1 // SPDX-License-Identifier: GPL-2.0
5 #include "util/evsel.h"
6 #include "util/evlist.h"
9 #include "util/cache.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/parse-options.h>
16 #include "util/trace-event.h"
17 #include "util/debug.h"
18 #include "util/tool.h"
19 #include "util/stat.h"
21 #include "util/data.h"
22 #include "util/ordered-events.h"
24 #include <sys/prctl.h>
25 #ifdef HAVE_TIMERFD_SUPPORT
26 #include <sys/timerfd.h>
29 #include <sys/types.h>
33 #include <linux/kernel.h>
34 #include <linux/time64.h>
39 #include <semaphore.h>
43 static const char *get_filename_for_perf_kvm(void)
47 if (perf_host
&& !perf_guest
)
48 filename
= strdup("perf.data.host");
49 else if (!perf_host
&& perf_guest
)
50 filename
= strdup("perf.data.guest");
52 filename
= strdup("perf.data.kvm");
57 #ifdef HAVE_KVM_STAT_SUPPORT
58 #include "util/kvm-stat.h"
60 void exit_event_get_key(struct perf_evsel
*evsel
,
61 struct perf_sample
*sample
,
62 struct event_key
*key
)
65 key
->key
= perf_evsel__intval(evsel
, sample
, kvm_exit_reason
);
68 bool kvm_exit_event(struct perf_evsel
*evsel
)
70 return !strcmp(evsel
->name
, kvm_exit_trace
);
73 bool exit_event_begin(struct perf_evsel
*evsel
,
74 struct perf_sample
*sample
, struct event_key
*key
)
76 if (kvm_exit_event(evsel
)) {
77 exit_event_get_key(evsel
, sample
, key
);
84 bool kvm_entry_event(struct perf_evsel
*evsel
)
86 return !strcmp(evsel
->name
, kvm_entry_trace
);
89 bool exit_event_end(struct perf_evsel
*evsel
,
90 struct perf_sample
*sample __maybe_unused
,
91 struct event_key
*key __maybe_unused
)
93 return kvm_entry_event(evsel
);
96 static const char *get_exit_reason(struct perf_kvm_stat
*kvm
,
97 struct exit_reasons_table
*tbl
,
100 while (tbl
->reason
!= NULL
) {
101 if (tbl
->exit_code
== exit_code
)
106 pr_err("unknown kvm exit code:%lld on %s\n",
107 (unsigned long long)exit_code
, kvm
->exit_reasons_isa
);
111 void exit_event_decode_key(struct perf_kvm_stat
*kvm
,
112 struct event_key
*key
,
115 const char *exit_reason
= get_exit_reason(kvm
, key
->exit_reasons
,
118 scnprintf(decode
, decode_str_len
, "%s", exit_reason
);
121 static bool register_kvm_events_ops(struct perf_kvm_stat
*kvm
)
123 struct kvm_reg_events_ops
*events_ops
= kvm_reg_events_ops
;
125 for (events_ops
= kvm_reg_events_ops
; events_ops
->name
; events_ops
++) {
126 if (!strcmp(events_ops
->name
, kvm
->report_event
)) {
127 kvm
->events_ops
= events_ops
->ops
;
135 struct vcpu_event_record
{
138 struct kvm_event
*last_event
;
142 static void init_kvm_event_record(struct perf_kvm_stat
*kvm
)
146 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++)
147 INIT_LIST_HEAD(&kvm
->kvm_events_cache
[i
]);
150 #ifdef HAVE_TIMERFD_SUPPORT
151 static void clear_events_cache_stats(struct list_head
*kvm_events_cache
)
153 struct list_head
*head
;
154 struct kvm_event
*event
;
158 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++) {
159 head
= &kvm_events_cache
[i
];
160 list_for_each_entry(event
, head
, hash_entry
) {
161 /* reset stats for event */
162 event
->total
.time
= 0;
163 init_stats(&event
->total
.stats
);
165 for (j
= 0; j
< event
->max_vcpu
; ++j
) {
166 event
->vcpu
[j
].time
= 0;
167 init_stats(&event
->vcpu
[j
].stats
);
174 static int kvm_events_hash_fn(u64 key
)
176 return key
& (EVENTS_CACHE_SIZE
- 1);
179 static bool kvm_event_expand(struct kvm_event
*event
, int vcpu_id
)
181 int old_max_vcpu
= event
->max_vcpu
;
184 if (vcpu_id
< event
->max_vcpu
)
187 while (event
->max_vcpu
<= vcpu_id
)
188 event
->max_vcpu
+= DEFAULT_VCPU_NUM
;
191 event
->vcpu
= realloc(event
->vcpu
,
192 event
->max_vcpu
* sizeof(*event
->vcpu
));
195 pr_err("Not enough memory\n");
199 memset(event
->vcpu
+ old_max_vcpu
, 0,
200 (event
->max_vcpu
- old_max_vcpu
) * sizeof(*event
->vcpu
));
204 static struct kvm_event
*kvm_alloc_init_event(struct event_key
*key
)
206 struct kvm_event
*event
;
208 event
= zalloc(sizeof(*event
));
210 pr_err("Not enough memory\n");
215 init_stats(&event
->total
.stats
);
219 static struct kvm_event
*find_create_kvm_event(struct perf_kvm_stat
*kvm
,
220 struct event_key
*key
)
222 struct kvm_event
*event
;
223 struct list_head
*head
;
225 BUG_ON(key
->key
== INVALID_KEY
);
227 head
= &kvm
->kvm_events_cache
[kvm_events_hash_fn(key
->key
)];
228 list_for_each_entry(event
, head
, hash_entry
) {
229 if (event
->key
.key
== key
->key
&& event
->key
.info
== key
->info
)
233 event
= kvm_alloc_init_event(key
);
237 list_add(&event
->hash_entry
, head
);
241 static bool handle_begin_event(struct perf_kvm_stat
*kvm
,
242 struct vcpu_event_record
*vcpu_record
,
243 struct event_key
*key
, u64 timestamp
)
245 struct kvm_event
*event
= NULL
;
247 if (key
->key
!= INVALID_KEY
)
248 event
= find_create_kvm_event(kvm
, key
);
250 vcpu_record
->last_event
= event
;
251 vcpu_record
->start_time
= timestamp
;
256 kvm_update_event_stats(struct kvm_event_stats
*kvm_stats
, u64 time_diff
)
258 kvm_stats
->time
+= time_diff
;
259 update_stats(&kvm_stats
->stats
, time_diff
);
262 static double kvm_event_rel_stddev(int vcpu_id
, struct kvm_event
*event
)
264 struct kvm_event_stats
*kvm_stats
= &event
->total
;
267 kvm_stats
= &event
->vcpu
[vcpu_id
];
269 return rel_stddev_stats(stddev_stats(&kvm_stats
->stats
),
270 avg_stats(&kvm_stats
->stats
));
273 static bool update_kvm_event(struct kvm_event
*event
, int vcpu_id
,
277 kvm_update_event_stats(&event
->total
, time_diff
);
281 if (!kvm_event_expand(event
, vcpu_id
))
284 kvm_update_event_stats(&event
->vcpu
[vcpu_id
], time_diff
);
288 static bool is_child_event(struct perf_kvm_stat
*kvm
,
289 struct perf_evsel
*evsel
,
290 struct perf_sample
*sample
,
291 struct event_key
*key
)
293 struct child_event_ops
*child_ops
;
295 child_ops
= kvm
->events_ops
->child_ops
;
300 for (; child_ops
->name
; child_ops
++) {
301 if (!strcmp(evsel
->name
, child_ops
->name
)) {
302 child_ops
->get_key(evsel
, sample
, key
);
310 static bool handle_child_event(struct perf_kvm_stat
*kvm
,
311 struct vcpu_event_record
*vcpu_record
,
312 struct event_key
*key
,
313 struct perf_sample
*sample __maybe_unused
)
315 struct kvm_event
*event
= NULL
;
317 if (key
->key
!= INVALID_KEY
)
318 event
= find_create_kvm_event(kvm
, key
);
320 vcpu_record
->last_event
= event
;
325 static bool skip_event(const char *event
)
327 const char * const *skip_events
;
329 for (skip_events
= kvm_skip_events
; *skip_events
; skip_events
++)
330 if (!strcmp(event
, *skip_events
))
336 static bool handle_end_event(struct perf_kvm_stat
*kvm
,
337 struct vcpu_event_record
*vcpu_record
,
338 struct event_key
*key
,
339 struct perf_sample
*sample
)
341 struct kvm_event
*event
;
342 u64 time_begin
, time_diff
;
345 if (kvm
->trace_vcpu
== -1)
348 vcpu
= vcpu_record
->vcpu_id
;
350 event
= vcpu_record
->last_event
;
351 time_begin
= vcpu_record
->start_time
;
353 /* The begin event is not caught. */
358 * In some case, the 'begin event' only records the start timestamp,
359 * the actual event is recognized in the 'end event' (e.g. mmio-event).
362 /* Both begin and end events did not get the key. */
363 if (!event
&& key
->key
== INVALID_KEY
)
367 event
= find_create_kvm_event(kvm
, key
);
372 vcpu_record
->last_event
= NULL
;
373 vcpu_record
->start_time
= 0;
375 /* seems to happen once in a while during live mode */
376 if (sample
->time
< time_begin
) {
377 pr_debug("End time before begin time; skipping event.\n");
381 time_diff
= sample
->time
- time_begin
;
383 if (kvm
->duration
&& time_diff
> kvm
->duration
) {
384 char decode
[decode_str_len
];
386 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
387 if (!skip_event(decode
)) {
388 pr_info("%" PRIu64
" VM %d, vcpu %d: %s event took %" PRIu64
"usec\n",
389 sample
->time
, sample
->pid
, vcpu_record
->vcpu_id
,
390 decode
, time_diff
/ NSEC_PER_USEC
);
394 return update_kvm_event(event
, vcpu
, time_diff
);
398 struct vcpu_event_record
*per_vcpu_record(struct thread
*thread
,
399 struct perf_evsel
*evsel
,
400 struct perf_sample
*sample
)
402 /* Only kvm_entry records vcpu id. */
403 if (!thread__priv(thread
) && kvm_entry_event(evsel
)) {
404 struct vcpu_event_record
*vcpu_record
;
406 vcpu_record
= zalloc(sizeof(*vcpu_record
));
408 pr_err("%s: Not enough memory\n", __func__
);
412 vcpu_record
->vcpu_id
= perf_evsel__intval(evsel
, sample
,
414 thread__set_priv(thread
, vcpu_record
);
417 return thread__priv(thread
);
420 static bool handle_kvm_event(struct perf_kvm_stat
*kvm
,
421 struct thread
*thread
,
422 struct perf_evsel
*evsel
,
423 struct perf_sample
*sample
)
425 struct vcpu_event_record
*vcpu_record
;
426 struct event_key key
= { .key
= INVALID_KEY
,
427 .exit_reasons
= kvm
->exit_reasons
};
429 vcpu_record
= per_vcpu_record(thread
, evsel
, sample
);
433 /* only process events for vcpus user cares about */
434 if ((kvm
->trace_vcpu
!= -1) &&
435 (kvm
->trace_vcpu
!= vcpu_record
->vcpu_id
))
438 if (kvm
->events_ops
->is_begin_event(evsel
, sample
, &key
))
439 return handle_begin_event(kvm
, vcpu_record
, &key
, sample
->time
);
441 if (is_child_event(kvm
, evsel
, sample
, &key
))
442 return handle_child_event(kvm
, vcpu_record
, &key
, sample
);
444 if (kvm
->events_ops
->is_end_event(evsel
, sample
, &key
))
445 return handle_end_event(kvm
, vcpu_record
, &key
, sample
);
450 #define GET_EVENT_KEY(func, field) \
451 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
454 return event->total.field; \
456 if (vcpu >= event->max_vcpu) \
459 return event->vcpu[vcpu].field; \
462 #define COMPARE_EVENT_KEY(func, field) \
463 GET_EVENT_KEY(func, field) \
464 static int compare_kvm_event_ ## func(struct kvm_event *one, \
465 struct kvm_event *two, int vcpu)\
467 return get_event_ ##func(one, vcpu) > \
468 get_event_ ##func(two, vcpu); \
471 GET_EVENT_KEY(time
, time
);
472 COMPARE_EVENT_KEY(count
, stats
.n
);
473 COMPARE_EVENT_KEY(mean
, stats
.mean
);
474 GET_EVENT_KEY(max
, stats
.max
);
475 GET_EVENT_KEY(min
, stats
.min
);
477 #define DEF_SORT_NAME_KEY(name, compare_key) \
478 { #name, compare_kvm_event_ ## compare_key }
480 static struct kvm_event_key keys
[] = {
481 DEF_SORT_NAME_KEY(sample
, count
),
482 DEF_SORT_NAME_KEY(time
, mean
),
486 static bool select_key(struct perf_kvm_stat
*kvm
)
490 for (i
= 0; keys
[i
].name
; i
++) {
491 if (!strcmp(keys
[i
].name
, kvm
->sort_key
)) {
492 kvm
->compare
= keys
[i
].key
;
497 pr_err("Unknown compare key:%s\n", kvm
->sort_key
);
501 static void insert_to_result(struct rb_root
*result
, struct kvm_event
*event
,
502 key_cmp_fun bigger
, int vcpu
)
504 struct rb_node
**rb
= &result
->rb_node
;
505 struct rb_node
*parent
= NULL
;
509 p
= container_of(*rb
, struct kvm_event
, rb
);
512 if (bigger(event
, p
, vcpu
))
513 rb
= &(*rb
)->rb_left
;
515 rb
= &(*rb
)->rb_right
;
518 rb_link_node(&event
->rb
, parent
, rb
);
519 rb_insert_color(&event
->rb
, result
);
523 update_total_count(struct perf_kvm_stat
*kvm
, struct kvm_event
*event
)
525 int vcpu
= kvm
->trace_vcpu
;
527 kvm
->total_count
+= get_event_count(event
, vcpu
);
528 kvm
->total_time
+= get_event_time(event
, vcpu
);
531 static bool event_is_valid(struct kvm_event
*event
, int vcpu
)
533 return !!get_event_count(event
, vcpu
);
536 static void sort_result(struct perf_kvm_stat
*kvm
)
539 int vcpu
= kvm
->trace_vcpu
;
540 struct kvm_event
*event
;
542 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++) {
543 list_for_each_entry(event
, &kvm
->kvm_events_cache
[i
], hash_entry
) {
544 if (event_is_valid(event
, vcpu
)) {
545 update_total_count(kvm
, event
);
546 insert_to_result(&kvm
->result
, event
,
553 /* returns left most element of result, and erase it */
554 static struct kvm_event
*pop_from_result(struct rb_root
*result
)
556 struct rb_node
*node
= rb_first(result
);
561 rb_erase(node
, result
);
562 return container_of(node
, struct kvm_event
, rb
);
565 static void print_vcpu_info(struct perf_kvm_stat
*kvm
)
567 int vcpu
= kvm
->trace_vcpu
;
569 pr_info("Analyze events for ");
571 if (kvm
->opts
.target
.system_wide
)
572 pr_info("all VMs, ");
573 else if (kvm
->opts
.target
.pid
)
574 pr_info("pid(s) %s, ", kvm
->opts
.target
.pid
);
576 pr_info("dazed and confused on what is monitored, ");
579 pr_info("all VCPUs:\n\n");
581 pr_info("VCPU %d:\n\n", vcpu
);
584 static void show_timeofday(void)
590 gettimeofday(&tv
, NULL
);
591 if (localtime_r(&tv
.tv_sec
, <ime
)) {
592 strftime(date
, sizeof(date
), "%H:%M:%S", <ime
);
593 pr_info("%s.%06ld", date
, tv
.tv_usec
);
595 pr_info("00:00:00.000000");
600 static void print_result(struct perf_kvm_stat
*kvm
)
602 char decode
[decode_str_len
];
603 struct kvm_event
*event
;
604 int vcpu
= kvm
->trace_vcpu
;
612 print_vcpu_info(kvm
);
613 pr_info("%*s ", decode_str_len
, kvm
->events_ops
->name
);
614 pr_info("%10s ", "Samples");
615 pr_info("%9s ", "Samples%");
617 pr_info("%9s ", "Time%");
618 pr_info("%11s ", "Min Time");
619 pr_info("%11s ", "Max Time");
620 pr_info("%16s ", "Avg time");
623 while ((event
= pop_from_result(&kvm
->result
))) {
624 u64 ecount
, etime
, max
, min
;
626 ecount
= get_event_count(event
, vcpu
);
627 etime
= get_event_time(event
, vcpu
);
628 max
= get_event_max(event
, vcpu
);
629 min
= get_event_min(event
, vcpu
);
631 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
632 pr_info("%*s ", decode_str_len
, decode
);
633 pr_info("%10llu ", (unsigned long long)ecount
);
634 pr_info("%8.2f%% ", (double)ecount
/ kvm
->total_count
* 100);
635 pr_info("%8.2f%% ", (double)etime
/ kvm
->total_time
* 100);
636 pr_info("%9.2fus ", (double)min
/ NSEC_PER_USEC
);
637 pr_info("%9.2fus ", (double)max
/ NSEC_PER_USEC
);
638 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime
/ ecount
/ NSEC_PER_USEC
,
639 kvm_event_rel_stddev(vcpu
, event
));
643 pr_info("\nTotal Samples:%" PRIu64
", Total events handled time:%.2fus.\n\n",
644 kvm
->total_count
, kvm
->total_time
/ (double)NSEC_PER_USEC
);
646 if (kvm
->lost_events
)
647 pr_info("\nLost events: %" PRIu64
"\n\n", kvm
->lost_events
);
650 #ifdef HAVE_TIMERFD_SUPPORT
651 static int process_lost_event(struct perf_tool
*tool
,
652 union perf_event
*event __maybe_unused
,
653 struct perf_sample
*sample __maybe_unused
,
654 struct machine
*machine __maybe_unused
)
656 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
, tool
);
663 static bool skip_sample(struct perf_kvm_stat
*kvm
,
664 struct perf_sample
*sample
)
666 if (kvm
->pid_list
&& intlist__find(kvm
->pid_list
, sample
->pid
) == NULL
)
672 static int process_sample_event(struct perf_tool
*tool
,
673 union perf_event
*event
,
674 struct perf_sample
*sample
,
675 struct perf_evsel
*evsel
,
676 struct machine
*machine
)
679 struct thread
*thread
;
680 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
,
683 if (skip_sample(kvm
, sample
))
686 thread
= machine__findnew_thread(machine
, sample
->pid
, sample
->tid
);
687 if (thread
== NULL
) {
688 pr_debug("problem processing %d event, skipping it.\n",
693 if (!handle_kvm_event(kvm
, thread
, evsel
, sample
))
700 static int cpu_isa_config(struct perf_kvm_stat
*kvm
)
702 char buf
[64], *cpuid
;
706 err
= get_cpuid(buf
, sizeof(buf
));
708 pr_err("Failed to look up CPU type\n");
713 cpuid
= kvm
->session
->header
.env
.cpuid
;
716 pr_err("Failed to look up CPU type\n");
720 err
= cpu_isa_init(kvm
, cpuid
);
722 pr_err("CPU %s is not supported.\n", cpuid
);
727 static bool verify_vcpu(int vcpu
)
729 if (vcpu
!= -1 && vcpu
< 0) {
730 pr_err("Invalid vcpu:%d.\n", vcpu
);
737 #ifdef HAVE_TIMERFD_SUPPORT
738 /* keeping the max events to a modest level to keep
739 * the processing of samples per mmap smooth.
741 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
743 static s64
perf_kvm__mmap_read_idx(struct perf_kvm_stat
*kvm
, int idx
,
746 struct perf_evlist
*evlist
= kvm
->evlist
;
747 union perf_event
*event
;
748 struct perf_mmap
*md
;
753 *mmap_time
= ULLONG_MAX
;
754 md
= &evlist
->mmap
[idx
];
755 err
= perf_mmap__read_init(md
);
757 return (err
== -EAGAIN
) ? 0 : -1;
759 while ((event
= perf_mmap__read_event(md
)) != NULL
) {
760 err
= perf_evlist__parse_sample_timestamp(evlist
, event
, ×tamp
);
762 perf_mmap__consume(md
);
763 pr_err("Failed to parse sample\n");
767 err
= perf_session__queue_event(kvm
->session
, event
, timestamp
, 0);
769 * FIXME: Here we can't consume the event, as perf_session__queue_event will
770 * point to it, and it'll get possibly overwritten by the kernel.
772 perf_mmap__consume(md
);
775 pr_err("Failed to enqueue sample: %d\n", err
);
779 /* save time stamp of our first sample for this mmap */
781 *mmap_time
= timestamp
;
783 /* limit events per mmap handled all at once */
785 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
789 perf_mmap__read_done(md
);
793 static int perf_kvm__mmap_read(struct perf_kvm_stat
*kvm
)
795 int i
, err
, throttled
= 0;
797 u64 flush_time
= ULLONG_MAX
, mmap_time
;
799 for (i
= 0; i
< kvm
->evlist
->nr_mmaps
; i
++) {
800 n
= perf_kvm__mmap_read_idx(kvm
, i
, &mmap_time
);
804 /* flush time is going to be the minimum of all the individual
805 * mmap times. Essentially, we flush all the samples queued up
806 * from the last pass under our minimal start time -- that leaves
807 * a very small race for samples to come in with a lower timestamp.
808 * The ioctl to return the perf_clock timestamp should close the
811 if (mmap_time
< flush_time
)
812 flush_time
= mmap_time
;
815 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
819 /* flush queue after each round in which we processed events */
821 struct ordered_events
*oe
= &kvm
->session
->ordered_events
;
823 oe
->next_flush
= flush_time
;
824 err
= ordered_events__flush(oe
, OE_FLUSH__ROUND
);
826 if (kvm
->lost_events
)
827 pr_info("\nLost events: %" PRIu64
"\n\n",
836 static volatile int done
;
838 static void sig_handler(int sig __maybe_unused
)
843 static int perf_kvm__timerfd_create(struct perf_kvm_stat
*kvm
)
845 struct itimerspec new_value
;
848 kvm
->timerfd
= timerfd_create(CLOCK_MONOTONIC
, TFD_NONBLOCK
);
849 if (kvm
->timerfd
< 0) {
850 pr_err("timerfd_create failed\n");
854 new_value
.it_value
.tv_sec
= kvm
->display_time
;
855 new_value
.it_value
.tv_nsec
= 0;
856 new_value
.it_interval
.tv_sec
= kvm
->display_time
;
857 new_value
.it_interval
.tv_nsec
= 0;
859 if (timerfd_settime(kvm
->timerfd
, 0, &new_value
, NULL
) != 0) {
860 pr_err("timerfd_settime failed: %d\n", errno
);
870 static int perf_kvm__handle_timerfd(struct perf_kvm_stat
*kvm
)
875 rc
= read(kvm
->timerfd
, &c
, sizeof(uint64_t));
880 pr_err("Failed to read timer fd: %d\n", errno
);
884 if (rc
!= sizeof(uint64_t)) {
885 pr_err("Error reading timer fd - invalid size returned\n");
890 pr_debug("Missed timer beats: %" PRIu64
"\n", c
-1);
897 clear_events_cache_stats(kvm
->kvm_events_cache
);
898 kvm
->total_count
= 0;
900 kvm
->lost_events
= 0;
905 static int fd_set_nonblock(int fd
)
909 arg
= fcntl(fd
, F_GETFL
);
911 pr_err("Failed to get current flags for fd %d\n", fd
);
915 if (fcntl(fd
, F_SETFL
, arg
| O_NONBLOCK
) < 0) {
916 pr_err("Failed to set non-block option on fd %d\n", fd
);
923 static int perf_kvm__handle_stdin(void)
934 static int kvm_events_live_report(struct perf_kvm_stat
*kvm
)
936 int nr_stdin
, ret
, err
= -EINVAL
;
939 /* live flag must be set first */
942 ret
= cpu_isa_config(kvm
);
946 if (!verify_vcpu(kvm
->trace_vcpu
) ||
948 !register_kvm_events_ops(kvm
)) {
952 set_term_quiet_input(&save
);
953 init_kvm_event_record(kvm
);
955 signal(SIGINT
, sig_handler
);
956 signal(SIGTERM
, sig_handler
);
959 if (perf_kvm__timerfd_create(kvm
) < 0) {
964 if (perf_evlist__add_pollfd(kvm
->evlist
, kvm
->timerfd
) < 0)
967 nr_stdin
= perf_evlist__add_pollfd(kvm
->evlist
, fileno(stdin
));
971 if (fd_set_nonblock(fileno(stdin
)) != 0)
974 /* everything is good - enable the events and process */
975 perf_evlist__enable(kvm
->evlist
);
978 struct fdarray
*fda
= &kvm
->evlist
->pollfd
;
981 rc
= perf_kvm__mmap_read(kvm
);
985 err
= perf_kvm__handle_timerfd(kvm
);
989 if (fda
->entries
[nr_stdin
].revents
& POLLIN
)
990 done
= perf_kvm__handle_stdin();
993 err
= fdarray__poll(fda
, 100);
996 perf_evlist__disable(kvm
->evlist
);
1004 if (kvm
->timerfd
>= 0)
1005 close(kvm
->timerfd
);
1007 tcsetattr(0, TCSAFLUSH
, &save
);
1011 static int kvm_live_open_events(struct perf_kvm_stat
*kvm
)
1014 struct perf_evsel
*pos
;
1015 struct perf_evlist
*evlist
= kvm
->evlist
;
1016 char sbuf
[STRERR_BUFSIZE
];
1018 perf_evlist__config(evlist
, &kvm
->opts
, NULL
);
1021 * Note: exclude_{guest,host} do not apply here.
1022 * This command processes KVM tracepoints from host only
1024 evlist__for_each_entry(evlist
, pos
) {
1025 struct perf_event_attr
*attr
= &pos
->attr
;
1027 /* make sure these *are* set */
1028 perf_evsel__set_sample_bit(pos
, TID
);
1029 perf_evsel__set_sample_bit(pos
, TIME
);
1030 perf_evsel__set_sample_bit(pos
, CPU
);
1031 perf_evsel__set_sample_bit(pos
, RAW
);
1032 /* make sure these are *not*; want as small a sample as possible */
1033 perf_evsel__reset_sample_bit(pos
, PERIOD
);
1034 perf_evsel__reset_sample_bit(pos
, IP
);
1035 perf_evsel__reset_sample_bit(pos
, CALLCHAIN
);
1036 perf_evsel__reset_sample_bit(pos
, ADDR
);
1037 perf_evsel__reset_sample_bit(pos
, READ
);
1042 attr
->sample_period
= 1;
1044 attr
->watermark
= 0;
1045 attr
->wakeup_events
= 1000;
1047 /* will enable all once we are ready */
1051 err
= perf_evlist__open(evlist
);
1053 printf("Couldn't create the events: %s\n",
1054 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
1058 if (perf_evlist__mmap(evlist
, kvm
->opts
.mmap_pages
) < 0) {
1059 ui__error("Failed to mmap the events: %s\n",
1060 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
1061 perf_evlist__close(evlist
);
1072 static int read_events(struct perf_kvm_stat
*kvm
)
1076 struct perf_tool eops
= {
1077 .sample
= process_sample_event
,
1078 .comm
= perf_event__process_comm
,
1079 .namespaces
= perf_event__process_namespaces
,
1080 .ordered_events
= true,
1082 struct perf_data file
= {
1083 .path
= kvm
->file_name
,
1084 .mode
= PERF_DATA_MODE_READ
,
1085 .force
= kvm
->force
,
1089 kvm
->session
= perf_session__new(&file
, false, &kvm
->tool
);
1090 if (!kvm
->session
) {
1091 pr_err("Initializing perf session failed\n");
1095 symbol__init(&kvm
->session
->header
.env
);
1097 if (!perf_session__has_traces(kvm
->session
, "kvm record")) {
1103 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1104 * traced in the old kernel.
1106 ret
= cpu_isa_config(kvm
);
1110 ret
= perf_session__process_events(kvm
->session
);
1113 perf_session__delete(kvm
->session
);
1117 static int parse_target_str(struct perf_kvm_stat
*kvm
)
1119 if (kvm
->opts
.target
.pid
) {
1120 kvm
->pid_list
= intlist__new(kvm
->opts
.target
.pid
);
1121 if (kvm
->pid_list
== NULL
) {
1122 pr_err("Error parsing process id string\n");
1130 static int kvm_events_report_vcpu(struct perf_kvm_stat
*kvm
)
1133 int vcpu
= kvm
->trace_vcpu
;
1135 if (parse_target_str(kvm
) != 0)
1138 if (!verify_vcpu(vcpu
))
1141 if (!select_key(kvm
))
1144 if (!register_kvm_events_ops(kvm
))
1147 init_kvm_event_record(kvm
);
1150 ret
= read_events(kvm
);
1161 #define STRDUP_FAIL_EXIT(s) \
1169 int __weak
setup_kvm_events_tp(struct perf_kvm_stat
*kvm __maybe_unused
)
1175 kvm_events_record(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1177 unsigned int rec_argc
, i
, j
, events_tp_size
;
1178 const char **rec_argv
;
1179 const char * const record_args
[] = {
1185 const char * const kvm_stat_record_usage
[] = {
1186 "perf kvm stat record [<options>]",
1189 const char * const *events_tp
;
1193 ret
= setup_kvm_events_tp(kvm
);
1195 pr_err("Unable to setup the kvm tracepoints\n");
1199 for (events_tp
= kvm_events_tp
; *events_tp
; events_tp
++)
1202 rec_argc
= ARRAY_SIZE(record_args
) + argc
+ 2 +
1204 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1206 if (rec_argv
== NULL
)
1209 for (i
= 0; i
< ARRAY_SIZE(record_args
); i
++)
1210 rec_argv
[i
] = STRDUP_FAIL_EXIT(record_args
[i
]);
1212 for (j
= 0; j
< events_tp_size
; j
++) {
1213 rec_argv
[i
++] = "-e";
1214 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm_events_tp
[j
]);
1217 rec_argv
[i
++] = STRDUP_FAIL_EXIT("-o");
1218 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm
->file_name
);
1220 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
1221 rec_argv
[i
] = argv
[j
];
1223 set_option_flag(record_options
, 'e', "event", PARSE_OPT_HIDDEN
);
1224 set_option_flag(record_options
, 0, "filter", PARSE_OPT_HIDDEN
);
1225 set_option_flag(record_options
, 'R', "raw-samples", PARSE_OPT_HIDDEN
);
1227 set_option_flag(record_options
, 'F', "freq", PARSE_OPT_DISABLED
);
1228 set_option_flag(record_options
, 0, "group", PARSE_OPT_DISABLED
);
1229 set_option_flag(record_options
, 'g', NULL
, PARSE_OPT_DISABLED
);
1230 set_option_flag(record_options
, 0, "call-graph", PARSE_OPT_DISABLED
);
1231 set_option_flag(record_options
, 'd', "data", PARSE_OPT_DISABLED
);
1232 set_option_flag(record_options
, 'T', "timestamp", PARSE_OPT_DISABLED
);
1233 set_option_flag(record_options
, 'P', "period", PARSE_OPT_DISABLED
);
1234 set_option_flag(record_options
, 'n', "no-samples", PARSE_OPT_DISABLED
);
1235 set_option_flag(record_options
, 'N', "no-buildid-cache", PARSE_OPT_DISABLED
);
1236 set_option_flag(record_options
, 'B', "no-buildid", PARSE_OPT_DISABLED
);
1237 set_option_flag(record_options
, 'G', "cgroup", PARSE_OPT_DISABLED
);
1238 set_option_flag(record_options
, 'b', "branch-any", PARSE_OPT_DISABLED
);
1239 set_option_flag(record_options
, 'j', "branch-filter", PARSE_OPT_DISABLED
);
1240 set_option_flag(record_options
, 'W', "weight", PARSE_OPT_DISABLED
);
1241 set_option_flag(record_options
, 0, "transaction", PARSE_OPT_DISABLED
);
1243 record_usage
= kvm_stat_record_usage
;
1244 return cmd_record(i
, rec_argv
);
1248 kvm_events_report(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1250 const struct option kvm_events_report_options
[] = {
1251 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1252 "event for reporting: vmexit, "
1253 "mmio (x86 only), ioport (x86 only)"),
1254 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1255 "vcpu id to report"),
1256 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1257 "key for sorting: sample(sort by samples number)"
1258 " time (sort by avg time)"),
1259 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1260 "analyze events only for given process id(s)"),
1261 OPT_BOOLEAN('f', "force", &kvm
->force
, "don't complain, do it"),
1265 const char * const kvm_events_report_usage
[] = {
1266 "perf kvm stat report [<options>]",
1271 argc
= parse_options(argc
, argv
,
1272 kvm_events_report_options
,
1273 kvm_events_report_usage
, 0);
1275 usage_with_options(kvm_events_report_usage
,
1276 kvm_events_report_options
);
1279 if (!kvm
->opts
.target
.pid
)
1280 kvm
->opts
.target
.system_wide
= true;
1282 return kvm_events_report_vcpu(kvm
);
1285 #ifdef HAVE_TIMERFD_SUPPORT
1286 static struct perf_evlist
*kvm_live_event_list(void)
1288 struct perf_evlist
*evlist
;
1289 char *tp
, *name
, *sys
;
1291 const char * const *events_tp
;
1293 evlist
= perf_evlist__new();
1297 for (events_tp
= kvm_events_tp
; *events_tp
; events_tp
++) {
1299 tp
= strdup(*events_tp
);
1303 /* split tracepoint into subsystem and name */
1305 name
= strchr(tp
, ':');
1307 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1315 if (perf_evlist__add_newtp(evlist
, sys
, name
, NULL
)) {
1316 pr_err("Failed to add %s tracepoint to the list\n", *events_tp
);
1328 perf_evlist__delete(evlist
);
1335 static int kvm_events_live(struct perf_kvm_stat
*kvm
,
1336 int argc
, const char **argv
)
1338 char errbuf
[BUFSIZ
];
1341 const struct option live_options
[] = {
1342 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1343 "record events on existing process id"),
1344 OPT_CALLBACK('m', "mmap-pages", &kvm
->opts
.mmap_pages
, "pages",
1345 "number of mmap data pages",
1346 perf_evlist__parse_mmap_pages
),
1347 OPT_INCR('v', "verbose", &verbose
,
1348 "be more verbose (show counter open errors, etc)"),
1349 OPT_BOOLEAN('a', "all-cpus", &kvm
->opts
.target
.system_wide
,
1350 "system-wide collection from all CPUs"),
1351 OPT_UINTEGER('d', "display", &kvm
->display_time
,
1352 "time in seconds between display updates"),
1353 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1354 "event for reporting: "
1355 "vmexit, mmio (x86 only), ioport (x86 only)"),
1356 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1357 "vcpu id to report"),
1358 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1359 "key for sorting: sample(sort by samples number)"
1360 " time (sort by avg time)"),
1361 OPT_U64(0, "duration", &kvm
->duration
,
1362 "show events other than"
1363 " HLT (x86 only) or Wait state (s390 only)"
1364 " that take longer than duration usecs"),
1365 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout
,
1366 "per thread proc mmap processing timeout in ms"),
1369 const char * const live_usage
[] = {
1370 "perf kvm stat live [<options>]",
1373 struct perf_data data
= {
1374 .mode
= PERF_DATA_MODE_WRITE
,
1378 /* event handling */
1379 kvm
->tool
.sample
= process_sample_event
;
1380 kvm
->tool
.comm
= perf_event__process_comm
;
1381 kvm
->tool
.exit
= perf_event__process_exit
;
1382 kvm
->tool
.fork
= perf_event__process_fork
;
1383 kvm
->tool
.lost
= process_lost_event
;
1384 kvm
->tool
.namespaces
= perf_event__process_namespaces
;
1385 kvm
->tool
.ordered_events
= true;
1386 perf_tool__fill_defaults(&kvm
->tool
);
1389 kvm
->display_time
= 1;
1390 kvm
->opts
.user_interval
= 1;
1391 kvm
->opts
.mmap_pages
= 512;
1392 kvm
->opts
.target
.uses_mmap
= false;
1393 kvm
->opts
.target
.uid_str
= NULL
;
1394 kvm
->opts
.target
.uid
= UINT_MAX
;
1397 disable_buildid_cache();
1402 argc
= parse_options(argc
, argv
, live_options
,
1405 usage_with_options(live_usage
, live_options
);
1408 kvm
->duration
*= NSEC_PER_USEC
; /* convert usec to nsec */
1411 * target related setups
1413 err
= target__validate(&kvm
->opts
.target
);
1415 target__strerror(&kvm
->opts
.target
, err
, errbuf
, BUFSIZ
);
1416 ui__warning("%s", errbuf
);
1419 if (target__none(&kvm
->opts
.target
))
1420 kvm
->opts
.target
.system_wide
= true;
1424 * generate the event list
1426 err
= setup_kvm_events_tp(kvm
);
1428 pr_err("Unable to setup the kvm tracepoints\n");
1432 kvm
->evlist
= kvm_live_event_list();
1433 if (kvm
->evlist
== NULL
) {
1438 if (perf_evlist__create_maps(kvm
->evlist
, &kvm
->opts
.target
) < 0)
1439 usage_with_options(live_usage
, live_options
);
1444 kvm
->session
= perf_session__new(&data
, false, &kvm
->tool
);
1445 if (kvm
->session
== NULL
) {
1449 kvm
->session
->evlist
= kvm
->evlist
;
1450 perf_session__set_id_hdr_size(kvm
->session
);
1451 ordered_events__set_copy_on_queue(&kvm
->session
->ordered_events
, true);
1452 machine__synthesize_threads(&kvm
->session
->machines
.host
, &kvm
->opts
.target
,
1453 kvm
->evlist
->threads
, false, 1);
1454 err
= kvm_live_open_events(kvm
);
1458 err
= kvm_events_live_report(kvm
);
1461 perf_session__delete(kvm
->session
);
1462 kvm
->session
= NULL
;
1463 perf_evlist__delete(kvm
->evlist
);
1469 static void print_kvm_stat_usage(void)
1471 printf("Usage: perf kvm stat <command>\n\n");
1473 printf("# Available commands:\n");
1474 printf("\trecord: record kvm events\n");
1475 printf("\treport: report statistical data of kvm events\n");
1476 printf("\tlive: live reporting of statistical data of kvm events\n");
1478 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1481 static int kvm_cmd_stat(const char *file_name
, int argc
, const char **argv
)
1483 struct perf_kvm_stat kvm
= {
1484 .file_name
= file_name
,
1487 .report_event
= "vmexit",
1488 .sort_key
= "sample",
1493 print_kvm_stat_usage();
1497 if (!strncmp(argv
[1], "rec", 3))
1498 return kvm_events_record(&kvm
, argc
- 1, argv
+ 1);
1500 if (!strncmp(argv
[1], "rep", 3))
1501 return kvm_events_report(&kvm
, argc
- 1 , argv
+ 1);
1503 #ifdef HAVE_TIMERFD_SUPPORT
1504 if (!strncmp(argv
[1], "live", 4))
1505 return kvm_events_live(&kvm
, argc
- 1 , argv
+ 1);
1509 return cmd_stat(argc
, argv
);
1511 #endif /* HAVE_KVM_STAT_SUPPORT */
1513 static int __cmd_record(const char *file_name
, int argc
, const char **argv
)
1515 int rec_argc
, i
= 0, j
;
1516 const char **rec_argv
;
1518 rec_argc
= argc
+ 2;
1519 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1520 rec_argv
[i
++] = strdup("record");
1521 rec_argv
[i
++] = strdup("-o");
1522 rec_argv
[i
++] = strdup(file_name
);
1523 for (j
= 1; j
< argc
; j
++, i
++)
1524 rec_argv
[i
] = argv
[j
];
1526 BUG_ON(i
!= rec_argc
);
1528 return cmd_record(i
, rec_argv
);
1531 static int __cmd_report(const char *file_name
, int argc
, const char **argv
)
1533 int rec_argc
, i
= 0, j
;
1534 const char **rec_argv
;
1536 rec_argc
= argc
+ 2;
1537 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1538 rec_argv
[i
++] = strdup("report");
1539 rec_argv
[i
++] = strdup("-i");
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_report(i
, rec_argv
);
1550 __cmd_buildid_list(const char *file_name
, int argc
, const char **argv
)
1552 int rec_argc
, i
= 0, j
;
1553 const char **rec_argv
;
1555 rec_argc
= argc
+ 2;
1556 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1557 rec_argv
[i
++] = strdup("buildid-list");
1558 rec_argv
[i
++] = strdup("-i");
1559 rec_argv
[i
++] = strdup(file_name
);
1560 for (j
= 1; j
< argc
; j
++, i
++)
1561 rec_argv
[i
] = argv
[j
];
1563 BUG_ON(i
!= rec_argc
);
1565 return cmd_buildid_list(i
, rec_argv
);
1568 int cmd_kvm(int argc
, const char **argv
)
1570 const char *file_name
= NULL
;
1571 const struct option kvm_options
[] = {
1572 OPT_STRING('i', "input", &file_name
, "file",
1574 OPT_STRING('o', "output", &file_name
, "file",
1575 "Output file name"),
1576 OPT_BOOLEAN(0, "guest", &perf_guest
,
1577 "Collect guest os data"),
1578 OPT_BOOLEAN(0, "host", &perf_host
,
1579 "Collect host os data"),
1580 OPT_STRING(0, "guestmount", &symbol_conf
.guestmount
, "directory",
1581 "guest mount directory under which every guest os"
1582 " instance has a subdir"),
1583 OPT_STRING(0, "guestvmlinux", &symbol_conf
.default_guest_vmlinux_name
,
1584 "file", "file saving guest os vmlinux"),
1585 OPT_STRING(0, "guestkallsyms", &symbol_conf
.default_guest_kallsyms
,
1586 "file", "file saving guest os /proc/kallsyms"),
1587 OPT_STRING(0, "guestmodules", &symbol_conf
.default_guest_modules
,
1588 "file", "file saving guest os /proc/modules"),
1589 OPT_INCR('v', "verbose", &verbose
,
1590 "be more verbose (show counter open errors, etc)"),
1594 const char *const kvm_subcommands
[] = { "top", "record", "report", "diff",
1595 "buildid-list", "stat", NULL
};
1596 const char *kvm_usage
[] = { NULL
, NULL
};
1601 argc
= parse_options_subcommand(argc
, argv
, kvm_options
, kvm_subcommands
, kvm_usage
,
1602 PARSE_OPT_STOP_AT_NON_OPTION
);
1604 usage_with_options(kvm_usage
, kvm_options
);
1610 file_name
= get_filename_for_perf_kvm();
1613 pr_err("Failed to allocate memory for filename\n");
1618 if (!strncmp(argv
[0], "rec", 3))
1619 return __cmd_record(file_name
, argc
, argv
);
1620 else if (!strncmp(argv
[0], "rep", 3))
1621 return __cmd_report(file_name
, argc
, argv
);
1622 else if (!strncmp(argv
[0], "diff", 4))
1623 return cmd_diff(argc
, argv
);
1624 else if (!strncmp(argv
[0], "top", 3))
1625 return cmd_top(argc
, argv
);
1626 else if (!strncmp(argv
[0], "buildid-list", 12))
1627 return __cmd_buildid_list(file_name
, argc
, argv
);
1628 #ifdef HAVE_KVM_STAT_SUPPORT
1629 else if (!strncmp(argv
[0], "stat", 4))
1630 return kvm_cmd_stat(file_name
, argc
, argv
);
1633 usage_with_options(kvm_usage
, kvm_options
);