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