tools/perf/builtin-kvm.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "builtin.h"
   3 #include "perf.h"
   4
   5 #include "util/build-id.h"
   6 #include "util/evsel.h"
   7 #include "util/evlist.h"
   8 #include "util/term.h"
   9 #include "util/symbol.h"
  10 #include "util/thread.h"
  11 #include "util/header.h"
  12 #include "util/session.h"
  13 #include "util/intlist.h"
  14 #include <subcmd/pager.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"
  20 #include "util/synthetic-events.h"
  21 #include "util/top.h"
  22 #include "util/data.h"
  23 #include "util/ordered-events.h"
  24 #include "util/kvm-stat.h"
  25 #include "ui/ui.h"
  26
  27 #include <sys/prctl.h>
  28 #ifdef HAVE_TIMERFD_SUPPORT
  29 #include <sys/timerfd.h>
  30 #endif
  31 #include <sys/time.h>
  32 #include <sys/types.h>
  33 #include <sys/stat.h>
  34 #include <fcntl.h>
  35
  36 #include <linux/err.h>
  37 #include <linux/kernel.h>
  38 #include <linux/string.h>
  39 #include <linux/time64.h>
  40 #include <linux/zalloc.h>
  41 #include <errno.h>
  42 #include <inttypes.h>
  43 #include <poll.h>
  44 #include <termios.h>
  45 #include <semaphore.h>
  46 #include <signal.h>
  47 #include <math.h>
  48
  49 static const char *get_filename_for_perf_kvm(void)
  50 {
  51         const char *filename;
  52
  53         if (perf_host && !perf_guest)
  54                 filename = strdup("perf.data.host");
  55         else if (!perf_host && perf_guest)
  56                 filename = strdup("perf.data.guest");
  57         else
  58                 filename = strdup("perf.data.kvm");
  59
  60         return filename;
  61 }
  62
  63 #ifdef HAVE_KVM_STAT_SUPPORT
  64
  65 void exit_event_get_key(struct evsel *evsel,
  66                         struct perf_sample *sample,
  67                         struct event_key *key)
  68 {
  69         key->info = 0;
  70         key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
  71 }
  72
  73 bool kvm_exit_event(struct evsel *evsel)
  74 {
  75         return !strcmp(evsel->name, kvm_exit_trace);
  76 }
  77
  78 bool exit_event_begin(struct evsel *evsel,
  79                       struct perf_sample *sample, struct event_key *key)
  80 {
  81         if (kvm_exit_event(evsel)) {
  82                 exit_event_get_key(evsel, sample, key);
  83                 return true;
  84         }
  85
  86         return false;
  87 }
  88
  89 bool kvm_entry_event(struct evsel *evsel)
  90 {
  91         return !strcmp(evsel->name, kvm_entry_trace);
  92 }
  93
  94 bool exit_event_end(struct evsel *evsel,
  95                     struct perf_sample *sample __maybe_unused,
  96                     struct event_key *key __maybe_unused)
  97 {
  98         return kvm_entry_event(evsel);
  99 }
 100
 101 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
 102                                    struct exit_reasons_table *tbl,
 103                                    u64 exit_code)
 104 {
 105         while (tbl->reason != NULL) {
 106                 if (tbl->exit_code == exit_code)
 107                         return tbl->reason;
 108                 tbl++;
 109         }
 110
 111         pr_err("unknown kvm exit code:%lld on %s\n",
 112                 (unsigned long long)exit_code, kvm->exit_reasons_isa);
 113         return "UNKNOWN";
 114 }
 115
 116 void exit_event_decode_key(struct perf_kvm_stat *kvm,
 117                            struct event_key *key,
 118                            char *decode)
 119 {
 120         const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
 121                                                   key->key);
 122
 123         scnprintf(decode, decode_str_len, "%s", exit_reason);
 124 }
 125
 126 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
 127 {
 128         struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
 129
 130         for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
 131                 if (!strcmp(events_ops->name, kvm->report_event)) {
 132                         kvm->events_ops = events_ops->ops;
 133                         return true;
 134                 }
 135         }
 136
 137         return false;
 138 }
 139
 140 struct vcpu_event_record {
 141         int vcpu_id;
 142         u64 start_time;
 143         struct kvm_event *last_event;
 144 };
 145
 146
 147 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
 148 {
 149         unsigned int i;
 150
 151         for (i = 0; i < EVENTS_CACHE_SIZE; i++)
 152                 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
 153 }
 154
 155 #ifdef HAVE_TIMERFD_SUPPORT
 156 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
 157 {
 158         struct list_head *head;
 159         struct kvm_event *event;
 160         unsigned int i;
 161         int j;
 162
 163         for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 164                 head = &kvm_events_cache[i];
 165                 list_for_each_entry(event, head, hash_entry) {
 166                         /* reset stats for event */
 167                         event->total.time = 0;
 168                         init_stats(&event->total.stats);
 169
 170                         for (j = 0; j < event->max_vcpu; ++j) {
 171                                 event->vcpu[j].time = 0;
 172                                 init_stats(&event->vcpu[j].stats);
 173                         }
 174                 }
 175         }
 176 }
 177 #endif
 178
 179 static int kvm_events_hash_fn(u64 key)
 180 {
 181         return key & (EVENTS_CACHE_SIZE - 1);
 182 }
 183
 184 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 185 {
 186         int old_max_vcpu = event->max_vcpu;
 187         void *prev;
 188
 189         if (vcpu_id < event->max_vcpu)
 190                 return true;
 191
 192         while (event->max_vcpu <= vcpu_id)
 193                 event->max_vcpu += DEFAULT_VCPU_NUM;
 194
 195         prev = event->vcpu;
 196         event->vcpu = realloc(event->vcpu,
 197                               event->max_vcpu * sizeof(*event->vcpu));
 198         if (!event->vcpu) {
 199                 free(prev);
 200                 pr_err("Not enough memory\n");
 201                 return false;
 202         }
 203
 204         memset(event->vcpu + old_max_vcpu, 0,
 205                (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 206         return true;
 207 }
 208
 209 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 210 {
 211         struct kvm_event *event;
 212
 213         event = zalloc(sizeof(*event));
 214         if (!event) {
 215                 pr_err("Not enough memory\n");
 216                 return NULL;
 217         }
 218
 219         event->key = *key;
 220         init_stats(&event->total.stats);
 221         return event;
 222 }
 223
 224 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 225                                                struct event_key *key)
 226 {
 227         struct kvm_event *event;
 228         struct list_head *head;
 229
 230         BUG_ON(key->key == INVALID_KEY);
 231
 232         head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 233         list_for_each_entry(event, head, hash_entry) {
 234                 if (event->key.key == key->key && event->key.info == key->info)
 235                         return event;
 236         }
 237
 238         event = kvm_alloc_init_event(key);
 239         if (!event)
 240                 return NULL;
 241
 242         list_add(&event->hash_entry, head);
 243         return event;
 244 }
 245
 246 static bool handle_begin_event(struct perf_kvm_stat *kvm,
 247                                struct vcpu_event_record *vcpu_record,
 248                                struct event_key *key, u64 timestamp)
 249 {
 250         struct kvm_event *event = NULL;
 251
 252         if (key->key != INVALID_KEY)
 253                 event = find_create_kvm_event(kvm, key);
 254
 255         vcpu_record->last_event = event;
 256         vcpu_record->start_time = timestamp;
 257         return true;
 258 }
 259
 260 static void
 261 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 262 {
 263         kvm_stats->time += time_diff;
 264         update_stats(&kvm_stats->stats, time_diff);
 265 }
 266
 267 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 268 {
 269         struct kvm_event_stats *kvm_stats = &event->total;
 270
 271         if (vcpu_id != -1)
 272                 kvm_stats = &event->vcpu[vcpu_id];
 273
 274         return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 275                                 avg_stats(&kvm_stats->stats));
 276 }
 277
 278 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 279                              u64 time_diff)
 280 {
 281         if (vcpu_id == -1) {
 282                 kvm_update_event_stats(&event->total, time_diff);
 283                 return true;
 284         }
 285
 286         if (!kvm_event_expand(event, vcpu_id))
 287                 return false;
 288
 289         kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 290         return true;
 291 }
 292
 293 static bool is_child_event(struct perf_kvm_stat *kvm,
 294                            struct evsel *evsel,
 295                            struct perf_sample *sample,
 296                            struct event_key *key)
 297 {
 298         struct child_event_ops *child_ops;
 299
 300         child_ops = kvm->events_ops->child_ops;
 301
 302         if (!child_ops)
 303                 return false;
 304
 305         for (; child_ops->name; child_ops++) {
 306                 if (!strcmp(evsel->name, child_ops->name)) {
 307                         child_ops->get_key(evsel, sample, key);
 308                         return true;
 309                 }
 310         }
 311
 312         return false;
 313 }
 314
 315 static bool handle_child_event(struct perf_kvm_stat *kvm,
 316                                struct vcpu_event_record *vcpu_record,
 317                                struct event_key *key,
 318                                struct perf_sample *sample __maybe_unused)
 319 {
 320         struct kvm_event *event = NULL;
 321
 322         if (key->key != INVALID_KEY)
 323                 event = find_create_kvm_event(kvm, key);
 324
 325         vcpu_record->last_event = event;
 326
 327         return true;
 328 }
 329
 330 static bool skip_event(const char *event)
 331 {
 332         const char * const *skip_events;
 333
 334         for (skip_events = kvm_skip_events; *skip_events; skip_events++)
 335                 if (!strcmp(event, *skip_events))
 336                         return true;
 337
 338         return false;
 339 }
 340
 341 static bool handle_end_event(struct perf_kvm_stat *kvm,
 342                              struct vcpu_event_record *vcpu_record,
 343                              struct event_key *key,
 344                              struct perf_sample *sample)
 345 {
 346         struct kvm_event *event;
 347         u64 time_begin, time_diff;
 348         int vcpu;
 349
 350         if (kvm->trace_vcpu == -1)
 351                 vcpu = -1;
 352         else
 353                 vcpu = vcpu_record->vcpu_id;
 354
 355         event = vcpu_record->last_event;
 356         time_begin = vcpu_record->start_time;
 357
 358         /* The begin event is not caught. */
 359         if (!time_begin)
 360                 return true;
 361
 362         /*
 363          * In some case, the 'begin event' only records the start timestamp,
 364          * the actual event is recognized in the 'end event' (e.g. mmio-event).
 365          */
 366
 367         /* Both begin and end events did not get the key. */
 368         if (!event && key->key == INVALID_KEY)
 369                 return true;
 370
 371         if (!event)
 372                 event = find_create_kvm_event(kvm, key);
 373
 374         if (!event)
 375                 return false;
 376
 377         vcpu_record->last_event = NULL;
 378         vcpu_record->start_time = 0;
 379
 380         /* seems to happen once in a while during live mode */
 381         if (sample->time < time_begin) {
 382                 pr_debug("End time before begin time; skipping event.\n");
 383                 return true;
 384         }
 385
 386         time_diff = sample->time - time_begin;
 387
 388         if (kvm->duration && time_diff > kvm->duration) {
 389                 char decode[decode_str_len];
 390
 391                 kvm->events_ops->decode_key(kvm, &event->key, decode);
 392                 if (!skip_event(decode)) {
 393                         pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
 394                                  sample->time, sample->pid, vcpu_record->vcpu_id,
 395                                  decode, time_diff / NSEC_PER_USEC);
 396                 }
 397         }
 398
 399         return update_kvm_event(event, vcpu, time_diff);
 400 }
 401
 402 static
 403 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 404                                           struct evsel *evsel,
 405                                           struct perf_sample *sample)
 406 {
 407         /* Only kvm_entry records vcpu id. */
 408         if (!thread__priv(thread) && kvm_entry_event(evsel)) {
 409                 struct vcpu_event_record *vcpu_record;
 410
 411                 vcpu_record = zalloc(sizeof(*vcpu_record));
 412                 if (!vcpu_record) {
 413                         pr_err("%s: Not enough memory\n", __func__);
 414                         return NULL;
 415                 }
 416
 417                 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
 418                                                           vcpu_id_str);
 419                 thread__set_priv(thread, vcpu_record);
 420         }
 421
 422         return thread__priv(thread);
 423 }
 424
 425 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
 426                              struct thread *thread,
 427                              struct evsel *evsel,
 428                              struct perf_sample *sample)
 429 {
 430         struct vcpu_event_record *vcpu_record;
 431         struct event_key key = { .key = INVALID_KEY,
 432                                  .exit_reasons = kvm->exit_reasons };
 433
 434         vcpu_record = per_vcpu_record(thread, evsel, sample);
 435         if (!vcpu_record)
 436                 return true;
 437
 438         /* only process events for vcpus user cares about */
 439         if ((kvm->trace_vcpu != -1) &&
 440             (kvm->trace_vcpu != vcpu_record->vcpu_id))
 441                 return true;
 442
 443         if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 444                 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
 445
 446         if (is_child_event(kvm, evsel, sample, &key))
 447                 return handle_child_event(kvm, vcpu_record, &key, sample);
 448
 449         if (kvm->events_ops->is_end_event(evsel, sample, &key))
 450                 return handle_end_event(kvm, vcpu_record, &key, sample);
 451
 452         return true;
 453 }
 454
 455 #define GET_EVENT_KEY(func, field)                                      \
 456 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)         \
 457 {                                                                       \
 458         if (vcpu == -1)                                                 \
 459                 return event->total.field;                              \
 460                                                                         \
 461         if (vcpu >= event->max_vcpu)                                    \
 462                 return 0;                                               \
 463                                                                         \
 464         return event->vcpu[vcpu].field;                                 \
 465 }
 466
 467 #define COMPARE_EVENT_KEY(func, field)                                  \
 468 GET_EVENT_KEY(func, field)                                              \
 469 static int compare_kvm_event_ ## func(struct kvm_event *one,            \
 470                                         struct kvm_event *two, int vcpu)\
 471 {                                                                       \
 472         return get_event_ ##func(one, vcpu) >                           \
 473                                 get_event_ ##func(two, vcpu);           \
 474 }
 475
 476 GET_EVENT_KEY(time, time);
 477 COMPARE_EVENT_KEY(count, stats.n);
 478 COMPARE_EVENT_KEY(mean, stats.mean);
 479 GET_EVENT_KEY(max, stats.max);
 480 GET_EVENT_KEY(min, stats.min);
 481
 482 #define DEF_SORT_NAME_KEY(name, compare_key)                            \
 483         { #name, compare_kvm_event_ ## compare_key }
 484
 485 static struct kvm_event_key keys[] = {
 486         DEF_SORT_NAME_KEY(sample, count),
 487         DEF_SORT_NAME_KEY(time, mean),
 488         { NULL, NULL }
 489 };
 490
 491 static bool select_key(struct perf_kvm_stat *kvm)
 492 {
 493         int i;
 494
 495         for (i = 0; keys[i].name; i++) {
 496                 if (!strcmp(keys[i].name, kvm->sort_key)) {
 497                         kvm->compare = keys[i].key;
 498                         return true;
 499                 }
 500         }
 501
 502         pr_err("Unknown compare key:%s\n", kvm->sort_key);
 503         return false;
 504 }
 505
 506 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
 507                              key_cmp_fun bigger, int vcpu)
 508 {
 509         struct rb_node **rb = &result->rb_node;
 510         struct rb_node *parent = NULL;
 511         struct kvm_event *p;
 512
 513         while (*rb) {
 514                 p = container_of(*rb, struct kvm_event, rb);
 515                 parent = *rb;
 516
 517                 if (bigger(event, p, vcpu))
 518                         rb = &(*rb)->rb_left;
 519                 else
 520                         rb = &(*rb)->rb_right;
 521         }
 522
 523         rb_link_node(&event->rb, parent, rb);
 524         rb_insert_color(&event->rb, result);
 525 }
 526
 527 static void
 528 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
 529 {
 530         int vcpu = kvm->trace_vcpu;
 531
 532         kvm->total_count += get_event_count(event, vcpu);
 533         kvm->total_time += get_event_time(event, vcpu);
 534 }
 535
 536 static bool event_is_valid(struct kvm_event *event, int vcpu)
 537 {
 538         return !!get_event_count(event, vcpu);
 539 }
 540
 541 static void sort_result(struct perf_kvm_stat *kvm)
 542 {
 543         unsigned int i;
 544         int vcpu = kvm->trace_vcpu;
 545         struct kvm_event *event;
 546
 547         for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 548                 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
 549                         if (event_is_valid(event, vcpu)) {
 550                                 update_total_count(kvm, event);
 551                                 insert_to_result(&kvm->result, event,
 552                                                  kvm->compare, vcpu);
 553                         }
 554                 }
 555         }
 556 }
 557
 558 /* returns left most element of result, and erase it */
 559 static struct kvm_event *pop_from_result(struct rb_root *result)
 560 {
 561         struct rb_node *node = rb_first(result);
 562
 563         if (!node)
 564                 return NULL;
 565
 566         rb_erase(node, result);
 567         return container_of(node, struct kvm_event, rb);
 568 }
 569
 570 static void print_vcpu_info(struct perf_kvm_stat *kvm)
 571 {
 572         int vcpu = kvm->trace_vcpu;
 573
 574         pr_info("Analyze events for ");
 575
 576         if (kvm->opts.target.system_wide)
 577                 pr_info("all VMs, ");
 578         else if (kvm->opts.target.pid)
 579                 pr_info("pid(s) %s, ", kvm->opts.target.pid);
 580         else
 581                 pr_info("dazed and confused on what is monitored, ");
 582
 583         if (vcpu == -1)
 584                 pr_info("all VCPUs:\n\n");
 585         else
 586                 pr_info("VCPU %d:\n\n", vcpu);
 587 }
 588
 589 static void show_timeofday(void)
 590 {
 591         char date[64];
 592         struct timeval tv;
 593         struct tm ltime;
 594
 595         gettimeofday(&tv, NULL);
 596         if (localtime_r(&tv.tv_sec, &ltime)) {
 597                 strftime(date, sizeof(date), "%H:%M:%S", &ltime);
 598                 pr_info("%s.%06ld", date, tv.tv_usec);
 599         } else
 600                 pr_info("00:00:00.000000");
 601
 602         return;
 603 }
 604
 605 static void print_result(struct perf_kvm_stat *kvm)
 606 {
 607         char decode[decode_str_len];
 608         struct kvm_event *event;
 609         int vcpu = kvm->trace_vcpu;
 610
 611         if (kvm->live) {
 612                 puts(CONSOLE_CLEAR);
 613                 show_timeofday();
 614         }
 615
 616         pr_info("\n\n");
 617         print_vcpu_info(kvm);
 618         pr_info("%*s ", decode_str_len, kvm->events_ops->name);
 619         pr_info("%10s ", "Samples");
 620         pr_info("%9s ", "Samples%");
 621
 622         pr_info("%9s ", "Time%");
 623         pr_info("%11s ", "Min Time");
 624         pr_info("%11s ", "Max Time");
 625         pr_info("%16s ", "Avg time");
 626         pr_info("\n\n");
 627
 628         while ((event = pop_from_result(&kvm->result))) {
 629                 u64 ecount, etime, max, min;
 630
 631                 ecount = get_event_count(event, vcpu);
 632                 etime = get_event_time(event, vcpu);
 633                 max = get_event_max(event, vcpu);
 634                 min = get_event_min(event, vcpu);
 635
 636                 kvm->events_ops->decode_key(kvm, &event->key, decode);
 637                 pr_info("%*s ", decode_str_len, decode);
 638                 pr_info("%10llu ", (unsigned long long)ecount);
 639                 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
 640                 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
 641                 pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
 642                 pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
 643                 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
 644                         kvm_event_rel_stddev(vcpu, event));
 645                 pr_info("\n");
 646         }
 647
 648         pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
 649                 kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
 650
 651         if (kvm->lost_events)
 652                 pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
 653 }
 654
 655 #ifdef HAVE_TIMERFD_SUPPORT
 656 static int process_lost_event(struct perf_tool *tool,
 657                               union perf_event *event __maybe_unused,
 658                               struct perf_sample *sample __maybe_unused,
 659                               struct machine *machine __maybe_unused)
 660 {
 661         struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
 662
 663         kvm->lost_events++;
 664         return 0;
 665 }
 666 #endif
 667
 668 static bool skip_sample(struct perf_kvm_stat *kvm,
 669                         struct perf_sample *sample)
 670 {
 671         if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
 672                 return true;
 673
 674         return false;
 675 }
 676
 677 static int process_sample_event(struct perf_tool *tool,
 678                                 union perf_event *event,
 679                                 struct perf_sample *sample,
 680                                 struct evsel *evsel,
 681                                 struct machine *machine)
 682 {
 683         int err = 0;
 684         struct thread *thread;
 685         struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
 686                                                  tool);
 687
 688         if (skip_sample(kvm, sample))
 689                 return 0;
 690
 691         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 692         if (thread == NULL) {
 693                 pr_debug("problem processing %d event, skipping it.\n",
 694                         event->header.type);
 695                 return -1;
 696         }
 697
 698         if (!handle_kvm_event(kvm, thread, evsel, sample))
 699                 err = -1;
 700
 701         thread__put(thread);
 702         return err;
 703 }
 704
 705 static int cpu_isa_config(struct perf_kvm_stat *kvm)
 706 {
 707         char buf[64], *cpuid;
 708         int err;
 709
 710         if (kvm->live) {
 711                 err = get_cpuid(buf, sizeof(buf));
 712                 if (err != 0) {
 713                         pr_err("Failed to look up CPU type\n");
 714                         return err;
 715                 }
 716                 cpuid = buf;
 717         } else
 718                 cpuid = kvm->session->header.env.cpuid;
 719
 720         if (!cpuid) {
 721                 pr_err("Failed to look up CPU type\n");
 722                 return -EINVAL;
 723         }
 724
 725         err = cpu_isa_init(kvm, cpuid);
 726         if (err == -ENOTSUP)
 727                 pr_err("CPU %s is not supported.\n", cpuid);
 728
 729         return err;
 730 }
 731
 732 static bool verify_vcpu(int vcpu)
 733 {
 734         if (vcpu != -1 && vcpu < 0) {
 735                 pr_err("Invalid vcpu:%d.\n", vcpu);
 736                 return false;
 737         }
 738
 739         return true;
 740 }
 741
 742 #ifdef HAVE_TIMERFD_SUPPORT
 743 /* keeping the max events to a modest level to keep
 744  * the processing of samples per mmap smooth.
 745  */
 746 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
 747
 748 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
 749                                    u64 *mmap_time)
 750 {
 751         struct evlist *evlist = kvm->evlist;
 752         union perf_event *event;
 753         struct mmap *md;
 754         u64 timestamp;
 755         s64 n = 0;
 756         int err;
 757
 758         *mmap_time = ULLONG_MAX;
 759         md = &evlist->mmap[idx];
 760         err = perf_mmap__read_init(md);
 761         if (err < 0)
 762                 return (err == -EAGAIN) ? 0 : -1;
 763
 764         while ((event = perf_mmap__read_event(md)) != NULL) {
 765                 err = perf_evlist__parse_sample_timestamp(evlist, event, &timestamp);
 766                 if (err) {
 767                         perf_mmap__consume(md);
 768                         pr_err("Failed to parse sample\n");
 769                         return -1;
 770                 }
 771
 772                 err = perf_session__queue_event(kvm->session, event, timestamp, 0);
 773                 /*
 774                  * FIXME: Here we can't consume the event, as perf_session__queue_event will
 775                  *        point to it, and it'll get possibly overwritten by the kernel.
 776                  */
 777                 perf_mmap__consume(md);
 778
 779                 if (err) {
 780                         pr_err("Failed to enqueue sample: %d\n", err);
 781                         return -1;
 782                 }
 783
 784                 /* save time stamp of our first sample for this mmap */
 785                 if (n == 0)
 786                         *mmap_time = timestamp;
 787
 788                 /* limit events per mmap handled all at once */
 789                 n++;
 790                 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 791                         break;
 792         }
 793
 794         perf_mmap__read_done(md);
 795         return n;
 796 }
 797
 798 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
 799 {
 800         int i, err, throttled = 0;
 801         s64 n, ntotal = 0;
 802         u64 flush_time = ULLONG_MAX, mmap_time;
 803
 804         for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
 805                 n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
 806                 if (n < 0)
 807                         return -1;
 808
 809                 /* flush time is going to be the minimum of all the individual
 810                  * mmap times. Essentially, we flush all the samples queued up
 811                  * from the last pass under our minimal start time -- that leaves
 812                  * a very small race for samples to come in with a lower timestamp.
 813                  * The ioctl to return the perf_clock timestamp should close the
 814                  * race entirely.
 815                  */
 816                 if (mmap_time < flush_time)
 817                         flush_time = mmap_time;
 818
 819                 ntotal += n;
 820                 if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 821                         throttled = 1;
 822         }
 823
 824         /* flush queue after each round in which we processed events */
 825         if (ntotal) {
 826                 struct ordered_events *oe = &kvm->session->ordered_events;
 827
 828                 oe->next_flush = flush_time;
 829                 err = ordered_events__flush(oe, OE_FLUSH__ROUND);
 830                 if (err) {
 831                         if (kvm->lost_events)
 832                                 pr_info("\nLost events: %" PRIu64 "\n\n",
 833                                         kvm->lost_events);
 834                         return err;
 835                 }
 836         }
 837
 838         return throttled;
 839 }
 840
 841 static volatile int done;
 842
 843 static void sig_handler(int sig __maybe_unused)
 844 {
 845         done = 1;
 846 }
 847
 848 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
 849 {
 850         struct itimerspec new_value;
 851         int rc = -1;
 852
 853         kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
 854         if (kvm->timerfd < 0) {
 855                 pr_err("timerfd_create failed\n");
 856                 goto out;
 857         }
 858
 859         new_value.it_value.tv_sec = kvm->display_time;
 860         new_value.it_value.tv_nsec = 0;
 861         new_value.it_interval.tv_sec = kvm->display_time;
 862         new_value.it_interval.tv_nsec = 0;
 863
 864         if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
 865                 pr_err("timerfd_settime failed: %d\n", errno);
 866                 close(kvm->timerfd);
 867                 goto out;
 868         }
 869
 870         rc = 0;
 871 out:
 872         return rc;
 873 }
 874
 875 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
 876 {
 877         uint64_t c;
 878         int rc;
 879
 880         rc = read(kvm->timerfd, &c, sizeof(uint64_t));
 881         if (rc < 0) {
 882                 if (errno == EAGAIN)
 883                         return 0;
 884
 885                 pr_err("Failed to read timer fd: %d\n", errno);
 886                 return -1;
 887         }
 888
 889         if (rc != sizeof(uint64_t)) {
 890                 pr_err("Error reading timer fd - invalid size returned\n");
 891                 return -1;
 892         }
 893
 894         if (c != 1)
 895                 pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
 896
 897         /* update display */
 898         sort_result(kvm);
 899         print_result(kvm);
 900
 901         /* reset counts */
 902         clear_events_cache_stats(kvm->kvm_events_cache);
 903         kvm->total_count = 0;
 904         kvm->total_time = 0;
 905         kvm->lost_events = 0;
 906
 907         return 0;
 908 }
 909
 910 static int fd_set_nonblock(int fd)
 911 {
 912         long arg = 0;
 913
 914         arg = fcntl(fd, F_GETFL);
 915         if (arg < 0) {
 916                 pr_err("Failed to get current flags for fd %d\n", fd);
 917                 return -1;
 918         }
 919
 920         if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
 921                 pr_err("Failed to set non-block option on fd %d\n", fd);
 922                 return -1;
 923         }
 924
 925         return 0;
 926 }
 927
 928 static int perf_kvm__handle_stdin(void)
 929 {
 930         int c;
 931
 932         c = getc(stdin);
 933         if (c == 'q')
 934                 return 1;
 935
 936         return 0;
 937 }
 938
 939 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
 940 {
 941         int nr_stdin, ret, err = -EINVAL;
 942         struct termios save;
 943
 944         /* live flag must be set first */
 945         kvm->live = true;
 946
 947         ret = cpu_isa_config(kvm);
 948         if (ret < 0)
 949                 return ret;
 950
 951         if (!verify_vcpu(kvm->trace_vcpu) ||
 952             !select_key(kvm) ||
 953             !register_kvm_events_ops(kvm)) {
 954                 goto out;
 955         }
 956
 957         set_term_quiet_input(&save);
 958         init_kvm_event_record(kvm);
 959
 960         signal(SIGINT, sig_handler);
 961         signal(SIGTERM, sig_handler);
 962
 963         /* add timer fd */
 964         if (perf_kvm__timerfd_create(kvm) < 0) {
 965                 err = -1;
 966                 goto out;
 967         }
 968
 969         if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
 970                 goto out;
 971
 972         nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
 973         if (nr_stdin < 0)
 974                 goto out;
 975
 976         if (fd_set_nonblock(fileno(stdin)) != 0)
 977                 goto out;
 978
 979         /* everything is good - enable the events and process */
 980         evlist__enable(kvm->evlist);
 981
 982         while (!done) {
 983                 struct fdarray *fda = &kvm->evlist->pollfd;
 984                 int rc;
 985
 986                 rc = perf_kvm__mmap_read(kvm);
 987                 if (rc < 0)
 988                         break;
 989
 990                 err = perf_kvm__handle_timerfd(kvm);
 991                 if (err)
 992                         goto out;
 993
 994                 if (fda->entries[nr_stdin].revents & POLLIN)
 995                         done = perf_kvm__handle_stdin();
 996
 997                 if (!rc && !done)
 998                         err = fdarray__poll(fda, 100);
 999         }
1000
1001         evlist__disable(kvm->evlist);
1002
1003         if (err == 0) {
1004                 sort_result(kvm);
1005                 print_result(kvm);
1006         }
1007
1008 out:
1009         if (kvm->timerfd >= 0)
1010                 close(kvm->timerfd);
1011
1012         tcsetattr(0, TCSAFLUSH, &save);
1013         return err;
1014 }
1015
1016 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1017 {
1018         int err, rc = -1;
1019         struct evsel *pos;
1020         struct evlist *evlist = kvm->evlist;
1021         char sbuf[STRERR_BUFSIZE];
1022
1023         perf_evlist__config(evlist, &kvm->opts, NULL);
1024
1025         /*
1026          * Note: exclude_{guest,host} do not apply here.
1027          *       This command processes KVM tracepoints from host only
1028          */
1029         evlist__for_each_entry(evlist, pos) {
1030                 struct perf_event_attr *attr = &pos->core.attr;
1031
1032                 /* make sure these *are* set */
1033                 perf_evsel__set_sample_bit(pos, TID);
1034                 perf_evsel__set_sample_bit(pos, TIME);
1035                 perf_evsel__set_sample_bit(pos, CPU);
1036                 perf_evsel__set_sample_bit(pos, RAW);
1037                 /* make sure these are *not*; want as small a sample as possible */
1038                 perf_evsel__reset_sample_bit(pos, PERIOD);
1039                 perf_evsel__reset_sample_bit(pos, IP);
1040                 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1041                 perf_evsel__reset_sample_bit(pos, ADDR);
1042                 perf_evsel__reset_sample_bit(pos, READ);
1043                 attr->mmap = 0;
1044                 attr->comm = 0;
1045                 attr->task = 0;
1046
1047                 attr->sample_period = 1;
1048
1049                 attr->watermark = 0;
1050                 attr->wakeup_events = 1000;
1051
1052                 /* will enable all once we are ready */
1053                 attr->disabled = 1;
1054         }
1055
1056         err = evlist__open(evlist);
1057         if (err < 0) {
1058                 printf("Couldn't create the events: %s\n",
1059                        str_error_r(errno, sbuf, sizeof(sbuf)));
1060                 goto out;
1061         }
1062
1063         if (evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1064                 ui__error("Failed to mmap the events: %s\n",
1065                           str_error_r(errno, sbuf, sizeof(sbuf)));
1066                 evlist__close(evlist);
1067                 goto out;
1068         }
1069
1070         rc = 0;
1071
1072 out:
1073         return rc;
1074 }
1075 #endif
1076
1077 static int read_events(struct perf_kvm_stat *kvm)
1078 {
1079         int ret;
1080
1081         struct perf_tool eops = {
1082                 .sample                 = process_sample_event,
1083                 .comm                   = perf_event__process_comm,
1084                 .namespaces             = perf_event__process_namespaces,
1085                 .ordered_events         = true,
1086         };
1087         struct perf_data file = {
1088                 .path  = kvm->file_name,
1089                 .mode  = PERF_DATA_MODE_READ,
1090                 .force = kvm->force,
1091         };
1092
1093         kvm->tool = eops;
1094         kvm->session = perf_session__new(&file, false, &kvm->tool);
1095         if (IS_ERR(kvm->session)) {
1096                 pr_err("Initializing perf session failed\n");
1097                 return PTR_ERR(kvm->session);
1098         }
1099
1100         symbol__init(&kvm->session->header.env);
1101
1102         if (!perf_session__has_traces(kvm->session, "kvm record")) {
1103                 ret = -EINVAL;
1104                 goto out_delete;
1105         }
1106
1107         /*
1108          * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1109          * traced in the old kernel.
1110          */
1111         ret = cpu_isa_config(kvm);
1112         if (ret < 0)
1113                 goto out_delete;
1114
1115         ret = perf_session__process_events(kvm->session);
1116
1117 out_delete:
1118         perf_session__delete(kvm->session);
1119         return ret;
1120 }
1121
1122 static int parse_target_str(struct perf_kvm_stat *kvm)
1123 {
1124         if (kvm->opts.target.pid) {
1125                 kvm->pid_list = intlist__new(kvm->opts.target.pid);
1126                 if (kvm->pid_list == NULL) {
1127                         pr_err("Error parsing process id string\n");
1128                         return -EINVAL;
1129                 }
1130         }
1131
1132         return 0;
1133 }
1134
1135 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1136 {
1137         int ret = -EINVAL;
1138         int vcpu = kvm->trace_vcpu;
1139
1140         if (parse_target_str(kvm) != 0)
1141                 goto exit;
1142
1143         if (!verify_vcpu(vcpu))
1144                 goto exit;
1145
1146         if (!select_key(kvm))
1147                 goto exit;
1148
1149         if (!register_kvm_events_ops(kvm))
1150                 goto exit;
1151
1152         init_kvm_event_record(kvm);
1153         setup_pager();
1154
1155         ret = read_events(kvm);
1156         if (ret)
1157                 goto exit;
1158
1159         sort_result(kvm);
1160         print_result(kvm);
1161
1162 exit:
1163         return ret;
1164 }
1165
1166 #define STRDUP_FAIL_EXIT(s)             \
1167         ({      char *_p;               \
1168         _p = strdup(s);         \
1169                 if (!_p)                \
1170                         return -ENOMEM; \
1171                 _p;                     \
1172         })
1173
1174 int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1175 {
1176         return 0;
1177 }
1178
1179 static int
1180 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1181 {
1182         unsigned int rec_argc, i, j, events_tp_size;
1183         const char **rec_argv;
1184         const char * const record_args[] = {
1185                 "record",
1186                 "-R",
1187                 "-m", "1024",
1188                 "-c", "1",
1189         };
1190         const char * const kvm_stat_record_usage[] = {
1191                 "perf kvm stat record [<options>]",
1192                 NULL
1193         };
1194         const char * const *events_tp;
1195         int ret;
1196
1197         events_tp_size = 0;
1198         ret = setup_kvm_events_tp(kvm);
1199         if (ret < 0) {
1200                 pr_err("Unable to setup the kvm tracepoints\n");
1201                 return ret;
1202         }
1203
1204         for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1205                 events_tp_size++;
1206
1207         rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1208                    2 * events_tp_size;
1209         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1210
1211         if (rec_argv == NULL)
1212                 return -ENOMEM;
1213
1214         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1215                 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1216
1217         for (j = 0; j < events_tp_size; j++) {
1218                 rec_argv[i++] = "-e";
1219                 rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1220         }
1221
1222         rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1223         rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1224
1225         for (j = 1; j < (unsigned int)argc; j++, i++)
1226                 rec_argv[i] = argv[j];
1227
1228         set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1229         set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1230         set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1231
1232         set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1233         set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1234         set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1235         set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1236         set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1237         set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1238         set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1239         set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1240         set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1241         set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1242         set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1243         set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1244         set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1245         set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1246         set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1247
1248         record_usage = kvm_stat_record_usage;
1249         return cmd_record(i, rec_argv);
1250 }
1251
1252 static int
1253 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1254 {
1255         const struct option kvm_events_report_options[] = {
1256                 OPT_STRING(0, "event", &kvm->report_event, "report event",
1257                            "event for reporting: vmexit, "
1258                            "mmio (x86 only), ioport (x86 only)"),
1259                 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1260                             "vcpu id to report"),
1261                 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1262                             "key for sorting: sample(sort by samples number)"
1263                             " time (sort by avg time)"),
1264                 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1265                            "analyze events only for given process id(s)"),
1266                 OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1267                 OPT_END()
1268         };
1269
1270         const char * const kvm_events_report_usage[] = {
1271                 "perf kvm stat report [<options>]",
1272                 NULL
1273         };
1274
1275         if (argc) {
1276                 argc = parse_options(argc, argv,
1277                                      kvm_events_report_options,
1278                                      kvm_events_report_usage, 0);
1279                 if (argc)
1280                         usage_with_options(kvm_events_report_usage,
1281                                            kvm_events_report_options);
1282         }
1283
1284         if (!kvm->opts.target.pid)
1285                 kvm->opts.target.system_wide = true;
1286
1287         return kvm_events_report_vcpu(kvm);
1288 }
1289
1290 #ifdef HAVE_TIMERFD_SUPPORT
1291 static struct evlist *kvm_live_event_list(void)
1292 {
1293         struct evlist *evlist;
1294         char *tp, *name, *sys;
1295         int err = -1;
1296         const char * const *events_tp;
1297
1298         evlist = evlist__new();
1299         if (evlist == NULL)
1300                 return NULL;
1301
1302         for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1303
1304                 tp = strdup(*events_tp);
1305                 if (tp == NULL)
1306                         goto out;
1307
1308                 /* split tracepoint into subsystem and name */
1309                 sys = tp;
1310                 name = strchr(tp, ':');
1311                 if (name == NULL) {
1312                         pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1313                                *events_tp);
1314                         free(tp);
1315                         goto out;
1316                 }
1317                 *name = '\0';
1318                 name++;
1319
1320                 if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1321                         pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1322                         free(tp);
1323                         goto out;
1324                 }
1325
1326                 free(tp);
1327         }
1328
1329         err = 0;
1330
1331 out:
1332         if (err) {
1333                 evlist__delete(evlist);
1334                 evlist = NULL;
1335         }
1336
1337         return evlist;
1338 }
1339
1340 static int kvm_events_live(struct perf_kvm_stat *kvm,
1341                            int argc, const char **argv)
1342 {
1343         char errbuf[BUFSIZ];
1344         int err;
1345
1346         const struct option live_options[] = {
1347                 OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1348                         "record events on existing process id"),
1349                 OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1350                         "number of mmap data pages",
1351                         perf_evlist__parse_mmap_pages),
1352                 OPT_INCR('v', "verbose", &verbose,
1353                         "be more verbose (show counter open errors, etc)"),
1354                 OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1355                         "system-wide collection from all CPUs"),
1356                 OPT_UINTEGER('d', "display", &kvm->display_time,
1357                         "time in seconds between display updates"),
1358                 OPT_STRING(0, "event", &kvm->report_event, "report event",
1359                         "event for reporting: "
1360                         "vmexit, mmio (x86 only), ioport (x86 only)"),
1361                 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1362                         "vcpu id to report"),
1363                 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1364                         "key for sorting: sample(sort by samples number)"
1365                         " time (sort by avg time)"),
1366                 OPT_U64(0, "duration", &kvm->duration,
1367                         "show events other than"
1368                         " HLT (x86 only) or Wait state (s390 only)"
1369                         " that take longer than duration usecs"),
1370                 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1371                                 "per thread proc mmap processing timeout in ms"),
1372                 OPT_END()
1373         };
1374         const char * const live_usage[] = {
1375                 "perf kvm stat live [<options>]",
1376                 NULL
1377         };
1378         struct perf_data data = {
1379                 .mode = PERF_DATA_MODE_WRITE,
1380         };
1381
1382
1383         /* event handling */
1384         kvm->tool.sample = process_sample_event;
1385         kvm->tool.comm   = perf_event__process_comm;
1386         kvm->tool.exit   = perf_event__process_exit;
1387         kvm->tool.fork   = perf_event__process_fork;
1388         kvm->tool.lost   = process_lost_event;
1389         kvm->tool.namespaces  = perf_event__process_namespaces;
1390         kvm->tool.ordered_events = true;
1391         perf_tool__fill_defaults(&kvm->tool);
1392
1393         /* set defaults */
1394         kvm->display_time = 1;
1395         kvm->opts.user_interval = 1;
1396         kvm->opts.mmap_pages = 512;
1397         kvm->opts.target.uses_mmap = false;
1398         kvm->opts.target.uid_str = NULL;
1399         kvm->opts.target.uid = UINT_MAX;
1400
1401         symbol__init(NULL);
1402         disable_buildid_cache();
1403
1404         use_browser = 0;
1405
1406         if (argc) {
1407                 argc = parse_options(argc, argv, live_options,
1408                                      live_usage, 0);
1409                 if (argc)
1410                         usage_with_options(live_usage, live_options);
1411         }
1412
1413         kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1414
1415         /*
1416          * target related setups
1417          */
1418         err = target__validate(&kvm->opts.target);
1419         if (err) {
1420                 target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1421                 ui__warning("%s", errbuf);
1422         }
1423
1424         if (target__none(&kvm->opts.target))
1425                 kvm->opts.target.system_wide = true;
1426
1427
1428         /*
1429          * generate the event list
1430          */
1431         err = setup_kvm_events_tp(kvm);
1432         if (err < 0) {
1433                 pr_err("Unable to setup the kvm tracepoints\n");
1434                 return err;
1435         }
1436
1437         kvm->evlist = kvm_live_event_list();
1438         if (kvm->evlist == NULL) {
1439                 err = -1;
1440                 goto out;
1441         }
1442
1443         if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1444                 usage_with_options(live_usage, live_options);
1445
1446         /*
1447          * perf session
1448          */
1449         kvm->session = perf_session__new(&data, false, &kvm->tool);
1450         if (IS_ERR(kvm->session)) {
1451                 err = PTR_ERR(kvm->session);
1452                 goto out;
1453         }
1454         kvm->session->evlist = kvm->evlist;
1455         perf_session__set_id_hdr_size(kvm->session);
1456         ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1457         machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1458                                     kvm->evlist->core.threads, false, 1);
1459         err = kvm_live_open_events(kvm);
1460         if (err)
1461                 goto out;
1462
1463         err = kvm_events_live_report(kvm);
1464
1465 out:
1466         perf_session__delete(kvm->session);
1467         kvm->session = NULL;
1468         evlist__delete(kvm->evlist);
1469
1470         return err;
1471 }
1472 #endif
1473
1474 static void print_kvm_stat_usage(void)
1475 {
1476         printf("Usage: perf kvm stat <command>\n\n");
1477
1478         printf("# Available commands:\n");
1479         printf("\trecord: record kvm events\n");
1480         printf("\treport: report statistical data of kvm events\n");
1481         printf("\tlive:   live reporting of statistical data of kvm events\n");
1482
1483         printf("\nOtherwise, it is the alias of 'perf stat':\n");
1484 }
1485
1486 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1487 {
1488         struct perf_kvm_stat kvm = {
1489                 .file_name = file_name,
1490
1491                 .trace_vcpu     = -1,
1492                 .report_event   = "vmexit",
1493                 .sort_key       = "sample",
1494
1495         };
1496
1497         if (argc == 1) {
1498                 print_kvm_stat_usage();
1499                 goto perf_stat;
1500         }
1501
1502         if (!strncmp(argv[1], "rec", 3))
1503                 return kvm_events_record(&kvm, argc - 1, argv + 1);
1504
1505         if (!strncmp(argv[1], "rep", 3))
1506                 return kvm_events_report(&kvm, argc - 1 , argv + 1);
1507
1508 #ifdef HAVE_TIMERFD_SUPPORT
1509         if (!strncmp(argv[1], "live", 4))
1510                 return kvm_events_live(&kvm, argc - 1 , argv + 1);
1511 #endif
1512
1513 perf_stat:
1514         return cmd_stat(argc, argv);
1515 }
1516 #endif /* HAVE_KVM_STAT_SUPPORT */
1517
1518 int __weak kvm_add_default_arch_event(int *argc __maybe_unused,
1519                                         const char **argv __maybe_unused)
1520 {
1521         return 0;
1522 }
1523
1524 static int __cmd_record(const char *file_name, int argc, const char **argv)
1525 {
1526         int rec_argc, i = 0, j, ret;
1527         const char **rec_argv;
1528
1529         ret = kvm_add_default_arch_event(&argc, argv);
1530         if (ret)
1531                 return -EINVAL;
1532
1533         rec_argc = argc + 2;
1534         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1535         rec_argv[i++] = strdup("record");
1536         rec_argv[i++] = strdup("-o");
1537         rec_argv[i++] = strdup(file_name);
1538         for (j = 1; j < argc; j++, i++)
1539                 rec_argv[i] = argv[j];
1540
1541         BUG_ON(i != rec_argc);
1542
1543         return cmd_record(i, rec_argv);
1544 }
1545
1546 static int __cmd_report(const char *file_name, int argc, const char **argv)
1547 {
1548         int rec_argc, i = 0, j;
1549         const char **rec_argv;
1550
1551         rec_argc = argc + 2;
1552         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1553         rec_argv[i++] = strdup("report");
1554         rec_argv[i++] = strdup("-i");
1555         rec_argv[i++] = strdup(file_name);
1556         for (j = 1; j < argc; j++, i++)
1557                 rec_argv[i] = argv[j];
1558
1559         BUG_ON(i != rec_argc);
1560
1561         return cmd_report(i, rec_argv);
1562 }
1563
1564 static int
1565 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1566 {
1567         int rec_argc, i = 0, j;
1568         const char **rec_argv;
1569
1570         rec_argc = argc + 2;
1571         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1572         rec_argv[i++] = strdup("buildid-list");
1573         rec_argv[i++] = strdup("-i");
1574         rec_argv[i++] = strdup(file_name);
1575         for (j = 1; j < argc; j++, i++)
1576                 rec_argv[i] = argv[j];
1577
1578         BUG_ON(i != rec_argc);
1579
1580         return cmd_buildid_list(i, rec_argv);
1581 }
1582
1583 int cmd_kvm(int argc, const char **argv)
1584 {
1585         const char *file_name = NULL;
1586         const struct option kvm_options[] = {
1587                 OPT_STRING('i', "input", &file_name, "file",
1588                            "Input file name"),
1589                 OPT_STRING('o', "output", &file_name, "file",
1590                            "Output file name"),
1591                 OPT_BOOLEAN(0, "guest", &perf_guest,
1592                             "Collect guest os data"),
1593                 OPT_BOOLEAN(0, "host", &perf_host,
1594                             "Collect host os data"),
1595                 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1596                            "guest mount directory under which every guest os"
1597                            " instance has a subdir"),
1598                 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1599                            "file", "file saving guest os vmlinux"),
1600                 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1601                            "file", "file saving guest os /proc/kallsyms"),
1602                 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1603                            "file", "file saving guest os /proc/modules"),
1604                 OPT_INCR('v', "verbose", &verbose,
1605                             "be more verbose (show counter open errors, etc)"),
1606                 OPT_END()
1607         };
1608
1609         const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1610                                                 "buildid-list", "stat", NULL };
1611         const char *kvm_usage[] = { NULL, NULL };
1612
1613         perf_host  = 0;
1614         perf_guest = 1;
1615
1616         argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1617                                         PARSE_OPT_STOP_AT_NON_OPTION);
1618         if (!argc)
1619                 usage_with_options(kvm_usage, kvm_options);
1620
1621         if (!perf_host)
1622                 perf_guest = 1;
1623
1624         if (!file_name) {
1625                 file_name = get_filename_for_perf_kvm();
1626
1627                 if (!file_name) {
1628                         pr_err("Failed to allocate memory for filename\n");
1629                         return -ENOMEM;
1630                 }
1631         }
1632
1633         if (!strncmp(argv[0], "rec", 3))
1634                 return __cmd_record(file_name, argc, argv);
1635         else if (!strncmp(argv[0], "rep", 3))
1636                 return __cmd_report(file_name, argc, argv);
1637         else if (!strncmp(argv[0], "diff", 4))
1638                 return cmd_diff(argc, argv);
1639         else if (!strncmp(argv[0], "top", 3))
1640                 return cmd_top(argc, argv);
1641         else if (!strncmp(argv[0], "buildid-list", 12))
1642                 return __cmd_buildid_list(file_name, argc, argv);
1643 #ifdef HAVE_KVM_STAT_SUPPORT
1644         else if (!strncmp(argv[0], "stat", 4))
1645                 return kvm_cmd_stat(file_name, argc, argv);
1646 #endif
1647         else
1648                 usage_with_options(kvm_usage, kvm_options);
1649
1650         return 0;
1651 }