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 <lk/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         void *prev;
 332
 333         if (vcpu_id < event->max_vcpu)
 334                 return true;
 335
 336         while (event->max_vcpu <= vcpu_id)
 337                 event->max_vcpu += DEFAULT_VCPU_NUM;
 338
 339         prev = event->vcpu;
 340         event->vcpu = realloc(event->vcpu,
 341                               event->max_vcpu * sizeof(*event->vcpu));
 342         if (!event->vcpu) {
 343                 free(prev);
 344                 pr_err("Not enough memory\n");
 345                 return false;
 346         }
 347
 348         memset(event->vcpu + old_max_vcpu, 0,
 349                (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 350         return true;
 351 }
 352
 353 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 354 {
 355         struct kvm_event *event;
 356
 357         event = zalloc(sizeof(*event));
 358         if (!event) {
 359                 pr_err("Not enough memory\n");
 360                 return NULL;
 361         }
 362
 363         event->key = *key;
 364         return event;
 365 }
 366
 367 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 368                                                struct event_key *key)
 369 {
 370         struct kvm_event *event;
 371         struct list_head *head;
 372
 373         BUG_ON(key->key == INVALID_KEY);
 374
 375         head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 376         list_for_each_entry(event, head, hash_entry) {
 377                 if (event->key.key == key->key && event->key.info == key->info)
 378                         return event;
 379         }
 380
 381         event = kvm_alloc_init_event(key);
 382         if (!event)
 383                 return NULL;
 384
 385         list_add(&event->hash_entry, head);
 386         return event;
 387 }
 388
 389 static bool handle_begin_event(struct perf_kvm_stat *kvm,
 390                                struct vcpu_event_record *vcpu_record,
 391                                struct event_key *key, u64 timestamp)
 392 {
 393         struct kvm_event *event = NULL;
 394
 395         if (key->key != INVALID_KEY)
 396                 event = find_create_kvm_event(kvm, key);
 397
 398         vcpu_record->last_event = event;
 399         vcpu_record->start_time = timestamp;
 400         return true;
 401 }
 402
 403 static void
 404 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 405 {
 406         kvm_stats->time += time_diff;
 407         update_stats(&kvm_stats->stats, time_diff);
 408 }
 409
 410 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 411 {
 412         struct kvm_event_stats *kvm_stats = &event->total;
 413
 414         if (vcpu_id != -1)
 415                 kvm_stats = &event->vcpu[vcpu_id];
 416
 417         return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 418                                 avg_stats(&kvm_stats->stats));
 419 }
 420
 421 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 422                              u64 time_diff)
 423 {
 424         if (vcpu_id == -1) {
 425                 kvm_update_event_stats(&event->total, time_diff);
 426                 return true;
 427         }
 428
 429         if (!kvm_event_expand(event, vcpu_id))
 430                 return false;
 431
 432         kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 433         return true;
 434 }
 435
 436 static bool handle_end_event(struct perf_kvm_stat *kvm,
 437                              struct vcpu_event_record *vcpu_record,
 438                              struct event_key *key,
 439                              u64 timestamp)
 440 {
 441         struct kvm_event *event;
 442         u64 time_begin, time_diff;
 443         int vcpu;
 444
 445         if (kvm->trace_vcpu == -1)
 446                 vcpu = -1;
 447         else
 448                 vcpu = vcpu_record->vcpu_id;
 449
 450         event = vcpu_record->last_event;
 451         time_begin = vcpu_record->start_time;
 452
 453         /* The begin event is not caught. */
 454         if (!time_begin)
 455                 return true;
 456
 457         /*
 458          * In some case, the 'begin event' only records the start timestamp,
 459          * the actual event is recognized in the 'end event' (e.g. mmio-event).
 460          */
 461
 462         /* Both begin and end events did not get the key. */
 463         if (!event && key->key == INVALID_KEY)
 464                 return true;
 465
 466         if (!event)
 467                 event = find_create_kvm_event(kvm, key);
 468
 469         if (!event)
 470                 return false;
 471
 472         vcpu_record->last_event = NULL;
 473         vcpu_record->start_time = 0;
 474
 475         BUG_ON(timestamp < time_begin);
 476
 477         time_diff = timestamp - time_begin;
 478         return update_kvm_event(event, vcpu, time_diff);
 479 }
 480
 481 static
 482 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 483                                           struct perf_evsel *evsel,
 484                                           struct perf_sample *sample)
 485 {
 486         /* Only kvm_entry records vcpu id. */
 487         if (!thread->priv && kvm_entry_event(evsel)) {
 488                 struct vcpu_event_record *vcpu_record;
 489
 490                 vcpu_record = zalloc(sizeof(*vcpu_record));
 491                 if (!vcpu_record) {
 492                         pr_err("%s: Not enough memory\n", __func__);
 493                         return NULL;
 494                 }
 495
 496                 vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
 497                 thread->priv = vcpu_record;
 498         }
 499
 500         return thread->priv;
 501 }
 502
 503 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
 504                              struct thread *thread,
 505                              struct perf_evsel *evsel,
 506                              struct perf_sample *sample)
 507 {
 508         struct vcpu_event_record *vcpu_record;
 509         struct event_key key = {.key = INVALID_KEY};
 510
 511         vcpu_record = per_vcpu_record(thread, evsel, sample);
 512         if (!vcpu_record)
 513                 return true;
 514
 515         /* only process events for vcpus user cares about */
 516         if ((kvm->trace_vcpu != -1) &&
 517             (kvm->trace_vcpu != vcpu_record->vcpu_id))
 518                 return true;
 519
 520         if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 521                 return handle_begin_event(kvm, vcpu_record, &key, sample->time);
 522
 523         if (kvm->events_ops->is_end_event(evsel, sample, &key))
 524                 return handle_end_event(kvm, vcpu_record, &key, sample->time);
 525
 526         return true;
 527 }
 528
 529 #define GET_EVENT_KEY(func, field)                                      \
 530 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)         \
 531 {                                                                       \
 532         if (vcpu == -1)                                                 \
 533                 return event->total.field;                              \
 534                                                                         \
 535         if (vcpu >= event->max_vcpu)                                    \
 536                 return 0;                                               \
 537                                                                         \
 538         return event->vcpu[vcpu].field;                                 \
 539 }
 540
 541 #define COMPARE_EVENT_KEY(func, field)                                  \
 542 GET_EVENT_KEY(func, field)                                              \
 543 static int compare_kvm_event_ ## func(struct kvm_event *one,            \
 544                                         struct kvm_event *two, int vcpu)\
 545 {                                                                       \
 546         return get_event_ ##func(one, vcpu) >                           \
 547                                 get_event_ ##func(two, vcpu);           \
 548 }
 549
 550 GET_EVENT_KEY(time, time);
 551 COMPARE_EVENT_KEY(count, stats.n);
 552 COMPARE_EVENT_KEY(mean, stats.mean);
 553
 554 #define DEF_SORT_NAME_KEY(name, compare_key)                            \
 555         { #name, compare_kvm_event_ ## compare_key }
 556
 557 static struct kvm_event_key keys[] = {
 558         DEF_SORT_NAME_KEY(sample, count),
 559         DEF_SORT_NAME_KEY(time, mean),
 560         { NULL, NULL }
 561 };
 562
 563 static bool select_key(struct perf_kvm_stat *kvm)
 564 {
 565         int i;
 566
 567         for (i = 0; keys[i].name; i++) {
 568                 if (!strcmp(keys[i].name, kvm->sort_key)) {
 569                         kvm->compare = keys[i].key;
 570                         return true;
 571                 }
 572         }
 573
 574         pr_err("Unknown compare key:%s\n", kvm->sort_key);
 575         return false;
 576 }
 577
 578 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
 579                              key_cmp_fun bigger, int vcpu)
 580 {
 581         struct rb_node **rb = &result->rb_node;
 582         struct rb_node *parent = NULL;
 583         struct kvm_event *p;
 584
 585         while (*rb) {
 586                 p = container_of(*rb, struct kvm_event, rb);
 587                 parent = *rb;
 588
 589                 if (bigger(event, p, vcpu))
 590                         rb = &(*rb)->rb_left;
 591                 else
 592                         rb = &(*rb)->rb_right;
 593         }
 594
 595         rb_link_node(&event->rb, parent, rb);
 596         rb_insert_color(&event->rb, result);
 597 }
 598
 599 static void
 600 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
 601 {
 602         int vcpu = kvm->trace_vcpu;
 603
 604         kvm->total_count += get_event_count(event, vcpu);
 605         kvm->total_time += get_event_time(event, vcpu);
 606 }
 607
 608 static bool event_is_valid(struct kvm_event *event, int vcpu)
 609 {
 610         return !!get_event_count(event, vcpu);
 611 }
 612
 613 static void sort_result(struct perf_kvm_stat *kvm)
 614 {
 615         unsigned int i;
 616         int vcpu = kvm->trace_vcpu;
 617         struct kvm_event *event;
 618
 619         for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 620                 list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
 621                         if (event_is_valid(event, vcpu)) {
 622                                 update_total_count(kvm, event);
 623                                 insert_to_result(&kvm->result, event,
 624                                                  kvm->compare, vcpu);
 625                         }
 626                 }
 627         }
 628 }
 629
 630 /* returns left most element of result, and erase it */
 631 static struct kvm_event *pop_from_result(struct rb_root *result)
 632 {
 633         struct rb_node *node = rb_first(result);
 634
 635         if (!node)
 636                 return NULL;
 637
 638         rb_erase(node, result);
 639         return container_of(node, struct kvm_event, rb);
 640 }
 641
 642 static void print_vcpu_info(int vcpu)
 643 {
 644         pr_info("Analyze events for ");
 645
 646         if (vcpu == -1)
 647                 pr_info("all VCPUs:\n\n");
 648         else
 649                 pr_info("VCPU %d:\n\n", vcpu);
 650 }
 651
 652 static void print_result(struct perf_kvm_stat *kvm)
 653 {
 654         char decode[20];
 655         struct kvm_event *event;
 656         int vcpu = kvm->trace_vcpu;
 657
 658         pr_info("\n\n");
 659         print_vcpu_info(vcpu);
 660         pr_info("%20s ", kvm->events_ops->name);
 661         pr_info("%10s ", "Samples");
 662         pr_info("%9s ", "Samples%");
 663
 664         pr_info("%9s ", "Time%");
 665         pr_info("%16s ", "Avg time");
 666         pr_info("\n\n");
 667
 668         while ((event = pop_from_result(&kvm->result))) {
 669                 u64 ecount, etime;
 670
 671                 ecount = get_event_count(event, vcpu);
 672                 etime = get_event_time(event, vcpu);
 673
 674                 kvm->events_ops->decode_key(kvm, &event->key, decode);
 675                 pr_info("%20s ", decode);
 676                 pr_info("%10llu ", (unsigned long long)ecount);
 677                 pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
 678                 pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
 679                 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
 680                         kvm_event_rel_stddev(vcpu, event));
 681                 pr_info("\n");
 682         }
 683
 684         pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
 685                 kvm->total_count, kvm->total_time / 1e3);
 686 }
 687
 688 static int process_sample_event(struct perf_tool *tool,
 689                                 union perf_event *event,
 690                                 struct perf_sample *sample,
 691                                 struct perf_evsel *evsel,
 692                                 struct machine *machine)
 693 {
 694         struct thread *thread = machine__findnew_thread(machine, sample->tid);
 695         struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
 696                                                  tool);
 697
 698         if (thread == NULL) {
 699                 pr_debug("problem processing %d event, skipping it.\n",
 700                         event->header.type);
 701                 return -1;
 702         }
 703
 704         if (!handle_kvm_event(kvm, thread, evsel, sample))
 705                 return -1;
 706
 707         return 0;
 708 }
 709
 710 static int get_cpu_isa(struct perf_session *session)
 711 {
 712         char *cpuid = session->header.env.cpuid;
 713         int isa;
 714
 715         if (strstr(cpuid, "Intel"))
 716                 isa = 1;
 717         else if (strstr(cpuid, "AMD"))
 718                 isa = 0;
 719         else {
 720                 pr_err("CPU %s is not supported.\n", cpuid);
 721                 isa = -ENOTSUP;
 722         }
 723
 724         return isa;
 725 }
 726
 727 static int read_events(struct perf_kvm_stat *kvm)
 728 {
 729         int ret;
 730
 731         struct perf_tool eops = {
 732                 .sample                 = process_sample_event,
 733                 .comm                   = perf_event__process_comm,
 734                 .ordered_samples        = true,
 735         };
 736
 737         kvm->tool = eops;
 738         kvm->session = perf_session__new(kvm->file_name, O_RDONLY, 0, false,
 739                                          &kvm->tool);
 740         if (!kvm->session) {
 741                 pr_err("Initializing perf session failed\n");
 742                 return -EINVAL;
 743         }
 744
 745         if (!perf_session__has_traces(kvm->session, "kvm record"))
 746                 return -EINVAL;
 747
 748         /*
 749          * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
 750          * traced in the old kernel.
 751          */
 752         ret = get_cpu_isa(kvm->session);
 753
 754         if (ret < 0)
 755                 return ret;
 756
 757         if (ret == 1) {
 758                 kvm->exit_reasons = vmx_exit_reasons;
 759                 kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
 760                 kvm->exit_reasons_isa = "VMX";
 761         }
 762
 763         return perf_session__process_events(kvm->session, &kvm->tool);
 764 }
 765
 766 static bool verify_vcpu(int vcpu)
 767 {
 768         if (vcpu != -1 && vcpu < 0) {
 769                 pr_err("Invalid vcpu:%d.\n", vcpu);
 770                 return false;
 771         }
 772
 773         return true;
 774 }
 775
 776 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
 777 {
 778         int ret = -EINVAL;
 779         int vcpu = kvm->trace_vcpu;
 780
 781         if (!verify_vcpu(vcpu))
 782                 goto exit;
 783
 784         if (!select_key(kvm))
 785                 goto exit;
 786
 787         if (!register_kvm_events_ops(kvm))
 788                 goto exit;
 789
 790         init_kvm_event_record(kvm);
 791         setup_pager();
 792
 793         ret = read_events(kvm);
 794         if (ret)
 795                 goto exit;
 796
 797         sort_result(kvm);
 798         print_result(kvm);
 799
 800 exit:
 801         return ret;
 802 }
 803
 804 static const char * const record_args[] = {
 805         "record",
 806         "-R",
 807         "-f",
 808         "-m", "1024",
 809         "-c", "1",
 810         "-e", "kvm:kvm_entry",
 811         "-e", "kvm:kvm_exit",
 812         "-e", "kvm:kvm_mmio",
 813         "-e", "kvm:kvm_pio",
 814 };
 815
 816 #define STRDUP_FAIL_EXIT(s)             \
 817         ({      char *_p;               \
 818         _p = strdup(s);         \
 819                 if (!_p)                \
 820                         return -ENOMEM; \
 821                 _p;                     \
 822         })
 823
 824 static int
 825 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
 826 {
 827         unsigned int rec_argc, i, j;
 828         const char **rec_argv;
 829
 830         rec_argc = ARRAY_SIZE(record_args) + argc + 2;
 831         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 832
 833         if (rec_argv == NULL)
 834                 return -ENOMEM;
 835
 836         for (i = 0; i < ARRAY_SIZE(record_args); i++)
 837                 rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
 838
 839         rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
 840         rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
 841
 842         for (j = 1; j < (unsigned int)argc; j++, i++)
 843                 rec_argv[i] = argv[j];
 844
 845         return cmd_record(i, rec_argv, NULL);
 846 }
 847
 848 static int
 849 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
 850 {
 851         const struct option kvm_events_report_options[] = {
 852                 OPT_STRING(0, "event", &kvm->report_event, "report event",
 853                             "event for reporting: vmexit, mmio, ioport"),
 854                 OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
 855                             "vcpu id to report"),
 856                 OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
 857                             "key for sorting: sample(sort by samples number)"
 858                             " time (sort by avg time)"),
 859                 OPT_END()
 860         };
 861
 862         const char * const kvm_events_report_usage[] = {
 863                 "perf kvm stat report [<options>]",
 864                 NULL
 865         };
 866
 867         symbol__init();
 868
 869         if (argc) {
 870                 argc = parse_options(argc, argv,
 871                                      kvm_events_report_options,
 872                                      kvm_events_report_usage, 0);
 873                 if (argc)
 874                         usage_with_options(kvm_events_report_usage,
 875                                            kvm_events_report_options);
 876         }
 877
 878         return kvm_events_report_vcpu(kvm);
 879 }
 880
 881 static void print_kvm_stat_usage(void)
 882 {
 883         printf("Usage: perf kvm stat <command>\n\n");
 884
 885         printf("# Available commands:\n");
 886         printf("\trecord: record kvm events\n");
 887         printf("\treport: report statistical data of kvm events\n");
 888
 889         printf("\nOtherwise, it is the alias of 'perf stat':\n");
 890 }
 891
 892 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
 893 {
 894         struct perf_kvm_stat kvm = {
 895                 .file_name = file_name,
 896
 897                 .trace_vcpu     = -1,
 898                 .report_event   = "vmexit",
 899                 .sort_key       = "sample",
 900
 901                 .exit_reasons = svm_exit_reasons,
 902                 .exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
 903                 .exit_reasons_isa = "SVM",
 904         };
 905
 906         if (argc == 1) {
 907                 print_kvm_stat_usage();
 908                 goto perf_stat;
 909         }
 910
 911         if (!strncmp(argv[1], "rec", 3))
 912                 return kvm_events_record(&kvm, argc - 1, argv + 1);
 913
 914         if (!strncmp(argv[1], "rep", 3))
 915                 return kvm_events_report(&kvm, argc - 1 , argv + 1);
 916
 917 perf_stat:
 918         return cmd_stat(argc, argv, NULL);
 919 }
 920 #endif
 921
 922 static int __cmd_record(const char *file_name, int argc, const char **argv)
 923 {
 924         int rec_argc, i = 0, j;
 925         const char **rec_argv;
 926
 927         rec_argc = argc + 2;
 928         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 929         rec_argv[i++] = strdup("record");
 930         rec_argv[i++] = strdup("-o");
 931         rec_argv[i++] = strdup(file_name);
 932         for (j = 1; j < argc; j++, i++)
 933                 rec_argv[i] = argv[j];
 934
 935         BUG_ON(i != rec_argc);
 936
 937         return cmd_record(i, rec_argv, NULL);
 938 }
 939
 940 static int __cmd_report(const char *file_name, int argc, const char **argv)
 941 {
 942         int rec_argc, i = 0, j;
 943         const char **rec_argv;
 944
 945         rec_argc = argc + 2;
 946         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 947         rec_argv[i++] = strdup("report");
 948         rec_argv[i++] = strdup("-i");
 949         rec_argv[i++] = strdup(file_name);
 950         for (j = 1; j < argc; j++, i++)
 951                 rec_argv[i] = argv[j];
 952
 953         BUG_ON(i != rec_argc);
 954
 955         return cmd_report(i, rec_argv, NULL);
 956 }
 957
 958 static int
 959 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
 960 {
 961         int rec_argc, i = 0, j;
 962         const char **rec_argv;
 963
 964         rec_argc = argc + 2;
 965         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 966         rec_argv[i++] = strdup("buildid-list");
 967         rec_argv[i++] = strdup("-i");
 968         rec_argv[i++] = strdup(file_name);
 969         for (j = 1; j < argc; j++, i++)
 970                 rec_argv[i] = argv[j];
 971
 972         BUG_ON(i != rec_argc);
 973
 974         return cmd_buildid_list(i, rec_argv, NULL);
 975 }
 976
 977 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
 978 {
 979         const char *file_name = NULL;
 980         const struct option kvm_options[] = {
 981                 OPT_STRING('i', "input", &file_name, "file",
 982                            "Input file name"),
 983                 OPT_STRING('o', "output", &file_name, "file",
 984                            "Output file name"),
 985                 OPT_BOOLEAN(0, "guest", &perf_guest,
 986                             "Collect guest os data"),
 987                 OPT_BOOLEAN(0, "host", &perf_host,
 988                             "Collect host os data"),
 989                 OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
 990                            "guest mount directory under which every guest os"
 991                            " instance has a subdir"),
 992                 OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
 993                            "file", "file saving guest os vmlinux"),
 994                 OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
 995                            "file", "file saving guest os /proc/kallsyms"),
 996                 OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
 997                            "file", "file saving guest os /proc/modules"),
 998                 OPT_END()
 999         };
1000
1001
1002         const char * const kvm_usage[] = {
1003                 "perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
1004                 NULL
1005         };
1006
1007         perf_host  = 0;
1008         perf_guest = 1;
1009
1010         argc = parse_options(argc, argv, kvm_options, kvm_usage,
1011                         PARSE_OPT_STOP_AT_NON_OPTION);
1012         if (!argc)
1013                 usage_with_options(kvm_usage, kvm_options);
1014
1015         if (!perf_host)
1016                 perf_guest = 1;
1017
1018         if (!file_name) {
1019                 if (perf_host && !perf_guest)
1020                         file_name = strdup("perf.data.host");
1021                 else if (!perf_host && perf_guest)
1022                         file_name = strdup("perf.data.guest");
1023                 else
1024                         file_name = strdup("perf.data.kvm");
1025
1026                 if (!file_name) {
1027                         pr_err("Failed to allocate memory for filename\n");
1028                         return -ENOMEM;
1029                 }
1030         }
1031
1032         if (!strncmp(argv[0], "rec", 3))
1033                 return __cmd_record(file_name, argc, argv);
1034         else if (!strncmp(argv[0], "rep", 3))
1035                 return __cmd_report(file_name, argc, argv);
1036         else if (!strncmp(argv[0], "diff", 4))
1037                 return cmd_diff(argc, argv, NULL);
1038         else if (!strncmp(argv[0], "top", 3))
1039                 return cmd_top(argc, argv, NULL);
1040         else if (!strncmp(argv[0], "buildid-list", 12))
1041                 return __cmd_buildid_list(file_name, argc, argv);
1042 #if defined(__i386__) || defined(__x86_64__)
1043         else if (!strncmp(argv[0], "stat", 4))
1044                 return kvm_cmd_stat(file_name, argc, argv);
1045 #endif
1046         else
1047                 usage_with_options(kvm_usage, kvm_options);
1048
1049         return 0;
1050 }