tools/perf/builtin-lock.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/util.h"
   5 #include "util/cache.h"
   6 #include "util/symbol.h"
   7 #include "util/thread.h"
   8 #include "util/header.h"
   9
  10 #include "util/parse-options.h"
  11 #include "util/trace-event.h"
  12
  13 #include "util/debug.h"
  14 #include "util/session.h"
  15
  16 #include <sys/types.h>
  17 #include <sys/prctl.h>
  18 #include <semaphore.h>
  19 #include <pthread.h>
  20 #include <math.h>
  21 #include <limits.h>
  22
  23 #include <linux/list.h>
  24 #include <linux/hash.h>
  25
  26 static struct perf_session *session;
  27
  28 /* based on kernel/lockdep.c */
  29 #define LOCKHASH_BITS           12
  30 #define LOCKHASH_SIZE           (1UL << LOCKHASH_BITS)
  31
  32 static struct list_head lockhash_table[LOCKHASH_SIZE];
  33
  34 #define __lockhashfn(key)       hash_long((unsigned long)key, LOCKHASH_BITS)
  35 #define lockhashentry(key)      (lockhash_table + __lockhashfn((key)))
  36
  37 struct lock_stat {
  38         struct list_head        hash_entry;
  39         struct rb_node          rb;             /* used for sorting */
  40
  41         /*
  42          * FIXME: raw_field_value() returns unsigned long long,
  43          * so address of lockdep_map should be dealed as 64bit.
  44          * Is there more better solution?
  45          */
  46         void                    *addr;          /* address of lockdep_map, used as ID */
  47         char                    *name;          /* for strcpy(), we cannot use const */
  48
  49         unsigned int            nr_acquire;
  50         unsigned int            nr_acquired;
  51         unsigned int            nr_contended;
  52         unsigned int            nr_release;
  53
  54         unsigned int            nr_readlock;
  55         unsigned int            nr_trylock;
  56         /* these times are in nano sec. */
  57         u64                     wait_time_total;
  58         u64                     wait_time_min;
  59         u64                     wait_time_max;
  60
  61         int                     discard; /* flag of blacklist */
  62 };
  63
  64 /*
  65  * States of lock_seq_stat
  66  *
  67  * UNINITIALIZED is required for detecting first event of acquire.
  68  * As the nature of lock events, there is no guarantee
  69  * that the first event for the locks are acquire,
  70  * it can be acquired, contended or release.
  71  */
  72 #define SEQ_STATE_UNINITIALIZED      0         /* initial state */
  73 #define SEQ_STATE_RELEASED      1
  74 #define SEQ_STATE_ACQUIRING     2
  75 #define SEQ_STATE_ACQUIRED      3
  76 #define SEQ_STATE_READ_ACQUIRED 4
  77 #define SEQ_STATE_CONTENDED     5
  78
  79 /*
  80  * MAX_LOCK_DEPTH
  81  * Imported from include/linux/sched.h.
  82  * Should this be synchronized?
  83  */
  84 #define MAX_LOCK_DEPTH 48
  85
  86 /*
  87  * struct lock_seq_stat:
  88  * Place to put on state of one lock sequence
  89  * 1) acquire -> acquired -> release
  90  * 2) acquire -> contended -> acquired -> release
  91  * 3) acquire (with read or try) -> release
  92  * 4) Are there other patterns?
  93  */
  94 struct lock_seq_stat {
  95         struct list_head        list;
  96         int                     state;
  97         u64                     prev_event_time;
  98         void                    *addr;
  99
 100         int                     read_count;
 101 };
 102
 103 struct thread_stat {
 104         struct rb_node          rb;
 105
 106         u32                     tid;
 107         struct list_head        seq_list;
 108 };
 109
 110 static struct rb_root           thread_stats;
 111
 112 static struct thread_stat *thread_stat_find(u32 tid)
 113 {
 114         struct rb_node *node;
 115         struct thread_stat *st;
 116
 117         node = thread_stats.rb_node;
 118         while (node) {
 119                 st = container_of(node, struct thread_stat, rb);
 120                 if (st->tid == tid)
 121                         return st;
 122                 else if (tid < st->tid)
 123                         node = node->rb_left;
 124                 else
 125                         node = node->rb_right;
 126         }
 127
 128         return NULL;
 129 }
 130
 131 static void thread_stat_insert(struct thread_stat *new)
 132 {
 133         struct rb_node **rb = &thread_stats.rb_node;
 134         struct rb_node *parent = NULL;
 135         struct thread_stat *p;
 136
 137         while (*rb) {
 138                 p = container_of(*rb, struct thread_stat, rb);
 139                 parent = *rb;
 140
 141                 if (new->tid < p->tid)
 142                         rb = &(*rb)->rb_left;
 143                 else if (new->tid > p->tid)
 144                         rb = &(*rb)->rb_right;
 145                 else
 146                         BUG_ON("inserting invalid thread_stat\n");
 147         }
 148
 149         rb_link_node(&new->rb, parent, rb);
 150         rb_insert_color(&new->rb, &thread_stats);
 151 }
 152
 153 static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
 154 {
 155         struct thread_stat *st;
 156
 157         st = thread_stat_find(tid);
 158         if (st)
 159                 return st;
 160
 161         st = zalloc(sizeof(struct thread_stat));
 162         if (!st)
 163                 die("memory allocation failed\n");
 164
 165         st->tid = tid;
 166         INIT_LIST_HEAD(&st->seq_list);
 167
 168         thread_stat_insert(st);
 169
 170         return st;
 171 }
 172
 173 static struct thread_stat *thread_stat_findnew_first(u32 tid);
 174 static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
 175         thread_stat_findnew_first;
 176
 177 static struct thread_stat *thread_stat_findnew_first(u32 tid)
 178 {
 179         struct thread_stat *st;
 180
 181         st = zalloc(sizeof(struct thread_stat));
 182         if (!st)
 183                 die("memory allocation failed\n");
 184         st->tid = tid;
 185         INIT_LIST_HEAD(&st->seq_list);
 186
 187         rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
 188         rb_insert_color(&st->rb, &thread_stats);
 189
 190         thread_stat_findnew = thread_stat_findnew_after_first;
 191         return st;
 192 }
 193
 194 /* build simple key function one is bigger than two */
 195 #define SINGLE_KEY(member)                                              \
 196         static int lock_stat_key_ ## member(struct lock_stat *one,      \
 197                                          struct lock_stat *two)         \
 198         {                                                               \
 199                 return one->member > two->member;                       \
 200         }
 201
 202 SINGLE_KEY(nr_acquired)
 203 SINGLE_KEY(nr_contended)
 204 SINGLE_KEY(wait_time_total)
 205 SINGLE_KEY(wait_time_min)
 206 SINGLE_KEY(wait_time_max)
 207
 208 struct lock_key {
 209         /*
 210          * name: the value for specify by user
 211          * this should be simpler than raw name of member
 212          * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
 213          */
 214         const char              *name;
 215         int                     (*key)(struct lock_stat*, struct lock_stat*);
 216 };
 217
 218 static const char               *sort_key = "acquired";
 219
 220 static int                      (*compare)(struct lock_stat *, struct lock_stat *);
 221
 222 static struct rb_root           result; /* place to store sorted data */
 223
 224 #define DEF_KEY_LOCK(name, fn_suffix)   \
 225         { #name, lock_stat_key_ ## fn_suffix }
 226 struct lock_key keys[] = {
 227         DEF_KEY_LOCK(acquired, nr_acquired),
 228         DEF_KEY_LOCK(contended, nr_contended),
 229         DEF_KEY_LOCK(wait_total, wait_time_total),
 230         DEF_KEY_LOCK(wait_min, wait_time_min),
 231         DEF_KEY_LOCK(wait_max, wait_time_max),
 232
 233         /* extra comparisons much complicated should be here */
 234
 235         { NULL, NULL }
 236 };
 237
 238 static void select_key(void)
 239 {
 240         int i;
 241
 242         for (i = 0; keys[i].name; i++) {
 243                 if (!strcmp(keys[i].name, sort_key)) {
 244                         compare = keys[i].key;
 245                         return;
 246                 }
 247         }
 248
 249         die("Unknown compare key:%s\n", sort_key);
 250 }
 251
 252 static void insert_to_result(struct lock_stat *st,
 253                              int (*bigger)(struct lock_stat *, struct lock_stat *))
 254 {
 255         struct rb_node **rb = &result.rb_node;
 256         struct rb_node *parent = NULL;
 257         struct lock_stat *p;
 258
 259         while (*rb) {
 260                 p = container_of(*rb, struct lock_stat, rb);
 261                 parent = *rb;
 262
 263                 if (bigger(st, p))
 264                         rb = &(*rb)->rb_left;
 265                 else
 266                         rb = &(*rb)->rb_right;
 267         }
 268
 269         rb_link_node(&st->rb, parent, rb);
 270         rb_insert_color(&st->rb, &result);
 271 }
 272
 273 /* returns left most element of result, and erase it */
 274 static struct lock_stat *pop_from_result(void)
 275 {
 276         struct rb_node *node = result.rb_node;
 277
 278         if (!node)
 279                 return NULL;
 280
 281         while (node->rb_left)
 282                 node = node->rb_left;
 283
 284         rb_erase(node, &result);
 285         return container_of(node, struct lock_stat, rb);
 286 }
 287
 288 static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
 289 {
 290         struct list_head *entry = lockhashentry(addr);
 291         struct lock_stat *ret, *new;
 292
 293         list_for_each_entry(ret, entry, hash_entry) {
 294                 if (ret->addr == addr)
 295                         return ret;
 296         }
 297
 298         new = zalloc(sizeof(struct lock_stat));
 299         if (!new)
 300                 goto alloc_failed;
 301
 302         new->addr = addr;
 303         new->name = zalloc(sizeof(char) * strlen(name) + 1);
 304         if (!new->name)
 305                 goto alloc_failed;
 306         strcpy(new->name, name);
 307
 308         new->wait_time_min = ULLONG_MAX;
 309
 310         list_add(&new->hash_entry, entry);
 311         return new;
 312
 313 alloc_failed:
 314         die("memory allocation failed\n");
 315 }
 316
 317 static char                     const *input_name = "perf.data";
 318
 319 static int                      profile_cpu = -1;
 320
 321 struct raw_event_sample {
 322         u32                     size;
 323         char                    data[0];
 324 };
 325
 326 struct trace_acquire_event {
 327         void                    *addr;
 328         const char              *name;
 329         int                     flag;
 330 };
 331
 332 struct trace_acquired_event {
 333         void                    *addr;
 334         const char              *name;
 335 };
 336
 337 struct trace_contended_event {
 338         void                    *addr;
 339         const char              *name;
 340 };
 341
 342 struct trace_release_event {
 343         void                    *addr;
 344         const char              *name;
 345 };
 346
 347 struct trace_lock_handler {
 348         void (*acquire_event)(struct trace_acquire_event *,
 349                               struct event *,
 350                               int cpu,
 351                               u64 timestamp,
 352                               struct thread *thread);
 353
 354         void (*acquired_event)(struct trace_acquired_event *,
 355                                struct event *,
 356                                int cpu,
 357                                u64 timestamp,
 358                                struct thread *thread);
 359
 360         void (*contended_event)(struct trace_contended_event *,
 361                                 struct event *,
 362                                 int cpu,
 363                                 u64 timestamp,
 364                                 struct thread *thread);
 365
 366         void (*release_event)(struct trace_release_event *,
 367                               struct event *,
 368                               int cpu,
 369                               u64 timestamp,
 370                               struct thread *thread);
 371 };
 372
 373 static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
 374 {
 375         struct lock_seq_stat *seq;
 376
 377         list_for_each_entry(seq, &ts->seq_list, list) {
 378                 if (seq->addr == addr)
 379                         return seq;
 380         }
 381
 382         seq = zalloc(sizeof(struct lock_seq_stat));
 383         if (!seq)
 384                 die("Not enough memory\n");
 385         seq->state = SEQ_STATE_UNINITIALIZED;
 386         seq->addr = addr;
 387
 388         list_add(&seq->list, &ts->seq_list);
 389         return seq;
 390 }
 391
 392 static int bad_hist[4];
 393
 394 static void
 395 report_lock_acquire_event(struct trace_acquire_event *acquire_event,
 396                         struct event *__event __used,
 397                         int cpu __used,
 398                         u64 timestamp __used,
 399                         struct thread *thread __used)
 400 {
 401         struct lock_stat *ls;
 402         struct thread_stat *ts;
 403         struct lock_seq_stat *seq;
 404
 405         ls = lock_stat_findnew(acquire_event->addr, acquire_event->name);
 406         if (ls->discard)
 407                 return;
 408
 409         ts = thread_stat_findnew(thread->pid);
 410         seq = get_seq(ts, acquire_event->addr);
 411
 412         switch (seq->state) {
 413         case SEQ_STATE_UNINITIALIZED:
 414         case SEQ_STATE_RELEASED:
 415                 if (!acquire_event->flag) {
 416                         seq->state = SEQ_STATE_ACQUIRING;
 417                 } else {
 418                         if (acquire_event->flag & 1)
 419                                 ls->nr_trylock++;
 420                         if (acquire_event->flag & 2)
 421                                 ls->nr_readlock++;
 422                         seq->state = SEQ_STATE_READ_ACQUIRED;
 423                         seq->read_count = 1;
 424                         ls->nr_acquired++;
 425                 }
 426                 break;
 427         case SEQ_STATE_READ_ACQUIRED:
 428                 if (acquire_event->flag & 2) {
 429                         seq->read_count++;
 430                         ls->nr_acquired++;
 431                         goto end;
 432                 } else {
 433                         goto broken;
 434                 }
 435                 break;
 436         case SEQ_STATE_ACQUIRED:
 437         case SEQ_STATE_ACQUIRING:
 438         case SEQ_STATE_CONTENDED:
 439 broken:
 440                 /* broken lock sequence, discard it */
 441                 ls->discard = 1;
 442                 bad_hist[0]++;
 443                 list_del(&seq->list);
 444                 free(seq);
 445                 goto end;
 446                 break;
 447         default:
 448                 BUG_ON("Unknown state of lock sequence found!\n");
 449                 break;
 450         }
 451
 452         ls->nr_acquire++;
 453         seq->prev_event_time = timestamp;
 454 end:
 455         return;
 456 }
 457
 458 static void
 459 report_lock_acquired_event(struct trace_acquired_event *acquired_event,
 460                          struct event *__event __used,
 461                          int cpu __used,
 462                          u64 timestamp __used,
 463                          struct thread *thread __used)
 464 {
 465         struct lock_stat *ls;
 466         struct thread_stat *ts;
 467         struct lock_seq_stat *seq;
 468         u64 contended_term;
 469
 470         ls = lock_stat_findnew(acquired_event->addr, acquired_event->name);
 471         if (ls->discard)
 472                 return;
 473
 474         ts = thread_stat_findnew(thread->pid);
 475         seq = get_seq(ts, acquired_event->addr);
 476
 477         switch (seq->state) {
 478         case SEQ_STATE_UNINITIALIZED:
 479                 /* orphan event, do nothing */
 480                 return;
 481         case SEQ_STATE_ACQUIRING:
 482                 break;
 483         case SEQ_STATE_CONTENDED:
 484                 contended_term = timestamp - seq->prev_event_time;
 485                 ls->wait_time_total += contended_term;
 486
 487                 if (contended_term < ls->wait_time_min)
 488                         ls->wait_time_min = contended_term;
 489                 else if (ls->wait_time_max < contended_term)
 490                         ls->wait_time_max = contended_term;
 491                 break;
 492         case SEQ_STATE_RELEASED:
 493         case SEQ_STATE_ACQUIRED:
 494         case SEQ_STATE_READ_ACQUIRED:
 495                 /* broken lock sequence, discard it */
 496                 ls->discard = 1;
 497                 bad_hist[1]++;
 498                 list_del(&seq->list);
 499                 free(seq);
 500                 goto end;
 501                 break;
 502
 503         default:
 504                 BUG_ON("Unknown state of lock sequence found!\n");
 505                 break;
 506         }
 507
 508         seq->state = SEQ_STATE_ACQUIRED;
 509         ls->nr_acquired++;
 510         seq->prev_event_time = timestamp;
 511 end:
 512         return;
 513 }
 514
 515 static void
 516 report_lock_contended_event(struct trace_contended_event *contended_event,
 517                           struct event *__event __used,
 518                           int cpu __used,
 519                           u64 timestamp __used,
 520                           struct thread *thread __used)
 521 {
 522         struct lock_stat *ls;
 523         struct thread_stat *ts;
 524         struct lock_seq_stat *seq;
 525
 526         ls = lock_stat_findnew(contended_event->addr, contended_event->name);
 527         if (ls->discard)
 528                 return;
 529
 530         ts = thread_stat_findnew(thread->pid);
 531         seq = get_seq(ts, contended_event->addr);
 532
 533         switch (seq->state) {
 534         case SEQ_STATE_UNINITIALIZED:
 535                 /* orphan event, do nothing */
 536                 return;
 537         case SEQ_STATE_ACQUIRING:
 538                 break;
 539         case SEQ_STATE_RELEASED:
 540         case SEQ_STATE_ACQUIRED:
 541         case SEQ_STATE_READ_ACQUIRED:
 542         case SEQ_STATE_CONTENDED:
 543                 /* broken lock sequence, discard it */
 544                 ls->discard = 1;
 545                 bad_hist[2]++;
 546                 list_del(&seq->list);
 547                 free(seq);
 548                 goto end;
 549                 break;
 550         default:
 551                 BUG_ON("Unknown state of lock sequence found!\n");
 552                 break;
 553         }
 554
 555         seq->state = SEQ_STATE_CONTENDED;
 556         ls->nr_contended++;
 557         seq->prev_event_time = timestamp;
 558 end:
 559         return;
 560 }
 561
 562 static void
 563 report_lock_release_event(struct trace_release_event *release_event,
 564                         struct event *__event __used,
 565                         int cpu __used,
 566                         u64 timestamp __used,
 567                         struct thread *thread __used)
 568 {
 569         struct lock_stat *ls;
 570         struct thread_stat *ts;
 571         struct lock_seq_stat *seq;
 572
 573         ls = lock_stat_findnew(release_event->addr, release_event->name);
 574         if (ls->discard)
 575                 return;
 576
 577         ts = thread_stat_findnew(thread->pid);
 578         seq = get_seq(ts, release_event->addr);
 579
 580         switch (seq->state) {
 581         case SEQ_STATE_UNINITIALIZED:
 582                 goto end;
 583                 break;
 584         case SEQ_STATE_ACQUIRED:
 585                 break;
 586         case SEQ_STATE_READ_ACQUIRED:
 587                 seq->read_count--;
 588                 BUG_ON(seq->read_count < 0);
 589                 if (!seq->read_count) {
 590                         ls->nr_release++;
 591                         goto end;
 592                 }
 593                 break;
 594         case SEQ_STATE_ACQUIRING:
 595         case SEQ_STATE_CONTENDED:
 596         case SEQ_STATE_RELEASED:
 597                 /* broken lock sequence, discard it */
 598                 ls->discard = 1;
 599                 bad_hist[3]++;
 600                 goto free_seq;
 601                 break;
 602         default:
 603                 BUG_ON("Unknown state of lock sequence found!\n");
 604                 break;
 605         }
 606
 607         ls->nr_release++;
 608 free_seq:
 609         list_del(&seq->list);
 610         free(seq);
 611 end:
 612         return;
 613 }
 614
 615 /* lock oriented handlers */
 616 /* TODO: handlers for CPU oriented, thread oriented */
 617 static struct trace_lock_handler report_lock_ops  = {
 618         .acquire_event          = report_lock_acquire_event,
 619         .acquired_event         = report_lock_acquired_event,
 620         .contended_event        = report_lock_contended_event,
 621         .release_event          = report_lock_release_event,
 622 };
 623
 624 static struct trace_lock_handler *trace_handler;
 625
 626 static void
 627 process_lock_acquire_event(void *data,
 628                            struct event *event __used,
 629                            int cpu __used,
 630                            u64 timestamp __used,
 631                            struct thread *thread __used)
 632 {
 633         struct trace_acquire_event acquire_event;
 634         u64 tmp;                /* this is required for casting... */
 635
 636         tmp = raw_field_value(event, "lockdep_addr", data);
 637         memcpy(&acquire_event.addr, &tmp, sizeof(void *));
 638         acquire_event.name = (char *)raw_field_ptr(event, "name", data);
 639         acquire_event.flag = (int)raw_field_value(event, "flag", data);
 640
 641         if (trace_handler->acquire_event)
 642                 trace_handler->acquire_event(&acquire_event, event, cpu, timestamp, thread);
 643 }
 644
 645 static void
 646 process_lock_acquired_event(void *data,
 647                             struct event *event __used,
 648                             int cpu __used,
 649                             u64 timestamp __used,
 650                             struct thread *thread __used)
 651 {
 652         struct trace_acquired_event acquired_event;
 653         u64 tmp;                /* this is required for casting... */
 654
 655         tmp = raw_field_value(event, "lockdep_addr", data);
 656         memcpy(&acquired_event.addr, &tmp, sizeof(void *));
 657         acquired_event.name = (char *)raw_field_ptr(event, "name", data);
 658
 659         if (trace_handler->acquire_event)
 660                 trace_handler->acquired_event(&acquired_event, event, cpu, timestamp, thread);
 661 }
 662
 663 static void
 664 process_lock_contended_event(void *data,
 665                              struct event *event __used,
 666                              int cpu __used,
 667                              u64 timestamp __used,
 668                              struct thread *thread __used)
 669 {
 670         struct trace_contended_event contended_event;
 671         u64 tmp;                /* this is required for casting... */
 672
 673         tmp = raw_field_value(event, "lockdep_addr", data);
 674         memcpy(&contended_event.addr, &tmp, sizeof(void *));
 675         contended_event.name = (char *)raw_field_ptr(event, "name", data);
 676
 677         if (trace_handler->acquire_event)
 678                 trace_handler->contended_event(&contended_event, event, cpu, timestamp, thread);
 679 }
 680
 681 static void
 682 process_lock_release_event(void *data,
 683                            struct event *event __used,
 684                            int cpu __used,
 685                            u64 timestamp __used,
 686                            struct thread *thread __used)
 687 {
 688         struct trace_release_event release_event;
 689         u64 tmp;                /* this is required for casting... */
 690
 691         tmp = raw_field_value(event, "lockdep_addr", data);
 692         memcpy(&release_event.addr, &tmp, sizeof(void *));
 693         release_event.name = (char *)raw_field_ptr(event, "name", data);
 694
 695         if (trace_handler->acquire_event)
 696                 trace_handler->release_event(&release_event, event, cpu, timestamp, thread);
 697 }
 698
 699 static void
 700 process_raw_event(void *data, int cpu __used,
 701                   u64 timestamp __used, struct thread *thread __used)
 702 {
 703         struct event *event;
 704         int type;
 705
 706         type = trace_parse_common_type(data);
 707         event = trace_find_event(type);
 708
 709         if (!strcmp(event->name, "lock_acquire"))
 710                 process_lock_acquire_event(data, event, cpu, timestamp, thread);
 711         if (!strcmp(event->name, "lock_acquired"))
 712                 process_lock_acquired_event(data, event, cpu, timestamp, thread);
 713         if (!strcmp(event->name, "lock_contended"))
 714                 process_lock_contended_event(data, event, cpu, timestamp, thread);
 715         if (!strcmp(event->name, "lock_release"))
 716                 process_lock_release_event(data, event, cpu, timestamp, thread);
 717 }
 718
 719 struct raw_event_queue {
 720         u64                     timestamp;
 721         int                     cpu;
 722         void                    *data;
 723         struct thread           *thread;
 724         struct list_head        list;
 725 };
 726
 727 static LIST_HEAD(raw_event_head);
 728
 729 #define FLUSH_PERIOD    (5 * NSEC_PER_SEC)
 730
 731 static u64 flush_limit = ULLONG_MAX;
 732 static u64 last_flush = 0;
 733 struct raw_event_queue *last_inserted;
 734
 735 static void flush_raw_event_queue(u64 limit)
 736 {
 737         struct raw_event_queue *tmp, *iter;
 738
 739         list_for_each_entry_safe(iter, tmp, &raw_event_head, list) {
 740                 if (iter->timestamp > limit)
 741                         return;
 742
 743                 if (iter == last_inserted)
 744                         last_inserted = NULL;
 745
 746                 process_raw_event(iter->data, iter->cpu, iter->timestamp,
 747                                   iter->thread);
 748
 749                 last_flush = iter->timestamp;
 750                 list_del(&iter->list);
 751                 free(iter->data);
 752                 free(iter);
 753         }
 754 }
 755
 756 static void __queue_raw_event_end(struct raw_event_queue *new)
 757 {
 758         struct raw_event_queue *iter;
 759
 760         list_for_each_entry_reverse(iter, &raw_event_head, list) {
 761                 if (iter->timestamp < new->timestamp) {
 762                         list_add(&new->list, &iter->list);
 763                         return;
 764                 }
 765         }
 766
 767         list_add(&new->list, &raw_event_head);
 768 }
 769
 770 static void __queue_raw_event_before(struct raw_event_queue *new,
 771                                      struct raw_event_queue *iter)
 772 {
 773         list_for_each_entry_continue_reverse(iter, &raw_event_head, list) {
 774                 if (iter->timestamp < new->timestamp) {
 775                         list_add(&new->list, &iter->list);
 776                         return;
 777                 }
 778         }
 779
 780         list_add(&new->list, &raw_event_head);
 781 }
 782
 783 static void __queue_raw_event_after(struct raw_event_queue *new,
 784                                      struct raw_event_queue *iter)
 785 {
 786         list_for_each_entry_continue(iter, &raw_event_head, list) {
 787                 if (iter->timestamp > new->timestamp) {
 788                         list_add_tail(&new->list, &iter->list);
 789                         return;
 790                 }
 791         }
 792         list_add_tail(&new->list, &raw_event_head);
 793 }
 794
 795 /* The queue is ordered by time */
 796 static void __queue_raw_event(struct raw_event_queue *new)
 797 {
 798         if (!last_inserted) {
 799                 __queue_raw_event_end(new);
 800                 return;
 801         }
 802
 803         /*
 804          * Most of the time the current event has a timestamp
 805          * very close to the last event inserted, unless we just switched
 806          * to another event buffer. Having a sorting based on a list and
 807          * on the last inserted event that is close to the current one is
 808          * probably more efficient than an rbtree based sorting.
 809          */
 810         if (last_inserted->timestamp >= new->timestamp)
 811                 __queue_raw_event_before(new, last_inserted);
 812         else
 813                 __queue_raw_event_after(new, last_inserted);
 814 }
 815
 816 static void queue_raw_event(void *data, int raw_size, int cpu,
 817                             u64 timestamp, struct thread *thread)
 818 {
 819         struct raw_event_queue *new;
 820
 821         if (flush_limit == ULLONG_MAX)
 822                 flush_limit = timestamp + FLUSH_PERIOD;
 823
 824         if (timestamp < last_flush) {
 825                 printf("Warning: Timestamp below last timeslice flush\n");
 826                 return;
 827         }
 828
 829         new = malloc(sizeof(*new));
 830         if (!new)
 831                 die("Not enough memory\n");
 832
 833         new->timestamp = timestamp;
 834         new->cpu = cpu;
 835         new->thread = thread;
 836
 837         new->data = malloc(raw_size);
 838         if (!new->data)
 839                 die("Not enough memory\n");
 840
 841         memcpy(new->data, data, raw_size);
 842
 843         __queue_raw_event(new);
 844         last_inserted = new;
 845
 846         /*
 847          * We want to have a slice of events covering 2 * FLUSH_PERIOD
 848          * If FLUSH_PERIOD is big enough, it ensures every events that occured
 849          * in the first half of the timeslice have all been buffered and there
 850          * are none remaining (we need that because of the weakly ordered
 851          * event recording we have). Then once we reach the 2 * FLUSH_PERIOD
 852          * timeslice, we flush the first half to be gentle with the memory
 853          * (the second half can still get new events in the middle, so wait
 854          * another period to flush it)
 855          */
 856         if (new->timestamp > flush_limit &&
 857                 new->timestamp - flush_limit > FLUSH_PERIOD) {
 858                 flush_limit += FLUSH_PERIOD;
 859                 flush_raw_event_queue(flush_limit);
 860         }
 861 }
 862
 863 static int process_sample_event(event_t *event, struct perf_session *s)
 864 {
 865         struct thread *thread;
 866         struct sample_data data;
 867
 868         bzero(&data, sizeof(struct sample_data));
 869         event__parse_sample(event, s->sample_type, &data);
 870         /* CAUTION: using tid as thread.pid */
 871         thread = perf_session__findnew(s, data.tid);
 872
 873         if (thread == NULL) {
 874                 pr_debug("problem processing %d event, skipping it.\n",
 875                          event->header.type);
 876                 return -1;
 877         }
 878
 879         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 880
 881         if (profile_cpu != -1 && profile_cpu != (int) data.cpu)
 882                 return 0;
 883
 884         queue_raw_event(data.raw_data, data.raw_size, data.cpu, data.time, thread);
 885
 886         return 0;
 887 }
 888
 889 /* TODO: various way to print, coloring, nano or milli sec */
 890 static void print_result(void)
 891 {
 892         struct lock_stat *st;
 893         char cut_name[20];
 894         int bad, total;
 895
 896         printf("%20s ", "Name");
 897         printf("%10s ", "acquired");
 898         printf("%10s ", "contended");
 899
 900         printf("%15s ", "total wait (ns)");
 901         printf("%15s ", "max wait (ns)");
 902         printf("%15s ", "min wait (ns)");
 903
 904         printf("\n\n");
 905
 906         bad = total = 0;
 907         while ((st = pop_from_result())) {
 908                 total++;
 909                 if (st->discard) {
 910                         bad++;
 911                         continue;
 912                 }
 913                 bzero(cut_name, 20);
 914
 915                 if (strlen(st->name) < 16) {
 916                         /* output raw name */
 917                         printf("%20s ", st->name);
 918                 } else {
 919                         strncpy(cut_name, st->name, 16);
 920                         cut_name[16] = '.';
 921                         cut_name[17] = '.';
 922                         cut_name[18] = '.';
 923                         cut_name[19] = '\0';
 924                         /* cut off name for saving output style */
 925                         printf("%20s ", cut_name);
 926                 }
 927
 928                 printf("%10u ", st->nr_acquired);
 929                 printf("%10u ", st->nr_contended);
 930
 931                 printf("%15llu ", st->wait_time_total);
 932                 printf("%15llu ", st->wait_time_max);
 933                 printf("%15llu ", st->wait_time_min == ULLONG_MAX ?
 934                        0 : st->wait_time_min);
 935                 printf("\n");
 936         }
 937
 938         {
 939                 /* Output for debug, this have to be removed */
 940                 int i;
 941                 const char *name[4] =
 942                         { "acquire", "acquired", "contended", "release" };
 943
 944                 printf("\n=== output for debug===\n\n");
 945                 printf("bad:%d, total:%d\n", bad, total);
 946                 printf("bad rate:%f\n", (double)(bad / total));
 947
 948                 printf("histogram of events caused bad sequence\n");
 949                 for (i = 0; i < 4; i++)
 950                         printf(" %10s: %d\n", name[i], bad_hist[i]);
 951         }
 952 }
 953
 954 static void dump_map(void)
 955 {
 956         unsigned int i;
 957         struct lock_stat *st;
 958
 959         for (i = 0; i < LOCKHASH_SIZE; i++) {
 960                 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 961                         printf("%p: %s\n", st->addr, st->name);
 962                 }
 963         }
 964 }
 965
 966 static struct perf_event_ops eops = {
 967         .sample                 = process_sample_event,
 968         .comm                   = event__process_comm,
 969 };
 970
 971 static int read_events(void)
 972 {
 973         session = perf_session__new(input_name, O_RDONLY, 0);
 974         if (!session)
 975                 die("Initializing perf session failed\n");
 976
 977         return perf_session__process_events(session, &eops);
 978 }
 979
 980 static void sort_result(void)
 981 {
 982         unsigned int i;
 983         struct lock_stat *st;
 984
 985         for (i = 0; i < LOCKHASH_SIZE; i++) {
 986                 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 987                         insert_to_result(st, compare);
 988                 }
 989         }
 990 }
 991
 992 static void __cmd_report(void)
 993 {
 994         setup_pager();
 995         select_key();
 996         read_events();
 997         flush_raw_event_queue(ULLONG_MAX);
 998         sort_result();
 999         print_result();
1000 }
1001
1002 static const char * const report_usage[] = {
1003         "perf lock report [<options>]",
1004         NULL
1005 };
1006
1007 static const struct option report_options[] = {
1008         OPT_STRING('k', "key", &sort_key, "acquired",
1009                     "key for sorting"),
1010         /* TODO: type */
1011         OPT_END()
1012 };
1013
1014 static const char * const lock_usage[] = {
1015         "perf lock [<options>] {record|trace|report}",
1016         NULL
1017 };
1018
1019 static const struct option lock_options[] = {
1020         OPT_STRING('i', "input", &input_name, "file", "input file name"),
1021         OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
1022         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
1023         OPT_END()
1024 };
1025
1026 static const char *record_args[] = {
1027         "record",
1028         "-a",
1029         "-R",
1030         "-f",
1031         "-m", "1024",
1032         "-c", "1",
1033         "-e", "lock:lock_acquire:r",
1034         "-e", "lock:lock_acquired:r",
1035         "-e", "lock:lock_contended:r",
1036         "-e", "lock:lock_release:r",
1037 };
1038
1039 static int __cmd_record(int argc, const char **argv)
1040 {
1041         unsigned int rec_argc, i, j;
1042         const char **rec_argv;
1043
1044         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1045         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1046
1047         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1048                 rec_argv[i] = strdup(record_args[i]);
1049
1050         for (j = 1; j < (unsigned int)argc; j++, i++)
1051                 rec_argv[i] = argv[j];
1052
1053         BUG_ON(i != rec_argc);
1054
1055         return cmd_record(i, rec_argv, NULL);
1056 }
1057
1058 int cmd_lock(int argc, const char **argv, const char *prefix __used)
1059 {
1060         unsigned int i;
1061
1062         symbol__init();
1063         for (i = 0; i < LOCKHASH_SIZE; i++)
1064                 INIT_LIST_HEAD(lockhash_table + i);
1065
1066         argc = parse_options(argc, argv, lock_options, lock_usage,
1067                              PARSE_OPT_STOP_AT_NON_OPTION);
1068         if (!argc)
1069                 usage_with_options(lock_usage, lock_options);
1070
1071         if (!strncmp(argv[0], "rec", 3)) {
1072                 return __cmd_record(argc, argv);
1073         } else if (!strncmp(argv[0], "report", 6)) {
1074                 trace_handler = &report_lock_ops;
1075                 if (argc) {
1076                         argc = parse_options(argc, argv,
1077                                              report_options, report_usage, 0);
1078                         if (argc)
1079                                 usage_with_options(report_usage, report_options);
1080                 }
1081                 __cmd_report();
1082         } else if (!strcmp(argv[0], "trace")) {
1083                 /* Aliased to 'perf trace' */
1084                 return cmd_trace(argc, argv, prefix);
1085         } else if (!strcmp(argv[0], "map")) {
1086                 /* recycling report_lock_ops */
1087                 trace_handler = &report_lock_ops;
1088                 setup_pager();
1089                 read_events();
1090                 dump_map();
1091         } else {
1092                 usage_with_options(lock_usage, lock_options);
1093         }
1094
1095         return 0;
1096 }