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 skip_info {
  41         unsigned long start, end;
  42         struct skip_info *prev, *next;
  43 };
  44
  45 struct per_dev_info {
  46         dev_t dev;
  47         char *name;
  48
  49         int backwards;
  50         unsigned long long events;
  51         unsigned long long first_reported_time;
  52         unsigned long long last_reported_time;
  53         unsigned long long last_read_time;
  54         struct io_stats io_stats;
  55         unsigned long last_sequence;
  56         unsigned long skips, nskips;
  57         unsigned long long seq_skips, seq_nskips;
  58
  59         struct rb_root rb_last;
  60         unsigned long rb_last_entries;
  61
  62         struct rb_root rb_track;
  63
  64         int nfiles;
  65         int ncpus;
  66
  67         unsigned long *cpu_map;
  68         unsigned int cpu_map_max;
  69
  70         struct per_cpu_info *cpus;
  71         struct skip_info *skips_head;
  72         struct skip_info *skips_tail;
  73 };
  74
  75 struct per_process_info {
  76         char name[16];
  77         __u32 pid;
  78         struct io_stats io_stats;
  79         struct per_process_info *hash_next, *list_next;
  80         int more_than_one;
  81
  82         /*
  83          * individual io stats
  84          */
  85         unsigned long long longest_allocation_wait[2];
  86         unsigned long long longest_dispatch_wait[2];
  87         unsigned long long longest_completion_wait[2];
  88 };
  89
  90 #define PPI_HASH_SHIFT  (8)
  91 #define PPI_HASH_SIZE   (1 << PPI_HASH_SHIFT)
  92 #define PPI_HASH_MASK   (PPI_HASH_SIZE - 1)
  93 static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
  94 static struct per_process_info *ppi_list;
  95 static int ppi_list_entries;
  96
  97 #define S_OPTS  "a:A:i:o:b:stqw:f:F:vVhD:"
  98 static struct option l_opts[] = {
  99         {
 100                 .name = "act-mask",
 101                 .has_arg = required_argument,
 102                 .flag = NULL,
 103                 .val = 'a'
 104         },
 105         {
 106                 .name = "set-mask",
 107                 .has_arg = required_argument,
 108                 .flag = NULL,
 109                 .val = 'A'
 110         },
 111         {
 112                 .name = "input",
 113                 .has_arg = required_argument,
 114                 .flag = NULL,
 115                 .val = 'i'
 116         },
 117         {
 118                 .name = "output",
 119                 .has_arg = required_argument,
 120                 .flag = NULL,
 121                 .val = 'o'
 122         },
 123         {
 124                 .name = "batch",
 125                 .has_arg = required_argument,
 126                 .flag = NULL,
 127                 .val = 'b'
 128         },
 129         {
 130                 .name = "per-program-stats",
 131                 .has_arg = no_argument,
 132                 .flag = NULL,
 133                 .val = 's'
 134         },
 135         {
 136                 .name = "track-ios",
 137                 .has_arg = no_argument,
 138                 .flag = NULL,
 139                 .val = 't'
 140         },
 141         {
 142                 .name = "quiet",
 143                 .has_arg = no_argument,
 144                 .flag = NULL,
 145                 .val = 'q'
 146         },
 147         {
 148                 .name = "stopwatch",
 149                 .has_arg = required_argument,
 150                 .flag = NULL,
 151                 .val = 'w'
 152         },
 153         {
 154                 .name = "format",
 155                 .has_arg = required_argument,
 156                 .flag = NULL,
 157                 .val = 'f'
 158         },
 159         {
 160                 .name = "format-spec",
 161                 .has_arg = required_argument,
 162                 .flag = NULL,
 163                 .val = 'F'
 164         },
 165         {
 166                 .name = "hash-by-name",
 167                 .has_arg = no_argument,
 168                 .flag = NULL,
 169                 .val = 'h'
 170         },
 171         {
 172                 .name = "verbose",
 173                 .has_arg = no_argument,
 174                 .flag = NULL,
 175                 .val = 'v'
 176         },
 177         {
 178                 .name = "version",
 179                 .has_arg = no_argument,
 180                 .flag = NULL,
 181                 .val = 'V'
 182         },
 183         {
 184                 .name = "input-directory",
 185                 .has_arg = required_argument,
 186                 .flag = NULL,
 187                 .val = 'D'
 188         },
 189         {
 190                 .name = NULL,
 191         }
 192 };
 193
 194 /*
 195  * for sorting the displayed output
 196  */
 197 struct trace {
 198         struct blk_io_trace *bit;
 199         struct rb_node rb_node;
 200         struct trace *next;
 201 };
 202
 203 static struct rb_root rb_sort_root;
 204 static unsigned long rb_sort_entries;
 205
 206 static struct trace *trace_list;
 207
 208 /*
 209  * allocation cache
 210  */
 211 static struct blk_io_trace *bit_alloc_list;
 212 static struct trace *t_alloc_list;
 213
 214 /*
 215  * for tracking individual ios
 216  */
 217 struct io_track {
 218         struct rb_node rb_node;
 219
 220         __u64 sector;
 221         __u32 pid;
 222         char comm[16];
 223         unsigned long long allocation_time;
 224         unsigned long long queue_time;
 225         unsigned long long dispatch_time;
 226         unsigned long long completion_time;
 227 };
 228
 229 static int ndevices;
 230 static struct per_dev_info *devices;
 231 static char *get_dev_name(struct per_dev_info *, char *, int);
 232
 233 FILE *ofp = NULL;
 234 static char *output_name;
 235 static char *input_dir;
 236
 237 static unsigned long long genesis_time;
 238 static unsigned long long last_allowed_time;
 239 static unsigned int smallest_seq_read;
 240 static unsigned long long stopwatch_start;      /* start from zero by default */
 241 static unsigned long long stopwatch_end = -1ULL;        /* "infinity" */
 242
 243 static int per_process_stats;
 244 static int per_device_and_cpu_stats = 1;
 245 static int track_ios;
 246 static int ppi_hash_by_pid = 1;
 247 static int verbose;
 248 static unsigned int act_mask = -1U;
 249 static int stats_printed;
 250
 251 static unsigned int t_alloc_cache;
 252 static unsigned int bit_alloc_cache;
 253
 254 #define RB_BATCH_DEFAULT        (512)
 255 static unsigned int rb_batch = RB_BATCH_DEFAULT;
 256
 257 static int pipeline;
 258
 259 #define is_done()       (*(volatile int *)(&done))
 260 static volatile int done;
 261
 262 #define JHASH_RANDOM    (0x3af5f2ee)
 263
 264 #define CPUS_PER_LONG   (8 * sizeof(unsigned long))
 265 #define CPU_IDX(cpu)    ((cpu) / CPUS_PER_LONG)
 266 #define CPU_BIT(cpu)    ((cpu) & (CPUS_PER_LONG - 1))
 267
 268 static void insert_skip(struct per_dev_info *pdi, unsigned long start,
 269                         unsigned long end)
 270 {
 271         struct skip_info *sip;
 272
 273         for (sip = pdi->skips_tail; sip != NULL; sip = sip->prev) {
 274                 if (end == (sip->start - 1)) {
 275                         sip->start = start;
 276                         return;
 277                 } else if (start == (sip->end + 1)) {
 278                         sip->end = end;
 279                         return;
 280                 }
 281         }
 282
 283         sip = malloc(sizeof(struct skip_info));
 284         sip->start = start;
 285         sip->end = end;
 286         sip->prev = sip->next = NULL;
 287         if (pdi->skips_tail == NULL)
 288                 pdi->skips_head = pdi->skips_tail = sip;
 289         else {
 290                 sip->prev = pdi->skips_tail;
 291                 pdi->skips_tail->next = sip;
 292                 pdi->skips_tail = sip;
 293         }
 294 }
 295
 296 static void remove_sip(struct per_dev_info *pdi, struct skip_info *sip)
 297 {
 298         if (sip->prev == NULL) {
 299                 if (sip->next == NULL)
 300                         pdi->skips_head = pdi->skips_tail = NULL;
 301                 else {
 302                         pdi->skips_head = sip->next;
 303                         sip->next->prev = NULL;
 304                 }
 305         } else if (sip->next == NULL) {
 306                 pdi->skips_tail = sip->prev;
 307                 sip->prev->next = NULL;
 308         } else {
 309                 sip->prev->next = sip->next;
 310                 sip->next->prev = sip->prev;
 311         }
 312
 313         sip->prev = sip->next = NULL;
 314         free(sip);
 315 }
 316
 317 #define IN_SKIP(sip,seq) (((sip)->start <= (seq)) && ((seq) <= sip->end))
 318 static int check_current_skips(struct per_dev_info *pdi, unsigned long seq)
 319 {
 320         struct skip_info *sip;
 321
 322         for (sip = pdi->skips_tail; sip != NULL; sip = sip->prev) {
 323                 if (IN_SKIP(sip,seq)) {
 324                         if (sip->start == seq) {
 325                                 if (sip->end == seq)
 326                                         remove_sip(pdi,sip);
 327                                 else
 328                                         sip->start += 1;
 329                         } else if (sip->end == seq)
 330                                 sip->end -= 1;
 331                         else {
 332                                 sip->end = seq - 1;
 333                                 insert_skip(pdi,seq+1,sip->end);
 334                         }
 335                         return 1;
 336                 }
 337         }
 338         return 0;
 339 }
 340
 341 static void collect_pdi_skips(struct per_dev_info *pdi)
 342 {
 343         struct skip_info *sip;
 344
 345         pdi->skips = 0;
 346         pdi->seq_skips = 0;
 347         for (sip = pdi->skips_head; sip != NULL; sip = sip->next) {
 348                 pdi->skips += 1;
 349                 pdi->seq_skips += (sip->end - sip->start + 1);
 350                 if (verbose)
 351                         fprintf(stderr, "(%d,%d): skipping %lu -> %lu\n",
 352                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 353                                 sip->start, sip->end);
 354         }
 355 }
 356
 357 static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu)
 358 {
 359         if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) {
 360                 int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1);
 361                 unsigned long *map = malloc(new_max / sizeof(long));
 362
 363                 memset(map, 0, new_max / sizeof(long));
 364
 365                 if (pdi->cpu_map) {
 366                         memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long));
 367                         free(pdi->cpu_map);
 368                 }
 369
 370                 pdi->cpu_map = map;
 371                 pdi->cpu_map_max = new_max;
 372         }
 373
 374         pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
 375 }
 376
 377 static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu)
 378 {
 379         pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu));
 380 }
 381
 382 static inline int cpu_is_online(struct per_dev_info *pdi, int cpu)
 383 {
 384         return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0;
 385 }
 386
 387 static inline int ppi_hash_pid(__u32 pid)
 388 {
 389         return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
 390 }
 391
 392 static inline int ppi_hash_name(const char *name)
 393 {
 394         return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
 395 }
 396
 397 static inline int ppi_hash(struct per_process_info *ppi)
 398 {
 399         if (ppi_hash_by_pid)
 400                 return ppi_hash_pid(ppi->pid);
 401
 402         return ppi_hash_name(ppi->name);
 403 }
 404
 405 static inline void add_process_to_hash(struct per_process_info *ppi)
 406 {
 407         const int hash_idx = ppi_hash(ppi);
 408
 409         ppi->hash_next = ppi_hash_table[hash_idx];
 410         ppi_hash_table[hash_idx] = ppi;
 411 }
 412
 413 static inline void add_process_to_list(struct per_process_info *ppi)
 414 {
 415         ppi->list_next = ppi_list;
 416         ppi_list = ppi;
 417         ppi_list_entries++;
 418 }
 419
 420 static struct per_process_info *find_process_by_name(char *name)
 421 {
 422         const int hash_idx = ppi_hash_name(name);
 423         struct per_process_info *ppi;
 424
 425         ppi = ppi_hash_table[hash_idx];
 426         while (ppi) {
 427                 if (!strcmp(ppi->name, name))
 428                         return ppi;
 429
 430                 ppi = ppi->hash_next;
 431         }
 432
 433         return NULL;
 434 }
 435
 436 static struct per_process_info *find_process_by_pid(__u32 pid)
 437 {
 438         const int hash_idx = ppi_hash_pid(pid);
 439         struct per_process_info *ppi;
 440
 441         ppi = ppi_hash_table[hash_idx];
 442         while (ppi) {
 443                 if (ppi->pid == pid)
 444                         return ppi;
 445
 446                 ppi = ppi->hash_next;
 447         }
 448
 449         return NULL;
 450 }
 451
 452 static struct per_process_info *find_process(__u32 pid, char *name)
 453 {
 454         struct per_process_info *ppi;
 455
 456         if (ppi_hash_by_pid)
 457                 return find_process_by_pid(pid);
 458
 459         ppi = find_process_by_name(name);
 460         if (ppi && ppi->pid != pid)
 461                 ppi->more_than_one = 1;
 462
 463         return ppi;
 464 }
 465
 466 static inline int trace_rb_insert(struct trace *t, struct rb_root *root,
 467                                   int check_time)
 468 {
 469         struct rb_node **p = &root->rb_node;
 470         struct rb_node *parent = NULL;
 471         struct trace *__t;
 472
 473         while (*p) {
 474                 parent = *p;
 475
 476                 __t = rb_entry(parent, struct trace, rb_node);
 477
 478                 if (check_time) {
 479                         if (t->bit->time < __t->bit->time) {
 480                                 p = &(*p)->rb_left;
 481                                 continue;
 482                         } else if (t->bit->time > __t->bit->time) {
 483                                 p = &(*p)->rb_right;
 484                                 continue;
 485                         }
 486                 }
 487                 if (t->bit->device < __t->bit->device)
 488                         p = &(*p)->rb_left;
 489                 else if (t->bit->device > __t->bit->device)
 490                         p = &(*p)->rb_right;
 491                 else if (t->bit->sequence < __t->bit->sequence)
 492                         p = &(*p)->rb_left;
 493                 else    /* >= sequence */
 494                         p = &(*p)->rb_right;
 495         }
 496
 497         rb_link_node(&t->rb_node, parent, p);
 498         rb_insert_color(&t->rb_node, root);
 499         return 0;
 500 }
 501
 502 static inline int trace_rb_insert_sort(struct trace *t)
 503 {
 504         if (!trace_rb_insert(t, &rb_sort_root, 1)) {
 505                 rb_sort_entries++;
 506                 return 0;
 507         }
 508
 509         return 1;
 510 }
 511
 512 static inline int trace_rb_insert_last(struct per_dev_info *pdi,struct trace *t)
 513 {
 514         if (!trace_rb_insert(t, &pdi->rb_last, 1)) {
 515                 pdi->rb_last_entries++;
 516                 return 0;
 517         }
 518
 519         return 1;
 520 }
 521
 522 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
 523                                    struct rb_root *root, int order)
 524 {
 525         struct rb_node *n = root->rb_node;
 526         struct rb_node *prev = NULL;
 527         struct trace *__t;
 528
 529         while (n) {
 530                 __t = rb_entry(n, struct trace, rb_node);
 531                 prev = n;
 532
 533                 if (device < __t->bit->device)
 534                         n = n->rb_left;
 535                 else if (device > __t->bit->device)
 536                         n = n->rb_right;
 537                 else if (sequence < __t->bit->sequence)
 538                         n = n->rb_left;
 539                 else if (sequence > __t->bit->sequence)
 540                         n = n->rb_right;
 541                 else
 542                         return __t;
 543         }
 544
 545         /*
 546          * hack - the list may not be sequence ordered because some
 547          * events don't have sequence and time matched. so we end up
 548          * being a little off in the rb lookup here, because we don't
 549          * know the time we are looking for. compensate by browsing
 550          * a little ahead from the last entry to find the match
 551          */
 552         if (order && prev) {
 553                 int max = 5;
 554
 555                 while (((n = rb_next(prev)) != NULL) && max--) {
 556                         __t = rb_entry(n, struct trace, rb_node);
 557
 558                         if (__t->bit->device == device &&
 559                             __t->bit->sequence == sequence)
 560                                 return __t;
 561
 562                         prev = n;
 563                 }
 564         }
 565
 566         return NULL;
 567 }
 568
 569 static inline struct trace *trace_rb_find_sort(dev_t dev, unsigned long seq)
 570 {
 571         return trace_rb_find(dev, seq, &rb_sort_root, 1);
 572 }
 573
 574 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
 575                                                unsigned long seq)
 576 {
 577         return trace_rb_find(pdi->dev, seq, &pdi->rb_last, 0);
 578 }
 579
 580 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
 581 {
 582         struct rb_node **p = &pdi->rb_track.rb_node;
 583         struct rb_node *parent = NULL;
 584         struct io_track *__iot;
 585
 586         while (*p) {
 587                 parent = *p;
 588                 __iot = rb_entry(parent, struct io_track, rb_node);
 589
 590                 if (iot->sector < __iot->sector)
 591                         p = &(*p)->rb_left;
 592                 else if (iot->sector > __iot->sector)
 593                         p = &(*p)->rb_right;
 594                 else {
 595                         fprintf(stderr,
 596                                 "sector alias (%Lu) on device %d,%d!\n",
 597                                 (unsigned long long) iot->sector,
 598                                 MAJOR(pdi->dev), MINOR(pdi->dev));
 599                         return 1;
 600                 }
 601         }
 602
 603         rb_link_node(&iot->rb_node, parent, p);
 604         rb_insert_color(&iot->rb_node, &pdi->rb_track);
 605         return 0;
 606 }
 607
 608 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
 609 {
 610         struct rb_node *n = pdi->rb_track.rb_node;
 611         struct io_track *__iot;
 612
 613         while (n) {
 614                 __iot = rb_entry(n, struct io_track, rb_node);
 615
 616                 if (sector < __iot->sector)
 617                         n = n->rb_left;
 618                 else if (sector > __iot->sector)
 619                         n = n->rb_right;
 620                 else
 621                         return __iot;
 622         }
 623
 624         return NULL;
 625 }
 626
 627 static struct io_track *find_track(struct per_dev_info *pdi, __u32 pid,
 628                                    char *comm, __u64 sector)
 629 {
 630         struct io_track *iot;
 631
 632         iot = __find_track(pdi, sector);
 633         if (!iot) {
 634                 iot = malloc(sizeof(*iot));
 635                 iot->pid = pid;
 636                 memcpy(iot->comm, comm, sizeof(iot->comm));
 637                 iot->sector = sector;
 638                 track_rb_insert(pdi, iot);
 639         }
 640
 641         return iot;
 642 }
 643
 644 static void log_track_frontmerge(struct per_dev_info *pdi,
 645                                  struct blk_io_trace *t)
 646 {
 647         struct io_track *iot;
 648
 649         if (!track_ios)
 650                 return;
 651
 652         iot = __find_track(pdi, t->sector + t_sec(t));
 653         if (!iot) {
 654                 if (verbose)
 655                         fprintf(stderr, "merge not found for (%d,%d): %llu\n",
 656                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 657                                 (unsigned long long) t->sector + t_sec(t));
 658                 return;
 659         }
 660
 661         rb_erase(&iot->rb_node, &pdi->rb_track);
 662         iot->sector -= t_sec(t);
 663         track_rb_insert(pdi, iot);
 664 }
 665
 666 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
 667 {
 668         struct io_track *iot;
 669
 670         if (!track_ios)
 671                 return;
 672
 673         iot = find_track(pdi, t->pid, t->comm, t->sector);
 674         iot->allocation_time = t->time;
 675 }
 676
 677 /*
 678  * return time between rq allocation and insertion
 679  */
 680 static unsigned long long log_track_insert(struct per_dev_info *pdi,
 681                                            struct blk_io_trace *t)
 682 {
 683         unsigned long long elapsed;
 684         struct io_track *iot;
 685
 686         if (!track_ios)
 687                 return -1;
 688
 689         iot = find_track(pdi, t->pid, t->comm, t->sector);
 690         iot->queue_time = t->time;
 691
 692         if (!iot->allocation_time)
 693                 return -1;
 694
 695         elapsed = iot->queue_time - iot->allocation_time;
 696
 697         if (per_process_stats) {
 698                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 699                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 700
 701                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
 702                         ppi->longest_allocation_wait[w] = elapsed;
 703         }
 704
 705         return elapsed;
 706 }
 707
 708 /*
 709  * return time between queue and issue
 710  */
 711 static unsigned long long log_track_issue(struct per_dev_info *pdi,
 712                                           struct blk_io_trace *t)
 713 {
 714         unsigned long long elapsed;
 715         struct io_track *iot;
 716
 717         if (!track_ios)
 718                 return -1;
 719         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 720                 return -1;
 721
 722         iot = __find_track(pdi, t->sector);
 723         if (!iot) {
 724                 if (verbose)
 725                         fprintf(stderr, "issue not found for (%d,%d): %llu\n",
 726                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 727                                 (unsigned long long) t->sector);
 728                 return -1;
 729         }
 730
 731         iot->dispatch_time = t->time;
 732         elapsed = iot->dispatch_time - iot->queue_time;
 733
 734         if (per_process_stats) {
 735                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 736                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 737
 738                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
 739                         ppi->longest_dispatch_wait[w] = elapsed;
 740         }
 741
 742         return elapsed;
 743 }
 744
 745 /*
 746  * return time between dispatch and complete
 747  */
 748 static unsigned long long log_track_complete(struct per_dev_info *pdi,
 749                                              struct blk_io_trace *t)
 750 {
 751         unsigned long long elapsed;
 752         struct io_track *iot;
 753
 754         if (!track_ios)
 755                 return -1;
 756         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
 757                 return -1;
 758
 759         iot = __find_track(pdi, t->sector);
 760         if (!iot) {
 761                 if (verbose)
 762                         fprintf(stderr,"complete not found for (%d,%d): %llu\n",
 763                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 764                                 (unsigned long long) t->sector);
 765                 return -1;
 766         }
 767
 768         iot->completion_time = t->time;
 769         elapsed = iot->completion_time - iot->dispatch_time;
 770
 771         if (per_process_stats) {
 772                 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
 773                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 774
 775                 if (ppi && elapsed > ppi->longest_completion_wait[w])
 776                         ppi->longest_completion_wait[w] = elapsed;
 777         }
 778
 779         /*
 780          * kill the trace, we don't need it after completion
 781          */
 782         rb_erase(&iot->rb_node, &pdi->rb_track);
 783         free(iot);
 784
 785         return elapsed;
 786 }
 787
 788
 789 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
 790 {
 791         struct per_process_info *ppi = find_process(pid, name);
 792
 793         if (!ppi) {
 794                 ppi = malloc(sizeof(*ppi));
 795                 memset(ppi, 0, sizeof(*ppi));
 796                 memcpy(ppi->name, name, 16);
 797                 ppi->pid = pid;
 798                 add_process_to_hash(ppi);
 799                 add_process_to_list(ppi);
 800         }
 801
 802         return &ppi->io_stats;
 803 }
 804
 805 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
 806 {
 807         struct per_cpu_info *cpus = pdi->cpus;
 808         int ncpus = pdi->ncpus;
 809         int new_count = cpu + 1;
 810         int new_space, size;
 811         char *new_start;
 812
 813         size = new_count * sizeof(struct per_cpu_info);
 814         cpus = realloc(cpus, size);
 815         if (!cpus) {
 816                 char name[20];
 817                 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
 818                         get_dev_name(pdi, name, sizeof(name)), size);
 819                 exit(1);
 820         }
 821
 822         new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
 823         new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
 824         memset(new_start, 0, new_space);
 825
 826         pdi->ncpus = new_count;
 827         pdi->cpus = cpus;
 828 }
 829
 830 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
 831 {
 832         struct per_cpu_info *pci;
 833
 834         if (cpu >= pdi->ncpus)
 835                 resize_cpu_info(pdi, cpu);
 836
 837         pci = &pdi->cpus[cpu];
 838         pci->cpu = cpu;
 839         return pci;
 840 }
 841
 842
 843 static int resize_devices(char *name)
 844 {
 845         int size = (ndevices + 1) * sizeof(struct per_dev_info);
 846
 847         devices = realloc(devices, size);
 848         if (!devices) {
 849                 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
 850                 return 1;
 851         }
 852         memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
 853         devices[ndevices].name = name;
 854         ndevices++;
 855         return 0;
 856 }
 857
 858 static struct per_dev_info *get_dev_info(dev_t dev)
 859 {
 860         struct per_dev_info *pdi;
 861         int i;
 862
 863         for (i = 0; i < ndevices; i++) {
 864                 if (!devices[i].dev)
 865                         devices[i].dev = dev;
 866                 if (devices[i].dev == dev)
 867                         return &devices[i];
 868         }
 869
 870         if (resize_devices(NULL))
 871                 return NULL;
 872
 873         pdi = &devices[ndevices - 1];
 874         pdi->dev = dev;
 875         pdi->first_reported_time = 0;
 876         pdi->last_sequence = -1;
 877         pdi->last_read_time = 0;
 878         memset(&pdi->rb_last, 0, sizeof(pdi->rb_last));
 879         pdi->rb_last_entries = 0;
 880
 881         pdi->skips_head = pdi->skips_tail = NULL;
 882
 883         return pdi;
 884 }
 885
 886 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
 887 {
 888         if (pdi->name)
 889                 snprintf(buffer, size, "%s", pdi->name);
 890         else
 891                 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
 892         return buffer;
 893 }
 894
 895 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
 896 {
 897         unsigned long long this = bit->time;
 898         unsigned long long last = pdi->last_reported_time;
 899
 900         pdi->backwards = (this < last) ? 'B' : ' ';
 901         pdi->last_reported_time = this;
 902 }
 903
 904 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
 905                                int rw)
 906 {
 907         if (rw) {
 908                 ios->mwrites++;
 909                 ios->qwrite_kb += t_kb(t);
 910         } else {
 911                 ios->mreads++;
 912                 ios->qread_kb += t_kb(t);
 913         }
 914 }
 915
 916 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
 917                              int rw)
 918 {
 919         __account_m(&pci->io_stats, t, rw);
 920
 921         if (per_process_stats) {
 922                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 923
 924                 __account_m(ios, t, rw);
 925         }
 926 }
 927
 928 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
 929                                    int rw)
 930 {
 931         if (rw) {
 932                 ios->qwrites++;
 933                 ios->qwrite_kb += t_kb(t);
 934         } else {
 935                 ios->qreads++;
 936                 ios->qread_kb += t_kb(t);
 937         }
 938 }
 939
 940 static inline void account_queue(struct blk_io_trace *t,
 941                                  struct per_cpu_info *pci, int rw)
 942 {
 943         __account_queue(&pci->io_stats, t, rw);
 944
 945         if (per_process_stats) {
 946                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 947
 948                 __account_queue(ios, t, rw);
 949         }
 950 }
 951
 952 static inline void __account_c(struct io_stats *ios, int rw, int bytes)
 953 {
 954         if (rw) {
 955                 ios->cwrites++;
 956                 ios->cwrite_kb += bytes >> 10;
 957         } else {
 958                 ios->creads++;
 959                 ios->cread_kb += bytes >> 10;
 960         }
 961 }
 962
 963 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
 964                              int rw, int bytes)
 965 {
 966         __account_c(&pci->io_stats, rw, bytes);
 967
 968         if (per_process_stats) {
 969                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 970
 971                 __account_c(ios, rw, bytes);
 972         }
 973 }
 974
 975 static inline void __account_issue(struct io_stats *ios, int rw,
 976                                    unsigned int bytes)
 977 {
 978         if (rw) {
 979                 ios->iwrites++;
 980                 ios->iwrite_kb += bytes >> 10;
 981         } else {
 982                 ios->ireads++;
 983                 ios->iread_kb += bytes >> 10;
 984         }
 985 }
 986
 987 static inline void account_issue(struct blk_io_trace *t,
 988                                  struct per_cpu_info *pci, int rw)
 989 {
 990         __account_issue(&pci->io_stats, rw, t->bytes);
 991
 992         if (per_process_stats) {
 993                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 994
 995                 __account_issue(ios, rw, t->bytes);
 996         }
 997 }
 998
 999 static inline void __account_unplug(struct io_stats *ios, int timer)
1000 {
1001         if (timer)
1002                 ios->timer_unplugs++;
1003         else
1004                 ios->io_unplugs++;
1005 }
1006
1007 static inline void account_unplug(struct blk_io_trace *t,
1008                                   struct per_cpu_info *pci, int timer)
1009 {
1010         __account_unplug(&pci->io_stats, timer);
1011
1012         if (per_process_stats) {
1013                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
1014
1015                 __account_unplug(ios, timer);
1016         }
1017 }
1018
1019 static inline void __account_requeue(struct io_stats *ios,
1020                                      struct blk_io_trace *t, int rw)
1021 {
1022         if (rw) {
1023                 ios->wrqueue++;
1024                 ios->iwrite_kb -= t_kb(t);
1025         } else {
1026                 ios->rrqueue++;
1027                 ios->iread_kb -= t_kb(t);
1028         }
1029 }
1030
1031 static inline void account_requeue(struct blk_io_trace *t,
1032                                    struct per_cpu_info *pci, int rw)
1033 {
1034         __account_requeue(&pci->io_stats, t, rw);
1035
1036         if (per_process_stats) {
1037                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
1038
1039                 __account_requeue(ios, t, rw);
1040         }
1041 }
1042
1043 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
1044                          struct blk_io_trace *t, char *act)
1045 {
1046         process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
1047 }
1048
1049 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
1050                        struct blk_io_trace *t, char *act)
1051 {
1052         process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
1053 }
1054
1055 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
1056                       char *act)
1057 {
1058         process_fmt(act, pci, t, -1, 0, NULL);
1059 }
1060
1061 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
1062                       struct blk_io_trace *t, char *act)
1063 {
1064         process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
1065 }
1066
1067 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
1068                       struct blk_io_trace *t, char *act)
1069 {
1070         if (act[0] == 'F')
1071                 log_track_frontmerge(pdi, t);
1072
1073         process_fmt(act, pci, t, -1ULL, 0, NULL);
1074 }
1075
1076 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
1077                         char *act)
1078 {
1079         process_fmt(act, pci, t, -1ULL, 0, NULL);
1080 }
1081
1082 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
1083                         char *act)
1084 {
1085         process_fmt(act, pci, t, -1ULL, 0, NULL);
1086 }
1087
1088 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
1089                       char *act)
1090 {
1091         process_fmt(act, pci, t, -1ULL, 0, NULL);
1092 }
1093
1094 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
1095                       char *act)
1096 {
1097         process_fmt(act, pci, t, -1ULL, 0, NULL);
1098 }
1099
1100 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1101 {
1102         unsigned char *buf = (unsigned char *) t + sizeof(*t);
1103
1104         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1105 }
1106
1107 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
1108 {
1109         int act = t->action & 0xffff;
1110
1111         switch (act) {
1112                 case __BLK_TA_QUEUE:
1113                         log_generic(pci, t, "Q");
1114                         break;
1115                 case __BLK_TA_GETRQ:
1116                         log_generic(pci, t, "G");
1117                         break;
1118                 case __BLK_TA_SLEEPRQ:
1119                         log_generic(pci, t, "S");
1120                         break;
1121                 case __BLK_TA_REQUEUE:
1122                         log_generic(pci, t, "R");
1123                         break;
1124                 case __BLK_TA_ISSUE:
1125                         log_pc(pci, t, "D");
1126                         break;
1127                 case __BLK_TA_COMPLETE:
1128                         log_pc(pci, t, "C");
1129                         break;
1130                 case __BLK_TA_INSERT:
1131                         log_pc(pci, t, "I");
1132                         break;
1133                 default:
1134                         fprintf(stderr, "Bad pc action %x\n", act);
1135                         break;
1136         }
1137 }
1138
1139 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1140                           struct per_cpu_info *pci)
1141 {
1142         int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
1143         int act = t->action & 0xffff;
1144
1145         switch (act) {
1146                 case __BLK_TA_QUEUE:
1147                         account_queue(t, pci, w);
1148                         log_queue(pci, t, "Q");
1149                         break;
1150                 case __BLK_TA_INSERT:
1151                         log_insert(pdi, pci, t, "I");
1152                         break;
1153                 case __BLK_TA_BACKMERGE:
1154                         account_m(t, pci, w);
1155                         log_merge(pdi, pci, t, "M");
1156                         break;
1157                 case __BLK_TA_FRONTMERGE:
1158                         account_m(t, pci, w);
1159                         log_merge(pdi, pci, t, "F");
1160                         break;
1161                 case __BLK_TA_GETRQ:
1162                         log_track_getrq(pdi, t);
1163                         log_generic(pci, t, "G");
1164                         break;
1165                 case __BLK_TA_SLEEPRQ:
1166                         log_generic(pci, t, "S");
1167                         break;
1168                 case __BLK_TA_REQUEUE:
1169                         account_requeue(t, pci, w);
1170                         log_queue(pci, t, "R");
1171                         break;
1172                 case __BLK_TA_ISSUE:
1173                         account_issue(t, pci, w);
1174                         log_issue(pdi, pci, t, "D");
1175                         break;
1176                 case __BLK_TA_COMPLETE:
1177                         account_c(t, pci, w, t->bytes);
1178                         log_complete(pdi, pci, t, "C");
1179                         break;
1180                 case __BLK_TA_PLUG:
1181                         log_action(pci, t, "P");
1182                         break;
1183                 case __BLK_TA_UNPLUG_IO:
1184                         account_unplug(t, pci, 0);
1185                         log_unplug(pci, t, "U");
1186                         break;
1187                 case __BLK_TA_UNPLUG_TIMER:
1188                         account_unplug(t, pci, 1);
1189                         log_unplug(pci, t, "UT");
1190                         break;
1191                 case __BLK_TA_SPLIT:
1192                         log_split(pci, t, "X");
1193                         break;
1194                 case __BLK_TA_BOUNCE:
1195                         log_generic(pci, t, "B");
1196                         break;
1197                 case __BLK_TA_REMAP:
1198                         log_generic(pci, t, "A");
1199                         break;
1200                 default:
1201                         fprintf(stderr, "Bad fs action %x\n", t->action);
1202                         break;
1203         }
1204 }
1205
1206 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1207                        struct per_dev_info *pdi)
1208 {
1209         if (t->action & BLK_TC_ACT(BLK_TC_PC))
1210                 dump_trace_pc(t, pci);
1211         else
1212                 dump_trace_fs(t, pdi, pci);
1213
1214         if (!pdi->events)
1215                 pdi->first_reported_time = t->time;
1216
1217         pdi->events++;
1218 }
1219
1220 /*
1221  * print in a proper way, not too small and not too big. if more than
1222  * 1000,000K, turn into M and so on
1223  */
1224 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1225 {
1226         char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1227         unsigned int i = 0;
1228
1229         if (in_kb)
1230                 i++;
1231
1232         while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1233                 i++;
1234                 num /= 1000;
1235         }
1236
1237         sprintf(dst, "%'8Lu%c", num, suff[i]);
1238         return dst;
1239 }
1240
1241 static void dump_io_stats(struct io_stats *ios, char *msg)
1242 {
1243         static char x[256], y[256];
1244
1245         fprintf(ofp, "%s\n", msg);
1246
1247         fprintf(ofp, " Reads Queued:    %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1248         fprintf(ofp, " Writes Queued:    %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1249
1250         fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1251         fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1252         fprintf(ofp, " Reads Requeued:  %s\t\t", size_cnv(x, ios->rrqueue, 0));
1253         fprintf(ofp, " Writes Requeued:  %s\n", size_cnv(x, ios->wrqueue, 0));
1254         fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1255         fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1256         fprintf(ofp, " Read Merges:     %'8lu%8c\t", ios->mreads, ' ');
1257         fprintf(ofp, " Write Merges:     %'8lu\n", ios->mwrites);
1258         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
1259         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
1260 }
1261
1262 static void dump_wait_stats(struct per_process_info *ppi)
1263 {
1264         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1265         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1266         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1267         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1268         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1269         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1270
1271         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1272         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
1273         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
1274         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
1275         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1276         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
1277 }
1278
1279 static int ppi_name_compare(const void *p1, const void *p2)
1280 {
1281         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1282         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1283         int res;
1284
1285         res = strverscmp(ppi1->name, ppi2->name);
1286         if (!res)
1287                 res = ppi1->pid > ppi2->pid;
1288
1289         return res;
1290 }
1291
1292 static void sort_process_list(void)
1293 {
1294         struct per_process_info **ppis;
1295         struct per_process_info *ppi;
1296         int i = 0;
1297
1298         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1299
1300         ppi = ppi_list;
1301         while (ppi) {
1302                 ppis[i++] = ppi;
1303                 ppi = ppi->list_next;
1304         }
1305
1306         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1307
1308         i = ppi_list_entries - 1;
1309         ppi_list = NULL;
1310         while (i >= 0) {
1311                 ppi = ppis[i];
1312
1313                 ppi->list_next = ppi_list;
1314                 ppi_list = ppi;
1315                 i--;
1316         }
1317
1318         free(ppis);
1319 }
1320
1321 static void show_process_stats(void)
1322 {
1323         struct per_process_info *ppi;
1324
1325         sort_process_list();
1326
1327         ppi = ppi_list;
1328         while (ppi) {
1329                 char name[64];
1330
1331                 if (ppi->more_than_one)
1332                         sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid);
1333                 else
1334                         sprintf(name, "%s (%u)", ppi->name, ppi->pid);
1335
1336                 dump_io_stats(&ppi->io_stats, name);
1337                 dump_wait_stats(ppi);
1338                 ppi = ppi->list_next;
1339         }
1340
1341         fprintf(ofp, "\n");
1342 }
1343
1344 static void show_device_and_cpu_stats(void)
1345 {
1346         struct per_dev_info *pdi;
1347         struct per_cpu_info *pci;
1348         struct io_stats total, *ios;
1349         unsigned long long rrate, wrate, msec;
1350         int i, j, pci_events;
1351         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1352         char name[32];
1353
1354         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1355
1356                 memset(&total, 0, sizeof(total));
1357                 pci_events = 0;
1358
1359                 if (i > 0)
1360                         fprintf(ofp, "\n");
1361
1362                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1363                         if (!pci->nelems)
1364                                 continue;
1365
1366                         ios = &pci->io_stats;
1367                         total.qreads += ios->qreads;
1368                         total.qwrites += ios->qwrites;
1369                         total.creads += ios->creads;
1370                         total.cwrites += ios->cwrites;
1371                         total.mreads += ios->mreads;
1372                         total.mwrites += ios->mwrites;
1373                         total.ireads += ios->ireads;
1374                         total.iwrites += ios->iwrites;
1375                         total.rrqueue += ios->rrqueue;
1376                         total.wrqueue += ios->wrqueue;
1377                         total.qread_kb += ios->qread_kb;
1378                         total.qwrite_kb += ios->qwrite_kb;
1379                         total.cread_kb += ios->cread_kb;
1380                         total.cwrite_kb += ios->cwrite_kb;
1381                         total.iread_kb += ios->iread_kb;
1382                         total.iwrite_kb += ios->iwrite_kb;
1383                         total.timer_unplugs += ios->timer_unplugs;
1384                         total.io_unplugs += ios->io_unplugs;
1385
1386                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1387                                  j, get_dev_name(pdi, name, sizeof(name)));
1388                         dump_io_stats(ios, line);
1389                         pci_events++;
1390                 }
1391
1392                 if (pci_events > 1) {
1393                         fprintf(ofp, "\n");
1394                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1395                                  get_dev_name(pdi, name, sizeof(name)));
1396                         dump_io_stats(&total, line);
1397                 }
1398
1399                 wrate = rrate = 0;
1400                 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1401                 if (msec) {
1402                         rrate = 1000 * total.cread_kb / msec;
1403                         wrate = 1000 * total.cwrite_kb / msec;
1404                 }
1405
1406                 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n",
1407                         rrate, wrate);
1408                 fprintf(ofp, "Events (%s): %'Lu entries\n",
1409                         get_dev_name(pdi, line, sizeof(line)), pdi->events);
1410
1411                 collect_pdi_skips(pdi);
1412                 fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%)\n",
1413                         pdi->skips,pdi->seq_skips,
1414                         100.0 * ((double)pdi->seq_skips /
1415                                 (double)(pdi->events + pdi->seq_skips)));
1416         }
1417 }
1418
1419 /*
1420  * struct trace and blktrace allocation cache, we do potentially
1421  * millions of mallocs for these structures while only using at most
1422  * a few thousand at the time
1423  */
1424 static inline void t_free(struct trace *t)
1425 {
1426         if (t_alloc_cache < 1024) {
1427                 t->next = t_alloc_list;
1428                 t_alloc_list = t;
1429                 t_alloc_cache++;
1430         } else
1431                 free(t);
1432 }
1433
1434 static inline struct trace *t_alloc(void)
1435 {
1436         struct trace *t = t_alloc_list;
1437
1438         if (t) {
1439                 t_alloc_list = t->next;
1440                 t_alloc_cache--;
1441                 return t;
1442         }
1443
1444         return malloc(sizeof(*t));
1445 }
1446
1447 static inline void bit_free(struct blk_io_trace *bit)
1448 {
1449         if (bit_alloc_cache < 1024 && !bit->pdu_len) {
1450                 /*
1451                  * abuse a 64-bit field for a next pointer for the free item
1452                  */
1453                 bit->time = (__u64) (unsigned long) bit_alloc_list;
1454                 bit_alloc_list = (struct blk_io_trace *) bit;
1455                 bit_alloc_cache++;
1456         } else
1457                 free(bit);
1458 }
1459
1460 static inline struct blk_io_trace *bit_alloc(void)
1461 {
1462         struct blk_io_trace *bit = bit_alloc_list;
1463
1464         if (bit) {
1465                 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
1466                                  bit->time;
1467                 bit_alloc_cache--;
1468                 return bit;
1469         }
1470
1471         return malloc(sizeof(*bit));
1472 }
1473
1474 static void find_genesis(void)
1475 {
1476         struct trace *t = trace_list;
1477
1478         genesis_time = -1ULL;
1479         while (t != NULL) {
1480                 if (t->bit->time < genesis_time)
1481                         genesis_time = t->bit->time;
1482
1483                 t = t->next;
1484         }
1485 }
1486
1487 static inline int check_stopwatch(struct blk_io_trace *bit)
1488 {
1489         if (bit->time < stopwatch_end &&
1490             bit->time >= stopwatch_start)
1491                 return 0;
1492
1493         return 1;
1494 }
1495
1496 /*
1497  * return youngest entry read
1498  */
1499 static int sort_entries(unsigned long long *youngest)
1500 {
1501         struct trace *t;
1502
1503         if (!genesis_time)
1504                 find_genesis();
1505
1506         *youngest = 0;
1507         while ((t = trace_list) != NULL) {
1508                 struct blk_io_trace *bit = t->bit;
1509
1510                 trace_list = t->next;
1511
1512                 bit->time -= genesis_time;
1513
1514                 if (bit->time < *youngest || !*youngest)
1515                         *youngest = bit->time;
1516
1517                 if (bit->sequence < smallest_seq_read)
1518                         smallest_seq_read = bit->sequence;
1519
1520                 if (check_stopwatch(bit)) {
1521                         bit_free(bit);
1522                         t_free(t);
1523                         continue;
1524                 }
1525
1526                 if (trace_rb_insert_sort(t))
1527                         return -1;
1528         }
1529
1530         return 0;
1531 }
1532
1533 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
1534 {
1535         rb_erase(&t->rb_node, &pdi->rb_last);
1536         pdi->rb_last_entries--;
1537
1538         bit_free(t->bit);
1539         t_free(t);
1540 }
1541
1542 static void put_trace(struct per_dev_info *pdi, struct trace *t)
1543 {
1544         rb_erase(&t->rb_node, &rb_sort_root);
1545         rb_sort_entries--;
1546
1547         trace_rb_insert_last(pdi, t);
1548
1549         if (pdi->rb_last_entries > rb_batch * pdi->nfiles) {
1550                 struct rb_node *n = rb_first(&pdi->rb_last);
1551
1552                 t = rb_entry(n, struct trace, rb_node);
1553                 __put_trace_last(pdi, t);
1554         }
1555 }
1556
1557 /*
1558  * to continue, we must have traces from all online cpus in the tree
1559  */
1560 static int check_cpu_map(struct per_dev_info *pdi)
1561 {
1562         unsigned long *cpu_map;
1563         struct rb_node *n;
1564         struct trace *__t;
1565         unsigned int i;
1566         int ret, cpu;
1567
1568         /*
1569          * create a map of the cpus we have traces for
1570          */
1571         cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1572         n = rb_first(&rb_sort_root);
1573         while (n) {
1574                 __t = rb_entry(n, struct trace, rb_node);
1575                 cpu = __t->bit->cpu;
1576
1577                 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1578                 n = rb_next(n);
1579         }
1580
1581         /*
1582          * we can't continue if pdi->cpu_map has entries set that we don't
1583          * have in the sort rbtree. the opposite is not a problem, though
1584          */
1585         ret = 0;
1586         for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1587                 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1588                         ret = 1;
1589                         break;
1590                 }
1591         }
1592
1593         free(cpu_map);
1594         return ret;
1595 }
1596
1597 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1598 {
1599         unsigned long expected_sequence = pdi->last_sequence + 1;
1600         struct blk_io_trace *bit = t->bit;
1601         struct trace *__t;
1602
1603         if (!expected_sequence) {
1604                 /*
1605                  * 1 should be the first entry, just allow it
1606                  */
1607                 if (bit->sequence == 1)
1608                         return 0;
1609
1610                 return check_cpu_map(pdi);
1611         }
1612
1613         if (bit->sequence == expected_sequence)
1614                 return 0;
1615
1616         /*
1617          * we may not have seen that sequence yet. if we are not doing
1618          * the final run, break and wait for more entries.
1619          */
1620         if (expected_sequence < smallest_seq_read) {
1621                 __t = trace_rb_find_last(pdi, expected_sequence);
1622                 if (!__t)
1623                         goto skip;
1624
1625                 __put_trace_last(pdi, __t);
1626                 return 0;
1627         } else if (!force) {
1628                 return 1;
1629         } else {
1630 skip:
1631                 if (check_current_skips(pdi,bit->sequence))
1632                         return 0;
1633
1634                 insert_skip(pdi,expected_sequence,bit->sequence-1);
1635                 return 0;
1636         }
1637 }
1638
1639 static void show_entries_rb(int force)
1640 {
1641         struct per_dev_info *pdi = NULL;
1642         struct per_cpu_info *pci = NULL;
1643         struct blk_io_trace *bit;
1644         struct rb_node *n;
1645         struct trace *t;
1646
1647         while ((n = rb_first(&rb_sort_root)) != NULL) {
1648                 if (is_done() && !force && !pipeline)
1649                         break;
1650
1651                 t = rb_entry(n, struct trace, rb_node);
1652                 bit = t->bit;
1653
1654                 if (!pdi || pdi->dev != bit->device)
1655                         pdi = get_dev_info(bit->device);
1656
1657                 if (!pdi) {
1658                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1659                                 MAJOR(bit->device), MINOR(bit->device));
1660                         break;
1661                 }
1662
1663                 if (check_sequence(pdi, t, force))
1664                         break;
1665
1666                 if (!force && bit->time > last_allowed_time)
1667                         break;
1668
1669                 pdi->last_sequence = bit->sequence;
1670
1671                 check_time(pdi, bit);
1672
1673                 if (!pci || pci->cpu != bit->cpu)
1674                         pci = get_cpu_info(pdi, bit->cpu);
1675
1676                 pci->nelems++;
1677
1678                 if (bit->action & (act_mask << BLK_TC_SHIFT))
1679                         dump_trace(bit, pci, pdi);
1680
1681                 put_trace(pdi, t);
1682         }
1683 }
1684
1685 static int read_data(int fd, void *buffer, int bytes, int block)
1686 {
1687         int ret, bytes_left, fl;
1688         void *p;
1689
1690         fl = fcntl(fd, F_GETFL);
1691
1692         if (!block)
1693                 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1694         else
1695                 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1696
1697         bytes_left = bytes;
1698         p = buffer;
1699         while (bytes_left > 0) {
1700                 ret = read(fd, p, bytes_left);
1701                 if (!ret)
1702                         return 1;
1703                 else if (ret < 0) {
1704                         if (errno != EAGAIN)
1705                                 perror("read");
1706
1707                         return -1;
1708                 } else {
1709                         p += ret;
1710                         bytes_left -= ret;
1711                 }
1712         }
1713
1714         return 0;
1715 }
1716
1717 static int read_events(int fd, int always_block)
1718 {
1719         struct per_dev_info *pdi = NULL;
1720         unsigned int events = 0;
1721
1722         while (!is_done() && events < rb_batch) {
1723                 struct blk_io_trace *bit;
1724                 struct trace *t;
1725                 int pdu_len;
1726                 __u32 magic;
1727
1728                 bit = bit_alloc();
1729
1730                 if (read_data(fd, bit, sizeof(*bit), !events || always_block)) {
1731                         bit_free(bit);
1732                         break;
1733                 }
1734
1735                 magic = be32_to_cpu(bit->magic);
1736                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1737                         fprintf(stderr, "Bad magic %x\n", magic);
1738                         break;
1739                 }
1740
1741                 pdu_len = be16_to_cpu(bit->pdu_len);
1742                 if (pdu_len) {
1743                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
1744
1745                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1)) {
1746                                 bit_free(ptr);
1747                                 break;
1748                         }
1749
1750                         bit = ptr;
1751                 }
1752
1753                 trace_to_cpu(bit);
1754
1755                 if (verify_trace(bit)) {
1756                         bit_free(bit);
1757                         continue;
1758                 }
1759
1760                 t = t_alloc();
1761                 memset(t, 0, sizeof(*t));
1762                 t->bit = bit;
1763
1764                 t->next = trace_list;
1765                 trace_list = t;
1766
1767                 if (!pdi || pdi->dev != bit->device)
1768                         pdi = get_dev_info(bit->device);
1769
1770                 if (bit->time > pdi->last_read_time)
1771                         pdi->last_read_time = bit->time;
1772
1773                 events++;
1774         }
1775
1776         return events;
1777 }
1778
1779 static int do_file(void)
1780 {
1781         struct per_cpu_info *pci;
1782         struct per_dev_info *pdi;
1783         int i, j, events, events_added;
1784
1785         /*
1786          * first prepare all files for reading
1787          */
1788         for (i = 0; i < ndevices; i++) {
1789                 pdi = &devices[i];
1790                 pdi->nfiles = 0;
1791                 pdi->last_sequence = -1;
1792
1793                 for (j = 0;; j++) {
1794                         struct stat st;
1795                         int len = 0;
1796                         char *p, *dname;
1797
1798                         pci = get_cpu_info(pdi, j);
1799                         pci->cpu = j;
1800                         pci->fd = -1;
1801
1802                         p = strdup(pdi->name);
1803                         dname = dirname(p);
1804                         if (strcmp(dname, ".")) {
1805                                 input_dir = dname;
1806                                 p = strdup(pdi->name);
1807                                 strcpy(pdi->name, basename(p));
1808                         }
1809                         free(p);
1810
1811                         if (input_dir)
1812                                 len = sprintf(pci->fname, "%s/", input_dir);
1813
1814                         snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
1815                                  "%s.blktrace.%d", pdi->name, pci->cpu);
1816                         if (stat(pci->fname, &st) < 0)
1817                                 break;
1818                         if (st.st_size) {
1819                                 pci->fd = open(pci->fname, O_RDONLY);
1820                                 if (pci->fd < 0) {
1821                                         perror(pci->fname);
1822                                         continue;
1823                                 }
1824                         }
1825
1826                         printf("Input file %s added\n", pci->fname);
1827                         pdi->nfiles++;
1828                         cpu_mark_online(pdi, pci->cpu);
1829                 }
1830         }
1831
1832         /*
1833          * now loop over the files reading in the data
1834          */
1835         do {
1836                 unsigned long long youngest;
1837
1838                 events_added = 0;
1839                 last_allowed_time = -1ULL;
1840                 smallest_seq_read = -1U;
1841
1842                 for (i = 0; i < ndevices; i++) {
1843                         pdi = &devices[i];
1844
1845                         for (j = 0; j < pdi->nfiles; j++) {
1846
1847                                 pci = get_cpu_info(pdi, j);
1848
1849                                 if (pci->fd == -1)
1850                                         continue;
1851
1852                                 events = read_events(pci->fd, 1);
1853                                 if (!events) {
1854                                         cpu_mark_offline(pdi, pci->cpu);
1855                                         close(pci->fd);
1856                                         pci->fd = -1;
1857                                         continue;
1858                                 }
1859
1860                                 if (pdi->last_read_time < last_allowed_time)
1861                                         last_allowed_time = pdi->last_read_time;
1862
1863                                 events_added += events;
1864                         }
1865                 }
1866
1867                 if (sort_entries(&youngest))
1868                         break;
1869
1870                 if (youngest > stopwatch_end)
1871                         break;
1872
1873                 show_entries_rb(0);
1874
1875         } while (events_added);
1876
1877         if (rb_sort_entries)
1878                 show_entries_rb(1);
1879
1880         return 0;
1881 }
1882
1883 static int do_stdin(void)
1884 {
1885         unsigned long long youngest;
1886         int fd, events;
1887
1888         last_allowed_time = -1ULL;
1889         fd = dup(STDIN_FILENO);
1890         if (fd == -1) {
1891                 perror("dup stdin");
1892                 return -1;
1893         }
1894
1895         while ((events = read_events(fd, 0)) != 0) {
1896
1897                 smallest_seq_read = -1U;
1898
1899                 if (sort_entries(&youngest))
1900                         break;
1901
1902                 if (youngest > stopwatch_end)
1903                         break;
1904
1905                 show_entries_rb(0);
1906         }
1907
1908         if (rb_sort_entries)
1909                 show_entries_rb(1);
1910
1911         close(fd);
1912         return 0;
1913 }
1914
1915 static void show_stats(void)
1916 {
1917         if (!ofp)
1918                 return;
1919         if (stats_printed)
1920                 return;
1921
1922         stats_printed = 1;
1923
1924         if (per_process_stats)
1925                 show_process_stats();
1926
1927         if (per_device_and_cpu_stats)
1928                 show_device_and_cpu_stats();
1929
1930         fflush(ofp);
1931 }
1932
1933 static void handle_sigint(__attribute__((__unused__)) int sig)
1934 {
1935         done = 1;
1936         show_stats();
1937 }
1938
1939 /*
1940  * Extract start and duration times from a string, allowing
1941  * us to specify a time interval of interest within a trace.
1942  * Format: "duration" (start is zero) or "start:duration".
1943  */
1944 static int find_stopwatch_interval(char *string)
1945 {
1946         double value;
1947         char *sp;
1948
1949         value = strtod(string, &sp);
1950         if (sp == string) {
1951                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1952                 return 1;
1953         }
1954         if (*sp == ':') {
1955                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1956                 string = sp + 1;
1957                 value = strtod(string, &sp);
1958                 if (sp == string || *sp != '\0') {
1959                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1960                                 string);
1961                         return 1;
1962                 }
1963         } else if (*sp != '\0') {
1964                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
1965                 return 1;
1966         }
1967         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
1968         if (stopwatch_end <= stopwatch_start) {
1969                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
1970                         stopwatch_start, stopwatch_end);
1971                 return 1;
1972         }
1973
1974         return 0;
1975 }
1976
1977 static char usage_str[] = \
1978         "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
1979         "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
1980         "\t-i Input file containing trace data, or '-' for stdin\n" \
1981         "\t-D Directory to prepend to input file names\n" \
1982         "\t-o Output file. If not given, output is stdout\n" \
1983         "\t-b stdin read batching\n" \
1984         "\t-s Show per-program io statistics\n" \
1985         "\t-h Hash processes by name, not pid\n" \
1986         "\t-t Track individual ios. Will tell you the time a request took\n" \
1987         "\t   to get queued, to get dispatched, and to get completed\n" \
1988         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
1989         "\t-w Only parse data between the given time interval in seconds.\n" \
1990         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
1991         "\t-f Output format. Customize the output format. The format field\n" \
1992         "\t   identifies can be found in the documentation\n" \
1993         "\t-F Format specification. Can be found in the documentation\n" \
1994         "\t-v More verbose for marginal errors\n" \
1995         "\t-V Print program version info\n\n";
1996
1997 static void usage(char *prog)
1998 {
1999         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
2000 }
2001
2002 int main(int argc, char *argv[])
2003 {
2004         char *ofp_buffer;
2005         int i, c, ret, mode;
2006         int act_mask_tmp = 0;
2007
2008         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
2009                 switch (c) {
2010                 case 'a':
2011                         i = find_mask_map(optarg);
2012                         if (i < 0) {
2013                                 fprintf(stderr,"Invalid action mask %s\n",
2014                                         optarg);
2015                                 return 1;
2016                         }
2017                         act_mask_tmp |= i;
2018                         break;
2019
2020                 case 'A':
2021                         if ((sscanf(optarg, "%x", &i) != 1) ||
2022                                                         !valid_act_opt(i)) {
2023                                 fprintf(stderr,
2024                                         "Invalid set action mask %s/0x%x\n",
2025                                         optarg, i);
2026                                 return 1;
2027                         }
2028                         act_mask_tmp = i;
2029                         break;
2030                 case 'i':
2031                         if (!strcmp(optarg, "-") && !pipeline)
2032                                 pipeline = 1;
2033                         else if (resize_devices(optarg) != 0)
2034                                 return 1;
2035                         break;
2036                 case 'D':
2037                         input_dir = optarg;
2038                         break;
2039                 case 'o':
2040                         output_name = optarg;
2041                         break;
2042                 case 'b':
2043                         rb_batch = atoi(optarg);
2044                         if (rb_batch <= 0)
2045                                 rb_batch = RB_BATCH_DEFAULT;
2046                         break;
2047                 case 's':
2048                         per_process_stats = 1;
2049                         break;
2050                 case 't':
2051                         track_ios = 1;
2052                         break;
2053                 case 'q':
2054                         per_device_and_cpu_stats = 0;
2055                         break;
2056                 case 'w':
2057                         if (find_stopwatch_interval(optarg) != 0)
2058                                 return 1;
2059                         break;
2060                 case 'f':
2061                         set_all_format_specs(optarg);
2062                         break;
2063                 case 'F':
2064                         if (add_format_spec(optarg) != 0)
2065                                 return 1;
2066                         break;
2067                 case 'h':
2068                         ppi_hash_by_pid = 0;
2069                         break;
2070                 case 'v':
2071                         verbose++;
2072                         break;
2073                 case 'V':
2074                         printf("%s version %s\n", argv[0], blkparse_version);
2075                         return 0;
2076                 default:
2077                         usage(argv[0]);
2078                         return 1;
2079                 }
2080         }
2081
2082         while (optind < argc) {
2083                 if (!strcmp(argv[optind], "-") && !pipeline)
2084                         pipeline = 1;
2085                 else if (resize_devices(argv[optind]) != 0)
2086                         return 1;
2087                 optind++;
2088         }
2089
2090         if (!pipeline && !ndevices) {
2091                 usage(argv[0]);
2092                 return 1;
2093         }
2094
2095         if (act_mask_tmp != 0)
2096                 act_mask = act_mask_tmp;
2097
2098         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
2099
2100         signal(SIGINT, handle_sigint);
2101         signal(SIGHUP, handle_sigint);
2102         signal(SIGTERM, handle_sigint);
2103
2104         setlocale(LC_NUMERIC, "en_US");
2105
2106         if (!output_name) {
2107                 ofp = fdopen(STDOUT_FILENO, "w");
2108                 mode = _IOLBF;
2109         } else {
2110                 char ofname[128];
2111
2112                 snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
2113                 ofp = fopen(ofname, "w");
2114                 mode = _IOFBF;
2115         }
2116
2117         if (!ofp) {
2118                 perror("fopen");
2119                 return 1;
2120         }
2121
2122         ofp_buffer = malloc(4096);
2123         if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
2124                 perror("setvbuf");
2125                 return 1;
2126         }
2127
2128         if (pipeline)
2129                 ret = do_stdin();
2130         else
2131                 ret = do_file();
2132
2133         show_stats();
2134         free(ofp_buffer);
2135         return ret;
2136 }