tools/perf/builtin-kvm.c

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