_subcommands
[] = { "top", "record", "report", "diff",
1613 "buildid-list", "stat", NULL
};
1614 const char *kvm_usage
[] = { NULL
, NULL
};
1619 argc
= parse_options_subcommand(argc
, argv
, kvm_options
, kvm_subcommands
, kvm_usage
,
1620 PARSE_OPT_STOP_AT_NON_OPTION
);
1622 usage_with_options(kvm_usage
, kvm_options
);
1628 file_name
= get_filename_for_perf_kvm();
1631 pr_err("Failed to allocate memory for filename\n");
1636 if (!strncmp(argv
[0], "rec", 3))
1637 return __cmd_record(file_name
, argc
, argv
);
1638 else if (!strncmp(argv
[0], "rep", 3))
1639 return __cmd_report(file_name
, argc
, argv
);
1640 else if (!strncmp(argv
[0], "diff", 4))
1641 return cmd_diff(argc
, argv
);
1642 else if (!strncmp(argv
[0], "top", 3))
1643 return cmd_top(argc
, argv
);
1644 else if (!strncmp(argv
[0], "buildid-list", 12))
1645 return __cmd_buildid_list(file_name
, argc
, argv
);
1646 #ifdef HAVE_KVM_STAT_SUPPORT
1647 else if (!strncmp(argv
[0], "stat", 4))
1648 return kvm_cmd_stat(file_name
, argc
, argv
);
1651 usage_with_options(kvm_usage
, kvm_options
);