tools/perf/builtin-kvm.c

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