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