blkparse.c

   1 /*
   2  * block queue tracing parse application
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; either version 2 of the License, or
   9  *  (at your option) any later version.
  10  *
  11  *  This program is distributed in the hope that it will be useful,
  12  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  *  GNU General Public License for more details.
  15  *
  16  *  You should have received a copy of the GNU General Public License
  17  *  along with this program; if not, write to the Free Software
  18  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  *
  20  */
  21 #include <sys/types.h>
  22 #include <sys/stat.h>
  23 #include <unistd.h>
  24 #include <stdio.h>
  25 #include <fcntl.h>
  26 #include <stdlib.h>
  27 #include <string.h>
  28 #include <getopt.h>
  29 #include <errno.h>
  30 #include <signal.h>
  31 #include <locale.h>
  32 #include <limits.h>
  33
  34 #include "blktrace.h"
  35 #include "rbtree.h"
  36 #include "jhash.h"
  37
  38 static char blkparse_version[] = "0.90";
  39
  40 struct per_dev_info {
  41         dev_t dev;
  42         char *name;
  43
  44         int backwards;
  45         unsigned long long events;
  46         unsigned long long first_reported_time;
  47         unsigned long long last_reported_time;
  48         unsigned long long last_read_time;
  49         struct io_stats io_stats;
  50         unsigned long last_sequence;
  51         unsigned long skips;
  52
  53         struct rb_root rb_last;
  54         unsigned long rb_last_entries;
  55
  56         struct rb_root rb_track;
  57
  58         int nfiles;
  59         int ncpus;
  60
  61         unsigned long *cpu_map;
  62         unsigned int cpu_map_max;
  63
  64         struct per_cpu_info *cpus;
  65 };
  66
  67 struct per_process_info {
  68         char name[16];
  69         __u32 pid;
  70         struct io_stats io_stats;
  71         struct per_process_info *hash_next, *list_next;
  72         int more_than_one;
  73
  74         /*
  75          * individual io stats
  76          */
  77         unsigned long long longest_allocation_wait[2];
  78         unsigned long long longest_dispatch_wait[2];
  79         unsigned long long longest_completion_wait[2];
  80 };
  81
  82 #define PPI_HASH_SHIFT  (8)
  83 #define PPI_HASH_SIZE   (1 << PPI_HASH_SHIFT)
  84 #define PPI_HASH_MASK   (PPI_HASH_SIZE - 1)
  85 static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
  86 static struct per_process_info *ppi_list;
  87 static int ppi_list_entries;
  88
  89 #define S_OPTS  "a:A:i:o:b:stqw:f:F:vVhD:"
  90 static struct option l_opts[] = {
  91         {
  92                 .name = "act-mask",
  93                 .has_arg = required_argument,
  94                 .flag = NULL,
  95                 .val = 'a'
  96         },
  97         {
  98                 .name = "set-mask",
  99                 .has_arg = required_argument,
 100                 .flag = NULL,
 101                 .val = 'A'
 102         },
 103         {
 104                 .name = "input",
 105                 .has_arg = required_argument,
 106                 .flag = NULL,
 107                 .val = 'i'
 108         },
 109         {
 110                 .name = "output",
 111                 .has_arg = required_argument,
 112                 .flag = NULL,
 113                 .val = 'o'
 114         },
 115         {
 116                 .name = "batch",
 117                 .has_arg = required_argument,
 118                 .flag = NULL,
 119                 .val = 'b'
 120         },
 121         {
 122                 .name = "per-program-stats",
 123                 .has_arg = no_argument,
 124                 .flag = NULL,
 125                 .val = 's'
 126         },
 127         {
 128                 .name = "track-ios",
 129                 .has_arg = no_argument,
 130                 .flag = NULL,
 131                 .val = 't'
 132         },
 133         {
 134                 .name = "quiet",
 135                 .has_arg = no_argument,
 136                 .flag = NULL,
 137                 .val = 'q'
 138         },
 139         {
 140                 .name = "stopwatch",
 141                 .has_arg = required_argument,
 142                 .flag = NULL,
 143                 .val = 'w'
 144         },
 145         {
 146                 .name = "format",
 147                 .has_arg = required_argument,
 148                 .flag = NULL,
 149                 .val = 'f'
 150         },
 151         {
 152                 .name = "format-spec",
 153                 .has_arg = required_argument,
 154                 .flag = NULL,
 155                 .val = 'F'
 156         },
 157         {
 158                 .name = "hash-by-name",
 159                 .has_arg = no_argument,
 160                 .flag = NULL,
 161                 .val = 'h'
 162         },
 163         {
 164                 .name = "verbose",
 165                 .has_arg = no_argument,
 166                 .flag = NULL,
 167                 .val = 'v'
 168         },
 169         {
 170                 .name = "version",
 171                 .has_arg = no_argument,
 172                 .flag = NULL,
 173                 .val = 'V'
 174         },
 175         {
 176                 .name = "input-directory",
 177                 .has_arg = required_argument,
 178                 .flag = NULL,
 179                 .val = 'D'
 180         },
 181         {
 182                 .name = NULL,
 183         }
 184 };
 185
 186 /*
 187  * for sorting the displayed output
 188  */
 189 struct trace {
 190         struct blk_io_trace *bit;
 191         struct rb_node rb_node;
 192         struct trace *next;
 193 };
 194
 195 static struct rb_root rb_sort_root;
 196 static unsigned long rb_sort_entries;
 197
 198 static struct trace *trace_list;
 199
 200 /*
 201  * allocation cache
 202  */
 203 static struct blk_io_trace *bit_alloc_list;
 204 static struct trace *t_alloc_list;
 205
 206 /*
 207  * for tracking individual ios
 208  */
 209 struct io_track {
 210         struct rb_node rb_node;
 211
 212         __u64 sector;
 213         __u32 pid;
 214         char comm[16];
 215         unsigned long long allocation_time;
 216         unsigned long long queue_time;
 217         unsigned long long dispatch_time;
 218         unsigned long long completion_time;
 219 };
 220
 221 static int ndevices;
 222 static struct per_dev_info *devices;
 223 static char *get_dev_name(struct per_dev_info *, char *, int);
 224
 225 FILE *ofp = NULL;
 226 static char *output_name;
 227 static char *input_dir;
 228
 229 static unsigned long long genesis_time;
 230 static unsigned long long last_allowed_time;
 231 static unsigned int smallest_seq_read;
 232 static unsigned long long stopwatch_start;      /* start from zero by default */
 233 static unsigned long long stopwatch_end = ULONG_LONG_MAX;       /* "infinity" */
 234
 235 static int per_process_stats;
 236 static int per_device_and_cpu_stats = 1;
 237 static int track_ios;
 238 static int ppi_hash_by_pid = 1;
 239 static int verbose;
 240 static unsigned int act_mask = -1U;
 241 static int stats_printed;
 242
 243 static unsigned int t_alloc_cache;
 244 static unsigned int bit_alloc_cache;
 245
 246 #define RB_BATCH_DEFAULT        (512)
 247 static unsigned int rb_batch = RB_BATCH_DEFAULT;
 248
 249 static int pipeline;
 250
 251 #define is_done()       (*(volatile int *)(&done))
 252 static volatile int done;
 253
 254 #define JHASH_RANDOM    (0x3af5f2ee)
 255
 256 #define CPUS_PER_LONG   (8 * sizeof(unsigned long))
 257 #define CPU_IDX(cpu)    ((cpu) / CPUS_PER_LONG)
 258 #define CPU_BIT(cpu)    ((cpu) & (CPUS_PER_LONG - 1))
 259
 260 static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu)
 261 {
 262         if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) {
 263                 int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1);
 264                 unsigned long *map = malloc(new_max / sizeof(long));
 265
 266                 memset(map, 0, new_max / sizeof(long));
 267
 268                 if (pdi->cpu_map) {
 269                         memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long));
 270                         free(pdi->cpu_map);
 271                 }
 272
 273                 pdi->cpu_map = map;
 274                 pdi->cpu_map_max = new_max;
 275         }
 276
 277         pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
 278 }
 279
 280 static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu)
 281 {
 282         pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu));
 283 }
 284
 285 static inline int cpu_is_online(struct per_dev_info *pdi, int cpu)
 286 {
 287         return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0;
 288 }
 289
 290 static inline int ppi_hash_pid(__u32 pid)
 291 {
 292         return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
 293 }
 294
 295 static inline int ppi_hash_name(const char *name)
 296 {
 297         return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
 298 }
 299
 300 static inline int ppi_hash(struct per_process_info *ppi)
 301 {
 302         if (ppi_hash_by_pid)
 303                 return ppi_hash_pid(ppi->pid);
 304
 305         return ppi_hash_name(ppi->name);
 306 }
 307
 308 static inline void add_process_to_hash(struct per_process_info *ppi)
 309 {
 310         const int hash_idx = ppi_hash(ppi);
 311
 312         ppi->hash_next = ppi_hash_table[hash_idx];
 313         ppi_hash_table[hash_idx] = ppi;
 314 }
 315
 316 static inline void add_process_to_list(struct per_process_info *ppi)
 317 {
 318         ppi->list_next = ppi_list;
 319         ppi_list = ppi;
 320         ppi_list_entries++;
 321 }
 322
 323 static struct per_process_info *find_process_by_name(char *name)
 324 {
 325         const int hash_idx = ppi_hash_name(name);
 326         struct per_process_info *ppi;
 327
 328         ppi = ppi_hash_table[hash_idx];
 329         while (ppi) {
 330                 if (!strcmp(ppi->name, name))
 331                         return ppi;
 332
 333                 ppi = ppi->hash_next;
 334         }
 335
 336         return NULL;
 337 }
 338
 339 static struct per_process_info *find_process_by_pid(__u32 pid)
 340 {
 341         const int hash_idx = ppi_hash_pid(pid);
 342         struct per_process_info *ppi;
 343
 344         ppi = ppi_hash_table[hash_idx];
 345         while (ppi) {
 346                 if (ppi->pid == pid)
 347                         return ppi;
 348
 349                 ppi = ppi->hash_next;
 350         }
 351
 352         return NULL;
 353 }
 354
 355 static struct per_process_info *find_process(__u32 pid, char *name)
 356 {
 357         struct per_process_info *ppi;
 358
 359         if (ppi_hash_by_pid)
 360                 return find_process_by_pid(pid);
 361
 362         ppi = find_process_by_name(name);
 363         if (ppi && ppi->pid != pid)
 364                 ppi->more_than_one = 1;
 365
 366         return ppi;
 367 }
 368
 369 static inline int trace_rb_insert(struct trace *t, struct rb_root *root,
 370                                   int check_time)
 371 {
 372         struct rb_node **p = &root->rb_node;
 373         struct rb_node *parent = NULL;
 374         struct trace *__t;
 375
 376         while (*p) {
 377                 parent = *p;
 378
 379                 __t = rb_entry(parent, struct trace, rb_node);
 380
 381                 if (check_time) {
 382                         if (t->bit->time < __t->bit->time) {
 383                                 p = &(*p)->rb_left;
 384                                 continue;
 385                         } else if (t->bit->time > __t->bit->time) {
 386                                 p = &(*p)->rb_right;
 387                                 continue;
 388                         }
 389                 }
 390                 if (t->bit->device < __t->bit->device)
 391                         p = &(*p)->rb_left;
 392                 else if (t->bit->device > __t->bit->device)
 393                         p = &(*p)->rb_right;
 394                 else if (t->bit->sequence < __t->bit->sequence)
 395                         p = &(*p)->rb_left;
 396                 else    /* >= sequence */
 397                         p = &(*p)->rb_right;
 398         }
 399
 400         rb_link_node(&t->rb_node, parent, p);
 401         rb_insert_color(&t->rb_node, root);
 402         return 0;
 403 }
 404
 405 static inline int trace_rb_insert_sort(struct trace *t)
 406 {
 407         if (!trace_rb_insert(t, &rb_sort_root, 1)) {
 408                 rb_sort_entries++;
 409                 return 0;
 410         }
 411
 412         return 1;
 413 }
 414
 415 static inline int trace_rb_insert_last(struct per_dev_info *pdi,struct trace *t)
 416 {
 417         if (!trace_rb_insert(t, &pdi->rb_last, 1)) {
 418                 pdi->rb_last_entries++;
 419                 return 0;
 420         }
 421
 422         return 1;
 423 }
 424
 425 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
 426                                    struct rb_root *root, int order)
 427 {
 428         struct rb_node *n = root->rb_node;
 429         struct rb_node *prev = NULL;
 430         struct trace *__t;
 431
 432         while (n) {
 433                 __t = rb_entry(n, struct trace, rb_node);
 434                 prev = n;
 435
 436                 if (device < __t->bit->device)
 437                         n = n->rb_left;
 438                 else if (device > __t->bit->device)
 439                         n = n->rb_right;
 440                 else if (sequence < __t->bit->sequence)
 441                         n = n->rb_left;
 442                 else if (sequence > __t->bit->sequence)
 443                         n = n->rb_right;
 444                 else
 445                         return __t;
 446         }
 447
 448         /*
 449          * hack - the list may not be sequence ordered because some
 450          * events don't have sequence and time matched. so we end up
 451          * being a little off in the rb lookup here, because we don't
 452          * know the time we are looking for. compensate by browsing
 453          * a little ahead from the last entry to find the match
 454          */
 455         if (order && prev) {
 456                 int max = 5;
 457
 458                 while (((n = rb_next(prev)) != NULL) && max--) {
 459                         __t = rb_entry(n, struct trace, rb_node);
 460
 461                         if (__t->bit->device == device &&
 462                             __t->bit->sequence == sequence)
 463                                 return __t;
 464
 465                         prev = n;
 466                 }
 467         }
 468
 469         return NULL;
 470 }
 471
 472 static inline struct trace *trace_rb_find_sort(dev_t dev, unsigned long seq)
 473 {
 474         return trace_rb_find(dev, seq, &rb_sort_root, 1);
 475 }
 476
 477 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
 478                                                unsigned long seq)
 479 {
 480         return trace_rb_find(pdi->dev, seq, &pdi->rb_last, 0);
 481 }
 482
 483 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
 484 {
 485         struct rb_node **p = &pdi->rb_track.rb_node;
 486         struct rb_node *parent = NULL;
 487         struct io_track *__iot;
 488
 489         while (*p) {
 490                 parent = *p;
 491                 __iot = rb_entry(parent, struct io_track, rb_node);
 492
 493                 if (iot->sector < __iot->sector)
 494                         p = &(*p)->rb_left;
 495                 else if (iot->sector > __iot->sector)
 496                         p = &(*p)->rb_right;
 497                 else {
 498                         fprintf(stderr,
 499                                 "sector alias (%Lu) on device %d,%d!\n",
 500                                 (unsigned long long) iot->sector,
 501                                 MAJOR(pdi->dev), MINOR(pdi->dev));
 502                         return 1;
 503                 }
 504         }
 505
 506         rb_link_node(&iot->rb_node, parent, p);
 507         rb_insert_color(&iot->rb_node, &pdi->rb_track);
 508         return 0;
 509 }
 510
 511 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
 512 {
 513         struct rb_node *n = pdi->rb_track.rb_node;
 514         struct io_track *__iot;
 515
 516         while (n) {
 517                 __iot = rb_entry(n, struct io_track, rb_node);
 518
 519                 if (sector < __iot->sector)
 520                         n = n->rb_left;
 521                 else if (sector > __iot->sector)
 522                         n = n->rb_right;
 523                 else
 524                         return __iot;
 525         }
 526
 527         return NULL;
 528 }
 529
 530 static struct io_track *find_track(struct per_dev_info *pdi, __u32 pid,
 531                                    char *comm, __u64 sector)
 532 {
 533         struct io_track *iot;
 534
 535         iot = __find_track(pdi, sector);
 536         if (!iot) {
 537                 iot = malloc(sizeof(*iot));
 538                 iot->pid = pid;
 539                 memcpy(iot->comm, comm, sizeof(iot->comm));
 540                 iot->sector = sector;
 541                 track_rb_insert(pdi, iot);
 542         }
 543
 544         return iot;
 545 }
 546
 547 static void log_track_frontmerge(struct per_dev_info *pdi,
 548                                  struct blk_io_trace *t)
 549 {
 550         struct io_track *iot;
 551
 552         if (!track_ios)
 553                 return;
 554
 555         iot = __find_track(pdi, t->sector + t_sec(t));
 556         if (!iot) {
 557                 if (verbose)
 558                         fprintf(stderr, "merge not found for (%d,%d): %llu\n",
 559                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 560                                 (unsigned long long) t->sector + t_sec(t));
 561                 return;
 562         }
 563
 564         rb_erase(&iot->rb_node, &pdi->rb_track);
 565         iot->sector -= t_sec(t);
 566         track_rb_insert(pdi, iot);
 567 }
 568
 569 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
 570 {
 571         struct io_track *iot;
 572
 573         if (!track_ios)
 574                 return;
 575
 576         iot = find_track(pdi, t->pid, t->comm, t->sector);
 577         iot->allocation_time = t->time;
 578 }
 579
 580 /*
 581  * return time between rq allocation and insertion
 582  */
 583 static unsigned long long log_track_insert(struct per_dev_info *pdi,
 584                                            struct blk_io_trace *t)
 585 {
 586         unsigned long long elapsed;
 587         struct io_track *iot;
 588
 589         if (!track_ios)
 590                 return -1;
 591
 592         iot = find_track(pdi, t->pid, t->comm, t->sector);
 593         iot->queue_time = t->time;
 594
 595         if (!iot->allocation_time)
 596                 return -1;
 597
 598         elapsed = iot->queue_time - iot->allocation_time;
 599
 600         if (per_process_stats) {
 601                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 602                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 603
 604                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
 605                         ppi->longest_allocation_wait[w] = elapsed;
 606         }
 607
 608         return elapsed;
 609 }
 610
 611 /*
 612  * return time between queue and issue
 613  */
 614 static unsigned long long log_track_issue(struct per_dev_info *pdi,
 615                                           struct blk_io_trace *t)
 616 {
 617         unsigned long long elapsed;
 618         struct io_track *iot;
 619
 620         if (!track_ios)
 621                 return -1;
 622         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 623                 return -1;
 624
 625         iot = __find_track(pdi, t->sector);
 626         if (!iot) {
 627                 if (verbose)
 628                         fprintf(stderr, "issue not found for (%d,%d): %llu\n",
 629                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 630                                 (unsigned long long) t->sector);
 631                 return -1;
 632         }
 633
 634         iot->dispatch_time = t->time;
 635         elapsed = iot->dispatch_time - iot->queue_time;
 636
 637         if (per_process_stats) {
 638                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 639                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 640
 641                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
 642                         ppi->longest_dispatch_wait[w] = elapsed;
 643         }
 644
 645         return elapsed;
 646 }
 647
 648 /*
 649  * return time between dispatch and complete
 650  */
 651 static unsigned long long log_track_complete(struct per_dev_info *pdi,
 652                                              struct blk_io_trace *t)
 653 {
 654         unsigned long long elapsed;
 655         struct io_track *iot;
 656
 657         if (!track_ios)
 658                 return -1;
 659         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 660                 return -1;
 661
 662         iot = __find_track(pdi, t->sector);
 663         if (!iot) {
 664                 if (verbose)
 665                         fprintf(stderr,"complete not found for (%d,%d): %llu\n",
 666                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 667                                 (unsigned long long) t->sector);
 668                 return -1;
 669         }
 670
 671         iot->completion_time = t->time;
 672         elapsed = iot->completion_time - iot->dispatch_time;
 673
 674         if (per_process_stats) {
 675                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 676                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 677
 678                 if (ppi && elapsed > ppi->longest_completion_wait[w])
 679                         ppi->longest_completion_wait[w] = elapsed;
 680         }
 681
 682         /*
 683          * kill the trace, we don't need it after completion
 684          */
 685         rb_erase(&iot->rb_node, &pdi->rb_track);
 686         free(iot);
 687
 688         return elapsed;
 689 }
 690
 691
 692 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
 693 {
 694         struct per_process_info *ppi = find_process(pid, name);
 695
 696         if (!ppi) {
 697                 ppi = malloc(sizeof(*ppi));
 698                 memset(ppi, 0, sizeof(*ppi));
 699                 memcpy(ppi->name, name, 16);
 700                 ppi->pid = pid;
 701                 add_process_to_hash(ppi);
 702                 add_process_to_list(ppi);
 703         }
 704
 705         return &ppi->io_stats;
 706 }
 707
 708 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
 709 {
 710         struct per_cpu_info *cpus = pdi->cpus;
 711         int ncpus = pdi->ncpus;
 712         int new_count = cpu + 1;
 713         int new_space, size;
 714         char *new_start;
 715
 716         size = new_count * sizeof(struct per_cpu_info);
 717         cpus = realloc(cpus, size);
 718         if (!cpus) {
 719                 char name[20];
 720                 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
 721                         get_dev_name(pdi, name, sizeof(name)), size);
 722                 exit(1);
 723         }
 724
 725         new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
 726         new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
 727         memset(new_start, 0, new_space);
 728
 729         pdi->ncpus = new_count;
 730         pdi->cpus = cpus;
 731 }
 732
 733 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
 734 {
 735         struct per_cpu_info *pci;
 736
 737         if (cpu >= pdi->ncpus)
 738                 resize_cpu_info(pdi, cpu);
 739
 740         pci = &pdi->cpus[cpu];
 741         pci->cpu = cpu;
 742         return pci;
 743 }
 744
 745
 746 static int resize_devices(char *name)
 747 {
 748         int size = (ndevices + 1) * sizeof(struct per_dev_info);
 749
 750         devices = realloc(devices, size);
 751         if (!devices) {
 752                 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
 753                 return 1;
 754         }
 755         memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
 756         devices[ndevices].name = name;
 757         ndevices++;
 758         return 0;
 759 }
 760
 761 static struct per_dev_info *get_dev_info(dev_t dev)
 762 {
 763         struct per_dev_info *pdi;
 764         int i;
 765
 766         for (i = 0; i < ndevices; i++) {
 767                 if (!devices[i].dev)
 768                         devices[i].dev = dev;
 769                 if (devices[i].dev == dev)
 770                         return &devices[i];
 771         }
 772
 773         if (resize_devices(NULL))
 774                 return NULL;
 775
 776         pdi = &devices[ndevices - 1];
 777         pdi->dev = dev;
 778         pdi->first_reported_time = 0;
 779         pdi->last_sequence = -1;
 780         pdi->last_read_time = 0;
 781         memset(&pdi->rb_last, 0, sizeof(pdi->rb_last));
 782         pdi->rb_last_entries = 0;
 783         return pdi;
 784 }
 785
 786 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
 787 {
 788         if (pdi->name)
 789                 snprintf(buffer, size, "%s", pdi->name);
 790         else
 791                 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
 792         return buffer;
 793 }
 794
 795 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
 796 {
 797         unsigned long long this = bit->time;
 798         unsigned long long last = pdi->last_reported_time;
 799
 800         pdi->backwards = (this < last) ? 'B' : ' ';
 801         pdi->last_reported_time = this;
 802 }
 803
 804 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
 805                                int rw)
 806 {
 807         if (rw) {
 808                 ios->mwrites++;
 809                 ios->qwrite_kb += t_kb(t);
 810         } else {
 811                 ios->mreads++;
 812                 ios->qread_kb += t_kb(t);
 813         }
 814 }
 815
 816 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
 817                              int rw)
 818 {
 819         __account_m(&pci->io_stats, t, rw);
 820
 821         if (per_process_stats) {
 822                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 823
 824                 __account_m(ios, t, rw);
 825         }
 826 }
 827
 828 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
 829                                    int rw)
 830 {
 831         if (rw) {
 832                 ios->qwrites++;
 833                 ios->qwrite_kb += t_kb(t);
 834         } else {
 835                 ios->qreads++;
 836                 ios->qread_kb += t_kb(t);
 837         }
 838 }
 839
 840 static inline void account_queue(struct blk_io_trace *t,
 841                                  struct per_cpu_info *pci, int rw)
 842 {
 843         __account_queue(&pci->io_stats, t, rw);
 844
 845         if (per_process_stats) {
 846                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 847
 848                 __account_queue(ios, t, rw);
 849         }
 850 }
 851
 852 static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
 853 {
 854         if (rw) {
 855                 ios->cwrites++;
 856                 ios->cwrite_kb += bytes >> 10;
 857         } else {
 858                 ios->creads++;
 859                 ios->cread_kb += bytes >> 10;
 860         }
 861 }
 862
 863 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
 864                              int rw, int bytes)
 865 {
 866         __account_c(&pci->io_stats, rw, bytes);
 867
 868         if (per_process_stats) {
 869                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 870
 871                 __account_c(ios, rw, bytes);
 872         }
 873 }
 874
 875 static inline void __account_issue(struct io_stats *ios, int rw,
 876                                    unsigned int bytes)
 877 {
 878         if (rw) {
 879                 ios->iwrites++;
 880                 ios->iwrite_kb += bytes >> 10;
 881         } else {
 882                 ios->ireads++;
 883                 ios->iread_kb += bytes >> 10;
 884         }
 885 }
 886
 887 static inline void account_issue(struct blk_io_trace *t,
 888                                  struct per_cpu_info *pci, int rw)
 889 {
 890         __account_issue(&pci->io_stats, rw, t->bytes);
 891
 892         if (per_process_stats) {
 893                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 894
 895                 __account_issue(ios, rw, t->bytes);
 896         }
 897 }
 898
 899 static inline void __account_unplug(struct io_stats *ios, int timer)
 900 {
 901         if (timer)
 902                 ios->timer_unplugs++;
 903         else
 904                 ios->io_unplugs++;
 905 }
 906
 907 static inline void account_unplug(struct blk_io_trace *t,
 908                                   struct per_cpu_info *pci, int timer)
 909 {
 910         __account_unplug(&pci->io_stats, timer);
 911
 912         if (per_process_stats) {
 913                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 914
 915                 __account_unplug(ios, timer);
 916         }
 917 }
 918
 919 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
 920                          struct blk_io_trace *t, char *act)
 921 {
 922         process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
 923 }
 924
 925 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
 926                        struct blk_io_trace *t, char *act)
 927 {
 928         process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
 929 }
 930
 931 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
 932                       char *act)
 933 {
 934         process_fmt(act, pci, t, -1, 0, NULL);
 935 }
 936
 937 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
 938                       struct blk_io_trace *t, char *act)
 939 {
 940         process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
 941 }
 942
 943 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
 944                       struct blk_io_trace *t, char *act)
 945 {
 946         if (act[0] == 'F')
 947                 log_track_frontmerge(pdi, t);
 948
 949         process_fmt(act, pci, t, -1ULL, 0, NULL);
 950 }
 951
 952 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
 953                         char *act)
 954 {
 955         process_fmt(act, pci, t, -1ULL, 0, NULL);
 956 }
 957
 958 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
 959                         char *act)
 960 {
 961         process_fmt(act, pci, t, -1ULL, 0, NULL);
 962 }
 963
 964 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
 965                       char *act)
 966 {
 967         process_fmt(act, pci, t, -1ULL, 0, NULL);
 968 }
 969
 970 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
 971                       char *act)
 972 {
 973         process_fmt(act, pci, t, -1ULL, 0, NULL);
 974 }
 975
 976 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
 977 {
 978         unsigned char *buf = (unsigned char *) t + sizeof(*t);
 979
 980         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
 981 }
 982
 983 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
 984 {
 985         int act = t->action & 0xffff;
 986
 987         switch (act) {
 988                 case __BLK_TA_QUEUE:
 989                         log_generic(pci, t, "Q");
 990                         break;
 991                 case __BLK_TA_GETRQ:
 992                         log_generic(pci, t, "G");
 993                         break;
 994                 case __BLK_TA_SLEEPRQ:
 995                         log_generic(pci, t, "S");
 996                         break;
 997                 case __BLK_TA_REQUEUE:
 998                         log_generic(pci, t, "R");
 999                         break;
1000                 case __BLK_TA_ISSUE:
1001                         log_pc(pci, t, "D");
1002                         break;
1003                 case __BLK_TA_COMPLETE:
1004                         log_pc(pci, t, "C");
1005                         break;
1006                 case __BLK_TA_INSERT:
1007                         log_pc(pci, t, "I");
1008                         break;
1009                 default:
1010                         fprintf(stderr, "Bad pc action %x\n", act);
1011                         break;
1012         }
1013 }
1014
1015 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1016                           struct per_cpu_info *pci)
1017 {
1018         int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
1019         int act = t->action & 0xffff;
1020
1021         switch (act) {
1022                 case __BLK_TA_QUEUE:
1023                         account_queue(t, pci, w);
1024                         log_queue(pci, t, "Q");
1025                         break;
1026                 case __BLK_TA_INSERT:
1027                         log_insert(pdi, pci, t, "I");
1028                         break;
1029                 case __BLK_TA_BACKMERGE:
1030                         account_m(t, pci, w);
1031                         log_merge(pdi, pci, t, "M");
1032                         break;
1033                 case __BLK_TA_FRONTMERGE:
1034                         account_m(t, pci, w);
1035                         log_merge(pdi, pci, t, "F");
1036                         break;
1037                 case __BLK_TA_GETRQ:
1038                         log_track_getrq(pdi, t);
1039                         log_generic(pci, t, "G");
1040                         break;
1041                 case __BLK_TA_SLEEPRQ:
1042                         log_generic(pci, t, "S");
1043                         break;
1044                 case __BLK_TA_REQUEUE:
1045                         account_c(t, pci, w, -t->bytes);
1046                         log_queue(pci, t, "R");
1047                         break;
1048                 case __BLK_TA_ISSUE:
1049                         account_issue(t, pci, w);
1050                         log_issue(pdi, pci, t, "D");
1051                         break;
1052                 case __BLK_TA_COMPLETE:
1053                         account_c(t, pci, w, t->bytes);
1054                         log_complete(pdi, pci, t, "C");
1055                         break;
1056                 case __BLK_TA_PLUG:
1057                         log_action(pci, t, "P");
1058                         break;
1059                 case __BLK_TA_UNPLUG_IO:
1060                         account_unplug(t, pci, 0);
1061                         log_unplug(pci, t, "U");
1062                         break;
1063                 case __BLK_TA_UNPLUG_TIMER:
1064                         account_unplug(t, pci, 1);
1065                         log_unplug(pci, t, "UT");
1066                         break;
1067                 case __BLK_TA_SPLIT:
1068                         log_split(pci, t, "X");
1069                         break;
1070                 case __BLK_TA_BOUNCE:
1071                         log_generic(pci, t, "B");
1072                         break;
1073                 case __BLK_TA_REMAP:
1074                         log_generic(pci, t, "A");
1075                         break;
1076                 default:
1077                         fprintf(stderr, "Bad fs action %x\n", t->action);
1078                         break;
1079         }
1080 }
1081
1082 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1083                        struct per_dev_info *pdi)
1084 {
1085         if (t->action & BLK_TC_ACT(BLK_TC_PC))
1086                 dump_trace_pc(t, pci);
1087         else
1088                 dump_trace_fs(t, pdi, pci);
1089
1090         if (!pdi->events)
1091                 pdi->first_reported_time = t->time;
1092
1093         pdi->events++;
1094 }
1095
1096 /*
1097  * print in a proper way, not too small and not too big. if more than
1098  * 1000,000K, turn into M and so on
1099  */
1100 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1101 {
1102         static char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1103         unsigned int i = 0;
1104
1105         if (in_kb)
1106                 i++;
1107
1108         while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1109                 i++;
1110                 num /= 1000;
1111         }
1112
1113         sprintf(dst, "%'8Lu%c", num, suff[i]);
1114         return dst;
1115 }
1116
1117 static void dump_io_stats(struct io_stats *ios, char *msg)
1118 {
1119         static char x[256], y[256];
1120
1121         fprintf(ofp, "%s\n", msg);
1122
1123         fprintf(ofp, " Reads Queued:    %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1124         fprintf(ofp, " Writes Queued:    %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1125
1126         fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1127         fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1128         fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1129         fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1130         fprintf(ofp, " Read Merges:     %'8lu%8c\t", ios->mreads, ' ');
1131         fprintf(ofp, " Write Merges:     %'8lu\n", ios->mwrites);
1132         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
1133         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
1134 }
1135
1136 static void dump_wait_stats(struct per_process_info *ppi)
1137 {
1138         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1139         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1140         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1141         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1142         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1143         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1144
1145         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1146         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
1147         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
1148         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
1149         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1150         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
1151 }
1152
1153 static int ppi_name_compare(const void *p1, const void *p2)
1154 {
1155         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1156         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1157         int res;
1158
1159         res = strverscmp(ppi1->name, ppi2->name);
1160         if (!res)
1161                 res = ppi1->pid > ppi2->pid;
1162
1163         return res;
1164 }
1165
1166 static void sort_process_list(void)
1167 {
1168         struct per_process_info **ppis;
1169         struct per_process_info *ppi;
1170         int i = 0;
1171
1172         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1173
1174         ppi = ppi_list;
1175         while (ppi) {
1176                 ppis[i++] = ppi;
1177                 ppi = ppi->list_next;
1178         }
1179
1180         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1181
1182         i = ppi_list_entries - 1;
1183         ppi_list = NULL;
1184         while (i >= 0) {
1185                 ppi = ppis[i];
1186
1187                 ppi->list_next = ppi_list;
1188                 ppi_list = ppi;
1189                 i--;
1190         }
1191
1192         free(ppis);
1193 }
1194
1195 static void show_process_stats(void)
1196 {
1197         struct per_process_info *ppi;
1198
1199         sort_process_list();
1200
1201         ppi = ppi_list;
1202         while (ppi) {
1203                 char name[64];
1204
1205                 if (ppi->more_than_one)
1206                         sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid);
1207                 else
1208                         sprintf(name, "%s (%u)", ppi->name, ppi->pid);
1209
1210                 dump_io_stats(&ppi->io_stats, name);
1211                 dump_wait_stats(ppi);
1212                 ppi = ppi->list_next;
1213         }
1214
1215         fprintf(ofp, "\n");
1216 }
1217
1218 static void show_device_and_cpu_stats(void)
1219 {
1220         struct per_dev_info *pdi;
1221         struct per_cpu_info *pci;
1222         struct io_stats total, *ios;
1223         unsigned long long rrate, wrate, msec;
1224         int i, j, pci_events;
1225         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1226         char name[32];
1227
1228         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1229
1230                 memset(&total, 0, sizeof(total));
1231                 pci_events = 0;
1232
1233                 if (i > 0)
1234                         fprintf(ofp, "\n");
1235
1236                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1237                         if (!pci->nelems)
1238                                 continue;
1239
1240                         ios = &pci->io_stats;
1241                         total.qreads += ios->qreads;
1242                         total.qwrites += ios->qwrites;
1243                         total.creads += ios->creads;
1244                         total.cwrites += ios->cwrites;
1245                         total.mreads += ios->mreads;
1246                         total.mwrites += ios->mwrites;
1247                         total.ireads += ios->ireads;
1248                         total.iwrites += ios->iwrites;
1249                         total.qread_kb += ios->qread_kb;
1250                         total.qwrite_kb += ios->qwrite_kb;
1251                         total.cread_kb += ios->cread_kb;
1252                         total.cwrite_kb += ios->cwrite_kb;
1253                         total.iread_kb += ios->iread_kb;
1254                         total.iwrite_kb += ios->iwrite_kb;
1255                         total.timer_unplugs += ios->timer_unplugs;
1256                         total.io_unplugs += ios->io_unplugs;
1257
1258                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1259                                  j, get_dev_name(pdi, name, sizeof(name)));
1260                         dump_io_stats(ios, line);
1261                         pci_events++;
1262                 }
1263
1264                 if (pci_events > 1) {
1265                         fprintf(ofp, "\n");
1266                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1267                                  get_dev_name(pdi, name, sizeof(name)));
1268                         dump_io_stats(&total, line);
1269                 }
1270
1271                 wrate = rrate = 0;
1272                 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1273                 if (msec) {
1274                         rrate = 1000 * total.cread_kb / msec;
1275                         wrate = 1000 * total.cwrite_kb / msec;
1276                 }
1277
1278                 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n", rrate, wrate);
1279                 fprintf(ofp, "Events (%s): %'Lu entries, %'lu skips\n",
1280                         get_dev_name(pdi, line, sizeof(line)), pdi->events,
1281                         pdi->skips);
1282         }
1283 }
1284
1285 /*
1286  * struct trace and blktrace allocation cache, we do potentially
1287  * millions of mallocs for these structures while only using at most
1288  * a few thousand at the time
1289  */
1290 static inline void t_free(struct trace *t)
1291 {
1292         if (t_alloc_cache < 1024) {
1293                 t->next = t_alloc_list;
1294                 t_alloc_list = t;
1295                 t_alloc_cache++;
1296         } else
1297                 free(t);
1298 }
1299
1300 static inline struct trace *t_alloc(void)
1301 {
1302         struct trace *t = t_alloc_list;
1303
1304         if (t) {
1305                 t_alloc_list = t->next;
1306                 t_alloc_cache--;
1307                 return t;
1308         }
1309
1310         return malloc(sizeof(*t));
1311 }
1312
1313 static inline void bit_free(struct blk_io_trace *bit)
1314 {
1315         if (bit_alloc_cache < 1024) {
1316                 /*
1317                  * abuse a 64-bit field for a next pointer for the free item
1318                  */
1319                 bit->time = (__u64) (unsigned long) bit_alloc_list;
1320                 bit_alloc_list = (struct blk_io_trace *) bit;
1321                 bit_alloc_cache++;
1322         } else
1323                 free(bit);
1324 }
1325
1326 static inline struct blk_io_trace *bit_alloc(void)
1327 {
1328         struct blk_io_trace *bit = bit_alloc_list;
1329
1330         if (bit) {
1331                 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
1332                                  bit->time;
1333                 bit_alloc_cache--;
1334                 return bit;
1335         }
1336
1337         return malloc(sizeof(*bit));
1338 }
1339
1340 static void find_genesis(void)
1341 {
1342         struct trace *t = trace_list;
1343
1344         genesis_time = -1ULL;
1345         while (t != NULL) {
1346                 if (t->bit->time < genesis_time)
1347                         genesis_time = t->bit->time;
1348
1349                 t = t->next;
1350         }
1351 }
1352
1353 static inline int check_stopwatch(struct blk_io_trace *bit)
1354 {
1355         if (bit->time < stopwatch_end &&
1356             bit->time >= stopwatch_start)
1357                 return 0;
1358
1359         return 1;
1360 }
1361
1362 /*
1363  * return youngest entry read
1364  */
1365 static int sort_entries(unsigned long long *youngest)
1366 {
1367         struct trace *t;
1368
1369         if (!genesis_time)
1370                 find_genesis();
1371
1372         *youngest = 0;
1373         while ((t = trace_list) != NULL) {
1374                 struct blk_io_trace *bit = t->bit;
1375
1376                 trace_list = t->next;
1377
1378                 bit->time -= genesis_time;
1379
1380                 if (bit->time < *youngest || !*youngest)
1381                         *youngest = bit->time;
1382
1383                 if (bit->sequence < smallest_seq_read)
1384                         smallest_seq_read = bit->sequence;
1385
1386                 if (check_stopwatch(bit)) {
1387                         bit_free(bit);
1388                         t_free(t);
1389                         continue;
1390                 }
1391
1392                 if (trace_rb_insert_sort(t))
1393                         return -1;
1394         }
1395
1396         return 0;
1397 }
1398
1399 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
1400 {
1401         rb_erase(&t->rb_node, &pdi->rb_last);
1402         pdi->rb_last_entries--;
1403
1404         bit_free(t->bit);
1405         t_free(t);
1406 }
1407
1408 static void put_trace(struct per_dev_info *pdi, struct trace *t)
1409 {
1410         rb_erase(&t->rb_node, &rb_sort_root);
1411         rb_sort_entries--;
1412
1413         trace_rb_insert_last(pdi, t);
1414
1415         if (pdi->rb_last_entries > rb_batch * pdi->nfiles) {
1416                 struct rb_node *n = rb_first(&pdi->rb_last);
1417
1418                 t = rb_entry(n, struct trace, rb_node);
1419                 __put_trace_last(pdi, t);
1420         }
1421 }
1422
1423 /*
1424  * to continue, we must have traces from all online cpus in the tree
1425  */
1426 static int check_cpu_map(struct per_dev_info *pdi)
1427 {
1428         unsigned long *cpu_map;
1429         struct rb_node *n;
1430         struct trace *__t;
1431         unsigned int i;
1432         int ret, cpu;
1433
1434         /*
1435          * create a map of the cpus we have traces for
1436          */
1437         cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1438         n = rb_first(&rb_sort_root);
1439         while (n) {
1440                 __t = rb_entry(n, struct trace, rb_node);
1441                 cpu = __t->bit->cpu;
1442
1443                 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1444                 n = rb_next(n);
1445         }
1446
1447         /*
1448          * we can't continue if pdi->cpu_map has entries set that we don't
1449          * have in the sort rbtree. the opposite is not a problem, though
1450          */
1451         ret = 0;
1452         for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1453                 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1454                         ret = 1;
1455                         break;
1456                 }
1457         }
1458
1459         free(cpu_map);
1460         return ret;
1461 }
1462
1463 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1464 {
1465         unsigned long expected_sequence = pdi->last_sequence + 1;
1466         struct blk_io_trace *bit = t->bit;
1467         struct trace *__t;
1468
1469         if (!expected_sequence) {
1470                 /*
1471                  * 1 should be the first entry, just allow it
1472                  */
1473                 if (bit->sequence == 1)
1474                         return 0;
1475
1476                 return check_cpu_map(pdi);
1477         }
1478
1479         if (bit->sequence == expected_sequence)
1480                 return 0;
1481
1482         /*
1483          * we may not have seen that sequence yet. if we are not doing
1484          * the final run, break and wait for more entries.
1485          */
1486         if (expected_sequence < smallest_seq_read) {
1487                 __t = trace_rb_find_last(pdi, expected_sequence);
1488                 if (!__t)
1489                         goto skip;
1490
1491                 __put_trace_last(pdi, __t);
1492                 return 0;
1493         } else if (!force) {
1494                 return 1;
1495         } else {
1496 skip:
1497                 if (verbose) {
1498                         fprintf(stderr, "(%d,%d): skipping %lu -> %u\n",
1499                                 MAJOR(pdi->dev), MINOR(pdi->dev),
1500                                 pdi->last_sequence, bit->sequence);
1501                 }
1502                 pdi->skips++;
1503                 return 0;
1504         }
1505 }
1506
1507 static void show_entries_rb(int force)
1508 {
1509         struct per_dev_info *pdi = NULL;
1510         struct per_cpu_info *pci = NULL;
1511         struct blk_io_trace *bit;
1512         struct rb_node *n;
1513         struct trace *t;
1514
1515         while ((n = rb_first(&rb_sort_root)) != NULL) {
1516                 if (is_done() && !force && !pipeline)
1517                         break;
1518
1519                 t = rb_entry(n, struct trace, rb_node);
1520                 bit = t->bit;
1521
1522                 if (!pdi || pdi->dev != bit->device)
1523                         pdi = get_dev_info(bit->device);
1524
1525                 if (!pdi) {
1526                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1527                                 MAJOR(bit->device), MINOR(bit->device));
1528                         break;
1529                 }
1530
1531                 if (check_sequence(pdi, t, force))
1532                         break;
1533
1534                 if (!force && bit->time > last_allowed_time)
1535                         break;
1536
1537                 pdi->last_sequence = bit->sequence;
1538
1539                 check_time(pdi, bit);
1540
1541                 if (!pci || pci->cpu != bit->cpu)
1542                         pci = get_cpu_info(pdi, bit->cpu);
1543
1544                 pci->nelems++;
1545
1546                 if (bit->action & (act_mask << BLK_TC_SHIFT))
1547                         dump_trace(bit, pci, pdi);
1548
1549                 put_trace(pdi, t);
1550         }
1551 }
1552
1553 static int read_data(int fd, void *buffer, int bytes, int block)
1554 {
1555         int ret, bytes_left, fl;
1556         void *p;
1557
1558         fl = fcntl(fd, F_GETFL);
1559
1560         if (!block)
1561                 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1562         else
1563                 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1564
1565         bytes_left = bytes;
1566         p = buffer;
1567         while (bytes_left > 0) {
1568                 ret = read(fd, p, bytes_left);
1569                 if (!ret)
1570                         return 1;
1571                 else if (ret < 0) {
1572                         if (errno != EAGAIN)
1573                                 perror("read");
1574
1575                         return -1;
1576                 } else {
1577                         p += ret;
1578                         bytes_left -= ret;
1579                 }
1580         }
1581
1582         return 0;
1583 }
1584
1585 static int read_events(int fd, int always_block)
1586 {
1587         struct per_dev_info *pdi = NULL;
1588         unsigned int events = 0;
1589
1590         while (!is_done() && events < rb_batch) {
1591                 struct blk_io_trace *bit;
1592                 struct trace *t;
1593                 int pdu_len;
1594                 __u32 magic;
1595
1596                 bit = bit_alloc();
1597
1598                 if (read_data(fd, bit, sizeof(*bit), !events || always_block))
1599                         break;
1600
1601                 magic = be32_to_cpu(bit->magic);
1602                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1603                         fprintf(stderr, "Bad magic %x\n", magic);
1604                         break;
1605                 }
1606
1607                 pdu_len = be16_to_cpu(bit->pdu_len);
1608                 if (pdu_len) {
1609                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
1610
1611                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1))
1612                                 break;
1613
1614                         bit = ptr;
1615                 }
1616
1617                 trace_to_cpu(bit);
1618
1619                 if (verify_trace(bit)) {
1620                         bit_free(bit);
1621                         continue;
1622                 }
1623
1624                 t = t_alloc();
1625                 memset(t, 0, sizeof(*t));
1626                 t->bit = bit;
1627
1628                 t->next = trace_list;
1629                 trace_list = t;
1630
1631                 if (!pdi || pdi->dev != bit->device)
1632                         pdi = get_dev_info(bit->device);
1633
1634                 if (bit->time > pdi->last_read_time)
1635                         pdi->last_read_time = bit->time;
1636
1637                 events++;
1638         }
1639
1640         return events;
1641 }
1642
1643 static int do_file(void)
1644 {
1645         struct per_cpu_info *pci;
1646         struct per_dev_info *pdi;
1647         int i, j, events, events_added;
1648
1649         /*
1650          * first prepare all files for reading
1651          */
1652         for (i = 0; i < ndevices; i++) {
1653                 pdi = &devices[i];
1654                 pdi->nfiles = 0;
1655                 pdi->last_sequence = -1;
1656
1657                 for (j = 0;; j++) {
1658                         struct stat st;
1659                         int len = 0;
1660
1661                         pci = get_cpu_info(pdi, j);
1662                         pci->cpu = j;
1663                         pci->fd = -1;
1664
1665                         if (input_dir)
1666                                 len = sprintf(pci->fname, "%s/", input_dir);
1667
1668                         snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
1669                                  "%s.blktrace.%d", pdi->name, pci->cpu);
1670                         if (stat(pci->fname, &st) < 0)
1671                                 break;
1672                         if (st.st_size) {
1673                                 pci->fd = open(pci->fname, O_RDONLY);
1674                                 if (pci->fd < 0) {
1675                                         perror(pci->fname);
1676                                         continue;
1677                                 }
1678                         }
1679
1680                         printf("Input file %s added\n", pci->fname);
1681                         pdi->nfiles++;
1682                         cpu_mark_online(pdi, pci->cpu);
1683                 }
1684         }
1685
1686         /*
1687          * now loop over the files reading in the data
1688          */
1689         do {
1690                 unsigned long long youngest;
1691
1692                 events_added = 0;
1693                 last_allowed_time = -1ULL;
1694                 smallest_seq_read = -1U;
1695
1696                 for (i = 0; i < ndevices; i++) {
1697                         pdi = &devices[i];
1698
1699                         for (j = 0; j < pdi->nfiles; j++) {
1700
1701                                 pci = get_cpu_info(pdi, j);
1702
1703                                 if (pci->fd == -1)
1704                                         continue;
1705
1706                                 events = read_events(pci->fd, 1);
1707                                 if (!events) {
1708                                         cpu_mark_offline(pdi, pci->cpu);
1709                                         close(pci->fd);
1710                                         pci->fd = -1;
1711                                         continue;
1712                                 }
1713
1714                                 if (pdi->last_read_time < last_allowed_time)
1715                                         last_allowed_time = pdi->last_read_time;
1716
1717                                 events_added += events;
1718                         }
1719                 }
1720
1721                 if (sort_entries(&youngest))
1722                         break;
1723
1724                 if (youngest > stopwatch_end)
1725                         break;
1726
1727                 show_entries_rb(0);
1728
1729         } while (events_added);
1730
1731         if (rb_sort_entries)
1732                 show_entries_rb(1);
1733
1734         return 0;
1735 }
1736
1737 static int do_stdin(void)
1738 {
1739         unsigned long long youngest;
1740         int fd, events;
1741
1742         last_allowed_time = -1ULL;
1743         fd = dup(STDIN_FILENO);
1744         if (fd == -1) {
1745                 perror("dup stdin");
1746                 return -1;
1747         }
1748
1749         while ((events = read_events(fd, 0)) != 0) {
1750
1751                 smallest_seq_read = -1U;
1752
1753                 if (sort_entries(&youngest))
1754                         break;
1755
1756                 if (youngest > stopwatch_end)
1757                         break;
1758
1759                 show_entries_rb(0);
1760         }
1761
1762         if (rb_sort_entries)
1763                 show_entries_rb(1);
1764
1765         close(fd);
1766         return 0;
1767 }
1768
1769 static void show_stats(void)
1770 {
1771         if (!ofp)
1772                 return;
1773         if (stats_printed)
1774                 return;
1775
1776         stats_printed = 1;
1777
1778         if (per_process_stats)
1779                 show_process_stats();
1780
1781         if (per_device_and_cpu_stats)
1782                 show_device_and_cpu_stats();
1783
1784         fflush(ofp);
1785 }
1786
1787 static void handle_sigint(__attribute__((__unused__)) int sig)
1788 {
1789         done = 1;
1790         show_stats();
1791 }
1792
1793 /*
1794  * Extract start and duration times from a string, allowing
1795  * us to specify a time interval of interest within a trace.
1796  * Format: "duration" (start is zero) or "start:duration".
1797  */
1798 static int find_stopwatch_interval(char *string)
1799 {
1800         double value;
1801         char *sp;
1802
1803         value = strtod(string, &sp);
1804         if (sp == string) {
1805                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1806                 return 1;
1807         }
1808         if (*sp == ':') {
1809                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1810                 string = sp + 1;
1811                 value = strtod(string, &sp);
1812                 if (sp == string || *sp != '\0') {
1813                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1814                                 string);
1815                         return 1;
1816                 }
1817         } else if (*sp != '\0') {
1818                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
1819                 return 1;
1820         }
1821         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
1822         if (stopwatch_end <= stopwatch_start) {
1823                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
1824                         stopwatch_start, stopwatch_end);
1825                 return 1;
1826         }
1827
1828         return 0;
1829 }
1830
1831 static char usage_str[] = \
1832         "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
1833         "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
1834         "\t-i Input file containing trace data, or '-' for stdin\n" \
1835         "\t-D Directory to prepend to input file names\n" \
1836         "\t-o Output file. If not given, output is stdout\n" \
1837         "\t-b stdin read batching\n" \
1838         "\t-s Show per-program io statistics\n" \
1839         "\t-n Hash processes by name, not pid\n" \
1840         "\t-t Track individual ios. Will tell you the time a request took\n" \
1841         "\t   to get queued, to get dispatched, and to get completed\n" \
1842         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
1843         "\t-w Only parse data between the given time interval in seconds.\n" \
1844         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
1845         "\t-f Output format. Customize the output format. The format field\n" \
1846         "\t   identifies can be found in the documentation\n" \
1847         "\t-F Format specification. Can be found in the documentation\n" \
1848         "\t-v More verbose for marginal errors\n" \
1849         "\t-V Print program version info\n\n";
1850
1851 static void usage(char *prog)
1852 {
1853         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
1854 }
1855
1856 int main(int argc, char *argv[])
1857 {
1858         char *ofp_buffer;
1859         int i, c, ret, mode;
1860         int act_mask_tmp = 0;
1861
1862         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
1863                 switch (c) {
1864                 case 'a':
1865                         i = find_mask_map(optarg);
1866                         if (i < 0) {
1867                                 fprintf(stderr,"Invalid action mask %s\n",
1868                                         optarg);
1869                                 return 1;
1870                         }
1871                         act_mask_tmp |= i;
1872                         break;
1873
1874                 case 'A':
1875                         if ((sscanf(optarg, "%x", &i) != 1) ||
1876                                                         !valid_act_opt(i)) {
1877                                 fprintf(stderr,
1878                                         "Invalid set action mask %s/0x%x\n",
1879                                         optarg, i);
1880                                 return 1;
1881                         }
1882                         act_mask_tmp = i;
1883                         break;
1884                 case 'i':
1885                         if (!strcmp(optarg, "-") && !pipeline)
1886                                 pipeline = 1;
1887                         else if (resize_devices(optarg) != 0)
1888                                 return 1;
1889                         break;
1890                 case 'D':
1891                         input_dir = optarg;
1892                         break;
1893                 case 'o':
1894                         output_name = optarg;
1895                         break;
1896                 case 'b':
1897                         rb_batch = atoi(optarg);
1898                         if (rb_batch <= 0)
1899                                 rb_batch = RB_BATCH_DEFAULT;
1900                         break;
1901                 case 's':
1902                         per_process_stats = 1;
1903                         break;
1904                 case 't':
1905                         track_ios = 1;
1906                         break;
1907                 case 'q':
1908                         per_device_and_cpu_stats = 0;
1909                         break;
1910                 case 'w':
1911                         if (find_stopwatch_interval(optarg) != 0)
1912                                 return 1;
1913                         break;
1914                 case 'f':
1915                         set_all_format_specs(optarg);
1916                         break;
1917                 case 'F':
1918                         if (add_format_spec(optarg) != 0)
1919                                 return 1;
1920                         break;
1921                 case 'h':
1922                         ppi_hash_by_pid = 0;
1923                         break;
1924                 case 'v':
1925                         verbose++;
1926                         break;
1927                 case 'V':
1928                         printf("%s version %s\n", argv[0], blkparse_version);
1929                         return 0;
1930                 default:
1931                         usage(argv[0]);
1932                         return 1;
1933                 }
1934         }
1935
1936         while (optind < argc) {
1937                 if (!strcmp(argv[optind], "-") && !pipeline)
1938                         pipeline = 1;
1939                 else if (resize_devices(argv[optind]) != 0)
1940                         return 1;
1941                 optind++;
1942         }
1943
1944         if (!pipeline && !ndevices) {
1945                 usage(argv[0]);
1946                 return 1;
1947         }
1948
1949         if (act_mask_tmp != 0)
1950                 act_mask = act_mask_tmp;
1951
1952         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
1953
1954         signal(SIGINT, handle_sigint);
1955         signal(SIGHUP, handle_sigint);
1956         signal(SIGTERM, handle_sigint);
1957
1958         setlocale(LC_NUMERIC, "en_US");
1959
1960         if (!output_name) {
1961                 ofp = fdopen(STDOUT_FILENO, "w");
1962                 mode = _IOLBF;
1963         } else {
1964                 char ofname[128];
1965
1966                 snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
1967                 ofp = fopen(ofname, "w");
1968                 mode = _IOFBF;
1969         }
1970
1971         if (!ofp) {
1972                 perror("fopen");
1973                 return 1;
1974         }
1975
1976         ofp_buffer = malloc(4096);
1977         if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
1978                 perror("setvbuf");
1979                 return 1;
1980         }
1981
1982         if (pipeline)
1983                 ret = do_stdin();
1984         else
1985                 ret = do_file();
1986
1987         show_stats();
1988         return ret;
1989 }