blkparse.c

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