tools/perf/builtin-kvm.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/evsel.h"
   5 #include "util/util.h"
   6 #include "util/cache.h"
   7 #include "util/symbol.h"
   8 #include "util/thread.h"
   9 #include "util/header.h"
  10 #include "util/session.h"
  11
  12 #include "util/parse-options.h"
  13 #include "util/trace-event.h"
  14 #include "util/debug.h"
  15 #include "util/debugfs.h"
  16 #include "util/tool.h"
  17 #include "util/stat.h"
  18
  19 #include <sys/prctl.h>
  20
  21 #include <semaphore.h>
  22 #include <pthread.h>
  23 #include <math.h>
  24
  25 #include "../../arch/x86/include/asm/svm.h"
  26 #include "../../arch/x86/include/asm/vmx.h"
  27 #include "../../arch/x86/include/asm/kvm.h"
  28
  29 struct event_key {
  30         #define INVALID_KEY     (~0ULL)
  31         u64 key;
  32         int info;
  33 };
  34
  35 struct kvm_event_stats {
  36         u64 time;
  37         struct stats stats;
  38 };
  39
  40 struct kvm_event {
  41         struct list_head hash_entry;
  42         struct rb_node rb;
  43
  44         struct event_key key;
  45
  46         struct kvm_event_stats total;
  47
  48         #define DEFAULT_VCPU_NUM 8
  49         int max_vcpu;
  50         struct kvm_event_stats *vcpu;
  51 };
  52
  53 typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
  54
  55 struct kvm_event_key {
  56         const char *name;
  57         key_cmp_fun key;
  58 };
  59
  60
  61 struct perf_kvm;
  62
  63 struct kvm_events_ops {
  64         bool (*is_begin_event)(struct perf_evsel *evsel,
  65                                struct perf_sample *sample,
  66                                struct event_key *key);
  67         bool (*is_end_event)(struct perf_evsel *evsel,
  68                              struct perf_sample *sample, struct event_key *key);
  69         void (*decode_key)(struct perf_kvm *kvm, struct event_key *key,
  70                            char decode[20]);
  71         const char *name;
  72 };
  73
  74 struct exit_reasons_table {
  75         unsigned long exit_code;
  76         const char *reason;
  77 };
  78
  79 #define EVENTS_BITS             12
  80 #define EVENTS_CACHE_SIZE       (1UL << EVENTS_BITS)
  81
  82 struct perf_kvm {
  83         struct perf_tool    tool;
  84         struct perf_session *session;
  85
  86         const char *file_name;
  87         const char *report_event;
  88         const char *sort_key;
  89         int trace_vcpu;
  90
  91         struct exit_reasons_table *exit_reasons;
  92         int exit_reasons_size;
  93         const char *exit_reasons_isa;
  94
  95         struct kvm_events_ops *events_ops;
  96         key_cmp_fun compare;
  97         struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
  98         u64 total_time;
  99         u64 total_count;
 100
 101         struct rb_root result;
 102 };
 103
 104
 105 static void exit_event_get_key(struct perf_evsel *evsel,
 106                                struct perf_sample *sample,
 107                                struct event_key *key)
 108 {
 109         key->info = 0;
 110         key->key = perf_evsel__intval(evsel, sample, "exit_reason");
 111 }
 112
 113 static bool kvm_exit_event(struct perf_evsel *evsel)
 114 {
 115         return !strcmp(evsel->name, "kvm:kvm_exit");
 116 }
 117
 118 static bool exit_event_begin(struct perf_evsel *evsel,
 119                              struct perf_sample *sample, struct event_key *key)
 120 {
 121         if (kvm_exit_event(evsel)) {
 122                 exit_event_get_key(evsel, sample, key);
 123                 return true;
 124         }
 125
 126         return false;
 127 }
 128
 129 static bool kvm_entry_event(struct perf_evsel *evsel)
 130 {
 131         return !strcmp(evsel->name, "kvm:kvm_entry");
 132 }
 133
 134 static bool exit_event_end(struct perf_evsel *evsel,
 135                            struct perf_sample *sample __maybe_unused,
 136                            struct event_key *key __maybe_unused)
 137 {
 138         return kvm_entry_event(evsel);
 139 }
 140
 141 static struct exit_reasons_table vmx_exit_reasons[] = {
 142         VMX_EXIT_REASONS
 143 };
 144
 145 static struct exit_reasons_table svm_exit_reasons[] = {
 146         SVM_EXIT_REASONS
 147 };
 148
 149 static const char *get_exit_reason(struct perf_kvm *kvm, u64 exit_code)
 150 {
 151         int i = kvm->exit_reasons_size;
 152         struct exit_reasons_table *tbl = kvm->exit_reasons;
 153
 154         while (i--) {
 155                 if (tbl->exit_code == exit_code)
 156                         return tbl->reason;
 157                 tbl++;
 158         }
 159
 160         pr_err("unknown kvm exit code:%lld on %s\n",
 161                 (unsigned long long)exit_code, kvm->exit_reasons_isa);
 162         return "UNKNOWN";
 163 }
 164
 165 static void exit_event_decode_key(struct perf_kvm *kvm,
 166                                   struct event_key *key,
 167                                   char decode[20])
 168 {
 169         const char *exit_reason = get_exit_reason(kvm, key->key);
 170
 171         scnprintf(decode, 20, "%s", exit_reason);
 172 }
 173
 174 static struct kvm_events_ops exit_events = {
 175         .is_begin_event = exit_event_begin,
 176         .is_end_event = exit_event_end,
 177         .decode_key = exit_event_decode_key,
 178         .name = "VM-EXIT"
 179 };
 180
 181 /*
 182  * For the mmio events, we treat:
 183  * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
 184  * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
 185  */
 186 static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
 187                                struct event_key *key)
 188 {
 189         key->key  = perf_evsel__intval(evsel, sample, "gpa");
 190         key->info = perf_evsel__intval(evsel, sample, "type");
 191 }
 192
 193 #define KVM_TRACE_MMIO_READ_UNSATISFIED 0
 194 #define KVM_TRACE_MMIO_READ 1
 195 #define KVM_TRACE_MMIO_WRITE 2
 196
 197 static bool mmio_event_begin(struct perf_evsel *evsel,
 198                              struct perf_sample *sample, struct event_key *key)
 199 {
 200         /* MMIO read begin event in kernel. */
 201         if (kvm_exit_event(evsel))
 202                 return true;
 203
 204         /* MMIO write begin event in kernel. */
 205         if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
 206             perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
 207                 mmio_event_get_key(evsel, sample, key);
 208                 return true;
 209         }
 210
 211         return false;
 212 }
 213
 214 static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
 215                            struct event_key *key)
 216 {
 217         /* MMIO write end event in kernel. */
 218         if (kvm_entry_event(evsel))
 219                 return true;
 220
 221         /* MMIO read end event in kernel.*/
 222         if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
 223             perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
 224                 mmio_event_get_key(evsel, sample, key);
 225                 return true;
 226         }
 227
 228         return false;
 229 }
 230
 231 static void mmio_event_decode_key(struct perf_kvm *kvm __maybe_unused,
 232                                   struct event_key *key,
 233                                   char decode[20])
 234 {
 235         scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
 236                                 key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
 237 }
 238
 239 static struct kvm_events_ops mmio_events = {
 240         .is_begin_event = mmio_event_begin,
 241         .is_end_event = mmio_event_end,
 242         .decode_key = mmio_event_decode_key,
 243         .name = "MMIO Access"
 244 };
 245
 246  /* The time of emulation pio access is from kvm_pio to kvm_entry. */
 247 static void ioport_event_get_key(struct perf_evsel *evsel,
 248                                  struct perf_sample *sample,
 249                                  struct event_key *key)
 250 {
 251         key->key  = perf_evsel__intval(evsel, sample, "port");
 252         key->info = perf_evsel__intval(evsel, sample, "rw");
 253 }
 254
 255 static bool ioport_event_begin(struct perf_evsel *evsel,
 256                                struct perf_sample *sample,
 257                                struct event_key *key)
 258 {
 259         if (!strcmp(evsel->name, "kvm:kvm_pio")) {
 260                 ioport_event_get_key(evsel, sample, key);
 261                 return true;
 262         }
 263
 264         return false;
 265 }
 266
 267 static bool ioport_event_end(struct perf_evsel *evsel,
 268                              struct perf_sample *sample __maybe_unused,
 269                              struct event_key *key __maybe_unused)
 270 {
 271         return kvm_entry_event(evsel);
 272 }
 273
 274 static void ioport_event_decode_key(struct perf_kvm *kvm __maybe_unused,
 275                                     struct event_key *key,
 276                                     char decode[20])
 277 {
 278         scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
 279                                 key->info ? "POUT" : "PIN");
 280 }
 281
 282 static struct kvm_events_ops ioport_events = {
 283         .is_begin_event = ioport_event_begin,
 284         .is_end_event = ioport_event_end,
 285         .decode_key = ioport_event_decode_key,
 286         .name = "IO Port Access"
 287 };
 288
 289 static bool register_kvm_events_ops(struct perf_kvm *kvm)
 290 {
 291         bool ret = true;
 292
 293         if (!strcmp(kvm->report_event, "vmexit"))
 294                 kvm->events_ops = &exit_events;
 295         else if (!strcmp(kvm->report_event, "mmio"))
 296                 kvm->events_ops = &mmio_events;
 297         else if (!strcmp(kvm->report_event, "ioport"))
 298                 kvm->events_ops = &ioport_events;
 299         else {
 300                 pr_err("Unknown report event:%s\n", kvm->report_event);
 301                 ret = false;
 302         }
 303
 304         return ret;
 305 }
 306
 307 struct vcpu_event_record {
 308         int vcpu_id;
 309         u64 start_time;
 310         struct kvm_event *last_event;
 311 };
 312
 313
 314 static void init_kvm_event_record(struct perf_kvm *kvm)
 315 {
 316         unsigned int i;
 317
 318         for (i = 0; i < EVENTS_CACHE_SIZE; i++)
 319                 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
 320 }
 321
 322 static int kvm_events_hash_fn(u64 key)
 323 {
 324         return key & (EVENTS_CACHE_SIZE - 1);
 325 }
 326
 327 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 328 {
 329         int old_max_vcpu = event->max_vcpu;
 330
 331         if (vcpu_id < event->max_vcpu)
 332                 return true;
 333
 334         while (event->max_vcpu <= vcpu_id)
 335                 event->max_vcpu += DEFAULT_VCPU_NUM;
 336
 337         event->vcpu = realloc(event->vcpu,
 338                               event->max_vcpu * sizeof(*event->vcpu));
 339         if (!event->vcpu) {
 340                 pr_err("Not enough memory\n");
 341                 return false;
 342         }
 343
 344         memset(event->vcpu + old_max_vcpu, 0,
 345                (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 346         return true;
 347 }
 348
 349 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 350 {
 351         struct kvm_event *event;
 352
 353         event = zalloc(sizeof(*event));
 354         if (!event) {
 355                 pr_err("Not enough memory\n");
 356                 return NULL;
 357         }
 358
 359         event->key = *key;
 360         return event;
 361 }
 362
 363 static struct kvm_event *find_create_kvm_event(struct perf_kvm *kvm,
 364                                                struct event_key *key)
 365 {
 366         struct kvm_event *event;
 367         struct list_head *head;
 368
 369         BUG_ON(key->key == INVALID_KEY);
 370
 371         head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 372         list_for_each_entry(event, head, hash_entry) {
 373                 if (event->key.key == key->key && event->key.info == key->info)
 374                         return event;
 375         }
 376
 377         event = kvm_alloc_init_event(key);
 378         if (!event)
 379                 return NULL;
 380
 381         list_add(&event->hash_entry, head);
 382         return event;
 383 }
 384
 385 static bool handle_begin_event(struct perf_kvm *kvm,
 386                                struct vcpu_event_record *vcpu_record,
 387                                struct event_key *key, u64 timestamp)
 388 {
 389         struct kvm_event *event = NULL;
 390
 391         if (key->key != INVALID_KEY)
 392                 event = find_create_kvm_event(kvm, key);
 393
 394         vcpu_record->last_event = event;
 395         vcpu_record->start_time = timestamp;
 396         return true;
 397 }
 398
 399 static void
 400 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 401 {
 402         kvm_stats->time += time_diff;
 403         update_stats(&kvm_stats->stats, time_diff);
 404 }
 405
 406 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 407 {
 408         struct kvm_event_stats *kvm_stats = &event->total;
 409
 410         if (vcpu_id != -1)
 411                 kvm_stats = &event->vcpu[vcpu_id];
 412
 413         return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 414                                 avg_stats(&kvm_stats->stats));
 415 }
 416
 417 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 418                              u64 time_diff)
 419 {
 420         if (vcpu_id == -1) {
 421                 kvm_update_event_stats(&event->total, time_diff);
 422                 return true;
 423         }
 424
 425         if (!kvm_event_expand(event, vcpu_id))
 426                 return false;
 427
 428         kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 429         return true;
 430 }
 431
 432 static bool handle_end_event(struct perf_kvm *kvm,
 433                              struct vcpu_event_record *vcpu_record,
 434                              struct event_key *key,
 435                              u64 timestamp)
 436 {
 437         struct kvm_event *event;
 438         u64 time_begin, time_diff;
 439         int vcpu;
 440
 441         if (kvm->trace_vcpu == -1)
 442                 vcpu = -1;
 443         else
 444                 vcpu = vcpu_record->vcpu_id;
 445
 446         event = vcpu_record->last_event;
 447         time_begin = vcpu_record->start_time;
 448
 449         /* The begin event is not caught. */
 450         if (!time_begin)
 451                 return true;
 452
 453         /*
 454          * In some case, the 'begin event' only records the start timestamp,
 455          * the actual event is recognized in the 'end event' (e.g. mmio-event).
 456          */
 457
 458         /* Both begin and end events did not get the key. */
 459         if (!event && key->key == INVALID_KEY)
 460                 return true;
 461
 462         if (!event)
 463                 event = find_create_kvm_event(kvm, key);
 464
 465         if (!event)
 466                 return false;
 467
 468         vcpu_record->last_event = NULL;
 469         vcpu_record->start_time = 0;
 470
 471         BUG_ON(timestamp < time_begin);
 472
 473         time_diff = timestamp - time_begin;
 474         return update_kvm_event(event, vcpu, time_diff);
 475 }
 476
 477 static
 478 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 479                                           struct perf_evsel *evsel,
 480                                           struct perf_sample *sample)
 481 {
 482         /* Only kvm_entry records vcpu id. */
 483         if (!thread->priv && kvm_entry_event(evsel)) {
 484                 struct vcpu_event_record *vcpu_record;
 485
 486                 vcpu_record = zalloc(sizeof(*vcpu_record));
 487                 if (!vcpu_record) {
 488                         pr_err("%s: Not enough memory\n", __func__);
 489                         return NULL;
 490                 }
 491
 492                 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
 493                 thread->priv = vcpu_record;
 494         }
 495
 496         return thread->priv;
 497 }
 498
 499 static bool handle_kvm_event(struct perf_kvm *kvm,
 500                              struct thread *thread,
 501                              struct perf_evsel *evsel,
 502                              struct perf_sample *sample)
 503 {
 504         struct vcpu_event_record *vcpu_record;
 505         struct event_key key = {.key = INVALID_KEY};
 506
 507         vcpu_record = per_vcpu_record(thread, evsel, sample);
 508         if (!vcpu_record)
 509                 return true;
 510
 511         /* only process events for vcpus user cares about */
 512         if ((kvm->trace_vcpu != -1) &&
 513             (kvm->trace_vcpu != vcpu_record->vcpu_id))
 514                 return true;
 515
 516         if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 517                 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
 518
 519         if (kvm->events_ops->is_end_event(evsel, sample, &key))
 520                 return handle_end_event(kvm, vcpu_record, &key, sample->time);
 521
 522         return true;
 523 }
 524
 525 #define GET_EVENT_KEY(func, field)                                      \
 526 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)         \
 527 {                                                                       \
 528         if (vcpu == -1)                                                 \
 529                 return event->total.field;                              \
 530                                                                         \
 531         if (vcpu >= event->max_vcpu)                                    \
 532                 return 0;                                               \
 533                                                                         \
 534         return event->vcpu[vcpu].field;                                 \
 535 }
 536
 537 #define COMPARE_EVENT_KEY(func, field)                                  \
 538 GET_EVENT_KEY(func, field)                                              \
 539 static int compare_kvm_event_ ## func(struct kvm_event *one,            \
 540                                         struct kvm_event *two, int vcpu)\
 541 {                                                                       \
 542         return get_event_ ##func(one, vcpu) >                           \
 543                                 get_event_ ##func(two, vcpu);           \
 544 }
 545
 546 GET_EVENT_KEY(time, time);
 547 COMPARE_EVENT_KEY(count, stats.n);
 548 COMPARE_EVENT_KEY(mean, stats.mean);
 549
 550 #define DEF_SORT_NAME_KEY(name, compare_key)                            \
 551         { #name, compare_kvm_event_ ## compare_key }
 552
 553 static struct kvm_event_key keys[] = {
 554         DEF_SORT_NAME_KEY(sample, count),
 555         DEF_SORT_NAME_KEY(time, mean),
 556         { NULL, NULL }
 557 };
 558
 559 static bool select_key(struct perf_kvm *kvm)
 560 {
 561         int i;
 562
 563         for (i = 0; keys[i].name; i++) {
 564                 if (!strcmp(keys[i].name, kvm->sort_key)) {
 565                         kvm->compare = keys[i].key;
 566                         return true;
 567                 }
 568         }
 569
 570         pr_err("Unknown compare key:%s\n", kvm->sort_key);
 571         return false;
 572 }
 573
 574 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
 575                              key_cmp_fun bigger, int vcpu)
 576 {
 577         struct rb_node **rb = &result->rb_node;
 578         struct rb_node *parent = NULL;
 579         struct kvm_event *p;
 580
 581         while (*rb) {
 582                 p = container_of(*rb, struct kvm_event, rb);
 583                 parent = *rb;
 584
 585                 if (bigger(event, p, vcpu))
 586                         rb = &(*rb)->rb_left;
 587                 else
 588                         rb = &(*rb)->rb_right;
 589         }
 590
 591         rb_link_node(&event->rb, parent, rb);
 592         rb_insert_color(&event->rb, result);
 593 }
 594
 595 static void update_total_count(struct perf_kvm *kvm, struct kvm_event *event)
 596 {
 597         int vcpu = kvm->trace_vcpu;
 598
 599         kvm->total_count += get_event_count(event, vcpu);
 600         kvm->total_time += get_event_time(event, vcpu);
 601 }
 602
 603 static bool event_is_valid(struct kvm_event *event, int vcpu)
 604 {
 605         return !!get_event_count(event, vcpu);
 606 }
 607
 608 static void sort_result(struct perf_kvm *kvm)
 609 {
 610         unsigned int i;
 611         int vcpu = kvm->trace_vcpu;
 612         struct kvm_event *event;
 613
 614         for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 615                 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
 616                         if (event_is_valid(event, vcpu)) {
 617                                 update_total_count(kvm, event);
 618                                 insert_to_result(&kvm->result, event,
 619                                                  kvm->compare, vcpu);
 620                         }
 621                 }
 622         }
 623 }
 624
 625 /* returns left most element of result, and erase it */
 626 static struct kvm_event *pop_from_result(struct rb_root *result)
 627 {
 628         struct rb_node *node = rb_first(result);
 629
 630         if (!node)
 631                 return NULL;
 632
 633         rb_erase(node, result);
 634         return container_of(node, struct kvm_event, rb);
 635 }
 636
 637 static void print_vcpu_info(int vcpu)
 638 {
 639         pr_info("Analyze events for ");
 640
 641         if (vcpu == -1)
 642                 pr_info("all VCPUs:\n\n");
 643         else
 644                 pr_info("VCPU %d:\n\n", vcpu);
 645 }
 646
 647 static void print_result(struct perf_kvm *kvm)
 648 {
 649         char decode[20];
 650         struct kvm_event *event;
 651         int vcpu = kvm->trace_vcpu;
 652
 653         pr_info("\n\n");
 654         print_vcpu_info(vcpu);
 655         pr_info("%20s ", kvm->events_ops->name);
 656         pr_info("%10s ", "Samples");
 657         pr_info("%9s ", "Samples%");
 658
 659         pr_info("%9s ", "Time%");
 660         pr_info("%16s ", "Avg time");
 661         pr_info("\n\n");
 662
 663         while ((event = pop_from_result(&kvm->result))) {
 664                 u64 ecount, etime;
 665
 666                 ecount = get_event_count(event, vcpu);
 667                 etime = get_event_time(event, vcpu);
 668
 669                 kvm->events_ops->decode_key(kvm, &event->key, decode);
 670                 pr_info("%20s ", decode);
 671                 pr_info("%10llu ", (unsigned long long)ecount);
 672                 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
 673                 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
 674                 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
 675                         kvm_event_rel_stddev(vcpu, event));
 676                 pr_info("\n");
 677         }
 678
 679         pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
 680                 kvm->total_count, kvm->total_time / 1e3);
 681 }
 682
 683 static int process_sample_event(struct perf_tool *tool,
 684                                 union perf_event *event,
 685                                 struct perf_sample *sample,
 686                                 struct perf_evsel *evsel,
 687                                 struct machine *machine)
 688 {
 689         struct thread *thread = machine__findnew_thread(machine, sample->tid);
 690         struct perf_kvm *kvm = container_of(tool, struct perf_kvm, tool);
 691
 692         if (thread == NULL) {
 693                 pr_debug("problem processing %d event, skipping it.\n",
 694                         event->header.type);
 695                 return -1;
 696         }
 697
 698         if (!handle_kvm_event(kvm, thread, evsel, sample))
 699                 return -1;
 700
 701         return 0;
 702 }
 703
 704 static int get_cpu_isa(struct perf_session *session)
 705 {
 706         char *cpuid = session->header.env.cpuid;
 707         int isa;
 708
 709         if (strstr(cpuid, "Intel"))
 710                 isa = 1;
 711         else if (strstr(cpuid, "AMD"))
 712                 isa = 0;
 713         else {
 714                 pr_err("CPU %s is not supported.\n", cpuid);
 715                 isa = -ENOTSUP;
 716         }
 717
 718         return isa;
 719 }
 720
 721 static int read_events(struct perf_kvm *kvm)
 722 {
 723         int ret;
 724
 725         struct perf_tool eops = {
 726                 .sample                 = process_sample_event,
 727                 .comm                   = perf_event__process_comm,
 728                 .ordered_samples        = true,
 729         };
 730
 731         kvm->tool = eops;
 732         kvm->session = perf_session__new(kvm->file_name, O_RDONLY, 0, false,
 733                                          &kvm->tool);
 734         if (!kvm->session) {
 735                 pr_err("Initializing perf session failed\n");
 736                 return -EINVAL;
 737         }
 738
 739         if (!perf_session__has_traces(kvm->session, "kvm record"))
 740                 return -EINVAL;
 741
 742         /*
 743          * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
 744          * traced in the old kernel.
 745          */
 746         ret = get_cpu_isa(kvm->session);
 747
 748         if (ret < 0)
 749                 return ret;
 750
 751         if (ret == 1) {
 752                 kvm->exit_reasons = vmx_exit_reasons;
 753                 kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
 754                 kvm->exit_reasons_isa = "VMX";
 755         }
 756
 757         return perf_session__process_events(kvm->session, &kvm->tool);
 758 }
 759
 760 static bool verify_vcpu(int vcpu)
 761 {
 762         if (vcpu != -1 && vcpu < 0) {
 763                 pr_err("Invalid vcpu:%d.\n", vcpu);
 764                 return false;
 765         }
 766
 767         return true;
 768 }
 769
 770 static int kvm_events_report_vcpu(struct perf_kvm *kvm)
 771 {
 772         int ret = -EINVAL;
 773         int vcpu = kvm->trace_vcpu;
 774
 775         if (!verify_vcpu(vcpu))
 776                 goto exit;
 777
 778         if (!select_key(kvm))
 779                 goto exit;
 780
 781         if (!register_kvm_events_ops(kvm))
 782                 goto exit;
 783
 784         init_kvm_event_record(kvm);
 785         setup_pager();
 786
 787         ret = read_events(kvm);
 788         if (ret)
 789                 goto exit;
 790
 791         sort_result(kvm);
 792         print_result(kvm);
 793
 794 exit:
 795         return ret;
 796 }
 797
 798 static const char * const record_args[] = {
 799         "record",
 800         "-R",
 801         "-f",
 802         "-m", "1024",
 803         "-c", "1",
 804         "-e", "kvm:kvm_entry",
 805         "-e", "kvm:kvm_exit",
 806         "-e", "kvm:kvm_mmio",
 807         "-e", "kvm:kvm_pio",
 808 };
 809
 810 #define STRDUP_FAIL_EXIT(s)             \
 811         ({      char *_p;               \
 812         _p = strdup(s);         \
 813                 if (!_p)                \
 814                         return -ENOMEM; \
 815                 _p;                     \
 816         })
 817
 818 static int kvm_events_record(struct perf_kvm *kvm, int argc, const char **argv)
 819 {
 820         unsigned int rec_argc, i, j;
 821         const char **rec_argv;
 822
 823         rec_argc = ARRAY_SIZE(record_args) + argc + 2;
 824         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 825
 826         if (rec_argv == NULL)
 827                 return -ENOMEM;
 828
 829         for (i = 0; i < ARRAY_SIZE(record_args); i++)
 830                 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
 831
 832         rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
 833         rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
 834
 835         for (j = 1; j < (unsigned int)argc; j++, i++)
 836                 rec_argv[i] = argv[j];
 837
 838         return cmd_record(i, rec_argv, NULL);
 839 }
 840
 841 static int kvm_events_report(struct perf_kvm *kvm, int argc, const char **argv)
 842 {
 843         const struct option kvm_events_report_options[] = {
 844                 OPT_STRING(0, "event", &kvm->report_event, "report event",
 845                             "event for reporting: vmexit, mmio, ioport"),
 846                 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
 847                             "vcpu id to report"),
 848                 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
 849                             "key for sorting: sample(sort by samples number)"
 850                             " time (sort by avg time)"),
 851                 OPT_END()
 852         };
 853
 854         const char * const kvm_events_report_usage[] = {
 855                 "perf kvm stat report [<options>]",
 856                 NULL
 857         };
 858
 859         symbol__init();
 860
 861         if (argc) {
 862                 argc = parse_options(argc, argv,
 863                                      kvm_events_report_options,
 864                                      kvm_events_report_usage, 0);
 865                 if (argc)
 866                         usage_with_options(kvm_events_report_usage,
 867                                            kvm_events_report_options);
 868         }
 869
 870         return kvm_events_report_vcpu(kvm);
 871 }
 872
 873 static void print_kvm_stat_usage(void)
 874 {
 875         printf("Usage: perf kvm stat <command>\n\n");
 876
 877         printf("# Available commands:\n");
 878         printf("\trecord: record kvm events\n");
 879         printf("\treport: report statistical data of kvm events\n");
 880
 881         printf("\nOtherwise, it is the alias of 'perf stat':\n");
 882 }
 883
 884 static int kvm_cmd_stat(struct perf_kvm *kvm, int argc, const char **argv)
 885 {
 886         if (argc == 1) {
 887                 print_kvm_stat_usage();
 888                 goto perf_stat;
 889         }
 890
 891         if (!strncmp(argv[1], "rec", 3))
 892                 return kvm_events_record(kvm, argc - 1, argv + 1);
 893
 894         if (!strncmp(argv[1], "rep", 3))
 895                 return kvm_events_report(kvm, argc - 1 , argv + 1);
 896
 897 perf_stat:
 898         return cmd_stat(argc, argv, NULL);
 899 }
 900
 901 static int __cmd_record(struct perf_kvm *kvm, int argc, const char **argv)
 902 {
 903         int rec_argc, i = 0, j;
 904         const char **rec_argv;
 905
 906         rec_argc = argc + 2;
 907         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 908         rec_argv[i++] = strdup("record");
 909         rec_argv[i++] = strdup("-o");
 910         rec_argv[i++] = strdup(kvm->file_name);
 911         for (j = 1; j < argc; j++, i++)
 912                 rec_argv[i] = argv[j];
 913
 914         BUG_ON(i != rec_argc);
 915
 916         return cmd_record(i, rec_argv, NULL);
 917 }
 918
 919 static int __cmd_report(struct perf_kvm *kvm, int argc, const char **argv)
 920 {
 921         int rec_argc, i = 0, j;
 922         const char **rec_argv;
 923
 924         rec_argc = argc + 2;
 925         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 926         rec_argv[i++] = strdup("report");
 927         rec_argv[i++] = strdup("-i");
 928         rec_argv[i++] = strdup(kvm->file_name);
 929         for (j = 1; j < argc; j++, i++)
 930                 rec_argv[i] = argv[j];
 931
 932         BUG_ON(i != rec_argc);
 933
 934         return cmd_report(i, rec_argv, NULL);
 935 }
 936
 937 static int __cmd_buildid_list(struct perf_kvm *kvm, int argc, const char **argv)
 938 {
 939         int rec_argc, i = 0, j;
 940         const char **rec_argv;
 941
 942         rec_argc = argc + 2;
 943         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 944         rec_argv[i++] = strdup("buildid-list");
 945         rec_argv[i++] = strdup("-i");
 946         rec_argv[i++] = strdup(kvm->file_name);
 947         for (j = 1; j < argc; j++, i++)
 948                 rec_argv[i] = argv[j];
 949
 950         BUG_ON(i != rec_argc);
 951
 952         return cmd_buildid_list(i, rec_argv, NULL);
 953 }
 954
 955 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
 956 {
 957         struct perf_kvm kvm = {
 958                 .trace_vcpu     = -1,
 959                 .report_event   = "vmexit",
 960                 .sort_key       = "sample",
 961
 962                 .exit_reasons = svm_exit_reasons,
 963                 .exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
 964                 .exit_reasons_isa = "SVM",
 965         };
 966
 967         const struct option kvm_options[] = {
 968                 OPT_STRING('i', "input", &kvm.file_name, "file",
 969                            "Input file name"),
 970                 OPT_STRING('o', "output", &kvm.file_name, "file",
 971                            "Output file name"),
 972                 OPT_BOOLEAN(0, "guest", &perf_guest,
 973                             "Collect guest os data"),
 974                 OPT_BOOLEAN(0, "host", &perf_host,
 975                             "Collect host os data"),
 976                 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
 977                            "guest mount directory under which every guest os"
 978                            " instance has a subdir"),
 979                 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
 980                            "file", "file saving guest os vmlinux"),
 981                 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
 982                            "file", "file saving guest os /proc/kallsyms"),
 983                 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
 984                            "file", "file saving guest os /proc/modules"),
 985                 OPT_END()
 986         };
 987
 988
 989         const char * const kvm_usage[] = {
 990                 "perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
 991                 NULL
 992         };
 993
 994         perf_host  = 0;
 995         perf_guest = 1;
 996
 997         argc = parse_options(argc, argv, kvm_options, kvm_usage,
 998                         PARSE_OPT_STOP_AT_NON_OPTION);
 999         if (!argc)
1000                 usage_with_options(kvm_usage, kvm_options);
1001
1002         if (!perf_host)
1003                 perf_guest = 1;
1004
1005         if (!kvm.file_name) {
1006                 if (perf_host && !perf_guest)
1007                         kvm.file_name = strdup("perf.data.host");
1008                 else if (!perf_host && perf_guest)
1009                         kvm.file_name = strdup("perf.data.guest");
1010                 else
1011                         kvm.file_name = strdup("perf.data.kvm");
1012
1013                 if (!kvm.file_name) {
1014                         pr_err("Failed to allocate memory for filename\n");
1015                         return -ENOMEM;
1016                 }
1017         }
1018
1019         if (!strncmp(argv[0], "rec", 3))
1020                 return __cmd_record(&kvm, argc, argv);
1021         else if (!strncmp(argv[0], "rep", 3))
1022                 return __cmd_report(&kvm, argc, argv);
1023         else if (!strncmp(argv[0], "diff", 4))
1024                 return cmd_diff(argc, argv, NULL);
1025         else if (!strncmp(argv[0], "top", 3))
1026                 return cmd_top(argc, argv, NULL);
1027         else if (!strncmp(argv[0], "buildid-list", 12))
1028                 return __cmd_buildid_list(&kvm, argc, argv);
1029         else if (!strncmp(argv[0], "stat", 4))
1030                 return kvm_cmd_stat(&kvm, argc, argv);
1031         else
1032                 usage_with_options(kvm_usage, kvm_options);
1033
1034         return 0;
1035 }