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 #include <ctype.h>
  34
  35 #include "blktrace.h"
  36 #include "rbtree.h"
  37
  38 static char blkparse_version[] = "0.90";
  39
  40 #define SECONDS(x)              ((unsigned long long)(x) / 1000000000)
  41 #define NANO_SECONDS(x)         ((unsigned long long)(x) % 1000000000)
  42 #define DOUBLE_TO_NANO_ULL(d)   ((unsigned long long)((d) * 1000000000))
  43
  44 #define MINORBITS       20
  45 #define MINORMASK       ((1U << MINORBITS) - 1)
  46 #define MAJOR(dev)      ((unsigned int) ((dev) >> MINORBITS))
  47 #define MINOR(dev)      ((unsigned int) ((dev) & MINORMASK))
  48
  49 #define min(a, b)       ((a) < (b) ? (a) : (b))
  50
  51 #define HEADER          "%D %2c %8s %5T.%9t %5p %2a %3d "
  52
  53 struct io_stats {
  54         unsigned long qreads, qwrites, creads, cwrites, mreads, mwrites;
  55         unsigned long ireads, iwrites;
  56         unsigned long long qread_kb, qwrite_kb, cread_kb, cwrite_kb;
  57         unsigned long long iread_kb, iwrite_kb;
  58         unsigned long io_unplugs, timer_unplugs;
  59 };
  60
  61 struct per_cpu_info {
  62         int cpu;
  63         int nelems;
  64
  65         int fd;
  66         char fname[128];
  67
  68         struct io_stats io_stats;
  69 };
  70
  71 struct per_dev_info {
  72         dev_t id;
  73         char *name;
  74
  75         int backwards;
  76         unsigned long long events;
  77         unsigned long long last_reported_time;
  78         struct io_stats io_stats;
  79         unsigned long last_sequence;
  80         unsigned long skips;
  81
  82         int ncpus;
  83         struct per_cpu_info *cpus;
  84 };
  85
  86 struct per_process_info {
  87         char name[16];
  88         __u32 pid;
  89         struct io_stats io_stats;
  90         struct per_process_info *hash_next, *list_next;
  91
  92         /*
  93          * individual io stats
  94          */
  95         unsigned long long longest_allocation_wait[2];
  96         unsigned long long longest_dispatch_wait[2];
  97         unsigned long long longest_completion_wait[2];
  98 };
  99
 100 #define PPI_HASH_SHIFT  (8)
 101 static struct per_process_info *ppi_hash[1 << PPI_HASH_SHIFT];
 102 static struct per_process_info *ppi_list;
 103 static int ppi_list_entries;
 104
 105 #define S_OPTS  "i:o:b:stqw:f:F:v"
 106 static struct option l_opts[] = {
 107         {
 108                 .name = "input",
 109                 .has_arg = required_argument,
 110                 .flag = NULL,
 111                 .val = 'i'
 112         },
 113         {
 114                 .name = "output",
 115                 .has_arg = required_argument,
 116                 .flag = NULL,
 117                 .val = 'o'
 118         },
 119         {
 120                 .name = "batch",
 121                 .has_arg = required_argument,
 122                 .flag = NULL,
 123                 .val = 'b'
 124         },
 125         {
 126                 .name = "per program stats",
 127                 .has_arg = no_argument,
 128                 .flag = NULL,
 129                 .val = 's'
 130         },
 131         {
 132                 .name = "track ios",
 133                 .has_arg = no_argument,
 134                 .flag = NULL,
 135                 .val = 't'
 136         },
 137         {
 138                 .name = "quiet",
 139                 .has_arg = no_argument,
 140                 .flag = NULL,
 141                 .val = 'q'
 142         },
 143         {
 144                 .name = "stopwatch",
 145                 .has_arg = required_argument,
 146                 .flag = NULL,
 147                 .val = 'w'
 148         },
 149         {
 150                 .name = "format",
 151                 .has_arg = required_argument,
 152                 .flag = NULL,
 153                 .val = 'f'
 154         },
 155         {
 156                 .name = "format-spec",
 157                 .has_arg = required_argument,
 158                 .flag = NULL,
 159                 .val = 'F'
 160         },
 161         {
 162                 .name = "version",
 163                 .has_arg = no_argument,
 164                 .flag = NULL,
 165                 .val = 'v'
 166         },
 167 };
 168
 169 /*
 170  * for sorting the displayed output
 171  */
 172 struct trace {
 173         struct blk_io_trace *bit;
 174         struct rb_node rb_node;
 175         struct trace *next;
 176         int skipped;
 177 };
 178
 179 static struct rb_root rb_sort_root;
 180 static struct rb_root rb_track_root;
 181
 182 static struct trace *trace_list;
 183
 184 /*
 185  * allocation cache
 186  */
 187 static struct blk_io_trace *bit_alloc_list;
 188 static struct trace *t_alloc_list;
 189
 190 /*
 191  * for tracking individual ios
 192  */
 193 struct io_track {
 194         struct rb_node rb_node;
 195
 196         dev_t device;
 197         __u64 sector;
 198         __u32 pid;
 199         unsigned long long allocation_time;
 200         unsigned long long queue_time;
 201         unsigned long long dispatch_time;
 202         unsigned long long completion_time;
 203 };
 204
 205 static int ndevices;
 206 static struct per_dev_info *devices;
 207 static char *get_dev_name(struct per_dev_info *, char *, int);
 208
 209 static FILE *ofp;
 210 static char *output_name;
 211
 212 static unsigned long long genesis_time;
 213 static unsigned long long stopwatch_start;      /* start from zero by default */
 214 static unsigned long long stopwatch_end = ULONG_LONG_MAX;       /* "infinity" */
 215
 216 static int per_process_stats;
 217 static int track_ios;
 218
 219 #define RB_BATCH_DEFAULT        (512)
 220 static int rb_batch = RB_BATCH_DEFAULT;
 221
 222 static int pipeline;
 223
 224 #define is_done()       (*(volatile int *)(&done))
 225 static volatile int done;
 226
 227 static inline unsigned long hash_long(unsigned long val)
 228 {
 229 #if __WORDSIZE == 32
 230         val *= 0x9e370001UL;
 231 #elif __WORDSIZE == 64
 232         val *= 0x9e37fffffffc0001UL;
 233 #else
 234 #error unknown word size
 235 #endif
 236
 237         return val >> (__WORDSIZE - PPI_HASH_SHIFT);
 238 }
 239
 240 static inline void add_process_to_hash(struct per_process_info *ppi)
 241 {
 242         const int hash_idx = hash_long(ppi->pid);
 243
 244         ppi->hash_next = ppi_hash[hash_idx];
 245         ppi_hash[hash_idx] = ppi;
 246 }
 247
 248 static inline void add_process_to_list(struct per_process_info *ppi)
 249 {
 250         ppi->list_next = ppi_list;
 251         ppi_list = ppi;
 252         ppi_list_entries++;
 253 }
 254
 255 static struct per_process_info *find_process_by_pid(__u32 pid)
 256 {
 257         const int hash_idx = hash_long(pid);
 258         struct per_process_info *ppi;
 259
 260         ppi = ppi_hash[hash_idx];
 261         while (ppi) {
 262                 if (ppi->pid == pid)
 263                         return ppi;
 264
 265                 ppi = ppi->hash_next;
 266         }
 267
 268         return NULL;
 269 }
 270
 271 static inline int trace_rb_insert(struct trace *t)
 272 {
 273         struct rb_node **p = &rb_sort_root.rb_node;
 274         struct rb_node *parent = NULL;
 275         struct trace *__t;
 276
 277         if (genesis_time == 0 || t->bit->time < genesis_time)
 278                 genesis_time = t->bit->time;
 279
 280         while (*p) {
 281                 parent = *p;
 282                 __t = rb_entry(parent, struct trace, rb_node);
 283
 284                 if (t->bit->time < __t->bit->time)
 285                         p = &(*p)->rb_left;
 286                 else if (t->bit->time > __t->bit->time)
 287                         p = &(*p)->rb_right;
 288                 else if (t->bit->device < __t->bit->device)
 289                         p = &(*p)->rb_left;
 290                 else if (t->bit->device > __t->bit->device)
 291                         p = &(*p)->rb_right;
 292                 else if (t->bit->sequence < __t->bit->sequence)
 293                         p = &(*p)->rb_left;
 294                 else if (t->bit->sequence > __t->bit->sequence)
 295                         p = &(*p)->rb_right;
 296                 else if (t->bit->device == __t->bit->device) {
 297                         fprintf(stderr,
 298                                 "sequence alias (%d) on device %d,%d!\n",
 299                                 t->bit->sequence,
 300                                 MAJOR(t->bit->device), MINOR(t->bit->device));
 301                         return 1;
 302                 }
 303         }
 304
 305         rb_link_node(&t->rb_node, parent, p);
 306         rb_insert_color(&t->rb_node, &rb_sort_root);
 307         return 0;
 308 }
 309
 310 static inline int track_rb_insert(struct io_track *iot)
 311 {
 312         struct rb_node **p = &rb_track_root.rb_node;
 313         struct rb_node *parent = NULL;
 314         struct io_track *__iot;
 315
 316         while (*p) {
 317                 parent = *p;
 318
 319                 __iot = rb_entry(parent, struct io_track, rb_node);
 320
 321                 if (iot->device < __iot->device)
 322                         p = &(*p)->rb_left;
 323                 else if (iot->device > __iot->device)
 324                         p = &(*p)->rb_right;
 325                 else if (iot->sector < __iot->sector)
 326                         p = &(*p)->rb_left;
 327                 else if (iot->sector > __iot->sector)
 328                         p = &(*p)->rb_right;
 329                 else {
 330                         fprintf(stderr,
 331                                 "sector alias (%Lu) on device %d,%d!\n",
 332                                 (unsigned long long) iot->sector,
 333                                 MAJOR(iot->device), MINOR(iot->device));
 334                         return 1;
 335                 }
 336         }
 337
 338         rb_link_node(&iot->rb_node, parent, p);
 339         rb_insert_color(&iot->rb_node, &rb_track_root);
 340         return 0;
 341 }
 342
 343 static struct io_track *__find_track(dev_t device, __u64 sector)
 344 {
 345         struct rb_node **p = &rb_track_root.rb_node;
 346         struct rb_node *parent = NULL;
 347         struct io_track *__iot;
 348
 349         while (*p) {
 350                 parent = *p;
 351
 352                 __iot = rb_entry(parent, struct io_track, rb_node);
 353
 354                 if (device < __iot->device)
 355                         p = &(*p)->rb_left;
 356                 else if (device > __iot->device)
 357                         p = &(*p)->rb_right;
 358                 else if (sector < __iot->sector)
 359                         p = &(*p)->rb_left;
 360                 else if (sector > __iot->sector)
 361                         p = &(*p)->rb_right;
 362                 else
 363                         return __iot;
 364         }
 365
 366         return NULL;
 367 }
 368
 369 static struct io_track *find_track(__u32 pid, dev_t device, __u64 sector)
 370 {
 371         struct io_track *iot;
 372
 373         iot = __find_track(device, sector);
 374         if (!iot) {
 375                 iot = malloc(sizeof(*iot));
 376                 iot->pid = pid;
 377                 iot->device = device;
 378                 iot->sector = sector;
 379                 track_rb_insert(iot);
 380         }
 381
 382         return iot;
 383 }
 384
 385 static void log_track_frontmerge(struct blk_io_trace *t)
 386 {
 387         struct io_track *iot;
 388
 389         if (!track_ios)
 390                 return;
 391
 392         iot = __find_track(t->device, t->sector + (t->bytes >> 9));
 393         if (!iot) {
 394                 fprintf(stderr, "failed to find mergeable event\n");
 395                 return;
 396         }
 397
 398         rb_erase(&iot->rb_node, &rb_track_root);
 399         iot->sector -= t->bytes >> 9;
 400         track_rb_insert(iot);
 401 }
 402
 403 static void log_track_getrq(struct blk_io_trace *t)
 404 {
 405         struct io_track *iot;
 406
 407         if (!track_ios)
 408                 return;
 409
 410         iot = find_track(t->pid, t->device, t->sector);
 411         iot->allocation_time = t->time;
 412 }
 413
 414
 415 /*
 416  * return time between rq allocation and insertion
 417  */
 418 static unsigned long long log_track_insert(struct blk_io_trace *t)
 419 {
 420         unsigned long long elapsed;
 421         struct io_track *iot;
 422
 423         if (!track_ios)
 424                 return -1;
 425
 426         iot = find_track(t->pid, t->device, t->sector);
 427         iot->queue_time = t->time;
 428         elapsed = iot->queue_time - iot->allocation_time;
 429
 430         if (per_process_stats) {
 431                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
 432                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 433
 434                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
 435                         ppi->longest_allocation_wait[w] = elapsed;
 436         }
 437
 438         return elapsed;
 439 }
 440
 441 /*
 442  * return time between queue and issue
 443  */
 444 static unsigned long long log_track_issue(struct blk_io_trace *t)
 445 {
 446         unsigned long long elapsed;
 447         struct io_track *iot;
 448
 449         if (!track_ios)
 450                 return -1;
 451         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 452                 return -1;
 453
 454         iot = __find_track(t->device, t->sector);
 455         if (!iot) {
 456                 fprintf(stderr, "failed to find issue event\n");
 457                 return -1;
 458         }
 459
 460         iot->dispatch_time = t->time;
 461         elapsed = iot->dispatch_time - iot->queue_time;
 462
 463         if (per_process_stats) {
 464                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
 465                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 466
 467                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
 468                         ppi->longest_dispatch_wait[w] = elapsed;
 469         }
 470
 471         return elapsed;
 472 }
 473
 474 /*
 475  * return time between dispatch and complete
 476  */
 477 static unsigned long long log_track_complete(struct blk_io_trace *t)
 478 {
 479         unsigned long long elapsed;
 480         struct io_track *iot;
 481
 482         if (!track_ios)
 483                 return -1;
 484         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 485                 return -1;
 486
 487         iot = __find_track(t->device, t->sector);
 488         if (!iot) {
 489                 fprintf(stderr, "failed to find complete event\n");
 490                 return -1;
 491         }
 492
 493         iot->completion_time = t->time;
 494         elapsed = iot->completion_time - iot->dispatch_time;
 495
 496         if (per_process_stats) {
 497                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
 498                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 499
 500                 if (ppi && elapsed > ppi->longest_completion_wait[w])
 501                         ppi->longest_completion_wait[w] = elapsed;
 502         }
 503
 504         /*
 505          * kill the trace, we don't need it after completion
 506          */
 507         rb_erase(&iot->rb_node, &rb_track_root);
 508         free(iot);
 509
 510         return elapsed;
 511 }
 512
 513
 514 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
 515 {
 516         struct per_process_info *ppi = find_process_by_pid(pid);
 517
 518         if (!ppi) {
 519                 ppi = malloc(sizeof(*ppi));
 520                 memset(ppi, 0, sizeof(*ppi));
 521                 strncpy(ppi->name, name, sizeof(ppi->name));
 522                 ppi->pid = pid;
 523                 add_process_to_hash(ppi);
 524                 add_process_to_list(ppi);
 525         }
 526
 527         return &ppi->io_stats;
 528 }
 529
 530
 531 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
 532 {
 533         struct per_cpu_info *cpus = pdi->cpus;
 534         int ncpus = pdi->ncpus;
 535         int new_count = cpu + 1;
 536         int new_space, size;
 537         char *new_start;
 538
 539         size = new_count * sizeof(struct per_cpu_info);
 540         cpus = realloc(cpus, size);
 541         if (!cpus) {
 542                 char name[20];
 543                 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
 544                         get_dev_name(pdi, name, sizeof(name)), size);
 545                 exit(1);
 546         }
 547
 548         new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
 549         new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
 550         memset(new_start, 0, new_space);
 551
 552         pdi->ncpus = new_count;
 553         pdi->cpus = cpus;
 554 }
 555
 556 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
 557 {
 558         struct per_cpu_info *pci;
 559
 560         if (cpu >= pdi->ncpus)
 561                 resize_cpu_info(pdi, cpu);
 562
 563         pci = &pdi->cpus[cpu];
 564         pci->cpu = cpu;
 565         return pci;
 566 }
 567
 568
 569 static int resize_devices(char *name)
 570 {
 571         int size = (ndevices + 1) * sizeof(struct per_dev_info);
 572
 573         devices = realloc(devices, size);
 574         if (!devices) {
 575                 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
 576                 return 1;
 577         }
 578         memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
 579         devices[ndevices].name = name;
 580         ndevices++;
 581         return 0;
 582 }
 583
 584 static struct per_dev_info *get_dev_info(dev_t id)
 585 {
 586         struct per_dev_info *pdi;
 587         int i;
 588
 589         for (i = 0; i < ndevices; i++) {
 590                 if (!devices[i].id)
 591                         devices[i].id = id;
 592                 if (devices[i].id == id)
 593                         return &devices[i];
 594         }
 595
 596         if (resize_devices(NULL) != 0)
 597                 return NULL;
 598
 599         pdi = &devices[ndevices - 1];
 600         pdi->id = id;
 601         pdi->last_sequence = -1;
 602         return pdi;
 603 }
 604
 605 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
 606 {
 607         if (pdi->name)
 608                 snprintf(buffer, size, "%s", pdi->name);
 609         else
 610                 snprintf(buffer, size, "%d,%d", MAJOR(pdi->id), MINOR(pdi->id));
 611         return buffer;
 612 }
 613
 614 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
 615 {
 616         unsigned long long this = bit->time;
 617         unsigned long long last = pdi->last_reported_time;
 618
 619         pdi->backwards = (this < last) ? 'B' : ' ';
 620         pdi->last_reported_time = this;
 621 }
 622
 623 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
 624                                int rw)
 625 {
 626         if (rw) {
 627                 ios->mwrites++;
 628                 ios->qwrite_kb += t->bytes >> 10;
 629         } else {
 630                 ios->mreads++;
 631                 ios->qread_kb += t->bytes >> 10;
 632         }
 633 }
 634
 635 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
 636                              int rw)
 637 {
 638         __account_m(&pci->io_stats, t, rw);
 639
 640         if (per_process_stats) {
 641                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 642
 643                 __account_m(ios, t, rw);
 644         }
 645 }
 646
 647 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
 648                                    int rw)
 649 {
 650         if (rw) {
 651                 ios->qwrites++;
 652                 ios->qwrite_kb += t->bytes >> 10;
 653         } else {
 654                 ios->qreads++;
 655                 ios->qread_kb += t->bytes >> 10;
 656         }
 657 }
 658
 659 static inline void account_queue(struct blk_io_trace *t,
 660                                  struct per_cpu_info *pci, int rw)
 661 {
 662         __account_queue(&pci->io_stats, t, rw);
 663
 664         if (per_process_stats) {
 665                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 666
 667                 __account_queue(ios, t, rw);
 668         }
 669 }
 670
 671 static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
 672 {
 673         if (rw) {
 674                 ios->cwrites++;
 675                 ios->cwrite_kb += bytes >> 10;
 676         } else {
 677                 ios->creads++;
 678                 ios->cread_kb += bytes >> 10;
 679         }
 680 }
 681
 682 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
 683                              int rw, int bytes)
 684 {
 685         __account_c(&pci->io_stats, rw, bytes);
 686
 687         if (per_process_stats) {
 688                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 689
 690                 __account_c(ios, rw, bytes);
 691         }
 692 }
 693
 694 static inline void __account_issue(struct io_stats *ios, int rw,
 695                                    unsigned int bytes)
 696 {
 697         if (rw) {
 698                 ios->iwrites++;
 699                 ios->iwrite_kb += bytes >> 10;
 700         } else {
 701                 ios->ireads++;
 702                 ios->iread_kb += bytes >> 10;
 703         }
 704 }
 705
 706 static inline void account_issue(struct blk_io_trace *t,
 707                                  struct per_cpu_info *pci, int rw)
 708 {
 709         __account_issue(&pci->io_stats, rw, t->bytes);
 710
 711         if (per_process_stats) {
 712                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 713
 714                 __account_issue(ios, rw, t->bytes);
 715         }
 716 }
 717
 718 static inline void __account_unplug(struct io_stats *ios, int timer)
 719 {
 720         if (timer)
 721                 ios->timer_unplugs++;
 722         else
 723                 ios->io_unplugs++;
 724 }
 725
 726 static inline void account_unplug(struct blk_io_trace *t,
 727                                   struct per_cpu_info *pci, int timer)
 728 {
 729         __account_unplug(&pci->io_stats, timer);
 730
 731         if (per_process_stats) {
 732                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 733
 734                 __account_unplug(ios, timer);
 735         }
 736 }
 737
 738 #define VALID_SPECS     "BCDFGMPQRSTU"
 739 static char *override_format[256];
 740 static inline int valid_spec(int spec)
 741 {
 742         return strchr(VALID_SPECS, spec) != NULL;
 743 }
 744
 745 static void set_all_format_specs(char *optarg)
 746 {
 747         char *p;
 748
 749         for (p = VALID_SPECS; *p; p++)
 750                 if (override_format[(int)(*p)] == NULL)
 751                         override_format[(int)(*p)] = strdup(optarg);
 752 }
 753
 754 static int add_format_spec(char *optarg)
 755 {
 756         int spec = optarg[0];
 757
 758         if (!valid_spec(spec)) {
 759                 fprintf(stderr,"Bad format specifier %c\n", spec);
 760                 return 1;
 761         }
 762         if (optarg[1] != ',') {
 763                 fprintf(stderr,"Bad format specifier - need ',' %s\n", optarg);
 764                 return 1;
 765         }
 766         optarg += 2;
 767         if (*optarg == '\0') {
 768                 fprintf(stderr,"Bad format specifier - need fmt %s\n", optarg);
 769                 return 1;
 770         }
 771
 772         /*
 773          * Set both merges (front and back)
 774          */
 775         if (spec == 'M') {
 776                 override_format['B'] = strdup(optarg);
 777                 override_format['M'] = strdup(optarg);
 778         } else
 779                 override_format[spec] = strdup(optarg);
 780
 781         return 0;
 782 }
 783
 784 static void print_field(char *act, struct per_cpu_info *pci,
 785                         struct blk_io_trace *t, unsigned long long elapsed,
 786                         int pdu_len, unsigned char *pdu_buf, char field,
 787                         int minus, int has_w, int width)
 788 {
 789         char format[64];
 790
 791         if (has_w) {
 792                 if (minus)
 793                         sprintf(format, "%%-%d", width);
 794                 else
 795                         sprintf(format, "%%%d", width);
 796         } else
 797                 sprintf(format, "%%");
 798
 799         switch (field) {
 800         case 'a':
 801                 fprintf(ofp, strcat(format, "s"), act);
 802                 break;
 803         case 'c':
 804                 fprintf(ofp, strcat(format, "d"), pci->cpu);
 805                 break;
 806         case 'C':
 807                 fprintf(ofp, strcat(format, "s"), t->comm);
 808                 break;
 809         case 'd': {
 810                 char rwbs[4];
 811                 int i = 0;
 812                 int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
 813                 int b = t->action & BLK_TC_ACT(BLK_TC_BARRIER);
 814                 int s = t->action & BLK_TC_ACT(BLK_TC_SYNC);
 815                 if (w)
 816                         rwbs[i++] = 'W';
 817                 else
 818                         rwbs[i++] = 'R';
 819                 if (b)
 820                         rwbs[i++] = 'B';
 821                 if (s)
 822                         rwbs[i++] = 'S';
 823                 rwbs[i] = '\0';
 824                 fprintf(ofp, strcat(format, "s"), rwbs);
 825                 break;
 826         }
 827         case 'D':       /* format width ignored */
 828                 fprintf(ofp,"%3d,%-3d", MAJOR(t->device), MINOR(t->device));
 829                 break;
 830         case 'e':
 831                 fprintf(ofp, strcat(format, "d"), t->error);
 832                 break;
 833         case 'M':
 834                 fprintf(ofp, strcat(format, "d"), MAJOR(t->device));
 835                 break;
 836         case 'm':
 837                 fprintf(ofp, strcat(format, "d"), MINOR(t->device));
 838                 break;
 839         case 'n':
 840                 fprintf(ofp, strcat(format, "u"), t->bytes >> 9);
 841                 break;
 842         case 'p':
 843                 fprintf(ofp, strcat(format, "u"), t->pid);
 844                 break;
 845         case 'P':       /* format width ignored */
 846                 if ((pdu_len > 0) && (pdu_buf != NULL)) {
 847                         int i;
 848                         unsigned char *p = pdu_buf;
 849                         for (i = 0; i < pdu_len; i++) {
 850                                 if (i)
 851                                         fprintf(ofp, " ");
 852
 853                                 fprintf(ofp, "%02x", *p++);
 854                         }
 855                 }
 856                 break;
 857         case 's':
 858                 fprintf(ofp, strcat(format, "ld"), t->sequence);
 859                 break;
 860         case 'S':
 861                 fprintf(ofp, strcat(format, "lu"), t->sector);
 862                 break;
 863         case 't':
 864                 sprintf(format, "%%0%dlu", has_w ? width : 9);
 865                 fprintf(ofp, format, NANO_SECONDS(t->time));
 866                 break;
 867         case 'T':
 868                 fprintf(ofp, strcat(format, "d"), SECONDS(t->time));
 869                 break;
 870         case 'u':
 871                 if (elapsed == -1ULL) {
 872                         fprintf(stderr, "Expecting elapsed value\n");
 873                         exit(1);
 874                 }
 875                 fprintf(ofp, strcat(format, "llu"), elapsed / 1000);
 876                 break;
 877         case 'U': {
 878                 __u64 *depth = (__u64 *) ((char *) t + sizeof(*t));
 879                 fprintf(ofp, strcat(format, "u"),
 880                                         (unsigned int) be64_to_cpu(*depth));
 881                 break;
 882         }
 883         default:
 884                 fprintf(ofp,strcat(format, "c"), field);
 885                 break;
 886         }
 887 }
 888
 889 static char *parse_field(char *act, struct per_cpu_info *pci,
 890                          struct blk_io_trace *t, unsigned long long elapsed,
 891                          int pdu_len, unsigned char *pdu_buf,
 892                          char *master_format)
 893 {
 894         int minus = 0;
 895         int has_w = 0;
 896         int width = 0;
 897         char *p = master_format;
 898
 899         if (*p == '-') {
 900                 minus = 1;
 901                 p++;
 902         }
 903         if (isdigit(*p)) {
 904                 has_w = 1;
 905                 do {
 906                         width = (width * 10) + (*p++ - '0');
 907                 } while ((*p) && (isdigit(*p)));
 908         }
 909         if (*p) {
 910                 print_field(act, pci, t, elapsed, pdu_len, pdu_buf, *p++,
 911                             minus, has_w, width);
 912         }
 913         return p;
 914 }
 915
 916 static char *fmt_select(int fmt_spec, struct blk_io_trace *t,
 917                         unsigned long long elapsed)
 918 {
 919         char *fmt;
 920         char scratch_format[1024];
 921
 922         if (override_format[fmt_spec] != NULL)
 923                 return override_format[fmt_spec];
 924
 925         switch (fmt_spec) {
 926         case 'C':       /* Complete */
 927                 if (t->action & BLK_TC_ACT(BLK_TC_PC))
 928                         strcpy(scratch_format, HEADER "(%P) [%e]\n");
 929                 else {
 930                         if (elapsed != -1ULL) {
 931                                 strcpy(scratch_format,
 932                                         HEADER "%S +%n (%8u) [%e]\n");
 933                         } else
 934                                 strcpy(scratch_format, HEADER "%S + %n [%e]\n");
 935                 }
 936                 fmt = scratch_format;
 937                 break;
 938
 939         case 'D':       /* Issue */
 940                 if (t->action & BLK_TC_ACT(BLK_TC_PC))
 941                         strcpy(scratch_format, HEADER "%n (%P) [%C]\n");
 942                 else {
 943                         if (elapsed != -1ULL) {
 944                                 strcpy(scratch_format,
 945                                         HEADER "%S + %n (%8u) [%C]\n");
 946                         } else
 947                                 strcpy(scratch_format, HEADER "%S + %n [%C]\n");
 948                 }
 949                 fmt = scratch_format;
 950                 break;
 951
 952         case 'I':       /* Insert */
 953                 if (t->action & BLK_TC_ACT(BLK_TC_PC))
 954                         strcpy(scratch_format, HEADER "%n (%P) [%C]\n");
 955                 else {
 956                         if (elapsed != -1ULL) {
 957                                 strcpy(scratch_format,
 958                                         HEADER "%S + %n (%8u) [%C]\n");
 959                         } else
 960                                 strcpy(scratch_format, HEADER "%S + %n [%C]\n");
 961                 }
 962                 fmt = scratch_format;
 963                 break;
 964
 965         case 'Q':       /* Queue */
 966         case 'W':       /* Bounce */
 967                 if (elapsed != -1ULL) {
 968                         strcpy(scratch_format, HEADER "%S + %n (%8u) [%C]\n");
 969                 } else
 970                         strcpy(scratch_format, HEADER "%S + %n [%C]\n");
 971                 fmt = scratch_format;
 972                 break;
 973
 974         case 'B':       /* Back merge */
 975         case 'F':       /* Front merge */
 976         case 'M':       /* Front or back merge */
 977                 fmt = HEADER "%S + %n [%C]\n";
 978                 break;
 979
 980         case 'P':       /* Plug */
 981                 fmt = HEADER "[%C]\n";
 982                 break;
 983
 984         case 'G':       /* Get request */
 985         case 'S':       /* Sleep request */
 986                 fmt = HEADER "%S + %n [%C]\n";
 987                 break;
 988
 989         case 'U':       /* Unplug IO */
 990         case 'T':       /* Unplug timer */
 991                 fmt = HEADER "[%C] %U\n";
 992                 break;
 993
 994         case 'X':       /* Split */
 995                 strcpy(scratch_format, HEADER "%S / %U [%C]\n");
 996                 fmt = scratch_format;
 997                 break;
 998
 999         default:
1000                 fprintf(stderr,"FATAL: Invalid format spec %c\n", fmt_spec);
1001                 exit(1);
1002                 /*NOTREACHED*/
1003         }
1004
1005         return fmt;
1006 }
1007
1008 static void process_fmt(char *act, struct per_cpu_info *pci,
1009                            struct blk_io_trace *t, unsigned long long elapsed,
1010                            int pdu_len, unsigned char *pdu_buf)
1011 {
1012         char *p = fmt_select(act[0], t, elapsed);
1013
1014         while (*p) {
1015                 switch (*p) {
1016                 case '%':       /* Field specifier */
1017                         p++;
1018                         if (*p == '%')
1019                                 fprintf(ofp, "%c", *p++);
1020                         else if (!*p)
1021                                 fprintf(ofp, "%c", '%');
1022                         else
1023                                 p = parse_field(act, pci, t, elapsed,
1024                                                 pdu_len, pdu_buf, p);
1025                         break;
1026                 case '\\': {    /* escape */
1027                         switch (p[1]) {
1028                         case 'b': fprintf(ofp, "\b"); break;
1029                         case 'n': fprintf(ofp, "\n"); break;
1030                         case 'r': fprintf(ofp, "\r"); break;
1031                         case 't': fprintf(ofp, "\t"); break;
1032                         default:
1033                                 fprintf(stderr,
1034                                         "Invalid escape char in format %c\n",
1035                                         p[1]);
1036                                 exit(1);
1037                                 /*NOTREACHED*/
1038                         }
1039                         p += 2;
1040                         break;
1041                 }
1042                 default:
1043                         fprintf(ofp, "%c", *p++);
1044                         break;
1045                 }
1046         }
1047 }
1048
1049 static void log_complete(struct per_cpu_info *pci, struct blk_io_trace *t,
1050                          char *act)
1051 {
1052         process_fmt(act, pci, t, log_track_complete(t), 0, NULL);
1053 }
1054
1055 static void log_insert(struct per_cpu_info *pci, struct blk_io_trace *t,
1056                       char *act)
1057 {
1058         process_fmt(act, pci, t, log_track_insert(t), 0, NULL);
1059 }
1060
1061 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
1062                       char *act)
1063 {
1064         process_fmt(act, pci, t, -1, 0, NULL);
1065 }
1066
1067 static void log_issue(struct per_cpu_info *pci, struct blk_io_trace *t,
1068                       char *act)
1069 {
1070         process_fmt(act, pci, t, log_track_issue(t), 0, NULL);
1071 }
1072
1073 static void log_merge(struct per_cpu_info *pci, struct blk_io_trace *t,
1074                       char *act)
1075 {
1076         if (act[0] == 'F')
1077                 log_track_frontmerge(t);
1078
1079         process_fmt(act, pci, t, -1ULL, 0, NULL);
1080 }
1081
1082 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
1083                         char *act)
1084 {
1085         process_fmt(act, pci, t, -1ULL, 0, NULL);
1086 }
1087
1088 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
1089                         char *act)
1090 {
1091         process_fmt(act, pci, t, -1ULL, 0, NULL);
1092 }
1093
1094 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
1095                       char *act)
1096 {
1097         process_fmt(act, pci, t, -1ULL, 0, NULL);
1098 }
1099
1100 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
1101                       char *act)
1102 {
1103         process_fmt(act, pci, t, -1ULL, 0, NULL);
1104 }
1105
1106 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1107 {
1108         unsigned char *buf = (unsigned char *) t + sizeof(*t);
1109
1110         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1111 }
1112
1113 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
1114 {
1115         int act = t->action & 0xffff;
1116
1117         switch (act) {
1118                 case __BLK_TA_QUEUE:
1119                         log_generic(pci, t, "Q");
1120                         break;
1121                 case __BLK_TA_GETRQ:
1122                         log_generic(pci, t, "G");
1123                         break;
1124                 case __BLK_TA_SLEEPRQ:
1125                         log_generic(pci, t, "S");
1126                         break;
1127                 case __BLK_TA_REQUEUE:
1128                         log_generic(pci, t, "R");
1129                         break;
1130                 case __BLK_TA_ISSUE:
1131                         log_pc(pci, t, "D");
1132                         break;
1133                 case __BLK_TA_COMPLETE:
1134                         log_pc(pci, t, "C");
1135                         break;
1136                 case __BLK_TA_INSERT:
1137                         log_pc(pci, t, "I");
1138                         break;
1139                 default:
1140                         fprintf(stderr, "Bad pc action %x\n", act);
1141                         break;
1142         }
1143 }
1144
1145 static void dump_trace_fs(struct blk_io_trace *t, struct per_cpu_info *pci)
1146 {
1147         int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
1148         int act = t->action & 0xffff;
1149
1150         switch (act) {
1151                 case __BLK_TA_QUEUE:
1152                         account_queue(t, pci, w);
1153                         log_queue(pci, t, "Q");
1154                         break;
1155                 case __BLK_TA_INSERT:
1156                         log_insert(pci, t, "I");
1157                         break;
1158                 case __BLK_TA_BACKMERGE:
1159                         account_m(t, pci, w);
1160                         log_merge(pci, t, "M");
1161                         break;
1162                 case __BLK_TA_FRONTMERGE:
1163                         account_m(t, pci, w);
1164                         log_merge(pci, t, "F");
1165                         break;
1166                 case __BLK_TA_GETRQ:
1167                         log_track_getrq(t);
1168                         log_generic(pci, t, "G");
1169                         break;
1170                 case __BLK_TA_SLEEPRQ:
1171                         log_generic(pci, t, "S");
1172                         break;
1173                 case __BLK_TA_REQUEUE:
1174                         account_c(t, pci, w, -t->bytes);
1175                         log_queue(pci, t, "R");
1176                         break;
1177                 case __BLK_TA_ISSUE:
1178                         account_issue(t, pci, w);
1179                         log_issue(pci, t, "D");
1180                         break;
1181                 case __BLK_TA_COMPLETE:
1182                         account_c(t, pci, w, t->bytes);
1183                         log_complete(pci, t, "C");
1184                         break;
1185                 case __BLK_TA_PLUG:
1186                         log_action(pci, t, "P");
1187                         break;
1188                 case __BLK_TA_UNPLUG_IO:
1189                         account_unplug(t, pci, 0);
1190                         log_unplug(pci, t, "U");
1191                         break;
1192                 case __BLK_TA_UNPLUG_TIMER:
1193                         account_unplug(t, pci, 1);
1194                         log_unplug(pci, t, "UT");
1195                         break;
1196                 case __BLK_TA_SPLIT:
1197                         log_split(pci, t, "X");
1198                         break;
1199                 case __BLK_TA_BOUNCE:
1200                         log_generic(pci, t, "B");
1201                         break;
1202                 default:
1203                         fprintf(stderr, "Bad fs action %x\n", t->action);
1204                         break;
1205         }
1206 }
1207
1208 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1209                        struct per_dev_info *pdi)
1210 {
1211         if (t->action & BLK_TC_ACT(BLK_TC_PC))
1212                 dump_trace_pc(t, pci);
1213         else
1214                 dump_trace_fs(t, pci);
1215
1216         pdi->events++;
1217 }
1218
1219 static void dump_io_stats(struct io_stats *ios, char *msg)
1220 {
1221         fprintf(ofp, "%s\n", msg);
1222
1223         fprintf(ofp, " Reads Queued:    %'8lu, %'8LuKiB\t", ios->qreads, ios->qread_kb);
1224         fprintf(ofp, " Writes Queued:    %'8lu, %'8LuKiB\n", ios->qwrites,ios->qwrite_kb);
1225
1226         fprintf(ofp, " Read Dispatches: %'8lu, %'8LuKiB\t", ios->ireads, ios->iread_kb);
1227         fprintf(ofp, " Write Dispatches: %'8lu, %'8LuKiB\n", ios->iwrites,ios->iwrite_kb);
1228         fprintf(ofp, " Reads Completed: %'8lu, %'8LuKiB\t", ios->creads, ios->cread_kb);
1229         fprintf(ofp, " Writes Completed: %'8lu, %'8LuKiB\n", ios->cwrites,ios->cwrite_kb);
1230         fprintf(ofp, " Read Merges:     %'8lu%8c\t", ios->mreads, ' ');
1231         fprintf(ofp, " Write Merges:     %'8lu\n", ios->mwrites);
1232         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
1233         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
1234 }
1235
1236 static void dump_wait_stats(struct per_process_info *ppi)
1237 {
1238         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1239         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1240         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1241         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1242         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1243         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1244
1245         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1246         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
1247         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
1248         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
1249         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1250         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
1251 }
1252
1253 static int ppi_name_compare(const void *p1, const void *p2)
1254 {
1255         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1256         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1257         int res;
1258
1259         res = strverscmp(ppi1->name, ppi2->name);
1260         if (!res)
1261                 res = ppi1->pid > ppi2->pid;
1262
1263         return res;
1264 }
1265
1266 static void sort_process_list(void)
1267 {
1268         struct per_process_info **ppis;
1269         struct per_process_info *ppi;
1270         int i = 0;
1271
1272         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1273
1274         ppi = ppi_list;
1275         while (ppi) {
1276                 ppis[i++] = ppi;
1277                 ppi = ppi->list_next;
1278         }
1279
1280         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1281
1282         i = ppi_list_entries - 1;
1283         ppi_list = NULL;
1284         while (i >= 0) {
1285                 ppi = ppis[i];
1286
1287                 ppi->list_next = ppi_list;
1288                 ppi_list = ppi;
1289                 i--;
1290         }
1291
1292         free(ppis);
1293 }
1294
1295 static void show_process_stats(void)
1296 {
1297         struct per_process_info *ppi;
1298
1299         sort_process_list();
1300
1301         ppi = ppi_list;
1302         while (ppi) {
1303                 char name[64];
1304
1305                 snprintf(name, sizeof(name)-1, "%s (%u)", ppi->name, ppi->pid);
1306                 dump_io_stats(&ppi->io_stats, name);
1307                 dump_wait_stats(ppi);
1308                 ppi = ppi->list_next;
1309         }
1310
1311         fprintf(ofp, "\n");
1312 }
1313
1314 static void show_device_and_cpu_stats(void)
1315 {
1316         struct per_dev_info *pdi;
1317         struct per_cpu_info *pci;
1318         struct io_stats total, *ios;
1319         int i, j, pci_events;
1320         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1321         char name[32];
1322
1323         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1324
1325                 memset(&total, 0, sizeof(total));
1326                 pci_events = 0;
1327
1328                 if (i > 0)
1329                         fprintf(ofp, "\n");
1330
1331                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1332                         if (!pci->nelems)
1333                                 continue;
1334
1335                         ios = &pci->io_stats;
1336                         total.qreads += ios->qreads;
1337                         total.qwrites += ios->qwrites;
1338                         total.creads += ios->creads;
1339                         total.cwrites += ios->cwrites;
1340                         total.mreads += ios->mreads;
1341                         total.mwrites += ios->mwrites;
1342                         total.ireads += ios->ireads;
1343                         total.iwrites += ios->iwrites;
1344                         total.qread_kb += ios->qread_kb;
1345                         total.qwrite_kb += ios->qwrite_kb;
1346                         total.cread_kb += ios->cread_kb;
1347                         total.cwrite_kb += ios->cwrite_kb;
1348                         total.iread_kb += ios->iread_kb;
1349                         total.iwrite_kb += ios->iwrite_kb;
1350                         total.timer_unplugs += ios->timer_unplugs;
1351                         total.io_unplugs += ios->io_unplugs;
1352
1353                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1354                                  j, get_dev_name(pdi, name, sizeof(name)));
1355                         dump_io_stats(ios, line);
1356                         pci_events++;
1357                 }
1358
1359                 if (pci_events > 1) {
1360                         fprintf(ofp, "\n");
1361                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1362                                  get_dev_name(pdi, name, sizeof(name)));
1363                         dump_io_stats(&total, line);
1364                 }
1365
1366                 fprintf(ofp, "\nEvents (%s): %'Lu entries, %'lu skips\n",
1367                         get_dev_name(pdi, line, sizeof(line)), pdi->events,
1368                         pdi->skips);
1369         }
1370 }
1371
1372 static int sort_entries(void)
1373 {
1374         struct trace *t;
1375         int nr = 0;
1376
1377         while ((t = trace_list) != NULL) {
1378                 trace_list = t->next;
1379
1380                 if (verify_trace(t->bit))
1381                         continue;
1382                 if (trace_rb_insert(t))
1383                         break;
1384                 nr++;
1385         }
1386
1387         return nr;
1388 }
1389
1390 /*
1391  * struct trace and blktrace allocation cache, we do potentially
1392  * millions of mallocs for these structures while only using at most
1393  * a few thousand at the time
1394  */
1395 static inline void t_free(struct trace *t)
1396 {
1397         t->next = t_alloc_list;
1398         t_alloc_list = t;
1399 }
1400
1401 static inline struct trace *t_alloc(void)
1402 {
1403         struct trace *t = t_alloc_list;
1404
1405         if (t) {
1406                 t_alloc_list = t->next;
1407                 return t;
1408         }
1409
1410         return malloc(sizeof(*t));
1411 }
1412
1413 static inline void bit_free(struct blk_io_trace *bit)
1414 {
1415         /*
1416          * abuse a 64-bit field for a next pointer for the free item
1417          */
1418         bit->time = (__u64) bit_alloc_list;
1419         bit_alloc_list = (struct blk_io_trace *) bit;
1420 }
1421
1422 static inline struct blk_io_trace *bit_alloc(void)
1423 {
1424         struct blk_io_trace *bit = bit_alloc_list;
1425
1426         if (bit) {
1427                 bit_alloc_list = (struct blk_io_trace *) bit->time;
1428                 return bit;
1429         }
1430
1431         return malloc(sizeof(*bit));
1432 }
1433
1434 static void show_entries_rb(void)
1435 {
1436         struct per_dev_info *pdi = NULL;
1437         struct per_cpu_info *pci = NULL;
1438         struct blk_io_trace *bit;
1439         struct rb_node *n;
1440         struct trace *t;
1441         __u32 device = 0;
1442         int cpu = 0;
1443
1444         while ((n = rb_first(&rb_sort_root)) != NULL) {
1445
1446                 if (done)
1447                         break;
1448
1449                 t = rb_entry(n, struct trace, rb_node);
1450                 bit = t->bit;
1451
1452                 if (!pdi || device != bit->device) {
1453                         device = bit->device;
1454                         pdi = get_dev_info(device);
1455                 }
1456
1457                 if (!pdi) {
1458                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1459                                 MAJOR(bit->device), MINOR(bit->device));
1460                         break;
1461                 }
1462
1463                 if (bit->cpu > pdi->ncpus) {
1464                         fprintf(stderr, "Unknown CPU ID? (%d, device %d,%d)\n",
1465                                 bit->cpu, MAJOR(bit->device),
1466                                 MINOR(bit->device));
1467                         break;
1468                 }
1469
1470                 if (!pci || cpu != bit->cpu) {
1471                         cpu = bit->cpu;
1472                         pci = get_cpu_info(pdi, cpu);
1473                 }
1474
1475                 /*
1476                  * back off displaying more info if we are out of sync
1477                  * on SMP systems. to prevent stalling on lost events,
1478                  * only allow an event to us a few times
1479                  */
1480                 if (bit->sequence != (pdi->last_sequence + 1)
1481                     && pdi->last_sequence != -1) {
1482                         if (t->skipped < 5) {
1483                                 t->skipped++;
1484                                 break;
1485                         } else {
1486                                 fprintf(stderr, "skipping from %lu to %u\n", pdi->last_sequence, bit->sequence);
1487                                 pdi->skips++;
1488                         }
1489                 }
1490
1491                 pdi->last_sequence = bit->sequence;
1492
1493                 bit->time -= genesis_time;
1494                 if (bit->time >= stopwatch_end)
1495                         break;
1496
1497                 if (bit->time >= stopwatch_start) {
1498                         check_time(pdi, bit);
1499
1500                         dump_trace(bit, pci, pdi);
1501                 }
1502
1503                 rb_erase(&t->rb_node, &rb_sort_root);
1504                 bit_free(bit);
1505                 t_free(t);
1506         }
1507 }
1508
1509 static int read_data(int fd, void *buffer, int bytes, int block)
1510 {
1511         int ret, bytes_left, fl;
1512         void *p;
1513
1514         fl = fcntl(fd, F_GETFL);
1515
1516         if (!block)
1517                 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1518         else
1519                 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1520
1521         bytes_left = bytes;
1522         p = buffer;
1523         while (bytes_left > 0) {
1524                 ret = read(fd, p, bytes_left);
1525                 if (!ret)
1526                         return 1;
1527                 else if (ret < 0) {
1528                         if (errno != EAGAIN)
1529                                 perror("read");
1530                         return -1;
1531                 } else {
1532                         p += ret;
1533                         bytes_left -= ret;
1534                 }
1535         }
1536
1537         return 0;
1538 }
1539
1540 static int read_sort_events(int fd)
1541 {
1542         int events = 0;
1543
1544         while (!is_done() && events < rb_batch) {
1545                 struct blk_io_trace *bit;
1546                 struct trace *t;
1547                 int pdu_len;
1548                 __u32 magic;
1549
1550                 bit = bit_alloc();
1551
1552                 if (read_data(fd, bit, sizeof(*bit), !events))
1553                         break;
1554
1555                 magic = be32_to_cpu(bit->magic);
1556                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1557                         fprintf(stderr, "Bad magic %x\n", magic);
1558                         break;
1559                 }
1560
1561                 pdu_len = be16_to_cpu(bit->pdu_len);
1562                 if (pdu_len) {
1563                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
1564
1565                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1))
1566                                 break;
1567
1568                         bit = ptr;
1569                 }
1570
1571                 t = t_alloc();
1572                 memset(t, 0, sizeof(*t));
1573                 t->bit = bit;
1574
1575                 trace_to_cpu(bit);
1576                 t->next = trace_list;
1577                 trace_list = t;
1578
1579                 events++;
1580         }
1581
1582         return events;
1583 }
1584
1585 static int do_file(void)
1586 {
1587         struct per_cpu_info *pci;
1588         int i, j, *nfiles, events, events_added;
1589
1590         nfiles = malloc(ndevices * sizeof(int));
1591
1592         /*
1593          * first prepare all files for reading
1594          */
1595         for (i = 0; i < ndevices; i++) {
1596                 nfiles[i] = 0;
1597                 for (j = 0;; j++) {
1598                         struct per_dev_info *pdi;
1599                         struct stat st;
1600
1601                         pdi = &devices[i];
1602                         pdi->last_sequence = -1;
1603                         pci = get_cpu_info(pdi, j);
1604                         pci->cpu = j;
1605                         pci->fd = -1;
1606
1607                         snprintf(pci->fname, sizeof(pci->fname)-1,
1608                                  "%s.blktrace.%d", pdi->name, pci->cpu);
1609                         if (stat(pci->fname, &st) < 0)
1610                                 break;
1611                         if (st.st_size) {
1612                                 pci->fd = open(pci->fname, O_RDONLY);
1613                                 if (pci->fd < 0) {
1614                                         perror(pci->fname);
1615                                         continue;
1616                                 }
1617                         }
1618
1619                         printf("Input file %s added\n", pci->fname);
1620                         nfiles[i] += 1;
1621                 }
1622         }
1623
1624         /*
1625          * now loop over the files reading in the data
1626          */
1627         do {
1628                 events_added = 0;
1629
1630                 for (i = 0; i < ndevices; i++) {
1631                         for (j = 0; j < nfiles[i]; j++) {
1632
1633                                 pci = get_cpu_info(&devices[i], j);
1634
1635                                 if (pci->fd == -1)
1636                                         continue;
1637
1638                                 events = read_sort_events(pci->fd);
1639                                 if (!events) {
1640                                         close(pci->fd);
1641                                         pci->fd = -1;
1642                                         continue;
1643                                 }
1644
1645                                 if (sort_entries() == -1)
1646                                         continue;
1647
1648                                 events_added += events;
1649                         }
1650                 }
1651
1652                 show_entries_rb();
1653
1654         } while (events_added);
1655
1656         return 0;
1657 }
1658
1659 static int do_stdin(void)
1660 {
1661         int fd;
1662
1663         fd = dup(STDIN_FILENO);
1664         do {
1665                 int events;
1666
1667                 events = read_sort_events(fd);
1668                 if (!events)
1669                         break;
1670
1671                 if (sort_entries() == -1)
1672                         break;
1673
1674                 show_entries_rb();
1675         } while (1);
1676
1677         close(fd);
1678         return 0;
1679 }
1680
1681 static void flush_output(void)
1682 {
1683         fflush(ofp);
1684 }
1685
1686 static void handle_sigint(int sig)
1687 {
1688         done = 1;
1689         flush_output();
1690 }
1691
1692 /*
1693  * Extract start and duration times from a string, allowing
1694  * us to specify a time interval of interest within a trace.
1695  * Format: "duration" (start is zero) or "start:duration".
1696  */
1697 static int find_stopwatch_interval(char *string)
1698 {
1699         double value;
1700         char *sp;
1701
1702         value = strtod(string, &sp);
1703         if (sp == string) {
1704                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1705                 return 1;
1706         }
1707         if (*sp == ':') {
1708                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1709                 string = sp + 1;
1710                 value = strtod(string, &sp);
1711                 if (sp == string || *sp != '\0') {
1712                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1713                                 string);
1714                         return 1;
1715                 }
1716         } else if (*sp != '\0') {
1717                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
1718                 return 1;
1719         }
1720         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
1721         if (stopwatch_end <= stopwatch_start) {
1722                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
1723                         stopwatch_start, stopwatch_end);
1724                 return 1;
1725         }
1726
1727         return 0;
1728 }
1729
1730 static char usage_str[] = \
1731         "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
1732         "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
1733         "\t-i Input file containing trace data, or '-' for stdin\n" \
1734         "\t-o Output file. If not given, output is stdout\n" \
1735         "\t-b stdin read batching\n" \
1736         "\t-s Show per-program io statistics\n" \
1737         "\t-t Track individual ios. Will tell you the time a request took\n" \
1738         "\t   to get queued, to get dispatched, and to get completed\n" \
1739         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
1740         "\t-w Only parse data between the given time interval in seconds.\n" \
1741         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
1742         "\t -f Output format. Customize the output format. The format field\n" \
1743         "\t    identifies can be found in the documentation\n" \
1744         "\t-F Format specification. Can be found in the documentation\n" \
1745         "\t-v Print program version info\n\n";
1746
1747 static void usage(char *prog)
1748 {
1749         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
1750 }
1751
1752 int main(int argc, char *argv[])
1753 {
1754         char *ofp_buffer;
1755         int c, ret, mode;
1756         int per_device_and_cpu_stats = 1;
1757
1758         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
1759                 switch (c) {
1760                 case 'i':
1761                         if (!strcmp(optarg, "-") && !pipeline)
1762                                 pipeline = 1;
1763                         else if (resize_devices(optarg) != 0)
1764                                 return 1;
1765                         break;
1766                 case 'o':
1767                         output_name = optarg;
1768                         break;
1769                 case 'b':
1770                         rb_batch = atoi(optarg);
1771                         if (rb_batch <= 0)
1772                                 rb_batch = RB_BATCH_DEFAULT;
1773                         break;
1774                 case 's':
1775                         per_process_stats = 1;
1776                         break;
1777                 case 't':
1778                         track_ios = 1;
1779                         break;
1780                 case 'q':
1781                         per_device_and_cpu_stats = 0;
1782                         break;
1783                 case 'w':
1784                         if (find_stopwatch_interval(optarg) != 0)
1785                                 return 1;
1786                         break;
1787                 case 'f':
1788                         set_all_format_specs(optarg);
1789                         break;
1790                 case 'F':
1791                         if (add_format_spec(optarg) != 0)
1792                                 return 1;
1793                         break;
1794                 case 'v':
1795                         printf("%s version %s\n", argv[0], blkparse_version);
1796                         return 0;
1797                 default:
1798                         usage(argv[0]);
1799                         return 1;
1800                 }
1801         }
1802
1803         while (optind < argc) {
1804                 if (!strcmp(argv[optind], "-") && !pipeline)
1805                         pipeline = 1;
1806                 else if (resize_devices(argv[optind]) != 0)
1807                         return 1;
1808                 optind++;
1809         }
1810
1811         if (!pipeline && !ndevices) {
1812                 usage(argv[0]);
1813                 return 1;
1814         }
1815
1816         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
1817         memset(&rb_track_root, 0, sizeof(rb_track_root));
1818
1819         signal(SIGINT, handle_sigint);
1820         signal(SIGHUP, handle_sigint);
1821         signal(SIGTERM, handle_sigint);
1822
1823         setlocale(LC_NUMERIC, "en_US");
1824
1825         if (!output_name) {
1826                 ofp = fdopen(STDOUT_FILENO, "w");
1827                 mode = _IOLBF;
1828         } else {
1829                 char ofname[128];
1830
1831                 snprintf(ofname, sizeof(ofname) - 1, "%s.log", output_name);
1832                 ofp = fopen(ofname, "w");
1833                 mode = _IOFBF;
1834         }
1835
1836         if (!ofp) {
1837                 perror("fopen");
1838                 return 1;
1839         }
1840
1841         ofp_buffer = malloc(4096);
1842         if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
1843                 perror("setvbuf");
1844                 return 1;
1845         }
1846
1847         if (pipeline)
1848                 ret = do_stdin();
1849         else
1850                 ret = do_file();
1851
1852         if (per_process_stats)
1853                 show_process_stats();
1854
1855         if (per_device_and_cpu_stats)
1856                 show_device_and_cpu_stats();
1857
1858         flush_output();
1859         return ret;
1860 }