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 <libgen.h>
  33
  34 #include "blktrace.h"
  35 #include "rbtree.h"
  36 #include "jhash.h"
  37
  38 static char blkparse_version[] = "0.99";
  39
  40 struct per_dev_info {
  41         dev_t dev;
  42         char *name;
  43
  44         int backwards;
  45         unsigned long long events;
  46         unsigned long long first_reported_time;
  47         unsigned long long last_reported_time;
  48         unsigned long long last_read_time;
  49         struct io_stats io_stats;
  50         unsigned long last_sequence;
  51         unsigned long skips;
  52
  53         struct rb_root rb_last;
  54         unsigned long rb_last_entries;
  55
  56         struct rb_root rb_track;
  57
  58         int nfiles;
  59         int ncpus;
  60
  61         unsigned long *cpu_map;
  62         unsigned int cpu_map_max;
  63
  64         struct per_cpu_info *cpus;
  65 };
  66
  67 struct per_process_info {
  68         char name[16];
  69         __u32 pid;
  70         struct io_stats io_stats;
  71         struct per_process_info *hash_next, *list_next;
  72         int more_than_one;
  73
  74         /*
  75          * individual io stats
  76          */
  77         unsigned long long longest_allocation_wait[2];
  78         unsigned long long longest_dispatch_wait[2];
  79         unsigned long long longest_completion_wait[2];
  80 };
  81
  82 #define PPI_HASH_SHIFT  (8)
  83 #define PPI_HASH_SIZE   (1 << PPI_HASH_SHIFT)
  84 #define PPI_HASH_MASK   (PPI_HASH_SIZE - 1)
  85 static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
  86 static struct per_process_info *ppi_list;
  87 static int ppi_list_entries;
  88
  89 #define S_OPTS  "a:A:i:o:b:stqw:f:F:vVhD:"
  90 static struct option l_opts[] = {
  91         {
  92                 .name = "act-mask",
  93                 .has_arg = required_argument,
  94                 .flag = NULL,
  95                 .val = 'a'
  96         },
  97         {
  98                 .name = "set-mask",
  99                 .has_arg = required_argument,
 100                 .flag = NULL,
 101                 .val = 'A'
 102         },
 103         {
 104                 .name = "input",
 105                 .has_arg = required_argument,
 106                 .flag = NULL,
 107                 .val = 'i'
 108         },
 109         {
 110                 .name = "output",
 111                 .has_arg = required_argument,
 112                 .flag = NULL,
 113                 .val = 'o'
 114         },
 115         {
 116                 .name = "batch",
 117                 .has_arg = required_argument,
 118                 .flag = NULL,
 119                 .val = 'b'
 120         },
 121         {
 122                 .name = "per-program-stats",
 123                 .has_arg = no_argument,
 124                 .flag = NULL,
 125                 .val = 's'
 126         },
 127         {
 128                 .name = "track-ios",
 129                 .has_arg = no_argument,
 130                 .flag = NULL,
 131                 .val = 't'
 132         },
 133         {
 134                 .name = "quiet",
 135                 .has_arg = no_argument,
 136                 .flag = NULL,
 137                 .val = 'q'
 138         },
 139         {
 140                 .name = "stopwatch",
 141                 .has_arg = required_argument,
 142                 .flag = NULL,
 143                 .val = 'w'
 144         },
 145         {
 146                 .name = "format",
 147                 .has_arg = required_argument,
 148                 .flag = NULL,
 149                 .val = 'f'
 150         },
 151         {
 152                 .name = "format-spec",
 153                 .has_arg = required_argument,
 154                 .flag = NULL,
 155                 .val = 'F'
 156         },
 157         {
 158                 .name = "hash-by-name",
 159                 .has_arg = no_argument,
 160                 .flag = NULL,
 161                 .val = 'h'
 162         },
 163         {
 164                 .name = "verbose",
 165                 .has_arg = no_argument,
 166                 .flag = NULL,
 167                 .val = 'v'
 168         },
 169         {
 170                 .name = "version",
 171                 .has_arg = no_argument,
 172                 .flag = NULL,
 173                 .val = 'V'
 174         },
 175         {
 176                 .name = "input-directory",
 177                 .has_arg = required_argument,
 178                 .flag = NULL,
 179                 .val = 'D'
 180         },
 181         {
 182                 .name = NULL,
 183         }
 184 };
 185
 186 /*
 187  * for sorting the displayed output
 188  */
 189 struct trace {
 190         struct blk_io_trace *bit;
 191         struct rb_node rb_node;
 192         struct trace *next;
 193 };
 194
 195 static struct rb_root rb_sort_root;
 196 static unsigned long rb_sort_entries;
 197
 198 static struct trace *trace_list;
 199
 200 /*
 201  * allocation cache
 202  */
 203 static struct blk_io_trace *bit_alloc_list;
 204 static struct trace *t_alloc_list;
 205
 206 /*
 207  * for tracking individual ios
 208  */
 209 struct io_track {
 210         struct rb_node rb_node;
 211
 212         __u64 sector;
 213         __u32 pid;
 214         char comm[16];
 215         unsigned long long allocation_time;
 216         unsigned long long queue_time;
 217         unsigned long long dispatch_time;
 218         unsigned long long completion_time;
 219 };
 220
 221 static int ndevices;
 222 static struct per_dev_info *devices;
 223 static char *get_dev_name(struct per_dev_info *, char *, int);
 224
 225 FILE *ofp = NULL;
 226 static char *output_name;
 227 static char *input_dir;
 228
 229 static unsigned long long genesis_time;
 230 static unsigned long long last_allowed_time;
 231 static unsigned int smallest_seq_read;
 232 static unsigned long long stopwatch_start;      /* start from zero by default */
 233 static unsigned long long stopwatch_end = -1ULL;        /* "infinity" */
 234
 235 static int per_process_stats;
 236 static int per_device_and_cpu_stats = 1;
 237 static int track_ios;
 238 static int ppi_hash_by_pid = 1;
 239 static int verbose;
 240 static unsigned int act_mask = -1U;
 241 static int stats_printed;
 242
 243 static unsigned int t_alloc_cache;
 244 static unsigned int bit_alloc_cache;
 245
 246 #define RB_BATCH_DEFAULT        (512)
 247 static unsigned int rb_batch = RB_BATCH_DEFAULT;
 248
 249 static int pipeline;
 250
 251 #define is_done()       (*(volatile int *)(&done))
 252 static volatile int done;
 253
 254 #define JHASH_RANDOM    (0x3af5f2ee)
 255
 256 #define CPUS_PER_LONG   (8 * sizeof(unsigned long))
 257 #define CPU_IDX(cpu)    ((cpu) / CPUS_PER_LONG)
 258 #define CPU_BIT(cpu)    ((cpu) & (CPUS_PER_LONG - 1))
 259
 260 static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu)
 261 {
 262         if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) {
 263                 int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1);
 264                 unsigned long *map = malloc(new_max / sizeof(long));
 265
 266                 memset(map, 0, new_max / sizeof(long));
 267
 268                 if (pdi->cpu_map) {
 269                         memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long));
 270                         free(pdi->cpu_map);
 271                 }
 272
 273                 pdi->cpu_map = map;
 274                 pdi->cpu_map_max = new_max;
 275         }
 276
 277         pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
 278 }
 279
 280 static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu)
 281 {
 282         pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu));
 283 }
 284
 285 static inline int cpu_is_online(struct per_dev_info *pdi, int cpu)
 286 {
 287         return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0;
 288 }
 289
 290 static inline int ppi_hash_pid(__u32 pid)
 291 {
 292         return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
 293 }
 294
 295 static inline int ppi_hash_name(const char *name)
 296 {
 297         return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
 298 }
 299
 300 static inline int ppi_hash(struct per_process_info *ppi)
 301 {
 302         if (ppi_hash_by_pid)
 303                 return ppi_hash_pid(ppi->pid);
 304
 305         return ppi_hash_name(ppi->name);
 306 }
 307
 308 static inline void add_process_to_hash(struct per_process_info *ppi)
 309 {
 310         const int hash_idx = ppi_hash(ppi);
 311
 312         ppi->hash_next = ppi_hash_table[hash_idx];
 313         ppi_hash_table[hash_idx] = ppi;
 314 }
 315
 316 static inline void add_process_to_list(struct per_process_info *ppi)
 317 {
 318         ppi->list_next = ppi_list;
 319         ppi_list = ppi;
 320         ppi_list_entries++;
 321 }
 322
 323 static struct per_process_info *find_process_by_name(char *name)
 324 {
 325         const int hash_idx = ppi_hash_name(name);
 326         struct per_process_info *ppi;
 327
 328         ppi = ppi_hash_table[hash_idx];
 329         while (ppi) {
 330                 if (!strcmp(ppi->name, name))
 331                         return ppi;
 332
 333                 ppi = ppi->hash_next;
 334         }
 335
 336         return NULL;
 337 }
 338
 339 static struct per_process_info *find_process_by_pid(__u32 pid)
 340 {
 341         const int hash_idx = ppi_hash_pid(pid);
 342         struct per_process_info *ppi;
 343
 344         ppi = ppi_hash_table[hash_idx];
 345         while (ppi) {
 346                 if (ppi->pid == pid)
 347                         return ppi;
 348
 349                 ppi = ppi->hash_next;
 350         }
 351
 352         return NULL;
 353 }
 354
 355 static struct per_process_info *find_process(__u32 pid, char *name)
 356 {
 357         struct per_process_info *ppi;
 358
 359         if (ppi_hash_by_pid)
 360                 return find_process_by_pid(pid);
 361
 362         ppi = find_process_by_name(name);
 363         if (ppi && ppi->pid != pid)
 364                 ppi->more_than_one = 1;
 365
 366         return ppi;
 367 }
 368
 369 static inline int trace_rb_insert(struct trace *t, struct rb_root *root,
 370                                   int check_time)
 371 {
 372         struct rb_node **p = &root->rb_node;
 373         struct rb_node *parent = NULL;
 374         struct trace *__t;
 375
 376         while (*p) {
 377                 parent = *p;
 378
 379                 __t = rb_entry(parent, struct trace, rb_node);
 380
 381                 if (check_time) {
 382                         if (t->bit->time < __t->bit->time) {
 383                                 p = &(*p)->rb_left;
 384                                 continue;
 385                         } else if (t->bit->time > __t->bit->time) {
 386                                 p = &(*p)->rb_right;
 387                                 continue;
 388                         }
 389                 }
 390                 if (t->bit->device < __t->bit->device)
 391                         p = &(*p)->rb_left;
 392                 else if (t->bit->device > __t->bit->device)
 393                         p = &(*p)->rb_right;
 394                 else if (t->bit->sequence < __t->bit->sequence)
 395                         p = &(*p)->rb_left;
 396                 else    /* >= sequence */
 397                         p = &(*p)->rb_right;
 398         }
 399
 400         rb_link_node(&t->rb_node, parent, p);
 401         rb_insert_color(&t->rb_node, root);
 402         return 0;
 403 }
 404
 405 static inline int trace_rb_insert_sort(struct trace *t)
 406 {
 407         if (!trace_rb_insert(t, &rb_sort_root, 1)) {
 408                 rb_sort_entries++;
 409                 return 0;
 410         }
 411
 412         return 1;
 413 }
 414
 415 static inline int trace_rb_insert_last(struct per_dev_info *pdi,struct trace *t)
 416 {
 417         if (!trace_rb_insert(t, &pdi->rb_last, 1)) {
 418                 pdi->rb_last_entries++;
 419                 return 0;
 420         }
 421
 422         return 1;
 423 }
 424
 425 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
 426                                    struct rb_root *root, int order)
 427 {
 428         struct rb_node *n = root->rb_node;
 429         struct rb_node *prev = NULL;
 430         struct trace *__t;
 431
 432         while (n) {
 433                 __t = rb_entry(n, struct trace, rb_node);
 434                 prev = n;
 435
 436                 if (device < __t->bit->device)
 437                         n = n->rb_left;
 438                 else if (device > __t->bit->device)
 439                         n = n->rb_right;
 440                 else if (sequence < __t->bit->sequence)
 441                         n = n->rb_left;
 442                 else if (sequence > __t->bit->sequence)
 443                         n = n->rb_right;
 444                 else
 445                         return __t;
 446         }
 447
 448         /*
 449          * hack - the list may not be sequence ordered because some
 450          * events don't have sequence and time matched. so we end up
 451          * being a little off in the rb lookup here, because we don't
 452          * know the time we are looking for. compensate by browsing
 453          * a little ahead from the last entry to find the match
 454          */
 455         if (order && prev) {
 456                 int max = 5;
 457
 458                 while (((n = rb_next(prev)) != NULL) && max--) {
 459                         __t = rb_entry(n, struct trace, rb_node);
 460
 461                         if (__t->bit->device == device &&
 462                             __t->bit->sequence == sequence)
 463                                 return __t;
 464
 465                         prev = n;
 466                 }
 467         }
 468
 469         return NULL;
 470 }
 471
 472 static inline struct trace *trace_rb_find_sort(dev_t dev, unsigned long seq)
 473 {
 474         return trace_rb_find(dev, seq, &rb_sort_root, 1);
 475 }
 476
 477 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
 478                                                unsigned long seq)
 479 {
 480         return trace_rb_find(pdi->dev, seq, &pdi->rb_last, 0);
 481 }
 482
 483 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
 484 {
 485         struct rb_node **p = &pdi->rb_track.rb_node;
 486         struct rb_node *parent = NULL;
 487         struct io_track *__iot;
 488
 489         while (*p) {
 490                 parent = *p;
 491                 __iot = rb_entry(parent, struct io_track, rb_node);
 492
 493                 if (iot->sector < __iot->sector)
 494                         p = &(*p)->rb_left;
 495                 else if (iot->sector > __iot->sector)
 496                         p = &(*p)->rb_right;
 497                 else {
 498                         fprintf(stderr,
 499                                 "sector alias (%Lu) on device %d,%d!\n",
 500                                 (unsigned long long) iot->sector,
 501                                 MAJOR(pdi->dev), MINOR(pdi->dev));
 502                         return 1;
 503                 }
 504         }
 505
 506         rb_link_node(&iot->rb_node, parent, p);
 507         rb_insert_color(&iot->rb_node, &pdi->rb_track);
 508         return 0;
 509 }
 510
 511 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
 512 {
 513         struct rb_node *n = pdi->rb_track.rb_node;
 514         struct io_track *__iot;
 515
 516         while (n) {
 517                 __iot = rb_entry(n, struct io_track, rb_node);
 518
 519                 if (sector < __iot->sector)
 520                         n = n->rb_left;
 521                 else if (sector > __iot->sector)
 522                         n = n->rb_right;
 523                 else
 524                         return __iot;
 525         }
 526
 527         return NULL;
 528 }
 529
 530 static struct io_track *find_track(struct per_dev_info *pdi, __u32 pid,
 531                                    char *comm, __u64 sector)
 532 {
 533         struct io_track *iot;
 534
 535         iot = __find_track(pdi, sector);
 536         if (!iot) {
 537                 iot = malloc(sizeof(*iot));
 538                 iot->pid = pid;
 539                 memcpy(iot->comm, comm, sizeof(iot->comm));
 540                 iot->sector = sector;
 541                 track_rb_insert(pdi, iot);
 542         }
 543
 544         return iot;
 545 }
 546
 547 static void log_track_frontmerge(struct per_dev_info *pdi,
 548                                  struct blk_io_trace *t)
 549 {
 550         struct io_track *iot;
 551
 552         if (!track_ios)
 553                 return;
 554
 555         iot = __find_track(pdi, t->sector + t_sec(t));
 556         if (!iot) {
 557                 if (verbose)
 558                         fprintf(stderr, "merge not found for (%d,%d): %llu\n",
 559                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 560                                 (unsigned long long) t->sector + t_sec(t));
 561                 return;
 562         }
 563
 564         rb_erase(&iot->rb_node, &pdi->rb_track);
 565         iot->sector -= t_sec(t);
 566         track_rb_insert(pdi, iot);
 567 }
 568
 569 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
 570 {
 571         struct io_track *iot;
 572
 573         if (!track_ios)
 574                 return;
 575
 576         iot = find_track(pdi, t->pid, t->comm, t->sector);
 577         iot->allocation_time = t->time;
 578 }
 579
 580 /*
 581  * return time between rq allocation and insertion
 582  */
 583 static unsigned long long log_track_insert(struct per_dev_info *pdi,
 584                                            struct blk_io_trace *t)
 585 {
 586         unsigned long long elapsed;
 587         struct io_track *iot;
 588
 589         if (!track_ios)
 590                 return -1;
 591
 592         iot = find_track(pdi, t->pid, t->comm, t->sector);
 593         iot->queue_time = t->time;
 594
 595         if (!iot->allocation_time)
 596                 return -1;
 597
 598         elapsed = iot->queue_time - iot->allocation_time;
 599
 600         if (per_process_stats) {
 601                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 602                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 603
 604                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
 605                         ppi->longest_allocation_wait[w] = elapsed;
 606         }
 607
 608         return elapsed;
 609 }
 610
 611 /*
 612  * return time between queue and issue
 613  */
 614 static unsigned long long log_track_issue(struct per_dev_info *pdi,
 615                                           struct blk_io_trace *t)
 616 {
 617         unsigned long long elapsed;
 618         struct io_track *iot;
 619
 620         if (!track_ios)
 621                 return -1;
 622         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 623                 return -1;
 624
 625         iot = __find_track(pdi, t->sector);
 626         if (!iot) {
 627                 if (verbose)
 628                         fprintf(stderr, "issue not found for (%d,%d): %llu\n",
 629                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 630                                 (unsigned long long) t->sector);
 631                 return -1;
 632         }
 633
 634         iot->dispatch_time = t->time;
 635         elapsed = iot->dispatch_time - iot->queue_time;
 636
 637         if (per_process_stats) {
 638                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 639                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 640
 641                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
 642                         ppi->longest_dispatch_wait[w] = elapsed;
 643         }
 644
 645         return elapsed;
 646 }
 647
 648 /*
 649  * return time between dispatch and complete
 650  */
 651 static unsigned long long log_track_complete(struct per_dev_info *pdi,
 652                                              struct blk_io_trace *t)
 653 {
 654         unsigned long long elapsed;
 655         struct io_track *iot;
 656
 657         if (!track_ios)
 658                 return -1;
 659         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 660                 return -1;
 661
 662         iot = __find_track(pdi, t->sector);
 663         if (!iot) {
 664                 if (verbose)
 665                         fprintf(stderr,"complete not found for (%d,%d): %llu\n",
 666                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 667                                 (unsigned long long) t->sector);
 668                 return -1;
 669         }
 670
 671         iot->completion_time = t->time;
 672         elapsed = iot->completion_time - iot->dispatch_time;
 673
 674         if (per_process_stats) {
 675                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 676                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 677
 678                 if (ppi && elapsed > ppi->longest_completion_wait[w])
 679                         ppi->longest_completion_wait[w] = elapsed;
 680         }
 681
 682         /*
 683          * kill the trace, we don't need it after completion
 684          */
 685         rb_erase(&iot->rb_node, &pdi->rb_track);
 686         free(iot);
 687
 688         return elapsed;
 689 }
 690
 691
 692 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
 693 {
 694         struct per_process_info *ppi = find_process(pid, name);
 695
 696         if (!ppi) {
 697                 ppi = malloc(sizeof(*ppi));
 698                 memset(ppi, 0, sizeof(*ppi));
 699                 memcpy(ppi->name, name, 16);
 700                 ppi->pid = pid;
 701                 add_process_to_hash(ppi);
 702                 add_process_to_list(ppi);
 703         }
 704
 705         return &ppi->io_stats;
 706 }
 707
 708 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
 709 {
 710         struct per_cpu_info *cpus = pdi->cpus;
 711         int ncpus = pdi->ncpus;
 712         int new_count = cpu + 1;
 713         int new_space, size;
 714         char *new_start;
 715
 716         size = new_count * sizeof(struct per_cpu_info);
 717         cpus = realloc(cpus, size);
 718         if (!cpus) {
 719                 char name[20];
 720                 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
 721                         get_dev_name(pdi, name, sizeof(name)), size);
 722                 exit(1);
 723         }
 724
 725         new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
 726         new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
 727         memset(new_start, 0, new_space);
 728
 729         pdi->ncpus = new_count;
 730         pdi->cpus = cpus;
 731 }
 732
 733 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
 734 {
 735         struct per_cpu_info *pci;
 736
 737         if (cpu >= pdi->ncpus)
 738                 resize_cpu_info(pdi, cpu);
 739
 740         pci = &pdi->cpus[cpu];
 741         pci->cpu = cpu;
 742         return pci;
 743 }
 744
 745
 746 static int resize_devices(char *name)
 747 {
 748         int size = (ndevices + 1) * sizeof(struct per_dev_info);
 749
 750         devices = realloc(devices, size);
 751         if (!devices) {
 752                 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
 753                 return 1;
 754         }
 755         memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
 756         devices[ndevices].name = name;
 757         ndevices++;
 758         return 0;
 759 }
 760
 761 static struct per_dev_info *get_dev_info(dev_t dev)
 762 {
 763         struct per_dev_info *pdi;
 764         int i;
 765
 766         for (i = 0; i < ndevices; i++) {
 767                 if (!devices[i].dev)
 768                         devices[i].dev = dev;
 769                 if (devices[i].dev == dev)
 770                         return &devices[i];
 771         }
 772
 773         if (resize_devices(NULL))
 774                 return NULL;
 775
 776         pdi = &devices[ndevices - 1];
 777         pdi->dev = dev;
 778         pdi->first_reported_time = 0;
 779         pdi->last_sequence = -1;
 780         pdi->last_read_time = 0;
 781         memset(&pdi->rb_last, 0, sizeof(pdi->rb_last));
 782         pdi->rb_last_entries = 0;
 783         return pdi;
 784 }
 785
 786 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
 787 {
 788         if (pdi->name)
 789                 snprintf(buffer, size, "%s", pdi->name);
 790         else
 791                 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
 792         return buffer;
 793 }
 794
 795 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
 796 {
 797         unsigned long long this = bit->time;
 798         unsigned long long last = pdi->last_reported_time;
 799
 800         pdi->backwards = (this < last) ? 'B' : ' ';
 801         pdi->last_reported_time = this;
 802 }
 803
 804 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
 805                                int rw)
 806 {
 807         if (rw) {
 808                 ios->mwrites++;
 809                 ios->qwrite_kb += t_kb(t);
 810         } else {
 811                 ios->mreads++;
 812                 ios->qread_kb += t_kb(t);
 813         }
 814 }
 815
 816 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
 817                              int rw)
 818 {
 819         __account_m(&pci->io_stats, t, rw);
 820
 821         if (per_process_stats) {
 822                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 823
 824                 __account_m(ios, t, rw);
 825         }
 826 }
 827
 828 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
 829                                    int rw)
 830 {
 831         if (rw) {
 832                 ios->qwrites++;
 833                 ios->qwrite_kb += t_kb(t);
 834         } else {
 835                 ios->qreads++;
 836                 ios->qread_kb += t_kb(t);
 837         }
 838 }
 839
 840 static inline void account_queue(struct blk_io_trace *t,
 841                                  struct per_cpu_info *pci, int rw)
 842 {
 843         __account_queue(&pci->io_stats, t, rw);
 844
 845         if (per_process_stats) {
 846                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 847
 848                 __account_queue(ios, t, rw);
 849         }
 850 }
 851
 852 static inline void __account_c(struct io_stats *ios, int rw, int bytes)
 853 {
 854         if (rw) {
 855                 ios->cwrites++;
 856                 ios->cwrite_kb += bytes >> 10;
 857         } else {
 858                 ios->creads++;
 859                 ios->cread_kb += bytes >> 10;
 860         }
 861 }
 862
 863 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
 864                              int rw, int bytes)
 865 {
 866         __account_c(&pci->io_stats, rw, bytes);
 867
 868         if (per_process_stats) {
 869                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 870
 871                 __account_c(ios, rw, bytes);
 872         }
 873 }
 874
 875 static inline void __account_issue(struct io_stats *ios, int rw,
 876                                    unsigned int bytes)
 877 {
 878         if (rw) {
 879                 ios->iwrites++;
 880                 ios->iwrite_kb += bytes >> 10;
 881         } else {
 882                 ios->ireads++;
 883                 ios->iread_kb += bytes >> 10;
 884         }
 885 }
 886
 887 static inline void account_issue(struct blk_io_trace *t,
 888                                  struct per_cpu_info *pci, int rw)
 889 {
 890         __account_issue(&pci->io_stats, rw, t->bytes);
 891
 892         if (per_process_stats) {
 893                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 894
 895                 __account_issue(ios, rw, t->bytes);
 896         }
 897 }
 898
 899 static inline void __account_unplug(struct io_stats *ios, int timer)
 900 {
 901         if (timer)
 902                 ios->timer_unplugs++;
 903         else
 904                 ios->io_unplugs++;
 905 }
 906
 907 static inline void account_unplug(struct blk_io_trace *t,
 908                                   struct per_cpu_info *pci, int timer)
 909 {
 910         __account_unplug(&pci->io_stats, timer);
 911
 912         if (per_process_stats) {
 913                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 914
 915                 __account_unplug(ios, timer);
 916         }
 917 }
 918
 919 static inline void __account_requeue(struct io_stats *ios,
 920                                      struct blk_io_trace *t, int rw)
 921 {
 922         if (rw) {
 923                 ios->wrqueue++;
 924                 ios->iwrite_kb -= t_kb(t);
 925         } else {
 926                 ios->rrqueue++;
 927                 ios->iread_kb -= t_kb(t);
 928         }
 929 }
 930
 931 static inline void account_requeue(struct blk_io_trace *t,
 932                                    struct per_cpu_info *pci, int rw)
 933 {
 934         __account_requeue(&pci->io_stats, t, rw);
 935
 936         if (per_process_stats) {
 937                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 938
 939                 __account_requeue(ios, t, rw);
 940         }
 941 }
 942
 943 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
 944                          struct blk_io_trace *t, char *act)
 945 {
 946         process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
 947 }
 948
 949 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
 950                        struct blk_io_trace *t, char *act)
 951 {
 952         process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
 953 }
 954
 955 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
 956                       char *act)
 957 {
 958         process_fmt(act, pci, t, -1, 0, NULL);
 959 }
 960
 961 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
 962                       struct blk_io_trace *t, char *act)
 963 {
 964         process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
 965 }
 966
 967 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
 968                       struct blk_io_trace *t, char *act)
 969 {
 970         if (act[0] == 'F')
 971                 log_track_frontmerge(pdi, t);
 972
 973         process_fmt(act, pci, t, -1ULL, 0, NULL);
 974 }
 975
 976 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
 977                         char *act)
 978 {
 979         process_fmt(act, pci, t, -1ULL, 0, NULL);
 980 }
 981
 982 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
 983                         char *act)
 984 {
 985         process_fmt(act, pci, t, -1ULL, 0, NULL);
 986 }
 987
 988 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
 989                       char *act)
 990 {
 991         process_fmt(act, pci, t, -1ULL, 0, NULL);
 992 }
 993
 994 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
 995                       char *act)
 996 {
 997         process_fmt(act, pci, t, -1ULL, 0, NULL);
 998 }
 999
1000 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1001 {
1002         unsigned char *buf = (unsigned char *) t + sizeof(*t);
1003
1004         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1005 }
1006
1007 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
1008 {
1009         int act = t->action & 0xffff;
1010
1011         switch (act) {
1012                 case __BLK_TA_QUEUE:
1013                         log_generic(pci, t, "Q");
1014                         break;
1015                 case __BLK_TA_GETRQ:
1016                         log_generic(pci, t, "G");
1017                         break;
1018                 case __BLK_TA_SLEEPRQ:
1019                         log_generic(pci, t, "S");
1020                         break;
1021                 case __BLK_TA_REQUEUE:
1022                         log_generic(pci, t, "R");
1023                         break;
1024                 case __BLK_TA_ISSUE:
1025                         log_pc(pci, t, "D");
1026                         break;
1027                 case __BLK_TA_COMPLETE:
1028                         log_pc(pci, t, "C");
1029                         break;
1030                 case __BLK_TA_INSERT:
1031                         log_pc(pci, t, "I");
1032                         break;
1033                 default:
1034                         fprintf(stderr, "Bad pc action %x\n", act);
1035                         break;
1036         }
1037 }
1038
1039 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1040                           struct per_cpu_info *pci)
1041 {
1042         int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
1043         int act = t->action & 0xffff;
1044
1045         switch (act) {
1046                 case __BLK_TA_QUEUE:
1047                         account_queue(t, pci, w);
1048                         log_queue(pci, t, "Q");
1049                         break;
1050                 case __BLK_TA_INSERT:
1051                         log_insert(pdi, pci, t, "I");
1052                         break;
1053                 case __BLK_TA_BACKMERGE:
1054                         account_m(t, pci, w);
1055                         log_merge(pdi, pci, t, "M");
1056                         break;
1057                 case __BLK_TA_FRONTMERGE:
1058                         account_m(t, pci, w);
1059                         log_merge(pdi, pci, t, "F");
1060                         break;
1061                 case __BLK_TA_GETRQ:
1062                         log_track_getrq(pdi, t);
1063                         log_generic(pci, t, "G");
1064                         break;
1065                 case __BLK_TA_SLEEPRQ:
1066                         log_generic(pci, t, "S");
1067                         break;
1068                 case __BLK_TA_REQUEUE:
1069                         account_requeue(t, pci, w);
1070                         log_queue(pci, t, "R");
1071                         break;
1072                 case __BLK_TA_ISSUE:
1073                         account_issue(t, pci, w);
1074                         log_issue(pdi, pci, t, "D");
1075                         break;
1076                 case __BLK_TA_COMPLETE:
1077                         account_c(t, pci, w, t->bytes);
1078                         log_complete(pdi, pci, t, "C");
1079                         break;
1080                 case __BLK_TA_PLUG:
1081                         log_action(pci, t, "P");
1082                         break;
1083                 case __BLK_TA_UNPLUG_IO:
1084                         account_unplug(t, pci, 0);
1085                         log_unplug(pci, t, "U");
1086                         break;
1087                 case __BLK_TA_UNPLUG_TIMER:
1088                         account_unplug(t, pci, 1);
1089                         log_unplug(pci, t, "UT");
1090                         break;
1091                 case __BLK_TA_SPLIT:
1092                         log_split(pci, t, "X");
1093                         break;
1094                 case __BLK_TA_BOUNCE:
1095                         log_generic(pci, t, "B");
1096                         break;
1097                 case __BLK_TA_REMAP:
1098                         log_generic(pci, t, "A");
1099                         break;
1100                 default:
1101                         fprintf(stderr, "Bad fs action %x\n", t->action);
1102                         break;
1103         }
1104 }
1105
1106 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1107                        struct per_dev_info *pdi)
1108 {
1109         if (t->action & BLK_TC_ACT(BLK_TC_PC))
1110                 dump_trace_pc(t, pci);
1111         else
1112                 dump_trace_fs(t, pdi, pci);
1113
1114         if (!pdi->events)
1115                 pdi->first_reported_time = t->time;
1116
1117         pdi->events++;
1118 }
1119
1120 /*
1121  * print in a proper way, not too small and not too big. if more than
1122  * 1000,000K, turn into M and so on
1123  */
1124 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1125 {
1126         char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1127         unsigned int i = 0;
1128
1129         if (in_kb)
1130                 i++;
1131
1132         while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1133                 i++;
1134                 num /= 1000;
1135         }
1136
1137         sprintf(dst, "%'8Lu%c", num, suff[i]);
1138         return dst;
1139 }
1140
1141 static void dump_io_stats(struct io_stats *ios, char *msg)
1142 {
1143         static char x[256], y[256];
1144
1145         fprintf(ofp, "%s\n", msg);
1146
1147         fprintf(ofp, " Reads Queued:    %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1148         fprintf(ofp, " Writes Queued:    %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1149
1150         fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1151         fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1152         fprintf(ofp, " Reads Requeued:  %s\t\t", size_cnv(x, ios->rrqueue, 0));
1153         fprintf(ofp, " Writes Requeued:  %s\n", size_cnv(x, ios->wrqueue, 0));
1154         fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1155         fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1156         fprintf(ofp, " Read Merges:     %'8lu%8c\t", ios->mreads, ' ');
1157         fprintf(ofp, " Write Merges:     %'8lu\n", ios->mwrites);
1158         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
1159         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
1160 }
1161
1162 static void dump_wait_stats(struct per_process_info *ppi)
1163 {
1164         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1165         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1166         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1167         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1168         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1169         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1170
1171         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1172         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
1173         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
1174         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
1175         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1176         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
1177 }
1178
1179 static int ppi_name_compare(const void *p1, const void *p2)
1180 {
1181         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1182         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1183         int res;
1184
1185         res = strverscmp(ppi1->name, ppi2->name);
1186         if (!res)
1187                 res = ppi1->pid > ppi2->pid;
1188
1189         return res;
1190 }
1191
1192 static void sort_process_list(void)
1193 {
1194         struct per_process_info **ppis;
1195         struct per_process_info *ppi;
1196         int i = 0;
1197
1198         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1199
1200         ppi = ppi_list;
1201         while (ppi) {
1202                 ppis[i++] = ppi;
1203                 ppi = ppi->list_next;
1204         }
1205
1206         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1207
1208         i = ppi_list_entries - 1;
1209         ppi_list = NULL;
1210         while (i >= 0) {
1211                 ppi = ppis[i];
1212
1213                 ppi->list_next = ppi_list;
1214                 ppi_list = ppi;
1215                 i--;
1216         }
1217
1218         free(ppis);
1219 }
1220
1221 static void show_process_stats(void)
1222 {
1223         struct per_process_info *ppi;
1224
1225         sort_process_list();
1226
1227         ppi = ppi_list;
1228         while (ppi) {
1229                 char name[64];
1230
1231                 if (ppi->more_than_one)
1232                         sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid);
1233                 else
1234                         sprintf(name, "%s (%u)", ppi->name, ppi->pid);
1235
1236                 dump_io_stats(&ppi->io_stats, name);
1237                 dump_wait_stats(ppi);
1238                 ppi = ppi->list_next;
1239         }
1240
1241         fprintf(ofp, "\n");
1242 }
1243
1244 static void show_device_and_cpu_stats(void)
1245 {
1246         struct per_dev_info *pdi;
1247         struct per_cpu_info *pci;
1248         struct io_stats total, *ios;
1249         unsigned long long rrate, wrate, msec;
1250         int i, j, pci_events;
1251         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1252         char name[32];
1253
1254         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1255
1256                 memset(&total, 0, sizeof(total));
1257                 pci_events = 0;
1258
1259                 if (i > 0)
1260                         fprintf(ofp, "\n");
1261
1262                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1263                         if (!pci->nelems)
1264                                 continue;
1265
1266                         ios = &pci->io_stats;
1267                         total.qreads += ios->qreads;
1268                         total.qwrites += ios->qwrites;
1269                         total.creads += ios->creads;
1270                         total.cwrites += ios->cwrites;
1271                         total.mreads += ios->mreads;
1272                         total.mwrites += ios->mwrites;
1273                         total.ireads += ios->ireads;
1274                         total.iwrites += ios->iwrites;
1275                         total.rrqueue += ios->rrqueue;
1276                         total.wrqueue += ios->wrqueue;
1277                         total.qread_kb += ios->qread_kb;
1278                         total.qwrite_kb += ios->qwrite_kb;
1279                         total.cread_kb += ios->cread_kb;
1280                         total.cwrite_kb += ios->cwrite_kb;
1281                         total.iread_kb += ios->iread_kb;
1282                         total.iwrite_kb += ios->iwrite_kb;
1283                         total.timer_unplugs += ios->timer_unplugs;
1284                         total.io_unplugs += ios->io_unplugs;
1285
1286                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1287                                  j, get_dev_name(pdi, name, sizeof(name)));
1288                         dump_io_stats(ios, line);
1289                         pci_events++;
1290                 }
1291
1292                 if (pci_events > 1) {
1293                         fprintf(ofp, "\n");
1294                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1295                                  get_dev_name(pdi, name, sizeof(name)));
1296                         dump_io_stats(&total, line);
1297                 }
1298
1299                 wrate = rrate = 0;
1300                 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1301                 if (msec) {
1302                         rrate = 1000 * total.cread_kb / msec;
1303                         wrate = 1000 * total.cwrite_kb / msec;
1304                 }
1305
1306                 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n", rrate, wrate);
1307                 fprintf(ofp, "Events (%s): %'Lu entries, %'lu skips\n",
1308                         get_dev_name(pdi, line, sizeof(line)), pdi->events,
1309                         pdi->skips);
1310         }
1311 }
1312
1313 /*
1314  * struct trace and blktrace allocation cache, we do potentially
1315  * millions of mallocs for these structures while only using at most
1316  * a few thousand at the time
1317  */
1318 static inline void t_free(struct trace *t)
1319 {
1320         if (t_alloc_cache < 1024) {
1321                 t->next = t_alloc_list;
1322                 t_alloc_list = t;
1323                 t_alloc_cache++;
1324         } else
1325                 free(t);
1326 }
1327
1328 static inline struct trace *t_alloc(void)
1329 {
1330         struct trace *t = t_alloc_list;
1331
1332         if (t) {
1333                 t_alloc_list = t->next;
1334                 t_alloc_cache--;
1335                 return t;
1336         }
1337
1338         return malloc(sizeof(*t));
1339 }
1340
1341 static inline void bit_free(struct blk_io_trace *bit)
1342 {
1343         if (bit_alloc_cache < 1024 && !bit->pdu_len) {
1344                 /*
1345                  * abuse a 64-bit field for a next pointer for the free item
1346                  */
1347                 bit->time = (__u64) (unsigned long) bit_alloc_list;
1348                 bit_alloc_list = (struct blk_io_trace *) bit;
1349                 bit_alloc_cache++;
1350         } else
1351                 free(bit);
1352 }
1353
1354 static inline struct blk_io_trace *bit_alloc(void)
1355 {
1356         struct blk_io_trace *bit = bit_alloc_list;
1357
1358         if (bit) {
1359                 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
1360                                  bit->time;
1361                 bit_alloc_cache--;
1362                 return bit;
1363         }
1364
1365         return malloc(sizeof(*bit));
1366 }
1367
1368 static void find_genesis(void)
1369 {
1370         struct trace *t = trace_list;
1371
1372         genesis_time = -1ULL;
1373         while (t != NULL) {
1374                 if (t->bit->time < genesis_time)
1375                         genesis_time = t->bit->time;
1376
1377                 t = t->next;
1378         }
1379 }
1380
1381 static inline int check_stopwatch(struct blk_io_trace *bit)
1382 {
1383         if (bit->time < stopwatch_end &&
1384             bit->time >= stopwatch_start)
1385                 return 0;
1386
1387         return 1;
1388 }
1389
1390 /*
1391  * return youngest entry read
1392  */
1393 static int sort_entries(unsigned long long *youngest)
1394 {
1395         struct trace *t;
1396
1397         if (!genesis_time)
1398                 find_genesis();
1399
1400         *youngest = 0;
1401         while ((t = trace_list) != NULL) {
1402                 struct blk_io_trace *bit = t->bit;
1403
1404                 trace_list = t->next;
1405
1406                 bit->time -= genesis_time;
1407
1408                 if (bit->time < *youngest || !*youngest)
1409                         *youngest = bit->time;
1410
1411                 if (bit->sequence < smallest_seq_read)
1412                         smallest_seq_read = bit->sequence;
1413
1414                 if (check_stopwatch(bit)) {
1415                         bit_free(bit);
1416                         t_free(t);
1417                         continue;
1418                 }
1419
1420                 if (trace_rb_insert_sort(t))
1421                         return -1;
1422         }
1423
1424         return 0;
1425 }
1426
1427 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
1428 {
1429         rb_erase(&t->rb_node, &pdi->rb_last);
1430         pdi->rb_last_entries--;
1431
1432         bit_free(t->bit);
1433         t_free(t);
1434 }
1435
1436 static void put_trace(struct per_dev_info *pdi, struct trace *t)
1437 {
1438         rb_erase(&t->rb_node, &rb_sort_root);
1439         rb_sort_entries--;
1440
1441         trace_rb_insert_last(pdi, t);
1442
1443         if (pdi->rb_last_entries > rb_batch * pdi->nfiles) {
1444                 struct rb_node *n = rb_first(&pdi->rb_last);
1445
1446                 t = rb_entry(n, struct trace, rb_node);
1447                 __put_trace_last(pdi, t);
1448         }
1449 }
1450
1451 /*
1452  * to continue, we must have traces from all online cpus in the tree
1453  */
1454 static int check_cpu_map(struct per_dev_info *pdi)
1455 {
1456         unsigned long *cpu_map;
1457         struct rb_node *n;
1458         struct trace *__t;
1459         unsigned int i;
1460         int ret, cpu;
1461
1462         /*
1463          * create a map of the cpus we have traces for
1464          */
1465         cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1466         n = rb_first(&rb_sort_root);
1467         while (n) {
1468                 __t = rb_entry(n, struct trace, rb_node);
1469                 cpu = __t->bit->cpu;
1470
1471                 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1472                 n = rb_next(n);
1473         }
1474
1475         /*
1476          * we can't continue if pdi->cpu_map has entries set that we don't
1477          * have in the sort rbtree. the opposite is not a problem, though
1478          */
1479         ret = 0;
1480         for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1481                 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1482                         ret = 1;
1483                         break;
1484                 }
1485         }
1486
1487         free(cpu_map);
1488         return ret;
1489 }
1490
1491 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1492 {
1493         unsigned long expected_sequence = pdi->last_sequence + 1;
1494         struct blk_io_trace *bit = t->bit;
1495         struct trace *__t;
1496
1497         if (!expected_sequence) {
1498                 /*
1499                  * 1 should be the first entry, just allow it
1500                  */
1501                 if (bit->sequence == 1)
1502                         return 0;
1503
1504                 return check_cpu_map(pdi);
1505         }
1506
1507         if (bit->sequence == expected_sequence)
1508                 return 0;
1509
1510         /*
1511          * we may not have seen that sequence yet. if we are not doing
1512          * the final run, break and wait for more entries.
1513          */
1514         if (expected_sequence < smallest_seq_read) {
1515                 __t = trace_rb_find_last(pdi, expected_sequence);
1516                 if (!__t)
1517                         goto skip;
1518
1519                 __put_trace_last(pdi, __t);
1520                 return 0;
1521         } else if (!force) {
1522                 return 1;
1523         } else {
1524 skip:
1525                 if (verbose) {
1526                         fprintf(stderr, "(%d,%d): skipping %lu -> %u\n",
1527                                 MAJOR(pdi->dev), MINOR(pdi->dev),
1528                                 pdi->last_sequence, bit->sequence);
1529                 }
1530                 pdi->skips++;
1531                 return 0;
1532         }
1533 }
1534
1535 static void show_entries_rb(int force)
1536 {
1537         struct per_dev_info *pdi = NULL;
1538         struct per_cpu_info *pci = NULL;
1539         struct blk_io_trace *bit;
1540         struct rb_node *n;
1541         struct trace *t;
1542
1543         while ((n = rb_first(&rb_sort_root)) != NULL) {
1544                 if (is_done() && !force && !pipeline)
1545                         break;
1546
1547                 t = rb_entry(n, struct trace, rb_node);
1548                 bit = t->bit;
1549
1550                 if (!pdi || pdi->dev != bit->device)
1551                         pdi = get_dev_info(bit->device);
1552
1553                 if (!pdi) {
1554                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1555                                 MAJOR(bit->device), MINOR(bit->device));
1556                         break;
1557                 }
1558
1559                 if (check_sequence(pdi, t, force))
1560                         break;
1561
1562                 if (!force && bit->time > last_allowed_time)
1563                         break;
1564
1565                 pdi->last_sequence = bit->sequence;
1566
1567                 check_time(pdi, bit);
1568
1569                 if (!pci || pci->cpu != bit->cpu)
1570                         pci = get_cpu_info(pdi, bit->cpu);
1571
1572                 pci->nelems++;
1573
1574                 if (bit->action & (act_mask << BLK_TC_SHIFT))
1575                         dump_trace(bit, pci, pdi);
1576
1577                 put_trace(pdi, t);
1578         }
1579 }
1580
1581 static int read_data(int fd, void *buffer, int bytes, int block)
1582 {
1583         int ret, bytes_left, fl;
1584         void *p;
1585
1586         fl = fcntl(fd, F_GETFL);
1587
1588         if (!block)
1589                 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1590         else
1591                 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1592
1593         bytes_left = bytes;
1594         p = buffer;
1595         while (bytes_left > 0) {
1596                 ret = read(fd, p, bytes_left);
1597                 if (!ret)
1598                         return 1;
1599                 else if (ret < 0) {
1600                         if (errno != EAGAIN)
1601                                 perror("read");
1602
1603                         return -1;
1604                 } else {
1605                         p += ret;
1606                         bytes_left -= ret;
1607                 }
1608         }
1609
1610         return 0;
1611 }
1612
1613 static int read_events(int fd, int always_block)
1614 {
1615         struct per_dev_info *pdi = NULL;
1616         unsigned int events = 0;
1617
1618         while (!is_done() && events < rb_batch) {
1619                 struct blk_io_trace *bit;
1620                 struct trace *t;
1621                 int pdu_len;
1622                 __u32 magic;
1623
1624                 bit = bit_alloc();
1625
1626                 if (read_data(fd, bit, sizeof(*bit), !events || always_block)) {
1627                         bit_free(bit);
1628                         break;
1629                 }
1630
1631                 magic = be32_to_cpu(bit->magic);
1632                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1633                         fprintf(stderr, "Bad magic %x\n", magic);
1634                         break;
1635                 }
1636
1637                 pdu_len = be16_to_cpu(bit->pdu_len);
1638                 if (pdu_len) {
1639                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
1640
1641                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1)) {
1642                                 bit_free(ptr);
1643                                 break;
1644                         }
1645
1646                         bit = ptr;
1647                 }
1648
1649                 trace_to_cpu(bit);
1650
1651                 if (verify_trace(bit)) {
1652                         bit_free(bit);
1653                         continue;
1654                 }
1655
1656                 t = t_alloc();
1657                 memset(t, 0, sizeof(*t));
1658                 t->bit = bit;
1659
1660                 t->next = trace_list;
1661                 trace_list = t;
1662
1663                 if (!pdi || pdi->dev != bit->device)
1664                         pdi = get_dev_info(bit->device);
1665
1666                 if (bit->time > pdi->last_read_time)
1667                         pdi->last_read_time = bit->time;
1668
1669                 events++;
1670         }
1671
1672         return events;
1673 }
1674
1675 static int do_file(void)
1676 {
1677         struct per_cpu_info *pci;
1678         struct per_dev_info *pdi;
1679         int i, j, events, events_added;
1680
1681         /*
1682          * first prepare all files for reading
1683          */
1684         for (i = 0; i < ndevices; i++) {
1685                 pdi = &devices[i];
1686                 pdi->nfiles = 0;
1687                 pdi->last_sequence = -1;
1688
1689                 for (j = 0;; j++) {
1690                         struct stat st;
1691                         int len = 0;
1692                         char *p, *dname;
1693
1694                         pci = get_cpu_info(pdi, j);
1695                         pci->cpu = j;
1696                         pci->fd = -1;
1697
1698                         p = strdup(pdi->name);
1699                         dname = dirname(p);
1700                         if (strcmp(dname, ".")) {
1701                                 input_dir = dname;
1702                                 p = strdup(pdi->name);
1703                                 strcpy(pdi->name, basename(p));
1704                         }
1705                         free(p);
1706
1707                         if (input_dir)
1708                                 len = sprintf(pci->fname, "%s/", input_dir);
1709
1710                         snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
1711                                  "%s.blktrace.%d", pdi->name, pci->cpu);
1712                         if (stat(pci->fname, &st) < 0)
1713                                 break;
1714                         if (st.st_size) {
1715                                 pci->fd = open(pci->fname, O_RDONLY);
1716                                 if (pci->fd < 0) {
1717                                         perror(pci->fname);
1718                                         continue;
1719                                 }
1720                         }
1721
1722                         printf("Input file %s added\n", pci->fname);
1723                         pdi->nfiles++;
1724                         cpu_mark_online(pdi, pci->cpu);
1725                 }
1726         }
1727
1728         /*
1729          * now loop over the files reading in the data
1730          */
1731         do {
1732                 unsigned long long youngest;
1733
1734                 events_added = 0;
1735                 last_allowed_time = -1ULL;
1736                 smallest_seq_read = -1U;
1737
1738                 for (i = 0; i < ndevices; i++) {
1739                         pdi = &devices[i];
1740
1741                         for (j = 0; j < pdi->nfiles; j++) {
1742
1743                                 pci = get_cpu_info(pdi, j);
1744
1745                                 if (pci->fd == -1)
1746                                         continue;
1747
1748                                 events = read_events(pci->fd, 1);
1749                                 if (!events) {
1750                                         cpu_mark_offline(pdi, pci->cpu);
1751                                         close(pci->fd);
1752                                         pci->fd = -1;
1753                                         continue;
1754                                 }
1755
1756                                 if (pdi->last_read_time < last_allowed_time)
1757                                         last_allowed_time = pdi->last_read_time;
1758
1759                                 events_added += events;
1760                         }
1761                 }
1762
1763                 if (sort_entries(&youngest))
1764                         break;
1765
1766                 if (youngest > stopwatch_end)
1767                         break;
1768
1769                 show_entries_rb(0);
1770
1771         } while (events_added);
1772
1773         if (rb_sort_entries)
1774                 show_entries_rb(1);
1775
1776         return 0;
1777 }
1778
1779 static int do_stdin(void)
1780 {
1781         unsigned long long youngest;
1782         int fd, events;
1783
1784         last_allowed_time = -1ULL;
1785         fd = dup(STDIN_FILENO);
1786         if (fd == -1) {
1787                 perror("dup stdin");
1788                 return -1;
1789         }
1790
1791         while ((events = read_events(fd, 0)) != 0) {
1792
1793                 smallest_seq_read = -1U;
1794
1795                 if (sort_entries(&youngest))
1796                         break;
1797
1798                 if (youngest > stopwatch_end)
1799                         break;
1800
1801                 show_entries_rb(0);
1802         }
1803
1804         if (rb_sort_entries)
1805                 show_entries_rb(1);
1806
1807         close(fd);
1808         return 0;
1809 }
1810
1811 static void show_stats(void)
1812 {
1813         if (!ofp)
1814                 return;
1815         if (stats_printed)
1816                 return;
1817
1818         stats_printed = 1;
1819
1820         if (per_process_stats)
1821                 show_process_stats();
1822
1823         if (per_device_and_cpu_stats)
1824                 show_device_and_cpu_stats();
1825
1826         fflush(ofp);
1827 }
1828
1829 static void handle_sigint(__attribute__((__unused__)) int sig)
1830 {
1831         done = 1;
1832         show_stats();
1833 }
1834
1835 /*
1836  * Extract start and duration times from a string, allowing
1837  * us to specify a time interval of interest within a trace.
1838  * Format: "duration" (start is zero) or "start:duration".
1839  */
1840 static int find_stopwatch_interval(char *string)
1841 {
1842         double value;
1843         char *sp;
1844
1845         value = strtod(string, &sp);
1846         if (sp == string) {
1847                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1848                 return 1;
1849         }
1850         if (*sp == ':') {
1851                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1852                 string = sp + 1;
1853                 value = strtod(string, &sp);
1854                 if (sp == string || *sp != '\0') {
1855                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1856                                 string);
1857                         return 1;
1858                 }
1859         } else if (*sp != '\0') {
1860                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
1861                 return 1;
1862         }
1863         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
1864         if (stopwatch_end <= stopwatch_start) {
1865                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
1866                         stopwatch_start, stopwatch_end);
1867                 return 1;
1868         }
1869
1870         return 0;
1871 }
1872
1873 static char usage_str[] = \
1874         "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
1875         "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
1876         "\t-i Input file containing trace data, or '-' for stdin\n" \
1877         "\t-D Directory to prepend to input file names\n" \
1878         "\t-o Output file. If not given, output is stdout\n" \
1879         "\t-b stdin read batching\n" \
1880         "\t-s Show per-program io statistics\n" \
1881         "\t-n Hash processes by name, not pid\n" \
1882         "\t-t Track individual ios. Will tell you the time a request took\n" \
1883         "\t   to get queued, to get dispatched, and to get completed\n" \
1884         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
1885         "\t-w Only parse data between the given time interval in seconds.\n" \
1886         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
1887         "\t-f Output format. Customize the output format. The format field\n" \
1888         "\t   identifies can be found in the documentation\n" \
1889         "\t-F Format specification. Can be found in the documentation\n" \
1890         "\t-v More verbose for marginal errors\n" \
1891         "\t-V Print program version info\n\n";
1892
1893 static void usage(char *prog)
1894 {
1895         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
1896 }
1897
1898 int main(int argc, char *argv[])
1899 {
1900         char *ofp_buffer;
1901         int i, c, ret, mode;
1902         int act_mask_tmp = 0;
1903
1904         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
1905                 switch (c) {
1906                 case 'a':
1907                         i = find_mask_map(optarg);
1908                         if (i < 0) {
1909                                 fprintf(stderr,"Invalid action mask %s\n",
1910                                         optarg);
1911                                 return 1;
1912                         }
1913                         act_mask_tmp |= i;
1914                         break;
1915
1916                 case 'A':
1917                         if ((sscanf(optarg, "%x", &i) != 1) ||
1918                                                         !valid_act_opt(i)) {
1919                                 fprintf(stderr,
1920                                         "Invalid set action mask %s/0x%x\n",
1921                                         optarg, i);
1922                                 return 1;
1923                         }
1924                         act_mask_tmp = i;
1925                         break;
1926                 case 'i':
1927                         if (!strcmp(optarg, "-") && !pipeline)
1928                                 pipeline = 1;
1929                         else if (resize_devices(optarg) != 0)
1930                                 return 1;
1931                         break;
1932                 case 'D':
1933                         input_dir = optarg;
1934                         break;
1935                 case 'o':
1936                         output_name = optarg;
1937                         break;
1938                 case 'b':
1939                         rb_batch = atoi(optarg);
1940                         if (rb_batch <= 0)
1941                                 rb_batch = RB_BATCH_DEFAULT;
1942                         break;
1943                 case 's':
1944                         per_process_stats = 1;
1945                         break;
1946                 case 't':
1947                         track_ios = 1;
1948                         break;
1949                 case 'q':
1950                         per_device_and_cpu_stats = 0;
1951                         break;
1952                 case 'w':
1953                         if (find_stopwatch_interval(optarg) != 0)
1954                                 return 1;
1955                         break;
1956                 case 'f':
1957                         set_all_format_specs(optarg);
1958                         break;
1959                 case 'F':
1960                         if (add_format_spec(optarg) != 0)
1961                                 return 1;
1962                         break;
1963                 case 'h':
1964                         ppi_hash_by_pid = 0;
1965                         break;
1966                 case 'v':
1967                         verbose++;
1968                         break;
1969                 case 'V':
1970                         printf("%s version %s\n", argv[0], blkparse_version);
1971                         return 0;
1972                 default:
1973                         usage(argv[0]);
1974                         return 1;
1975                 }
1976         }
1977
1978         while (optind < argc) {
1979                 if (!strcmp(argv[optind], "-") && !pipeline)
1980                         pipeline = 1;
1981                 else if (resize_devices(argv[optind]) != 0)
1982                         return 1;
1983                 optind++;
1984         }
1985
1986         if (!pipeline && !ndevices) {
1987                 usage(argv[0]);
1988                 return 1;
1989         }
1990
1991         if (act_mask_tmp != 0)
1992                 act_mask = act_mask_tmp;
1993
1994         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
1995
1996         signal(SIGINT, handle_sigint);
1997         signal(SIGHUP, handle_sigint);
1998         signal(SIGTERM, handle_sigint);
1999
2000         setlocale(LC_NUMERIC, "en_US");
2001
2002         if (!output_name) {
2003                 ofp = fdopen(STDOUT_FILENO, "w");
2004                 mode = _IOLBF;
2005         } else {
2006                 char ofname[128];
2007
2008                 snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
2009                 ofp = fopen(ofname, "w");
2010                 mode = _IOFBF;
2011         }
2012
2013         if (!ofp) {
2014                 perror("fopen");
2015                 return 1;
2016         }
2017
2018         ofp_buffer = malloc(4096);
2019         if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
2020                 perror("setvbuf");
2021                 return 1;
2022         }
2023
2024         if (pipeline)
2025                 ret = do_stdin();
2026         else
2027                 ret = do_file();
2028
2029         show_stats();
2030         free(ofp_buffer);
2031         return ret;
2032 }