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
  37 static char blkparse_version[] = "0.90";
  38
  39 struct per_dev_info {
  40         dev_t id;
  41         char *name;
  42
  43         int backwards;
  44         unsigned long long events;
  45         unsigned long long last_reported_time;
  46         unsigned long long last_read_time;
  47         struct io_stats io_stats;
  48         unsigned long last_sequence;
  49         unsigned long skips;
  50
  51         int nfiles;
  52         int ncpus;
  53         struct per_cpu_info *cpus;
  54 };
  55
  56 struct per_process_info {
  57         char name[16];
  58         __u32 pid;
  59         struct io_stats io_stats;
  60         struct per_process_info *hash_next, *list_next;
  61
  62         /*
  63          * individual io stats
  64          */
  65         unsigned long long longest_allocation_wait[2];
  66         unsigned long long longest_dispatch_wait[2];
  67         unsigned long long longest_completion_wait[2];
  68 };
  69
  70 #define PPI_HASH_SHIFT  (8)
  71 static struct per_process_info *ppi_hash[1 << PPI_HASH_SHIFT];
  72 static struct per_process_info *ppi_list;
  73 static int ppi_list_entries;
  74
  75 #define S_OPTS  "i:o:b:stqw:f:F:v"
  76 static struct option l_opts[] = {
  77         {
  78                 .name = "input",
  79                 .has_arg = required_argument,
  80                 .flag = NULL,
  81                 .val = 'i'
  82         },
  83         {
  84                 .name = "output",
  85                 .has_arg = required_argument,
  86                 .flag = NULL,
  87                 .val = 'o'
  88         },
  89         {
  90                 .name = "batch",
  91                 .has_arg = required_argument,
  92                 .flag = NULL,
  93                 .val = 'b'
  94         },
  95         {
  96                 .name = "per program stats",
  97                 .has_arg = no_argument,
  98                 .flag = NULL,
  99                 .val = 's'
 100         },
 101         {
 102                 .name = "track ios",
 103                 .has_arg = no_argument,
 104                 .flag = NULL,
 105                 .val = 't'
 106         },
 107         {
 108                 .name = "quiet",
 109                 .has_arg = no_argument,
 110                 .flag = NULL,
 111                 .val = 'q'
 112         },
 113         {
 114                 .name = "stopwatch",
 115                 .has_arg = required_argument,
 116                 .flag = NULL,
 117                 .val = 'w'
 118         },
 119         {
 120                 .name = "format",
 121                 .has_arg = required_argument,
 122                 .flag = NULL,
 123                 .val = 'f'
 124         },
 125         {
 126                 .name = "format-spec",
 127                 .has_arg = required_argument,
 128                 .flag = NULL,
 129                 .val = 'F'
 130         },
 131         {
 132                 .name = "version",
 133                 .has_arg = no_argument,
 134                 .flag = NULL,
 135                 .val = 'v'
 136         },
 137 };
 138
 139 /*
 140  * for sorting the displayed output
 141  */
 142 struct trace {
 143         struct blk_io_trace *bit;
 144         struct rb_node rb_node;
 145         struct trace *next;
 146         int skipped;
 147 };
 148
 149 static struct rb_root rb_sort_root;
 150 static struct rb_root rb_track_root;
 151
 152 static struct trace *trace_list;
 153
 154 /*
 155  * allocation cache
 156  */
 157 static struct blk_io_trace *bit_alloc_list;
 158 static struct trace *t_alloc_list;
 159
 160 /*
 161  * for tracking individual ios
 162  */
 163 struct io_track {
 164         struct rb_node rb_node;
 165
 166         dev_t device;
 167         __u64 sector;
 168         __u32 pid;
 169         unsigned long long allocation_time;
 170         unsigned long long queue_time;
 171         unsigned long long dispatch_time;
 172         unsigned long long completion_time;
 173 };
 174
 175 static int ndevices;
 176 static struct per_dev_info *devices;
 177 static char *get_dev_name(struct per_dev_info *, char *, int);
 178
 179 FILE *ofp = NULL;
 180 static char *output_name;
 181
 182 static unsigned long long genesis_time;
 183 static unsigned long long last_allowed_time;
 184 static unsigned long long stopwatch_start;      /* start from zero by default */
 185 static unsigned long long stopwatch_end = ULONG_LONG_MAX;       /* "infinity" */
 186
 187 static int per_process_stats;
 188 static int track_ios;
 189
 190 #define RB_BATCH_DEFAULT        (512)
 191 static int rb_batch = RB_BATCH_DEFAULT;
 192
 193 static int pipeline;
 194
 195 #define is_done()       (*(volatile int *)(&done))
 196 static volatile int done;
 197
 198 static inline unsigned long hash_long(unsigned long val)
 199 {
 200 #if __WORDSIZE == 32
 201         val *= 0x9e370001UL;
 202 #elif __WORDSIZE == 64
 203         val *= 0x9e37fffffffc0001UL;
 204 #else
 205 #error unknown word size
 206 #endif
 207
 208         return val >> (__WORDSIZE - PPI_HASH_SHIFT);
 209 }
 210
 211 static inline void add_process_to_hash(struct per_process_info *ppi)
 212 {
 213         const int hash_idx = hash_long(ppi->pid);
 214
 215         ppi->hash_next = ppi_hash[hash_idx];
 216         ppi_hash[hash_idx] = ppi;
 217 }
 218
 219 static inline void add_process_to_list(struct per_process_info *ppi)
 220 {
 221         ppi->list_next = ppi_list;
 222         ppi_list = ppi;
 223         ppi_list_entries++;
 224 }
 225
 226 static struct per_process_info *find_process_by_pid(__u32 pid)
 227 {
 228         const int hash_idx = hash_long(pid);
 229         struct per_process_info *ppi;
 230
 231         ppi = ppi_hash[hash_idx];
 232         while (ppi) {
 233                 if (ppi->pid == pid)
 234                         return ppi;
 235
 236                 ppi = ppi->hash_next;
 237         }
 238
 239         return NULL;
 240 }
 241
 242 static inline int trace_rb_insert(struct trace *t)
 243 {
 244         struct rb_node **p = &rb_sort_root.rb_node;
 245         struct rb_node *parent = NULL;
 246         struct trace *__t;
 247
 248         while (*p) {
 249                 parent = *p;
 250                 __t = rb_entry(parent, struct trace, rb_node);
 251
 252                 if (t->bit->time < __t->bit->time)
 253                         p = &(*p)->rb_left;
 254                 else if (t->bit->time > __t->bit->time)
 255                         p = &(*p)->rb_right;
 256                 else if (t->bit->device < __t->bit->device)
 257                         p = &(*p)->rb_left;
 258                 else if (t->bit->device > __t->bit->device)
 259                         p = &(*p)->rb_right;
 260                 else if (t->bit->sequence < __t->bit->sequence)
 261                         p = &(*p)->rb_left;
 262                 else if (t->bit->sequence > __t->bit->sequence)
 263                         p = &(*p)->rb_right;
 264                 else if (t->bit->device == __t->bit->device) {
 265                         fprintf(stderr,
 266                                 "sequence alias (%d) on device %d,%d!\n",
 267                                 t->bit->sequence,
 268                                 MAJOR(t->bit->device), MINOR(t->bit->device));
 269                         return 1;
 270                 }
 271         }
 272
 273         rb_link_node(&t->rb_node, parent, p);
 274         rb_insert_color(&t->rb_node, &rb_sort_root);
 275         return 0;
 276 }
 277
 278 static inline int track_rb_insert(struct io_track *iot)
 279 {
 280         struct rb_node **p = &rb_track_root.rb_node;
 281         struct rb_node *parent = NULL;
 282         struct io_track *__iot;
 283
 284         while (*p) {
 285                 parent = *p;
 286
 287                 __iot = rb_entry(parent, struct io_track, rb_node);
 288
 289                 if (iot->device < __iot->device)
 290                         p = &(*p)->rb_left;
 291                 else if (iot->device > __iot->device)
 292                         p = &(*p)->rb_right;
 293                 else if (iot->sector < __iot->sector)
 294                         p = &(*p)->rb_left;
 295                 else if (iot->sector > __iot->sector)
 296                         p = &(*p)->rb_right;
 297                 else {
 298                         fprintf(stderr,
 299                                 "sector alias (%Lu) on device %d,%d!\n",
 300                                 (unsigned long long) iot->sector,
 301                                 MAJOR(iot->device), MINOR(iot->device));
 302                         return 1;
 303                 }
 304         }
 305
 306         rb_link_node(&iot->rb_node, parent, p);
 307         rb_insert_color(&iot->rb_node, &rb_track_root);
 308         return 0;
 309 }
 310
 311 static struct io_track *__find_track(dev_t device, __u64 sector)
 312 {
 313         struct rb_node **p = &rb_track_root.rb_node;
 314         struct rb_node *parent = NULL;
 315         struct io_track *__iot;
 316
 317         while (*p) {
 318                 parent = *p;
 319
 320                 __iot = rb_entry(parent, struct io_track, rb_node);
 321
 322                 if (device < __iot->device)
 323                         p = &(*p)->rb_left;
 324                 else if (device > __iot->device)
 325                         p = &(*p)->rb_right;
 326                 else if (sector < __iot->sector)
 327                         p = &(*p)->rb_left;
 328                 else if (sector > __iot->sector)
 329                         p = &(*p)->rb_right;
 330                 else
 331                         return __iot;
 332         }
 333
 334         return NULL;
 335 }
 336
 337 static struct io_track *find_track(__u32 pid, dev_t device, __u64 sector)
 338 {
 339         struct io_track *iot;
 340
 341         iot = __find_track(device, sector);
 342         if (!iot) {
 343                 iot = malloc(sizeof(*iot));
 344                 iot->pid = pid;
 345                 iot->device = device;
 346                 iot->sector = sector;
 347                 track_rb_insert(iot);
 348         }
 349
 350         return iot;
 351 }
 352
 353 static void log_track_frontmerge(struct blk_io_trace *t)
 354 {
 355         struct io_track *iot;
 356
 357         if (!track_ios)
 358                 return;
 359
 360         iot = __find_track(t->device, t->sector + (t->bytes >> 9));
 361         if (!iot) {
 362                 fprintf(stderr, "failed to find mergeable event\n");
 363                 return;
 364         }
 365
 366         rb_erase(&iot->rb_node, &rb_track_root);
 367         iot->sector -= t->bytes >> 9;
 368         track_rb_insert(iot);
 369 }
 370
 371 static void log_track_getrq(struct blk_io_trace *t)
 372 {
 373         struct io_track *iot;
 374
 375         if (!track_ios)
 376                 return;
 377
 378         iot = find_track(t->pid, t->device, t->sector);
 379         iot->allocation_time = t->time;
 380 }
 381
 382
 383 /*
 384  * return time between rq allocation and insertion
 385  */
 386 static unsigned long long log_track_insert(struct blk_io_trace *t)
 387 {
 388         unsigned long long elapsed;
 389         struct io_track *iot;
 390
 391         if (!track_ios)
 392                 return -1;
 393
 394         iot = find_track(t->pid, t->device, t->sector);
 395         iot->queue_time = t->time;
 396         elapsed = iot->queue_time - iot->allocation_time;
 397
 398         if (per_process_stats) {
 399                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
 400                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 401
 402                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
 403                         ppi->longest_allocation_wait[w] = elapsed;
 404         }
 405
 406         return elapsed;
 407 }
 408
 409 /*
 410  * return time between queue and issue
 411  */
 412 static unsigned long long log_track_issue(struct blk_io_trace *t)
 413 {
 414         unsigned long long elapsed;
 415         struct io_track *iot;
 416
 417         if (!track_ios)
 418                 return -1;
 419         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 420                 return -1;
 421
 422         iot = __find_track(t->device, t->sector);
 423         if (!iot) {
 424                 fprintf(stderr, "failed to find issue event\n");
 425                 return -1;
 426         }
 427
 428         iot->dispatch_time = t->time;
 429         elapsed = iot->dispatch_time - iot->queue_time;
 430
 431         if (per_process_stats) {
 432                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
 433                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 434
 435                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
 436                         ppi->longest_dispatch_wait[w] = elapsed;
 437         }
 438
 439         return elapsed;
 440 }
 441
 442 /*
 443  * return time between dispatch and complete
 444  */
 445 static unsigned long long log_track_complete(struct blk_io_trace *t)
 446 {
 447         unsigned long long elapsed;
 448         struct io_track *iot;
 449
 450         if (!track_ios)
 451                 return -1;
 452         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 453                 return -1;
 454
 455         iot = __find_track(t->device, t->sector);
 456         if (!iot) {
 457                 fprintf(stderr, "failed to find complete event\n");
 458                 return -1;
 459         }
 460
 461         iot->completion_time = t->time;
 462         elapsed = iot->completion_time - iot->dispatch_time;
 463
 464         if (per_process_stats) {
 465                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
 466                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 467
 468                 if (ppi && elapsed > ppi->longest_completion_wait[w])
 469                         ppi->longest_completion_wait[w] = elapsed;
 470         }
 471
 472         /*
 473          * kill the trace, we don't need it after completion
 474          */
 475         rb_erase(&iot->rb_node, &rb_track_root);
 476         free(iot);
 477
 478         return elapsed;
 479 }
 480
 481
 482 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
 483 {
 484         struct per_process_info *ppi = find_process_by_pid(pid);
 485
 486         if (!ppi) {
 487                 ppi = malloc(sizeof(*ppi));
 488                 memset(ppi, 0, sizeof(*ppi));
 489                 strncpy(ppi->name, name, sizeof(ppi->name));
 490                 ppi->pid = pid;
 491                 add_process_to_hash(ppi);
 492                 add_process_to_list(ppi);
 493         }
 494
 495         return &ppi->io_stats;
 496 }
 497
 498
 499 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
 500 {
 501         struct per_cpu_info *cpus = pdi->cpus;
 502         int ncpus = pdi->ncpus;
 503         int new_count = cpu + 1;
 504         int new_space, size;
 505         char *new_start;
 506
 507         size = new_count * sizeof(struct per_cpu_info);
 508         cpus = realloc(cpus, size);
 509         if (!cpus) {
 510                 char name[20];
 511                 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
 512                         get_dev_name(pdi, name, sizeof(name)), size);
 513                 exit(1);
 514         }
 515
 516         new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
 517         new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
 518         memset(new_start, 0, new_space);
 519
 520         pdi->ncpus = new_count;
 521         pdi->cpus = cpus;
 522 }
 523
 524 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
 525 {
 526         struct per_cpu_info *pci;
 527
 528         if (cpu >= pdi->ncpus)
 529                 resize_cpu_info(pdi, cpu);
 530
 531         pci = &pdi->cpus[cpu];
 532         pci->cpu = cpu;
 533         return pci;
 534 }
 535
 536
 537 static int resize_devices(char *name)
 538 {
 539         int size = (ndevices + 1) * sizeof(struct per_dev_info);
 540
 541         devices = realloc(devices, size);
 542         if (!devices) {
 543                 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
 544                 return 1;
 545         }
 546         memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
 547         devices[ndevices].name = name;
 548         ndevices++;
 549         return 0;
 550 }
 551
 552 static struct per_dev_info *get_dev_info(dev_t id)
 553 {
 554         struct per_dev_info *pdi;
 555         int i;
 556
 557         for (i = 0; i < ndevices; i++) {
 558                 if (!devices[i].id)
 559                         devices[i].id = id;
 560                 if (devices[i].id == id)
 561                         return &devices[i];
 562         }
 563
 564         if (resize_devices(NULL) != 0)
 565                 return NULL;
 566
 567         pdi = &devices[ndevices - 1];
 568         pdi->id = id;
 569         pdi->last_sequence = -1;
 570         pdi->last_read_time = 0;
 571         return pdi;
 572 }
 573
 574 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
 575 {
 576         if (pdi->name)
 577                 snprintf(buffer, size, "%s", pdi->name);
 578         else
 579                 snprintf(buffer, size, "%d,%d", MAJOR(pdi->id), MINOR(pdi->id));
 580         return buffer;
 581 }
 582
 583 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
 584 {
 585         unsigned long long this = bit->time;
 586         unsigned long long last = pdi->last_reported_time;
 587
 588         pdi->backwards = (this < last) ? 'B' : ' ';
 589         pdi->last_reported_time = this;
 590 }
 591
 592 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
 593                                int rw)
 594 {
 595         if (rw) {
 596                 ios->mwrites++;
 597                 ios->qwrite_kb += t->bytes >> 10;
 598         } else {
 599                 ios->mreads++;
 600                 ios->qread_kb += t->bytes >> 10;
 601         }
 602 }
 603
 604 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
 605                              int rw)
 606 {
 607         __account_m(&pci->io_stats, t, rw);
 608
 609         if (per_process_stats) {
 610                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 611
 612                 __account_m(ios, t, rw);
 613         }
 614 }
 615
 616 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
 617                                    int rw)
 618 {
 619         if (rw) {
 620                 ios->qwrites++;
 621                 ios->qwrite_kb += t->bytes >> 10;
 622         } else {
 623                 ios->qreads++;
 624                 ios->qread_kb += t->bytes >> 10;
 625         }
 626 }
 627
 628 static inline void account_queue(struct blk_io_trace *t,
 629                                  struct per_cpu_info *pci, int rw)
 630 {
 631         __account_queue(&pci->io_stats, t, rw);
 632
 633         if (per_process_stats) {
 634                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 635
 636                 __account_queue(ios, t, rw);
 637         }
 638 }
 639
 640 static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
 641 {
 642         if (rw) {
 643                 ios->cwrites++;
 644                 ios->cwrite_kb += bytes >> 10;
 645         } else {
 646                 ios->creads++;
 647                 ios->cread_kb += bytes >> 10;
 648         }
 649 }
 650
 651 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
 652                              int rw, int bytes)
 653 {
 654         __account_c(&pci->io_stats, rw, bytes);
 655
 656         if (per_process_stats) {
 657                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 658
 659                 __account_c(ios, rw, bytes);
 660         }
 661 }
 662
 663 static inline void __account_issue(struct io_stats *ios, int rw,
 664                                    unsigned int bytes)
 665 {
 666         if (rw) {
 667                 ios->iwrites++;
 668                 ios->iwrite_kb += bytes >> 10;
 669         } else {
 670                 ios->ireads++;
 671                 ios->iread_kb += bytes >> 10;
 672         }
 673 }
 674
 675 static inline void account_issue(struct blk_io_trace *t,
 676                                  struct per_cpu_info *pci, int rw)
 677 {
 678         __account_issue(&pci->io_stats, rw, t->bytes);
 679
 680         if (per_process_stats) {
 681                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 682
 683                 __account_issue(ios, rw, t->bytes);
 684         }
 685 }
 686
 687 static inline void __account_unplug(struct io_stats *ios, int timer)
 688 {
 689         if (timer)
 690                 ios->timer_unplugs++;
 691         else
 692                 ios->io_unplugs++;
 693 }
 694
 695 static inline void account_unplug(struct blk_io_trace *t,
 696                                   struct per_cpu_info *pci, int timer)
 697 {
 698         __account_unplug(&pci->io_stats, timer);
 699
 700         if (per_process_stats) {
 701                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 702
 703                 __account_unplug(ios, timer);
 704         }
 705 }
 706
 707 static void log_complete(struct per_cpu_info *pci, struct blk_io_trace *t,
 708                          char *act)
 709 {
 710         process_fmt(act, pci, t, log_track_complete(t), 0, NULL);
 711 }
 712
 713 static void log_insert(struct per_cpu_info *pci, struct blk_io_trace *t,
 714                       char *act)
 715 {
 716         process_fmt(act, pci, t, log_track_insert(t), 0, NULL);
 717 }
 718
 719 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
 720                       char *act)
 721 {
 722         process_fmt(act, pci, t, -1, 0, NULL);
 723 }
 724
 725 static void log_issue(struct per_cpu_info *pci, struct blk_io_trace *t,
 726                       char *act)
 727 {
 728         process_fmt(act, pci, t, log_track_issue(t), 0, NULL);
 729 }
 730
 731 static void log_merge(struct per_cpu_info *pci, struct blk_io_trace *t,
 732                       char *act)
 733 {
 734         if (act[0] == 'F')
 735                 log_track_frontmerge(t);
 736
 737         process_fmt(act, pci, t, -1ULL, 0, NULL);
 738 }
 739
 740 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
 741                         char *act)
 742 {
 743         process_fmt(act, pci, t, -1ULL, 0, NULL);
 744 }
 745
 746 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
 747                         char *act)
 748 {
 749         process_fmt(act, pci, t, -1ULL, 0, NULL);
 750 }
 751
 752 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
 753                       char *act)
 754 {
 755         process_fmt(act, pci, t, -1ULL, 0, NULL);
 756 }
 757
 758 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
 759                       char *act)
 760 {
 761         process_fmt(act, pci, t, -1ULL, 0, NULL);
 762 }
 763
 764 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
 765 {
 766         unsigned char *buf = (unsigned char *) t + sizeof(*t);
 767
 768         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
 769 }
 770
 771 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
 772 {
 773         int act = t->action & 0xffff;
 774
 775         switch (act) {
 776                 case __BLK_TA_QUEUE:
 777                         log_generic(pci, t, "Q");
 778                         break;
 779                 case __BLK_TA_GETRQ:
 780                         log_generic(pci, t, "G");
 781                         break;
 782                 case __BLK_TA_SLEEPRQ:
 783                         log_generic(pci, t, "S");
 784                         break;
 785                 case __BLK_TA_REQUEUE:
 786                         log_generic(pci, t, "R");
 787                         break;
 788                 case __BLK_TA_ISSUE:
 789                         log_pc(pci, t, "D");
 790                         break;
 791                 case __BLK_TA_COMPLETE:
 792                         log_pc(pci, t, "C");
 793                         break;
 794                 case __BLK_TA_INSERT:
 795                         log_pc(pci, t, "I");
 796                         break;
 797                 default:
 798                         fprintf(stderr, "Bad pc action %x\n", act);
 799                         break;
 800         }
 801 }
 802
 803 static void dump_trace_fs(struct blk_io_trace *t, struct per_cpu_info *pci)
 804 {
 805         int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
 806         int act = t->action & 0xffff;
 807
 808         switch (act) {
 809                 case __BLK_TA_QUEUE:
 810                         account_queue(t, pci, w);
 811                         log_queue(pci, t, "Q");
 812                         break;
 813                 case __BLK_TA_INSERT:
 814                         log_insert(pci, t, "I");
 815                         break;
 816                 case __BLK_TA_BACKMERGE:
 817                         account_m(t, pci, w);
 818                         log_merge(pci, t, "M");
 819                         break;
 820                 case __BLK_TA_FRONTMERGE:
 821                         account_m(t, pci, w);
 822                         log_merge(pci, t, "F");
 823                         break;
 824                 case __BLK_TA_GETRQ:
 825                         log_track_getrq(t);
 826                         log_generic(pci, t, "G");
 827                         break;
 828                 case __BLK_TA_SLEEPRQ:
 829                         log_generic(pci, t, "S");
 830                         break;
 831                 case __BLK_TA_REQUEUE:
 832                         account_c(t, pci, w, -t->bytes);
 833                         log_queue(pci, t, "R");
 834                         break;
 835                 case __BLK_TA_ISSUE:
 836                         account_issue(t, pci, w);
 837                         log_issue(pci, t, "D");
 838                         break;
 839                 case __BLK_TA_COMPLETE:
 840                         account_c(t, pci, w, t->bytes);
 841                         log_complete(pci, t, "C");
 842                         break;
 843                 case __BLK_TA_PLUG:
 844                         log_action(pci, t, "P");
 845                         break;
 846                 case __BLK_TA_UNPLUG_IO:
 847                         account_unplug(t, pci, 0);
 848                         log_unplug(pci, t, "U");
 849                         break;
 850                 case __BLK_TA_UNPLUG_TIMER:
 851                         account_unplug(t, pci, 1);
 852                         log_unplug(pci, t, "UT");
 853                         break;
 854                 case __BLK_TA_SPLIT:
 855                         log_split(pci, t, "X");
 856                         break;
 857                 case __BLK_TA_BOUNCE:
 858                         log_generic(pci, t, "B");
 859                         break;
 860                 default:
 861                         fprintf(stderr, "Bad fs action %x\n", t->action);
 862                         break;
 863         }
 864 }
 865
 866 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
 867                        struct per_dev_info *pdi)
 868 {
 869         if (t->action & BLK_TC_ACT(BLK_TC_PC))
 870                 dump_trace_pc(t, pci);
 871         else
 872                 dump_trace_fs(t, pci);
 873
 874         pdi->events++;
 875 }
 876
 877 static void dump_io_stats(struct io_stats *ios, char *msg)
 878 {
 879         fprintf(ofp, "%s\n", msg);
 880
 881         fprintf(ofp, " Reads Queued:    %'8lu, %'8LuKiB\t", ios->qreads, ios->qread_kb);
 882         fprintf(ofp, " Writes Queued:    %'8lu, %'8LuKiB\n", ios->qwrites,ios->qwrite_kb);
 883
 884         fprintf(ofp, " Read Dispatches: %'8lu, %'8LuKiB\t", ios->ireads, ios->iread_kb);
 885         fprintf(ofp, " Write Dispatches: %'8lu, %'8LuKiB\n", ios->iwrites,ios->iwrite_kb);
 886         fprintf(ofp, " Reads Completed: %'8lu, %'8LuKiB\t", ios->creads, ios->cread_kb);
 887         fprintf(ofp, " Writes Completed: %'8lu, %'8LuKiB\n", ios->cwrites,ios->cwrite_kb);
 888         fprintf(ofp, " Read Merges:     %'8lu%8c\t", ios->mreads, ' ');
 889         fprintf(ofp, " Write Merges:     %'8lu\n", ios->mwrites);
 890         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
 891         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
 892 }
 893
 894 static void dump_wait_stats(struct per_process_info *ppi)
 895 {
 896         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
 897         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
 898         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
 899         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
 900         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
 901         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
 902
 903         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
 904         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
 905         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
 906         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
 907         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
 908         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
 909 }
 910
 911 static int ppi_name_compare(const void *p1, const void *p2)
 912 {
 913         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
 914         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
 915         int res;
 916
 917         res = strverscmp(ppi1->name, ppi2->name);
 918         if (!res)
 919                 res = ppi1->pid > ppi2->pid;
 920
 921         return res;
 922 }
 923
 924 static void sort_process_list(void)
 925 {
 926         struct per_process_info **ppis;
 927         struct per_process_info *ppi;
 928         int i = 0;
 929
 930         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
 931
 932         ppi = ppi_list;
 933         while (ppi) {
 934                 ppis[i++] = ppi;
 935                 ppi = ppi->list_next;
 936         }
 937
 938         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
 939
 940         i = ppi_list_entries - 1;
 941         ppi_list = NULL;
 942         while (i >= 0) {
 943                 ppi = ppis[i];
 944
 945                 ppi->list_next = ppi_list;
 946                 ppi_list = ppi;
 947                 i--;
 948         }
 949
 950         free(ppis);
 951 }
 952
 953 static void show_process_stats(void)
 954 {
 955         struct per_process_info *ppi;
 956
 957         sort_process_list();
 958
 959         ppi = ppi_list;
 960         while (ppi) {
 961                 char name[64];
 962
 963                 snprintf(name, sizeof(name)-1, "%s (%u)", ppi->name, ppi->pid);
 964                 dump_io_stats(&ppi->io_stats, name);
 965                 dump_wait_stats(ppi);
 966                 ppi = ppi->list_next;
 967         }
 968
 969         fprintf(ofp, "\n");
 970 }
 971
 972 static void show_device_and_cpu_stats(void)
 973 {
 974         struct per_dev_info *pdi;
 975         struct per_cpu_info *pci;
 976         struct io_stats total, *ios;
 977         int i, j, pci_events;
 978         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
 979         char name[32];
 980
 981         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
 982
 983                 memset(&total, 0, sizeof(total));
 984                 pci_events = 0;
 985
 986                 if (i > 0)
 987                         fprintf(ofp, "\n");
 988
 989                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
 990                         if (!pci->nelems)
 991                                 continue;
 992
 993                         ios = &pci->io_stats;
 994                         total.qreads += ios->qreads;
 995                         total.qwrites += ios->qwrites;
 996                         total.creads += ios->creads;
 997                         total.cwrites += ios->cwrites;
 998                         total.mreads += ios->mreads;
 999                         total.mwrites += ios->mwrites;
1000                         total.ireads += ios->ireads;
1001                         total.iwrites += ios->iwrites;
1002                         total.qread_kb += ios->qread_kb;
1003                         total.qwrite_kb += ios->qwrite_kb;
1004                         total.cread_kb += ios->cread_kb;
1005                         total.cwrite_kb += ios->cwrite_kb;
1006                         total.iread_kb += ios->iread_kb;
1007                         total.iwrite_kb += ios->iwrite_kb;
1008                         total.timer_unplugs += ios->timer_unplugs;
1009                         total.io_unplugs += ios->io_unplugs;
1010
1011                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1012                                  j, get_dev_name(pdi, name, sizeof(name)));
1013                         dump_io_stats(ios, line);
1014                         pci_events++;
1015                 }
1016
1017                 if (pci_events > 1) {
1018                         fprintf(ofp, "\n");
1019                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1020                                  get_dev_name(pdi, name, sizeof(name)));
1021                         dump_io_stats(&total, line);
1022                 }
1023
1024                 fprintf(ofp, "\nEvents (%s): %'Lu entries, %'lu skips\n",
1025                         get_dev_name(pdi, line, sizeof(line)), pdi->events,
1026                         pdi->skips);
1027         }
1028 }
1029
1030 static int sort_entries(void)
1031 {
1032         struct trace *t;
1033         int nr = 0;
1034
1035         while ((t = trace_list) != NULL) {
1036                 trace_list = t->next;
1037
1038                 if (verify_trace(t->bit))
1039                         continue;
1040                 if (trace_rb_insert(t))
1041                         break;
1042                 nr++;
1043         }
1044
1045         return nr;
1046 }
1047
1048 /*
1049  * struct trace and blktrace allocation cache, we do potentially
1050  * millions of mallocs for these structures while only using at most
1051  * a few thousand at the time
1052  */
1053 static inline void t_free(struct trace *t)
1054 {
1055         t->next = t_alloc_list;
1056         t_alloc_list = t;
1057 }
1058
1059 static inline struct trace *t_alloc(void)
1060 {
1061         struct trace *t = t_alloc_list;
1062
1063         if (t) {
1064                 t_alloc_list = t->next;
1065                 return t;
1066         }
1067
1068         return malloc(sizeof(*t));
1069 }
1070
1071 static inline void bit_free(struct blk_io_trace *bit)
1072 {
1073         /*
1074          * abuse a 64-bit field for a next pointer for the free item
1075          */
1076         bit->time = (__u64) (unsigned long) bit_alloc_list;
1077         bit_alloc_list = (struct blk_io_trace *) bit;
1078 }
1079
1080 static inline struct blk_io_trace *bit_alloc(void)
1081 {
1082         struct blk_io_trace *bit = bit_alloc_list;
1083
1084         if (bit) {
1085                 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
1086                                  bit->time;
1087                 return bit;
1088         }
1089
1090         return malloc(sizeof(*bit));
1091 }
1092
1093 static void show_entries_rb(void)
1094 {
1095         struct per_dev_info *pdi = NULL;
1096         struct per_cpu_info *pci = NULL;
1097         struct blk_io_trace *bit;
1098         struct rb_node *n;
1099         struct trace *t;
1100
1101         while ((n = rb_first(&rb_sort_root)) != NULL) {
1102
1103                 if (done)
1104                         break;
1105
1106                 t = rb_entry(n, struct trace, rb_node);
1107                 bit = t->bit;
1108
1109                 if (!pdi || pdi->id != bit->device)
1110                         pdi = get_dev_info(bit->device);
1111
1112                 if (!pdi) {
1113                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1114                                 MAJOR(bit->device), MINOR(bit->device));
1115                         break;
1116                 }
1117
1118                 if (bit->cpu > pdi->ncpus) {
1119                         fprintf(stderr, "Unknown CPU ID? (%d, device %d,%d)\n",
1120                                 bit->cpu, MAJOR(bit->device),
1121                                 MINOR(bit->device));
1122                         break;
1123                 }
1124
1125                 /*
1126                  * back off displaying more info if we are out of sync
1127                  * on SMP systems. to prevent stalling on lost events,
1128                  * only allow an event to skip us a few times
1129                  */
1130                 if (bit->sequence != (pdi->last_sequence + 1)
1131                     && pdi->last_sequence != -1) {
1132                         if (t->skipped < 5) {
1133                                 t->skipped++;
1134                                 break;
1135                         } else {
1136                                 fprintf(stderr, "skipping from %lu to %u\n", pdi->last_sequence, bit->sequence);
1137                                 pdi->skips++;
1138                         }
1139                 }
1140
1141                 if (bit->time >= stopwatch_end || bit->time > last_allowed_time)
1142                         break;
1143
1144                 pdi->last_sequence = bit->sequence;
1145
1146                 if (bit->time >= stopwatch_start) {
1147                         check_time(pdi, bit);
1148
1149                         if (!pci || pci->cpu != bit->cpu)
1150                                 pci = get_cpu_info(pdi, bit->cpu);
1151
1152                         dump_trace(bit, pci, pdi);
1153                 }
1154
1155                 rb_erase(&t->rb_node, &rb_sort_root);
1156                 bit_free(bit);
1157                 t_free(t);
1158         }
1159 }
1160
1161 static int read_data(int fd, void *buffer, int bytes, int block)
1162 {
1163         int ret, bytes_left, fl;
1164         void *p;
1165
1166         fl = fcntl(fd, F_GETFL);
1167
1168         if (!block)
1169                 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1170         else
1171                 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1172
1173         bytes_left = bytes;
1174         p = buffer;
1175         while (bytes_left > 0) {
1176                 ret = read(fd, p, bytes_left);
1177                 if (!ret)
1178                         return 1;
1179                 else if (ret < 0) {
1180                         if (errno != EAGAIN)
1181                                 perror("read");
1182                         return -1;
1183                 } else {
1184                         p += ret;
1185                         bytes_left -= ret;
1186                 }
1187         }
1188
1189         return 0;
1190 }
1191
1192 static int read_events(int fd)
1193 {
1194         struct per_dev_info *pdi = NULL;
1195         int events = 0;
1196
1197         while (!is_done() && events < rb_batch) {
1198                 struct blk_io_trace *bit;
1199                 struct trace *t;
1200                 int pdu_len;
1201                 __u32 magic;
1202
1203                 bit = bit_alloc();
1204
1205                 if (read_data(fd, bit, sizeof(*bit), !events))
1206                         break;
1207
1208                 magic = be32_to_cpu(bit->magic);
1209                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1210                         fprintf(stderr, "Bad magic %x\n", magic);
1211                         break;
1212                 }
1213
1214                 pdu_len = be16_to_cpu(bit->pdu_len);
1215                 if (pdu_len) {
1216                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
1217
1218                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1))
1219                                 break;
1220
1221                         bit = ptr;
1222                 }
1223
1224                 t = t_alloc();
1225                 memset(t, 0, sizeof(*t));
1226                 t->bit = bit;
1227
1228                 trace_to_cpu(bit);
1229                 t->next = trace_list;
1230                 trace_list = t;
1231
1232                 if (genesis_time == 0 || t->bit->time < genesis_time)
1233                         genesis_time = t->bit->time;
1234
1235                 bit->time -= genesis_time;
1236
1237                 if (!pdi || pdi->id != bit->device)
1238                         pdi = get_dev_info(bit->device);
1239
1240                 if (bit->time > pdi->last_read_time)
1241                         pdi->last_read_time = bit->time;
1242
1243                 events++;
1244         }
1245
1246         return events;
1247 }
1248
1249 static int do_file(void)
1250 {
1251         struct per_cpu_info *pci;
1252         struct per_dev_info *pdi;
1253         int i, j, events, events_added;
1254
1255         /*
1256          * first prepare all files for reading
1257          */
1258         for (i = 0; i < ndevices; i++) {
1259                 pdi = &devices[i];
1260                 pdi->nfiles = 0;
1261                 pdi->last_sequence = -1;
1262
1263                 for (j = 0;; j++) {
1264                         struct stat st;
1265
1266                         pci = get_cpu_info(pdi, j);
1267                         pci->cpu = j;
1268                         pci->fd = -1;
1269
1270                         snprintf(pci->fname, sizeof(pci->fname)-1,
1271                                  "%s.blktrace.%d", pdi->name, pci->cpu);
1272                         if (stat(pci->fname, &st) < 0)
1273                                 break;
1274                         if (st.st_size) {
1275                                 pci->fd = open(pci->fname, O_RDONLY);
1276                                 if (pci->fd < 0) {
1277                                         perror(pci->fname);
1278                                         continue;
1279                                 }
1280                         }
1281
1282                         printf("Input file %s added\n", pci->fname);
1283                         pdi->nfiles++;
1284                 }
1285         }
1286
1287         /*
1288          * now loop over the files reading in the data
1289          */
1290         do {
1291                 events_added = 0;
1292                 last_allowed_time = -1ULL;
1293
1294                 for (i = 0; i < ndevices; i++) {
1295                         pdi = &devices[i];
1296
1297                         for (j = 0; j < pdi->nfiles; j++) {
1298
1299                                 pci = get_cpu_info(pdi, j);
1300
1301                                 if (pci->fd == -1)
1302                                         continue;
1303
1304                                 events = read_events(pci->fd);
1305                                 if (!events) {
1306                                         close(pci->fd);
1307                                         pci->fd = -1;
1308                                         continue;
1309                                 }
1310
1311                                 if (pdi->last_read_time < last_allowed_time)
1312                                         last_allowed_time = pdi->last_read_time;
1313
1314                                 events_added += events;
1315                         }
1316                 }
1317
1318                 if (sort_entries() == -1)
1319                         break;
1320
1321                 show_entries_rb();
1322
1323         } while (events_added);
1324
1325         return 0;
1326 }
1327
1328 static int do_stdin(void)
1329 {
1330         int fd;
1331
1332         last_allowed_time = -1ULL;
1333         fd = dup(STDIN_FILENO);
1334         do {
1335                 int events;
1336
1337                 events = read_events(fd);
1338                 if (!events)
1339                         break;
1340
1341                 if (sort_entries() == -1)
1342                         break;
1343
1344                 show_entries_rb();
1345         } while (1);
1346
1347         close(fd);
1348         return 0;
1349 }
1350
1351 static void flush_output(void)
1352 {
1353         fflush(ofp);
1354 }
1355
1356 static void handle_sigint(int sig)
1357 {
1358         done = 1;
1359         flush_output();
1360 }
1361
1362 /*
1363  * Extract start and duration times from a string, allowing
1364  * us to specify a time interval of interest within a trace.
1365  * Format: "duration" (start is zero) or "start:duration".
1366  */
1367 static int find_stopwatch_interval(char *string)
1368 {
1369         double value;
1370         char *sp;
1371
1372         value = strtod(string, &sp);
1373         if (sp == string) {
1374                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1375                 return 1;
1376         }
1377         if (*sp == ':') {
1378                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1379                 string = sp + 1;
1380                 value = strtod(string, &sp);
1381                 if (sp == string || *sp != '\0') {
1382                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1383                                 string);
1384                         return 1;
1385                 }
1386         } else if (*sp != '\0') {
1387                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
1388                 return 1;
1389         }
1390         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
1391         if (stopwatch_end <= stopwatch_start) {
1392                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
1393                         stopwatch_start, stopwatch_end);
1394                 return 1;
1395         }
1396
1397         return 0;
1398 }
1399
1400 static char usage_str[] = \
1401         "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
1402         "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
1403         "\t-i Input file containing trace data, or '-' for stdin\n" \
1404         "\t-o Output file. If not given, output is stdout\n" \
1405         "\t-b stdin read batching\n" \
1406         "\t-s Show per-program io statistics\n" \
1407         "\t-t Track individual ios. Will tell you the time a request took\n" \
1408         "\t   to get queued, to get dispatched, and to get completed\n" \
1409         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
1410         "\t-w Only parse data between the given time interval in seconds.\n" \
1411         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
1412         "\t -f Output format. Customize the output format. The format field\n" \
1413         "\t    identifies can be found in the documentation\n" \
1414         "\t-F Format specification. Can be found in the documentation\n" \
1415         "\t-v Print program version info\n\n";
1416
1417 static void usage(char *prog)
1418 {
1419         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
1420 }
1421
1422 int main(int argc, char *argv[])
1423 {
1424         char *ofp_buffer;
1425         int c, ret, mode;
1426         int per_device_and_cpu_stats = 1;
1427
1428         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
1429                 switch (c) {
1430                 case 'i':
1431                         if (!strcmp(optarg, "-") && !pipeline)
1432                                 pipeline = 1;
1433                         else if (resize_devices(optarg) != 0)
1434                                 return 1;
1435                         break;
1436                 case 'o':
1437                         output_name = optarg;
1438                         break;
1439                 case 'b':
1440                         rb_batch = atoi(optarg);
1441                         if (rb_batch <= 0)
1442                                 rb_batch = RB_BATCH_DEFAULT;
1443                         break;
1444                 case 's':
1445                         per_process_stats = 1;
1446                         break;
1447                 case 't':
1448                         track_ios = 1;
1449                         break;
1450                 case 'q':
1451                         per_device_and_cpu_stats = 0;
1452                         break;
1453                 case 'w':
1454                         if (find_stopwatch_interval(optarg) != 0)
1455                                 return 1;
1456                         break;
1457                 case 'f':
1458                         set_all_format_specs(optarg);
1459                         break;
1460                 case 'F':
1461                         if (add_format_spec(optarg) != 0)
1462                                 return 1;
1463                         break;
1464                 case 'v':
1465                         printf("%s version %s\n", argv[0], blkparse_version);
1466                         return 0;
1467                 default:
1468                         usage(argv[0]);
1469                         return 1;
1470                 }
1471         }
1472
1473         while (optind < argc) {
1474                 if (!strcmp(argv[optind], "-") && !pipeline)
1475                         pipeline = 1;
1476                 else if (resize_devices(argv[optind]) != 0)
1477                         return 1;
1478                 optind++;
1479         }
1480
1481         if (!pipeline && !ndevices) {
1482                 usage(argv[0]);
1483                 return 1;
1484         }
1485
1486         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
1487         memset(&rb_track_root, 0, sizeof(rb_track_root));
1488
1489         signal(SIGINT, handle_sigint);
1490         signal(SIGHUP, handle_sigint);
1491         signal(SIGTERM, handle_sigint);
1492
1493         setlocale(LC_NUMERIC, "en_US");
1494
1495         if (!output_name) {
1496                 ofp = fdopen(STDOUT_FILENO, "w");
1497                 mode = _IOLBF;
1498         } else {
1499                 char ofname[128];
1500
1501                 snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
1502                 ofp = fopen(ofname, "w");
1503                 mode = _IOFBF;
1504         }
1505
1506         if (!ofp) {
1507                 perror("fopen");
1508                 return 1;
1509         }
1510
1511         ofp_buffer = malloc(4096);
1512         if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
1513                 perror("setvbuf");
1514                 return 1;
1515         }
1516
1517         if (pipeline)
1518                 ret = do_stdin();
1519         else
1520                 ret = do_file();
1521
1522         if (per_process_stats)
1523                 show_process_stats();
1524
1525         if (per_device_and_cpu_stats)
1526                 show_device_and_cpu_stats();
1527
1528         flush_output();
1529         return ret;
1530 }