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 #if defined(__i386__) || defined(__x86_64__)
  26 #include <asm/svm.h>
  27 #include <asm/vmx.h>
  28 #include <asm/kvm.h>
  29
  30 struct event_key {
  31         #define INVALID_KEY     (~0ULL)
  32         u64 key;
  33         int info;
  34 };
  35
  36 struct kvm_event_stats {
  37         u64 time;
  38         struct stats stats;
  39 };
  40
  41 struct kvm_event {
  42         struct list_head hash_entry;
  43         struct rb_node rb;
  44
  45         struct event_key key;
  46
  47         struct kvm_event_stats total;
  48
  49         #define DEFAULT_VCPU_NUM 8
  50         int max_vcpu;
  51         struct kvm_event_stats *vcpu;
  52 };
  53
  54 typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
  55
  56 struct kvm_event_key {
  57         const char *name;
  58         key_cmp_fun key;
  59 };
  60
  61
  62 struct perf_kvm_stat;
  63
  64 struct kvm_events_ops {
  65         bool (*is_begin_event)(struct perf_evsel *evsel,
  66                                struct perf_sample *sample,
  67                                struct event_key *key);
  68         bool (*is_end_event)(struct perf_evsel *evsel,
  69                              struct perf_sample *sample, struct event_key *key);
  70         void (*decode_key)(struct perf_kvm_stat *kvm, struct event_key *key,
  71                            char decode[20]);
  72         const char *name;
  73 };
  74
  75 struct exit_reasons_table {
  76         unsigned long exit_code;
  77         const char *reason;
  78 };
  79
  80 #define EVENTS_BITS             12
  81 #define EVENTS_CACHE_SIZE       (1UL << EVENTS_BITS)
  82
  83 struct perf_kvm_stat {
  84         struct perf_tool    tool;
  85         struct perf_session *session;
  86
  87         const char *file_name;
  88         const char *report_event;
  89         const char *sort_key;
  90         int trace_vcpu;
  91
  92         struct exit_reasons_table *exit_reasons;
  93         int exit_reasons_size;
  94         const char *exit_reasons_isa;
  95
  96         struct kvm_events_ops *events_ops;
  97         key_cmp_fun compare;
  98         struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
  99         u64 total_time;
 100         u64 total_count;
 101
 102         struct rb_root result;
 103 };
 104
 105
 106 static void exit_event_get_key(struct perf_evsel *evsel,
 107                                struct perf_sample *sample,
 108                                struct event_key *key)
 109 {
 110         key->info = 0;
 111         key->key = perf_evsel__intval(evsel, sample, "exit_reason");
 112 }
 113
 114 static bool kvm_exit_event(struct perf_evsel *evsel)
 115 {
 116         return !strcmp(evsel->name, "kvm:kvm_exit");
 117 }
 118
 119 static bool exit_event_begin(struct perf_evsel *evsel,
 120                              struct perf_sample *sample, struct event_key *key)
 121 {
 122         if (kvm_exit_event(evsel)) {
 123                 exit_event_get_key(evsel, sample, key);
 124                 return true;
 125         }
 126
 127         return false;
 128 }
 129
 130 static bool kvm_entry_event(struct perf_evsel *evsel)
 131 {
 132         return !strcmp(evsel->name, "kvm:kvm_entry");
 133 }
 134
 135 static bool exit_event_end(struct perf_evsel *evsel,
 136                            struct perf_sample *sample __maybe_unused,
 137                            struct event_key *key __maybe_unused)
 138 {
 139         return kvm_entry_event(evsel);
 140 }
 141
 142 static struct exit_reasons_table vmx_exit_reasons[] = {
 143         VMX_EXIT_REASONS
 144 };
 145
 146 static struct exit_reasons_table svm_exit_reasons[] = {
 147         SVM_EXIT_REASONS
 148 };
 149
 150 static const char *get_exit_reason(struct perf_kvm_stat *kvm, u64 exit_code)
 151 {
 152         int i = kvm->exit_reasons_size;
 153         struct exit_reasons_table *tbl = kvm->exit_reasons;
 154
 155         while (i--) {
 156                 if (tbl->exit_code == exit_code)
 157                         return tbl->reason;
 158                 tbl++;
 159         }
 160
 161         pr_err("unknown kvm exit code:%lld on %s\n",
 162                 (unsigned long long)exit_code, kvm->exit_reasons_isa);
 163         return "UNKNOWN";
 164 }
 165
 166 static void exit_event_decode_key(struct perf_kvm_stat *kvm,
 167                                   struct event_key *key,
 168                                   char decode[20])
 169 {
 170         const char *exit_reason = get_exit_reason(kvm, key->key);
 171
 172         scnprintf(decode, 20, "%s", exit_reason);
 173 }
 174
 175 static struct kvm_events_ops exit_events = {
 176         .is_begin_event = exit_event_begin,
 177         .is_end_event = exit_event_end,
 178         .decode_key = exit_event_decode_key,
 179         .name = "VM-EXIT"
 180 };
 181
 182 /*
 183  * For the mmio events, we treat:
 184  * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
 185  * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
 186  */
 187 static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
 188                                struct event_key *key)
 189 {
 190         key->key  = perf_evsel__intval(evsel, sample, "gpa");
 191         key->info = perf_evsel__intval(evsel, sample, "type");
 192 }
 193
 194 #define KVM_TRACE_MMIO_READ_UNSATISFIED 0
 195 #define KVM_TRACE_MMIO_READ 1
 196 #define KVM_TRACE_MMIO_WRITE 2
 197
 198 static bool mmio_event_begin(struct perf_evsel *evsel,
 199                              struct perf_sample *sample, struct event_key *key)
 200 {
 201         /* MMIO read begin event in kernel. */
 202         if (kvm_exit_event(evsel))
 203                 return true;
 204
 205         /* MMIO write begin event in kernel. */
 206         if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
 207             perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
 208                 mmio_event_get_key(evsel, sample, key);
 209                 return true;
 210         }
 211
 212         return false;
 213 }
 214
 215 static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
 216                            struct event_key *key)
 217 {
 218         /* MMIO write end event in kernel. */
 219         if (kvm_entry_event(evsel))
 220                 return true;
 221
 222         /* MMIO read end event in kernel.*/
 223         if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
 224             perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
 225                 mmio_event_get_key(evsel, sample, key);
 226                 return true;
 227         }
 228
 229         return false;
 230 }
 231
 232 static void mmio_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
 233                                   struct event_key *key,
 234                                   char decode[20])
 235 {
 236         scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
 237                                 key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
 238 }
 239
 240 static struct kvm_events_ops mmio_events = {
 241         .is_begin_event = mmio_event_begin,
 242         .is_end_event = mmio_event_end,
 243         .decode_key = mmio_event_decode_key,
 244         .name = "MMIO Access"
 245 };
 246
 247  /* The time of emulation pio access is from kvm_pio to kvm_entry. */
 248 static void ioport_event_get_key(struct perf_evsel *evsel,
 249                                  struct perf_sample *sample,
 250                                  struct event_key *key)
 251 {
 252         key->key  = perf_evsel__intval(evsel, sample, "port");
 253         key->info = perf_evsel__intval(evsel, sample, "rw");
 254 }
 255
 256 static bool ioport_event_begin(struct perf_evsel *evsel,
 257                                struct perf_sample *sample,
 258                                struct event_key *key)
 259 {
 260         if (!strcmp(evsel->name, "kvm:kvm_pio")) {
 261                 ioport_event_get_key(evsel, sample, key);
 262                 return true;
 263         }
 264
 265         return false;
 266 }
 267
 268 static bool ioport_event_end(struct perf_evsel *evsel,
 269                              struct perf_sample *sample __maybe_unused,
 270                              struct event_key *key __maybe_unused)
 271 {
 272         return kvm_entry_event(evsel);
 273 }
 274
 275 static void ioport_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
 276                                     struct event_key *key,
 277                                     char decode[20])
 278 {
 279         scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
 280                                 key->info ? "POUT" : "PIN");
 281 }
 282
 283 static struct kvm_events_ops ioport_events = {
 284         .is_begin_event = ioport_event_begin,
 285         .is_end_event = ioport_event_end,
 286         .decode_key = ioport_event_decode_key,
 287         .name = "IO Port Access"
 288 };
 289
 290 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
 291 {
 292         bool ret = true;
 293
 294         if (!strcmp(kvm->report_event, "vmexit"))
 295                 kvm->events_ops = &exit_events;
 296         else if (!strcmp(kvm->report_event, "mmio"))
 297                 kvm->events_ops = &mmio_events;
 298         else if (!strcmp(kvm->report_event, "ioport"))
 299                 kvm->events_ops = &ioport_events;
 300         else {
 301                 pr_err("Unknown report event:%s\n", kvm->report_event);
 302                 ret = false;
 303         }
 304
 305         return ret;
 306 }
 307
 308 struct vcpu_event_record {
 309         int vcpu_id;
 310         u64 start_time;
 311         struct kvm_event *last_event;
 312 };
 313
 314
 315 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
 316 {
 317         unsigned int i;
 318
 319         for (i = 0; i < EVENTS_CACHE_SIZE; i++)
 320                 INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
 321 }
 322
 323 static int kvm_events_hash_fn(u64 key)
 324 {
 325         return key & (EVENTS_CACHE_SIZE - 1);
 326 }
 327
 328 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 329 {
 330         int old_max_vcpu = event->max_vcpu;
 331
 332         if (vcpu_id < event->max_vcpu)
 333                 return true;
 334
 335         while (event->max_vcpu <= vcpu_id)
 336                 event->max_vcpu += DEFAULT_VCPU_NUM;
 337
 338         event->vcpu = realloc(event->vcpu,
 339                               event->max_vcpu * sizeof(*event->vcpu));
 340         if (!event->vcpu) {
 341                 pr_err("Not enough memory\n");
 342                 return false;
 343         }
 344
 345         memset(event->vcpu + old_max_vcpu, 0,
 346                (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 347         return true;
 348 }
 349
 350 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 351 {
 352         struct kvm_event *event;
 353
 354         event = zalloc(sizeof(*event));
 355         if (!event) {
 356                 pr_err("Not enough memory\n");
 357                 return NULL;
 358         }
 359
 360         event->key = *key;
 361         return event;
 362 }
 363
 364 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 365                                                struct event_key *key)
 366 {
 367         struct kvm_event *event;
 368         struct list_head *head;
 369
 370         BUG_ON(key->key == INVALID_KEY);
 371
 372         head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 373         list_for_each_entry(event, head, hash_entry) {
 374                 if (event->key.key == key->key && event->key.info == key->info)
 375                         return event;
 376         }
 377
 378         event = kvm_alloc_init_event(key);
 379         if (!event)
 380                 return NULL;
 381
 382         list_add(&event->hash_entry, head);
 383         return event;
 384 }
 385
 386 static bool handle_begin_event(struct perf_kvm_stat *kvm,
 387                                struct vcpu_event_record *vcpu_record,
 388                                struct event_key *key, u64 timestamp)
 389 {
 390         struct kvm_event *event = NULL;
 391
 392         if (key->key != INVALID_KEY)
 393                 event = find_create_kvm_event(kvm, key);
 394
 395         vcpu_record->last_event = event;
 396         vcpu_record->start_time = timestamp;
 397         return true;
 398 }
 399
 400 static void
 401 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 402 {
 403         kvm_stats->time += time_diff;
 404         update_stats(&kvm_stats->stats, time_diff);
 405 }
 406
 407 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 408 {
 409         struct kvm_event_stats *kvm_stats = &event->total;
 410
 411         if (vcpu_id != -1)
 412                 kvm_stats = &event->vcpu[vcpu_id];
 413
 414         return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 415                                 avg_stats(&kvm_stats->stats));
 416 }
 417
 418 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 419                              u64 time_diff)
 420 {
 421         if (vcpu_id == -1) {
 422                 kvm_update_event_stats(&event->total, time_diff);
 423                 return true;
 424         }
 425
 426         if (!kvm_event_expand(event, vcpu_id))
 427                 return false;
 428
 429         kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 430         return true;
 431 }
 432
 433 static bool handle_end_event(struct perf_kvm_stat *kvm,
 434                              struct vcpu_event_record *vcpu_record,
 435                              struct event_key *key,
 436                              u64 timestamp)
 437 {
 438         struct kvm_event *event;
 439         u64 time_begin, time_diff;
 440         int vcpu;
 441
 442         if (kvm->trace_vcpu == -1)
 443                 vcpu = -1;
 444         else
 445                 vcpu = vcpu_record->vcpu_id;
 446
 447         event = vcpu_record->last_event;
 448         time_begin = vcpu_record->start_time;
 449
 450         /* The begin event is not caught. */
 451         if (!time_begin)
 452                 return true;
 453
 454         /*
 455          * In some case, the 'begin event' only records the start timestamp,
 456          * the actual event is recognized in the 'end event' (e.g. mmio-event).
 457          */
 458
 459         /* Both begin and end events did not get the key. */
 460         if (!event && key->key == INVALID_KEY)
 461                 return true;
 462
 463         if (!event)
 464                 event = find_create_kvm_event(kvm, key);
 465
 466         if (!event)
 467                 return false;
 468
 469         vcpu_record->last_event = NULL;
 470         vcpu_record->start_time = 0;
 471
 472         BUG_ON(timestamp < time_begin);
 473
 474         time_diff = timestamp - time_begin;
 475         return update_kvm_event(event, vcpu, time_diff);
 476 }
 477
 478 static
 479 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 480                                           struct perf_evsel *evsel,
 481                                           struct perf_sample *sample)
 482 {
 483         /* Only kvm_entry records vcpu id. */
 484         if (!thread->priv && kvm_entry_event(evsel)) {
 485                 struct vcpu_event_record *vcpu_record;
 486
 487                 vcpu_record = zalloc(sizeof(*vcpu_record));
 488                 if (!vcpu_record) {
 489                         pr_err("%s: Not enough memory\n", __func__);
 490                         return NULL;
 491                 }
 492
 493                 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
 494                 thread->priv = vcpu_record;
 495         }
 496
 497         return thread->priv;
 498 }
 499
 500 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
 501                              struct thread *thread,
 502                              struct perf_evsel *evsel,
 503                              struct perf_sample *sample)
 504 {
 505         struct vcpu_event_record *vcpu_record;
 506         struct event_key key = {.key = INVALID_KEY};
 507
 508         vcpu_record = per_vcpu_record(thread, evsel, sample);
 509         if (!vcpu_record)
 510                 return true;
 511
 512         /* only process events for vcpus user cares about */
 513         if ((kvm->trace_vcpu != -1) &&
 514             (kvm->trace_vcpu != vcpu_record->vcpu_id))
 515                 return true;
 516
 517         if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 518                 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
 519
 520         if (kvm->events_ops->is_end_event(evsel, sample, &key))
 521                 return handle_end_event(kvm, vcpu_record, &key, sample->time);
 522
 523         return true;
 524 }
 525
 526 #define GET_EVENT_KEY(func, field)                                      \
 527 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)         \
 528 {                                                                       \
 529         if (vcpu == -1)                                                 \
 530                 return event->total.field;                              \
 531                                                                         \
 532         if (vcpu >= event->max_vcpu)                                    \
 533                 return 0;                                               \
 534                                                                         \
 535         return event->vcpu[vcpu].field;                                 \
 536 }
 537
 538 #define COMPARE_EVENT_KEY(func, field)                                  \
 539 GET_EVENT_KEY(func, field)                                              \
 540 static int compare_kvm_event_ ## func(struct kvm_event *one,            \
 541                                         struct kvm_event *two, int vcpu)\
 542 {                                                                       \
 543         return get_event_ ##func(one, vcpu) >                           \
 544                                 get_event_ ##func(two, vcpu);           \
 545 }
 546
 547 GET_EVENT_KEY(time, time);
 548 COMPARE_EVENT_KEY(count, stats.n);
 549 COMPARE_EVENT_KEY(mean, stats.mean);
 550
 551 #define DEF_SORT_NAME_KEY(name, compare_key)                            \
 552         { #name, compare_kvm_event_ ## compare_key }
 553
 554 static struct kvm_event_key keys[] = {
 555         DEF_SORT_NAME_KEY(sample, count),
 556         DEF_SORT_NAME_KEY(time, mean),
 557         { NULL, NULL }
 558 };
 559
 560 static bool select_key(struct perf_kvm_stat *kvm)
 561 {
 562         int i;
 563
 564         for (i = 0; keys[i].name; i++) {
 565                 if (!strcmp(keys[i].name, kvm->sort_key)) {
 566                         kvm->compare = keys[i].key;
 567                         return true;
 568                 }
 569         }
 570
 571         pr_err("Unknown compare key:%s\n", kvm->sort_key);
 572         return false;
 573 }
 574
 575 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
 576                              key_cmp_fun bigger, int vcpu)
 577 {
 578         struct rb_node **rb = &result->rb_node;
 579         struct rb_node *parent = NULL;
 580         struct kvm_event *p;
 581
 582         while (*rb) {
 583                 p = container_of(*rb, struct kvm_event, rb);
 584                 parent = *rb;
 585
 586                 if (bigger(event, p, vcpu))
 587                         rb = &(*rb)->rb_left;
 588                 else
 589                         rb = &(*rb)->rb_right;
 590         }
 591
 592         rb_link_node(&event->rb, parent, rb);
 593         rb_insert_color(&event->rb, result);
 594 }
 595
 596 static void
 597 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
 598 {
 599         int vcpu = kvm->trace_vcpu;
 600
 601         kvm->total_count += get_event_count(event, vcpu);
 602         kvm->total_time += get_event_time(event, vcpu);
 603 }
 604
 605 static bool event_is_valid(struct kvm_event *event, int vcpu)
 606 {
 607         return !!get_event_count(event, vcpu);
 608 }
 609
 610 static void sort_result(struct perf_kvm_stat *kvm)
 611 {
 612         unsigned int i;
 613         int vcpu = kvm->trace_vcpu;
 614         struct kvm_event *event;
 615
 616         for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 617                 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
 618                         if (event_is_valid(event, vcpu)) {
 619                                 update_total_count(kvm, event);
 620                                 insert_to_result(&kvm->result, event,
 621                                                  kvm->compare, vcpu);
 622                         }
 623                 }
 624         }
 625 }
 626
 627 /* returns left most element of result, and erase it */
 628 static struct kvm_event *pop_from_result(struct rb_root *result)
 629 {
 630         struct rb_node *node = rb_first(result);
 631
 632         if (!node)
 633                 return NULL;
 634
 635         rb_erase(node, result);
 636         return container_of(node, struct kvm_event, rb);
 637 }
 638
 639 static void print_vcpu_info(int vcpu)
 640 {
 641         pr_info("Analyze events for ");
 642
 643         if (vcpu == -1)
 644                 pr_info("all VCPUs:\n\n");
 645         else
 646                 pr_info("VCPU %d:\n\n", vcpu);
 647 }
 648
 649 static void print_result(struct perf_kvm_stat *kvm)
 650 {
 651         char decode[20];
 652         struct kvm_event *event;
 653         int vcpu = kvm->trace_vcpu;
 654
 655         pr_info("\n\n");
 656         print_vcpu_info(vcpu);
 657         pr_info("%20s ", kvm->events_ops->name);
 658         pr_info("%10s ", "Samples");
 659         pr_info("%9s ", "Samples%");
 660
 661         pr_info("%9s ", "Time%");
 662         pr_info("%16s ", "Avg time");
 663         pr_info("\n\n");
 664
 665         while ((event = pop_from_result(&kvm->result))) {
 666                 u64 ecount, etime;
 667
 668                 ecount = get_event_count(event, vcpu);
 669                 etime = get_event_time(event, vcpu);
 670
 671                 kvm->events_ops->decode_key(kvm, &event->key, decode);
 672                 pr_info("%20s ", decode);
 673                 pr_info("%10llu ", (unsigned long long)ecount);
 674                 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
 675                 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
 676                 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
 677                         kvm_event_rel_stddev(vcpu, event));
 678                 pr_info("\n");
 679         }
 680
 681         pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
 682                 kvm->total_count, kvm->total_time / 1e3);
 683 }
 684
 685 static int process_sample_event(struct perf_tool *tool,
 686                                 union perf_event *event,
 687                                 struct perf_sample *sample,
 688                                 struct perf_evsel *evsel,
 689                                 struct machine *machine)
 690 {
 691         struct thread *thread = machine__findnew_thread(machine, sample->tid);
 692         struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
 693                                                  tool);
 694
 695         if (thread == NULL) {
 696                 pr_debug("problem processing %d event, skipping it.\n",
 697                         event->header.type);
 698                 return -1;
 699         }
 700
 701         if (!handle_kvm_event(kvm, thread, evsel, sample))
 702                 return -1;
 703
 704         return 0;
 705 }
 706
 707 static int get_cpu_isa(struct perf_session *session)
 708 {
 709         char *cpuid = session->header.env.cpuid;
 710         int isa;
 711
 712         if (strstr(cpuid, "Intel"))
 713                 isa = 1;
 714         else if (strstr(cpuid, "AMD"))
 715                 isa = 0;
 716         else {
 717                 pr_err("CPU %s is not supported.\n", cpuid);
 718                 isa = -ENOTSUP;
 719         }
 720
 721         return isa;
 722 }
 723
 724 static int read_events(struct perf_kvm_stat *kvm)
 725 {
 726         int ret;
 727
 728         struct perf_tool eops = {
 729                 .sample                 = process_sample_event,
 730                 .comm                   = perf_event__process_comm,
 731                 .ordered_samples        = true,
 732         };
 733
 734         kvm->tool = eops;
 735         kvm->session = perf_session__new(kvm->file_name, O_RDONLY, 0, false,
 736                                          &kvm->tool);
 737         if (!kvm->session) {
 738                 pr_err("Initializing perf session failed\n");
 739                 return -EINVAL;
 740         }
 741
 742         if (!perf_session__has_traces(kvm->session, "kvm record"))
 743                 return -EINVAL;
 744
 745         /*
 746          * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
 747          * traced in the old kernel.
 748          */
 749         ret = get_cpu_isa(kvm->session);
 750
 751         if (ret < 0)
 752                 return ret;
 753
 754         if (ret == 1) {
 755                 kvm->exit_reasons = vmx_exit_reasons;
 756                 kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
 757                 kvm->exit_reasons_isa = "VMX";
 758         }
 759
 760         return perf_session__process_events(kvm->session, &kvm->tool);
 761 }
 762
 763 static bool verify_vcpu(int vcpu)
 764 {
 765         if (vcpu != -1 && vcpu < 0) {
 766                 pr_err("Invalid vcpu:%d.\n", vcpu);
 767                 return false;
 768         }
 769
 770         return true;
 771 }
 772
 773 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
 774 {
 775         int ret = -EINVAL;
 776         int vcpu = kvm->trace_vcpu;
 777
 778         if (!verify_vcpu(vcpu))
 779                 goto exit;
 780
 781         if (!select_key(kvm))
 782                 goto exit;
 783
 784         if (!register_kvm_events_ops(kvm))
 785                 goto exit;
 786
 787         init_kvm_event_record(kvm);
 788         setup_pager();
 789
 790         ret = read_events(kvm);
 791         if (ret)
 792                 goto exit;
 793
 794         sort_result(kvm);
 795         print_result(kvm);
 796
 797 exit:
 798         return ret;
 799 }
 800
 801 static const char * const record_args[] = {
 802         "record",
 803         "-R",
 804         "-f",
 805         "-m", "1024",
 806         "-c", "1",
 807         "-e", "kvm:kvm_entry",
 808         "-e", "kvm:kvm_exit",
 809         "-e", "kvm:kvm_mmio",
 810         "-e", "kvm:kvm_pio",
 811 };
 812
 813 #define STRDUP_FAIL_EXIT(s)             \
 814         ({      char *_p;               \
 815         _p = strdup(s);         \
 816                 if (!_p)                \
 817                         return -ENOMEM; \
 818                 _p;                     \
 819         })
 820
 821 static int
 822 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
 823 {
 824         unsigned int rec_argc, i, j;
 825         const char **rec_argv;
 826
 827         rec_argc = ARRAY_SIZE(record_args) + argc + 2;
 828         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 829
 830         if (rec_argv == NULL)
 831                 return -ENOMEM;
 832
 833         for (i = 0; i < ARRAY_SIZE(record_args); i++)
 834                 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
 835
 836         rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
 837         rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
 838
 839         for (j = 1; j < (unsigned int)argc; j++, i++)
 840                 rec_argv[i] = argv[j];
 841
 842         return cmd_record(i, rec_argv, NULL);
 843 }
 844
 845 static int
 846 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
 847 {
 848         const struct option kvm_events_report_options[] = {
 849                 OPT_STRING(0, "event", &kvm->report_event, "report event",
 850                             "event for reporting: vmexit, mmio, ioport"),
 851                 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
 852                             "vcpu id to report"),
 853                 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
 854                             "key for sorting: sample(sort by samples number)"
 855                             " time (sort by avg time)"),
 856                 OPT_END()
 857         };
 858
 859         const char * const kvm_events_report_usage[] = {
 860                 "perf kvm stat report [<options>]",
 861                 NULL
 862         };
 863
 864         symbol__init();
 865
 866         if (argc) {
 867                 argc = parse_options(argc, argv,
 868                                      kvm_events_report_options,
 869                                      kvm_events_report_usage, 0);
 870                 if (argc)
 871                         usage_with_options(kvm_events_report_usage,
 872                                            kvm_events_report_options);
 873         }
 874
 875         return kvm_events_report_vcpu(kvm);
 876 }
 877
 878 static void print_kvm_stat_usage(void)
 879 {
 880         printf("Usage: perf kvm stat <command>\n\n");
 881
 882         printf("# Available commands:\n");
 883         printf("\trecord: record kvm events\n");
 884         printf("\treport: report statistical data of kvm events\n");
 885
 886         printf("\nOtherwise, it is the alias of 'perf stat':\n");
 887 }
 888
 889 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
 890 {
 891         struct perf_kvm_stat kvm = {
 892                 .file_name = file_name,
 893
 894                 .trace_vcpu     = -1,
 895                 .report_event   = "vmexit",
 896                 .sort_key       = "sample",
 897
 898                 .exit_reasons = svm_exit_reasons,
 899                 .exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
 900                 .exit_reasons_isa = "SVM",
 901         };
 902
 903         if (argc == 1) {
 904                 print_kvm_stat_usage();
 905                 goto perf_stat;
 906         }
 907
 908         if (!strncmp(argv[1], "rec", 3))
 909                 return kvm_events_record(&kvm, argc - 1, argv + 1);
 910
 911         if (!strncmp(argv[1], "rep", 3))
 912                 return kvm_events_report(&kvm, argc - 1 , argv + 1);
 913
 914 perf_stat:
 915         return cmd_stat(argc, argv, NULL);
 916 }
 917 #endif
 918
 919 static int __cmd_record(const char *file_name, 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("record");
 927         rec_argv[i++] = strdup("-o");
 928         rec_argv[i++] = strdup(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_record(i, rec_argv, NULL);
 935 }
 936
 937 static int __cmd_report(const char *file_name, 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("report");
 945         rec_argv[i++] = strdup("-i");
 946         rec_argv[i++] = strdup(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_report(i, rec_argv, NULL);
 953 }
 954
 955 static int
 956 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
 957 {
 958         int rec_argc, i = 0, j;
 959         const char **rec_argv;
 960
 961         rec_argc = argc + 2;
 962         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 963         rec_argv[i++] = strdup("buildid-list");
 964         rec_argv[i++] = strdup("-i");
 965         rec_argv[i++] = strdup(file_name);
 966         for (j = 1; j < argc; j++, i++)
 967                 rec_argv[i] = argv[j];
 968
 969         BUG_ON(i != rec_argc);
 970
 971         return cmd_buildid_list(i, rec_argv, NULL);
 972 }
 973
 974 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
 975 {
 976         const char *file_name;
 977
 978         const struct option kvm_options[] = {
 979                 OPT_STRING('i', "input", &file_name, "file",
 980                            "Input file name"),
 981                 OPT_STRING('o', "output", &file_name, "file",
 982                            "Output file name"),
 983                 OPT_BOOLEAN(0, "guest", &perf_guest,
 984                             "Collect guest os data"),
 985                 OPT_BOOLEAN(0, "host", &perf_host,
 986                             "Collect host os data"),
 987                 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
 988                            "guest mount directory under which every guest os"
 989                            " instance has a subdir"),
 990                 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
 991                            "file", "file saving guest os vmlinux"),
 992                 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
 993                            "file", "file saving guest os /proc/kallsyms"),
 994                 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
 995                            "file", "file saving guest os /proc/modules"),
 996                 OPT_END()
 997         };
 998
 999
1000         const char * const kvm_usage[] = {
1001                 "perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
1002                 NULL
1003         };
1004
1005         perf_host  = 0;
1006         perf_guest = 1;
1007
1008         argc = parse_options(argc, argv, kvm_options, kvm_usage,
1009                         PARSE_OPT_STOP_AT_NON_OPTION);
1010         if (!argc)
1011                 usage_with_options(kvm_usage, kvm_options);
1012
1013         if (!perf_host)
1014                 perf_guest = 1;
1015
1016         if (!file_name) {
1017                 if (perf_host && !perf_guest)
1018                         file_name = strdup("perf.data.host");
1019                 else if (!perf_host && perf_guest)
1020                         file_name = strdup("perf.data.guest");
1021                 else
1022                         file_name = strdup("perf.data.kvm");
1023
1024                 if (!file_name) {
1025                         pr_err("Failed to allocate memory for filename\n");
1026                         return -ENOMEM;
1027                 }
1028         }
1029
1030         if (!strncmp(argv[0], "rec", 3))
1031                 return __cmd_record(file_name, argc, argv);
1032         else if (!strncmp(argv[0], "rep", 3))
1033                 return __cmd_report(file_name, argc, argv);
1034         else if (!strncmp(argv[0], "diff", 4))
1035                 return cmd_diff(argc, argv, NULL);
1036         else if (!strncmp(argv[0], "top", 3))
1037                 return cmd_top(argc, argv, NULL);
1038         else if (!strncmp(argv[0], "buildid-list", 12))
1039                 return __cmd_buildid_list(file_name, argc, argv);
1040 #if defined(__i386__) || defined(__x86_64__)
1041         else if (!strncmp(argv[0], "stat", 4))
1042                 return kvm_cmd_stat(file_name, argc, argv);
1043 #endif
1044         else
1045                 usage_with_options(kvm_usage, kvm_options);
1046
1047         return 0;
1048 }