blkparse.c

   1 /*
   2  * block queue tracing parse application
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  *  This program is free software; you can redistribute it and/or modify
   8  *  it under the terms of the GNU General Public License as published by
   9  *  the Free Software Foundation; either version 2 of the License, or
  10  *  (at your option) any later version.
  11  *
  12  *  This program is distributed in the hope that it will be useful,
  13  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *  GNU General Public License for more details.
  16  *
  17  *  You should have received a copy of the GNU General Public License
  18  *  along with this program; if not, write to the Free Software
  19  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  20  *
  21  */
  22 #include <sys/types.h>
  23 #include <sys/stat.h>
  24 #include <unistd.h>
  25 #include <stdio.h>
  26 #include <fcntl.h>
  27 #include <stdlib.h>
  28 #include <string.h>
  29 #include <getopt.h>
  30 #include <errno.h>
  31 #include <signal.h>
  32 #include <locale.h>
  33 #include <libgen.h>
  34
  35 #include "blktrace.h"
  36 #include "rbtree.h"
  37 #include "jhash.h"
  38
  39 static char blkparse_version[] = "0.99.3";
  40
  41 struct skip_info {
  42         unsigned long start, end;
  43         struct skip_info *prev, *next;
  44 };
  45
  46 struct per_dev_info {
  47         dev_t dev;
  48         char *name;
  49
  50         int backwards;
  51         unsigned long long events;
  52         unsigned long long first_reported_time;
  53         unsigned long long last_reported_time;
  54         unsigned long long last_read_time;
  55         struct io_stats io_stats;
  56         unsigned long skips;
  57         unsigned long long seq_skips;
  58         unsigned int max_depth[2];
  59         unsigned int cur_depth[2];
  60
  61         struct rb_root rb_track;
  62
  63         int nfiles;
  64         int ncpus;
  65
  66         unsigned long *cpu_map;
  67         unsigned int cpu_map_max;
  68
  69         struct per_cpu_info *cpus;
  70 };
  71
  72 /*
  73  * some duplicated effort here, we can unify this hash and the ppi hash later
  74  */
  75 struct process_pid_map {
  76         pid_t pid;
  77         char comm[16];
  78         struct process_pid_map *hash_next, *list_next;
  79 };
  80
  81 #define PPM_HASH_SHIFT  (8)
  82 #define PPM_HASH_SIZE   (1 << PPM_HASH_SHIFT)
  83 #define PPM_HASH_MASK   (PPM_HASH_SIZE - 1)
  84 static struct process_pid_map *ppm_hash_table[PPM_HASH_SIZE];
  85
  86 struct per_process_info {
  87         struct process_pid_map *ppm;
  88         struct io_stats io_stats;
  89         struct per_process_info *hash_next, *list_next;
  90         int more_than_one;
  91
  92         /*
  93          * individual io stats
  94          */
  95         unsigned long long longest_allocation_wait[2];
  96         unsigned long long longest_dispatch_wait[2];
  97         unsigned long long longest_completion_wait[2];
  98 };
  99
 100 #define PPI_HASH_SHIFT  (8)
 101 #define PPI_HASH_SIZE   (1 << PPI_HASH_SHIFT)
 102 #define PPI_HASH_MASK   (PPI_HASH_SIZE - 1)
 103 static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
 104 static struct per_process_info *ppi_list;
 105 static int ppi_list_entries;
 106
 107 #define S_OPTS  "a:A:b:D:d:f:F:hi:o:Oqstw:vV"
 108 static struct option l_opts[] = {
 109         {
 110                 .name = "act-mask",
 111                 .has_arg = required_argument,
 112                 .flag = NULL,
 113                 .val = 'a'
 114         },
 115         {
 116                 .name = "set-mask",
 117                 .has_arg = required_argument,
 118                 .flag = NULL,
 119                 .val = 'A'
 120         },
 121         {
 122                 .name = "batch",
 123                 .has_arg = required_argument,
 124                 .flag = NULL,
 125                 .val = 'b'
 126         },
 127         {
 128                 .name = "input-directory",
 129                 .has_arg = required_argument,
 130                 .flag = NULL,
 131                 .val = 'D'
 132         },
 133         {
 134                 .name = "dump-binary",
 135                 .has_arg = required_argument,
 136                 .flag = NULL,
 137                 .val = 'd'
 138         },
 139         {
 140                 .name = "format",
 141                 .has_arg = required_argument,
 142                 .flag = NULL,
 143                 .val = 'f'
 144         },
 145         {
 146                 .name = "format-spec",
 147                 .has_arg = required_argument,
 148                 .flag = NULL,
 149                 .val = 'F'
 150         },
 151         {
 152                 .name = "hash-by-name",
 153                 .has_arg = no_argument,
 154                 .flag = NULL,
 155                 .val = 'h'
 156         },
 157         {
 158                 .name = "input",
 159                 .has_arg = required_argument,
 160                 .flag = NULL,
 161                 .val = 'i'
 162         },
 163         {
 164                 .name = "output",
 165                 .has_arg = required_argument,
 166                 .flag = NULL,
 167                 .val = 'o'
 168         },
 169         {
 170                 .name = "no-text-output",
 171                 .has_arg = no_argument,
 172                 .flag = NULL,
 173                 .val = 'O'
 174         },
 175         {
 176                 .name = "quiet",
 177                 .has_arg = no_argument,
 178                 .flag = NULL,
 179                 .val = 'q'
 180         },
 181         {
 182                 .name = "per-program-stats",
 183                 .has_arg = no_argument,
 184                 .flag = NULL,
 185                 .val = 's'
 186         },
 187         {
 188                 .name = "track-ios",
 189                 .has_arg = no_argument,
 190                 .flag = NULL,
 191                 .val = 't'
 192         },
 193         {
 194                 .name = "stopwatch",
 195                 .has_arg = required_argument,
 196                 .flag = NULL,
 197                 .val = 'w'
 198         },
 199         {
 200                 .name = "verbose",
 201                 .has_arg = no_argument,
 202                 .flag = NULL,
 203                 .val = 'v'
 204         },
 205         {
 206                 .name = "version",
 207                 .has_arg = no_argument,
 208                 .flag = NULL,
 209                 .val = 'V'
 210         },
 211         {
 212                 .name = NULL,
 213         }
 214 };
 215
 216 /*
 217  * for sorting the displayed output
 218  */
 219 struct trace {
 220         struct blk_io_trace *bit;
 221         struct rb_node rb_node;
 222         struct trace *next;
 223         unsigned long read_sequence;
 224 };
 225
 226 static struct rb_root rb_sort_root;
 227 static unsigned long rb_sort_entries;
 228
 229 static struct trace *trace_list;
 230
 231 /*
 232  * allocation cache
 233  */
 234 static struct blk_io_trace *bit_alloc_list;
 235 static struct trace *t_alloc_list;
 236
 237 /*
 238  * for tracking individual ios
 239  */
 240 struct io_track {
 241         struct rb_node rb_node;
 242
 243         struct process_pid_map *ppm;
 244         __u64 sector;
 245         unsigned long long allocation_time;
 246         unsigned long long queue_time;
 247         unsigned long long dispatch_time;
 248         unsigned long long completion_time;
 249 };
 250
 251 static int ndevices;
 252 static struct per_dev_info *devices;
 253 static char *get_dev_name(struct per_dev_info *, char *, int);
 254 static int trace_rb_insert_last(struct per_dev_info *, struct trace *);
 255
 256 FILE *ofp = NULL;
 257 static char *output_name;
 258 static char *input_dir;
 259
 260 static unsigned long long genesis_time;
 261 static unsigned long long last_allowed_time;
 262 static unsigned long long stopwatch_start;      /* start from zero by default */
 263 static unsigned long long stopwatch_end = -1ULL;        /* "infinity" */
 264 static unsigned long read_sequence;
 265
 266 static int per_process_stats;
 267 static int per_device_and_cpu_stats = 1;
 268 static int track_ios;
 269 static int ppi_hash_by_pid = 1;
 270 static int verbose;
 271 static unsigned int act_mask = -1U;
 272 static int stats_printed;
 273 int data_is_native = -1;
 274
 275 static FILE *dump_fp;
 276 static char *dump_binary;
 277
 278 static unsigned int t_alloc_cache;
 279 static unsigned int bit_alloc_cache;
 280
 281 #define RB_BATCH_DEFAULT        (512)
 282 static unsigned int rb_batch = RB_BATCH_DEFAULT;
 283
 284 static int pipeline;
 285 static char *pipename;
 286
 287 static int text_output = 1;
 288
 289 #define is_done()       (*(volatile int *)(&done))
 290 static volatile int done;
 291
 292 struct timespec         abs_start_time;
 293 static unsigned long long start_timestamp;
 294
 295 #define JHASH_RANDOM    (0x3af5f2ee)
 296
 297 #define CPUS_PER_LONG   (8 * sizeof(unsigned long))
 298 #define CPU_IDX(cpu)    ((cpu) / CPUS_PER_LONG)
 299 #define CPU_BIT(cpu)    ((cpu) & (CPUS_PER_LONG - 1))
 300
 301 static void output_binary(void *buf, int len)
 302 {
 303         if (dump_binary) {
 304                 size_t n = fwrite(buf, len, 1, dump_fp);
 305                 if (n != 1) {
 306                         perror(dump_binary);
 307                         fclose(dump_fp);
 308                         dump_binary = NULL;
 309                 }
 310         }
 311 }
 312
 313 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
 314 {
 315         struct per_cpu_info *cpus = pdi->cpus;
 316         int ncpus = pdi->ncpus;
 317         int new_count = cpu + 1;
 318         int new_space, size;
 319         char *new_start;
 320
 321         size = new_count * sizeof(struct per_cpu_info);
 322         cpus = realloc(cpus, size);
 323         if (!cpus) {
 324                 char name[20];
 325                 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
 326                         get_dev_name(pdi, name, sizeof(name)), size);
 327                 exit(1);
 328         }
 329
 330         new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
 331         new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
 332         memset(new_start, 0, new_space);
 333
 334         pdi->ncpus = new_count;
 335         pdi->cpus = cpus;
 336
 337         for (new_count = 0; new_count < pdi->ncpus; new_count++) {
 338                 struct per_cpu_info *pci = &pdi->cpus[new_count];
 339
 340                 if (!pci->fd) {
 341                         pci->fd = -1;
 342                         memset(&pci->rb_last, 0, sizeof(pci->rb_last));
 343                         pci->rb_last_entries = 0;
 344                         pci->last_sequence = -1;
 345                 }
 346         }
 347 }
 348
 349 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
 350 {
 351         struct per_cpu_info *pci;
 352
 353         if (cpu >= pdi->ncpus)
 354                 resize_cpu_info(pdi, cpu);
 355
 356         pci = &pdi->cpus[cpu];
 357         pci->cpu = cpu;
 358         return pci;
 359 }
 360
 361
 362 static int resize_devices(char *name)
 363 {
 364         int size = (ndevices + 1) * sizeof(struct per_dev_info);
 365
 366         devices = realloc(devices, size);
 367         if (!devices) {
 368                 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
 369                 return 1;
 370         }
 371         memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
 372         devices[ndevices].name = name;
 373         ndevices++;
 374         return 0;
 375 }
 376
 377 static struct per_dev_info *get_dev_info(dev_t dev)
 378 {
 379         struct per_dev_info *pdi;
 380         int i;
 381
 382         for (i = 0; i < ndevices; i++) {
 383                 if (!devices[i].dev)
 384                         devices[i].dev = dev;
 385                 if (devices[i].dev == dev)
 386                         return &devices[i];
 387         }
 388
 389         if (resize_devices(NULL))
 390                 return NULL;
 391
 392         pdi = &devices[ndevices - 1];
 393         pdi->dev = dev;
 394         pdi->first_reported_time = 0;
 395         pdi->last_read_time = 0;
 396
 397         return pdi;
 398 }
 399
 400 static void insert_skip(struct per_cpu_info *pci, unsigned long start,
 401                         unsigned long end)
 402 {
 403         struct skip_info *sip;
 404
 405         for (sip = pci->skips_tail; sip != NULL; sip = sip->prev) {
 406                 if (end == (sip->start - 1)) {
 407                         sip->start = start;
 408                         return;
 409                 } else if (start == (sip->end + 1)) {
 410                         sip->end = end;
 411                         return;
 412                 }
 413         }
 414
 415         sip = malloc(sizeof(struct skip_info));
 416         sip->start = start;
 417         sip->end = end;
 418         sip->prev = sip->next = NULL;
 419         if (pci->skips_tail == NULL)
 420                 pci->skips_head = pci->skips_tail = sip;
 421         else {
 422                 sip->prev = pci->skips_tail;
 423                 pci->skips_tail->next = sip;
 424                 pci->skips_tail = sip;
 425         }
 426 }
 427
 428 static void remove_sip(struct per_cpu_info *pci, struct skip_info *sip)
 429 {
 430         if (sip->prev == NULL) {
 431                 if (sip->next == NULL)
 432                         pci->skips_head = pci->skips_tail = NULL;
 433                 else {
 434                         pci->skips_head = sip->next;
 435                         sip->next->prev = NULL;
 436                 }
 437         } else if (sip->next == NULL) {
 438                 pci->skips_tail = sip->prev;
 439                 sip->prev->next = NULL;
 440         } else {
 441                 sip->prev->next = sip->next;
 442                 sip->next->prev = sip->prev;
 443         }
 444
 445         sip->prev = sip->next = NULL;
 446         free(sip);
 447 }
 448
 449 #define IN_SKIP(sip,seq) (((sip)->start <= (seq)) && ((seq) <= sip->end))
 450 static int check_current_skips(struct per_cpu_info *pci, unsigned long seq)
 451 {
 452         struct skip_info *sip;
 453
 454         for (sip = pci->skips_tail; sip != NULL; sip = sip->prev) {
 455                 if (IN_SKIP(sip, seq)) {
 456                         if (sip->start == seq) {
 457                                 if (sip->end == seq)
 458                                         remove_sip(pci, sip);
 459                                 else
 460                                         sip->start += 1;
 461                         } else if (sip->end == seq)
 462                                 sip->end -= 1;
 463                         else {
 464                                 sip->end = seq - 1;
 465                                 insert_skip(pci, seq + 1, sip->end);
 466                         }
 467                         return 1;
 468                 }
 469         }
 470
 471         return 0;
 472 }
 473
 474 static void collect_pdi_skips(struct per_dev_info *pdi)
 475 {
 476         struct skip_info *sip;
 477         int cpu;
 478
 479         pdi->skips = 0;
 480         pdi->seq_skips = 0;
 481
 482         for (cpu = 0; cpu < pdi->ncpus; cpu++) {
 483                 struct per_cpu_info *pci = &pdi->cpus[cpu];
 484
 485                 for (sip = pci->skips_head; sip != NULL; sip = sip->next) {
 486                         pdi->skips++;
 487                         pdi->seq_skips += (sip->end - sip->start + 1);
 488                         if (verbose)
 489                                 fprintf(stderr,"(%d,%d): skipping %lu -> %lu\n",
 490                                         MAJOR(pdi->dev), MINOR(pdi->dev),
 491                                         sip->start, sip->end);
 492                 }
 493         }
 494 }
 495
 496 static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu)
 497 {
 498         if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) {
 499                 int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1);
 500                 unsigned long *map = malloc(new_max / sizeof(long));
 501
 502                 memset(map, 0, new_max / sizeof(long));
 503
 504                 if (pdi->cpu_map) {
 505                         memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long));
 506                         free(pdi->cpu_map);
 507                 }
 508
 509                 pdi->cpu_map = map;
 510                 pdi->cpu_map_max = new_max;
 511         }
 512
 513         pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
 514 }
 515
 516 static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu)
 517 {
 518         pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu));
 519 }
 520
 521 static inline int cpu_is_online(struct per_dev_info *pdi, int cpu)
 522 {
 523         return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0;
 524 }
 525
 526 static inline int ppm_hash_pid(pid_t pid)
 527 {
 528         return jhash_1word(pid, JHASH_RANDOM) & PPM_HASH_MASK;
 529 }
 530
 531 static struct process_pid_map *find_ppm(pid_t pid)
 532 {
 533         const int hash_idx = ppm_hash_pid(pid);
 534         struct process_pid_map *ppm;
 535
 536         ppm = ppm_hash_table[hash_idx];
 537         while (ppm) {
 538                 if (ppm->pid == pid)
 539                         return ppm;
 540
 541                 ppm = ppm->hash_next;
 542         }
 543
 544         return NULL;
 545 }
 546
 547 static void add_ppm_hash(pid_t pid, const char *name)
 548 {
 549         const int hash_idx = ppm_hash_pid(pid);
 550         struct process_pid_map *ppm;
 551
 552         ppm = find_ppm(pid);
 553         if (!ppm) {
 554                 ppm = malloc(sizeof(*ppm));
 555                 memset(ppm, 0, sizeof(*ppm));
 556                 ppm->pid = pid;
 557                 strcpy(ppm->comm, name);
 558                 ppm->hash_next = ppm_hash_table[hash_idx];
 559                 ppm_hash_table[hash_idx] = ppm;
 560         }
 561 }
 562
 563 static void handle_notify(struct blk_io_trace *bit)
 564 {
 565         void    *payload = (caddr_t) bit + sizeof(*bit);
 566         __u32   two32[2];
 567
 568         switch (bit->action) {
 569         case BLK_TN_PROCESS:
 570                 add_ppm_hash(bit->pid, payload);
 571                 break;
 572
 573         case BLK_TN_TIMESTAMP:
 574                 if (bit->pdu_len != sizeof(two32))
 575                         return;
 576                 memcpy(two32, payload, sizeof(two32));
 577                 if (!data_is_native) {
 578                         two32[0] = be32_to_cpu(two32[0]);
 579                         two32[1] = be32_to_cpu(two32[1]);
 580                 }
 581                 start_timestamp = bit->time;
 582                 abs_start_time.tv_sec  = two32[0];
 583                 abs_start_time.tv_nsec = two32[1];
 584                 if (abs_start_time.tv_nsec < 0) {
 585                         abs_start_time.tv_sec--;
 586                         abs_start_time.tv_nsec += 1000000000;
 587                 }
 588
 589                 break;
 590
 591         default:
 592                 /* Ignore unknown notify events */
 593                 ;
 594         }
 595 }
 596
 597 char *find_process_name(pid_t pid)
 598 {
 599         struct process_pid_map *ppm = find_ppm(pid);
 600
 601         if (ppm)
 602                 return ppm->comm;
 603
 604         return NULL;
 605 }
 606
 607 static inline int ppi_hash_pid(pid_t pid)
 608 {
 609         return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
 610 }
 611
 612 static inline int ppi_hash_name(const char *name)
 613 {
 614         return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
 615 }
 616
 617 static inline int ppi_hash(struct per_process_info *ppi)
 618 {
 619         struct process_pid_map *ppm = ppi->ppm;
 620
 621         if (ppi_hash_by_pid)
 622                 return ppi_hash_pid(ppm->pid);
 623
 624         return ppi_hash_name(ppm->comm);
 625 }
 626
 627 static inline void add_ppi_to_hash(struct per_process_info *ppi)
 628 {
 629         const int hash_idx = ppi_hash(ppi);
 630
 631         ppi->hash_next = ppi_hash_table[hash_idx];
 632         ppi_hash_table[hash_idx] = ppi;
 633 }
 634
 635 static inline void add_ppi_to_list(struct per_process_info *ppi)
 636 {
 637         ppi->list_next = ppi_list;
 638         ppi_list = ppi;
 639         ppi_list_entries++;
 640 }
 641
 642 static struct per_process_info *find_ppi_by_name(char *name)
 643 {
 644         const int hash_idx = ppi_hash_name(name);
 645         struct per_process_info *ppi;
 646
 647         ppi = ppi_hash_table[hash_idx];
 648         while (ppi) {
 649                 struct process_pid_map *ppm = ppi->ppm;
 650
 651                 if (!strcmp(ppm->comm, name))
 652                         return ppi;
 653
 654                 ppi = ppi->hash_next;
 655         }
 656
 657         return NULL;
 658 }
 659
 660 static struct per_process_info *find_ppi_by_pid(pid_t pid)
 661 {
 662         const int hash_idx = ppi_hash_pid(pid);
 663         struct per_process_info *ppi;
 664
 665         ppi = ppi_hash_table[hash_idx];
 666         while (ppi) {
 667                 struct process_pid_map *ppm = ppi->ppm;
 668
 669                 if (ppm->pid == pid)
 670                         return ppi;
 671
 672                 ppi = ppi->hash_next;
 673         }
 674
 675         return NULL;
 676 }
 677
 678 static struct per_process_info *find_ppi(pid_t pid)
 679 {
 680         struct per_process_info *ppi;
 681         char *name;
 682
 683         if (ppi_hash_by_pid)
 684                 return find_ppi_by_pid(pid);
 685
 686         name = find_process_name(pid);
 687         if (!name)
 688                 return NULL;
 689
 690         ppi = find_ppi_by_name(name);
 691         if (ppi && ppi->ppm->pid != pid)
 692                 ppi->more_than_one = 1;
 693
 694         return ppi;
 695 }
 696
 697 /*
 698  * struct trace and blktrace allocation cache, we do potentially
 699  * millions of mallocs for these structures while only using at most
 700  * a few thousand at the time
 701  */
 702 static inline void t_free(struct trace *t)
 703 {
 704         if (t_alloc_cache < 1024) {
 705                 t->next = t_alloc_list;
 706                 t_alloc_list = t;
 707                 t_alloc_cache++;
 708         } else
 709                 free(t);
 710 }
 711
 712 static inline struct trace *t_alloc(void)
 713 {
 714         struct trace *t = t_alloc_list;
 715
 716         if (t) {
 717                 t_alloc_list = t->next;
 718                 t_alloc_cache--;
 719                 return t;
 720         }
 721
 722         return malloc(sizeof(*t));
 723 }
 724
 725 static inline void bit_free(struct blk_io_trace *bit)
 726 {
 727         if (bit_alloc_cache < 1024 && !bit->pdu_len) {
 728                 /*
 729                  * abuse a 64-bit field for a next pointer for the free item
 730                  */
 731                 bit->time = (__u64) (unsigned long) bit_alloc_list;
 732                 bit_alloc_list = (struct blk_io_trace *) bit;
 733                 bit_alloc_cache++;
 734         } else
 735                 free(bit);
 736 }
 737
 738 static inline struct blk_io_trace *bit_alloc(void)
 739 {
 740         struct blk_io_trace *bit = bit_alloc_list;
 741
 742         if (bit) {
 743                 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
 744                                  bit->time;
 745                 bit_alloc_cache--;
 746                 return bit;
 747         }
 748
 749         return malloc(sizeof(*bit));
 750 }
 751
 752 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
 753 {
 754         struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
 755
 756         rb_erase(&t->rb_node, &pci->rb_last);
 757         pci->rb_last_entries--;
 758
 759         bit_free(t->bit);
 760         t_free(t);
 761 }
 762
 763 static void put_trace(struct per_dev_info *pdi, struct trace *t)
 764 {
 765         rb_erase(&t->rb_node, &rb_sort_root);
 766         rb_sort_entries--;
 767
 768         trace_rb_insert_last(pdi, t);
 769 }
 770
 771 static inline int trace_rb_insert(struct trace *t, struct rb_root *root)
 772 {
 773         struct rb_node **p = &root->rb_node;
 774         struct rb_node *parent = NULL;
 775         struct trace *__t;
 776
 777         while (*p) {
 778                 parent = *p;
 779
 780                 __t = rb_entry(parent, struct trace, rb_node);
 781
 782                 if (t->bit->time < __t->bit->time)
 783                         p = &(*p)->rb_left;
 784                 else if (t->bit->time > __t->bit->time)
 785                         p = &(*p)->rb_right;
 786                 else if (t->bit->device < __t->bit->device)
 787                         p = &(*p)->rb_left;
 788                 else if (t->bit->device > __t->bit->device)
 789                         p = &(*p)->rb_right;
 790                 else if (t->bit->sequence < __t->bit->sequence)
 791                         p = &(*p)->rb_left;
 792                 else    /* >= sequence */
 793                         p = &(*p)->rb_right;
 794         }
 795
 796         rb_link_node(&t->rb_node, parent, p);
 797         rb_insert_color(&t->rb_node, root);
 798         return 0;
 799 }
 800
 801 static inline int trace_rb_insert_sort(struct trace *t)
 802 {
 803         if (!trace_rb_insert(t, &rb_sort_root)) {
 804                 rb_sort_entries++;
 805                 return 0;
 806         }
 807
 808         return 1;
 809 }
 810
 811 static int trace_rb_insert_last(struct per_dev_info *pdi, struct trace *t)
 812 {
 813         struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
 814
 815         if (trace_rb_insert(t, &pci->rb_last))
 816                 return 1;
 817
 818         pci->rb_last_entries++;
 819
 820         if (pci->rb_last_entries > rb_batch * pdi->nfiles) {
 821                 struct rb_node *n = rb_first(&pci->rb_last);
 822
 823                 t = rb_entry(n, struct trace, rb_node);
 824                 __put_trace_last(pdi, t);
 825         }
 826
 827         return 0;
 828 }
 829
 830 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
 831                                    struct rb_root *root, int order)
 832 {
 833         struct rb_node *n = root->rb_node;
 834         struct rb_node *prev = NULL;
 835         struct trace *__t;
 836
 837         while (n) {
 838                 __t = rb_entry(n, struct trace, rb_node);
 839                 prev = n;
 840
 841                 if (device < __t->bit->device)
 842                         n = n->rb_left;
 843                 else if (device > __t->bit->device)
 844                         n = n->rb_right;
 845                 else if (sequence < __t->bit->sequence)
 846                         n = n->rb_left;
 847                 else if (sequence > __t->bit->sequence)
 848                         n = n->rb_right;
 849                 else
 850                         return __t;
 851         }
 852
 853         /*
 854          * hack - the list may not be sequence ordered because some
 855          * events don't have sequence and time matched. so we end up
 856          * being a little off in the rb lookup here, because we don't
 857          * know the time we are looking for. compensate by browsing
 858          * a little ahead from the last entry to find the match
 859          */
 860         if (order && prev) {
 861                 int max = 5;
 862
 863                 while (((n = rb_next(prev)) != NULL) && max--) {
 864                         __t = rb_entry(n, struct trace, rb_node);
 865
 866                         if (__t->bit->device == device &&
 867                             __t->bit->sequence == sequence)
 868                                 return __t;
 869
 870                         prev = n;
 871                 }
 872         }
 873
 874         return NULL;
 875 }
 876
 877 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
 878                                                struct per_cpu_info *pci,
 879                                                unsigned long seq)
 880 {
 881         return trace_rb_find(pdi->dev, seq, &pci->rb_last, 0);
 882 }
 883
 884 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
 885 {
 886         struct rb_node **p = &pdi->rb_track.rb_node;
 887         struct rb_node *parent = NULL;
 888         struct io_track *__iot;
 889
 890         while (*p) {
 891                 parent = *p;
 892                 __iot = rb_entry(parent, struct io_track, rb_node);
 893
 894                 if (iot->sector < __iot->sector)
 895                         p = &(*p)->rb_left;
 896                 else if (iot->sector > __iot->sector)
 897                         p = &(*p)->rb_right;
 898                 else {
 899                         fprintf(stderr,
 900                                 "sector alias (%Lu) on device %d,%d!\n",
 901                                 (unsigned long long) iot->sector,
 902                                 MAJOR(pdi->dev), MINOR(pdi->dev));
 903                         return 1;
 904                 }
 905         }
 906
 907         rb_link_node(&iot->rb_node, parent, p);
 908         rb_insert_color(&iot->rb_node, &pdi->rb_track);
 909         return 0;
 910 }
 911
 912 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
 913 {
 914         struct rb_node *n = pdi->rb_track.rb_node;
 915         struct io_track *__iot;
 916
 917         while (n) {
 918                 __iot = rb_entry(n, struct io_track, rb_node);
 919
 920                 if (sector < __iot->sector)
 921                         n = n->rb_left;
 922                 else if (sector > __iot->sector)
 923                         n = n->rb_right;
 924                 else
 925                         return __iot;
 926         }
 927
 928         return NULL;
 929 }
 930
 931 static struct io_track *find_track(struct per_dev_info *pdi, pid_t pid,
 932                                    __u64 sector)
 933 {
 934         struct io_track *iot;
 935
 936         iot = __find_track(pdi, sector);
 937         if (!iot) {
 938                 iot = malloc(sizeof(*iot));
 939                 iot->ppm = find_ppm(pid);
 940                 iot->sector = sector;
 941                 track_rb_insert(pdi, iot);
 942         }
 943
 944         return iot;
 945 }
 946
 947 static void log_track_frontmerge(struct per_dev_info *pdi,
 948                                  struct blk_io_trace *t)
 949 {
 950         struct io_track *iot;
 951
 952         if (!track_ios)
 953                 return;
 954
 955         iot = __find_track(pdi, t->sector + t_sec(t));
 956         if (!iot) {
 957                 if (verbose)
 958                         fprintf(stderr, "merge not found for (%d,%d): %llu\n",
 959                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 960                                 (unsigned long long) t->sector + t_sec(t));
 961                 return;
 962         }
 963
 964         rb_erase(&iot->rb_node, &pdi->rb_track);
 965         iot->sector -= t_sec(t);
 966         track_rb_insert(pdi, iot);
 967 }
 968
 969 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
 970 {
 971         struct io_track *iot;
 972
 973         if (!track_ios)
 974                 return;
 975
 976         iot = find_track(pdi, t->pid, t->sector);
 977         iot->allocation_time = t->time;
 978 }
 979
 980 static inline int is_remapper(struct per_dev_info *pdi)
 981 {
 982         int major = MAJOR(pdi->dev);
 983
 984         return (major == 253 || major == 9);
 985 }
 986
 987 /*
 988  * for md/dm setups, the interesting cycle is Q -> C. So track queueing
 989  * time here, as dispatch time
 990  */
 991 static void log_track_queue(struct per_dev_info *pdi, struct blk_io_trace *t)
 992 {
 993         struct io_track *iot;
 994
 995         if (!track_ios)
 996                 return;
 997         if (!is_remapper(pdi))
 998                 return;
 999
1000         iot = find_track(pdi, t->pid, t->sector);
1001         iot->dispatch_time = t->time;
1002 }
1003
1004 /*
1005  * return time between rq allocation and insertion
1006  */
1007 static unsigned long long log_track_insert(struct per_dev_info *pdi,
1008                                            struct blk_io_trace *t)
1009 {
1010         unsigned long long elapsed;
1011         struct io_track *iot;
1012
1013         if (!track_ios)
1014                 return -1;
1015
1016         iot = find_track(pdi, t->pid, t->sector);
1017         iot->queue_time = t->time;
1018
1019         if (!iot->allocation_time)
1020                 return -1;
1021
1022         elapsed = iot->queue_time - iot->allocation_time;
1023
1024         if (per_process_stats) {
1025                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1026                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1027
1028                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
1029                         ppi->longest_allocation_wait[w] = elapsed;
1030         }
1031
1032         return elapsed;
1033 }
1034
1035 /*
1036  * return time between queue and issue
1037  */
1038 static unsigned long long log_track_issue(struct per_dev_info *pdi,
1039                                           struct blk_io_trace *t)
1040 {
1041         unsigned long long elapsed;
1042         struct io_track *iot;
1043
1044         if (!track_ios)
1045                 return -1;
1046         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
1047                 return -1;
1048
1049         iot = __find_track(pdi, t->sector);
1050         if (!iot) {
1051                 if (verbose)
1052                         fprintf(stderr, "issue not found for (%d,%d): %llu\n",
1053                                 MAJOR(pdi->dev), MINOR(pdi->dev),
1054                                 (unsigned long long) t->sector);
1055                 return -1;
1056         }
1057
1058         iot->dispatch_time = t->time;
1059         elapsed = iot->dispatch_time - iot->queue_time;
1060
1061         if (per_process_stats) {
1062                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1063                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1064
1065                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
1066                         ppi->longest_dispatch_wait[w] = elapsed;
1067         }
1068
1069         return elapsed;
1070 }
1071
1072 /*
1073  * return time between dispatch and complete
1074  */
1075 static unsigned long long log_track_complete(struct per_dev_info *pdi,
1076                                              struct blk_io_trace *t)
1077 {
1078         unsigned long long elapsed;
1079         struct io_track *iot;
1080
1081         if (!track_ios)
1082                 return -1;
1083
1084         iot = __find_track(pdi, t->sector);
1085         if (!iot) {
1086                 if (verbose)
1087                         fprintf(stderr,"complete not found for (%d,%d): %llu\n",
1088                                 MAJOR(pdi->dev), MINOR(pdi->dev),
1089                                 (unsigned long long) t->sector);
1090                 return -1;
1091         }
1092
1093         iot->completion_time = t->time;
1094         elapsed = iot->completion_time - iot->dispatch_time;
1095
1096         if (per_process_stats) {
1097                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1098                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1099
1100                 if (ppi && elapsed > ppi->longest_completion_wait[w])
1101                         ppi->longest_completion_wait[w] = elapsed;
1102         }
1103
1104         /*
1105          * kill the trace, we don't need it after completion
1106          */
1107         rb_erase(&iot->rb_node, &pdi->rb_track);
1108         free(iot);
1109
1110         return elapsed;
1111 }
1112
1113
1114 static struct io_stats *find_process_io_stats(pid_t pid)
1115 {
1116         struct per_process_info *ppi = find_ppi(pid);
1117
1118         if (!ppi) {
1119                 ppi = malloc(sizeof(*ppi));
1120                 memset(ppi, 0, sizeof(*ppi));
1121                 ppi->ppm = find_ppm(pid);
1122                 add_ppi_to_hash(ppi);
1123                 add_ppi_to_list(ppi);
1124         }
1125
1126         return &ppi->io_stats;
1127 }
1128
1129 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
1130 {
1131         if (pdi->name)
1132                 snprintf(buffer, size, "%s", pdi->name);
1133         else
1134                 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
1135         return buffer;
1136 }
1137
1138 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
1139 {
1140         unsigned long long this = bit->time;
1141         unsigned long long last = pdi->last_reported_time;
1142
1143         pdi->backwards = (this < last) ? 'B' : ' ';
1144         pdi->last_reported_time = this;
1145 }
1146
1147 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
1148                                int rw)
1149 {
1150         if (rw) {
1151                 ios->mwrites++;
1152                 ios->mwrite_kb += t_kb(t);
1153         } else {
1154                 ios->mreads++;
1155                 ios->mread_kb += t_kb(t);
1156         }
1157 }
1158
1159 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
1160                              int rw)
1161 {
1162         __account_m(&pci->io_stats, t, rw);
1163
1164         if (per_process_stats) {
1165                 struct io_stats *ios = find_process_io_stats(t->pid);
1166
1167                 __account_m(ios, t, rw);
1168         }
1169 }
1170
1171 static inline void __account_pc_queue(struct io_stats *ios,
1172                                       struct blk_io_trace *t, int rw)
1173 {
1174         if (rw) {
1175                 ios->qwrites_pc++;
1176                 ios->qwrite_kb_pc += t_kb(t);
1177         } else {
1178                 ios->qreads_pc++;
1179                 ios->qread_kb += t_kb(t);
1180         }
1181 }
1182
1183 static inline void account_pc_queue(struct blk_io_trace *t,
1184                                     struct per_cpu_info *pci, int rw)
1185 {
1186         __account_pc_queue(&pci->io_stats, t, rw);
1187
1188         if (per_process_stats) {
1189                 struct io_stats *ios = find_process_io_stats(t->pid);
1190
1191                 __account_pc_queue(ios, t, rw);
1192         }
1193 }
1194
1195 static inline void __account_pc_issue(struct io_stats *ios, int rw,
1196                                       unsigned int bytes)
1197 {
1198         if (rw) {
1199                 ios->iwrites_pc++;
1200                 ios->iwrite_kb_pc += bytes >> 10;
1201         } else {
1202                 ios->ireads_pc++;
1203                 ios->iread_kb_pc += bytes >> 10;
1204         }
1205 }
1206
1207 static inline void account_pc_issue(struct blk_io_trace *t,
1208                                     struct per_cpu_info *pci, int rw)
1209 {
1210         __account_pc_issue(&pci->io_stats, rw, t->bytes);
1211
1212         if (per_process_stats) {
1213                 struct io_stats *ios = find_process_io_stats(t->pid);
1214
1215                 __account_pc_issue(ios, rw, t->bytes);
1216         }
1217 }
1218
1219 static inline void __account_pc_requeue(struct io_stats *ios,
1220                                         struct blk_io_trace *t, int rw)
1221 {
1222         if (rw) {
1223                 ios->wrqueue_pc++;
1224                 ios->iwrite_kb_pc -= t_kb(t);
1225         } else {
1226                 ios->rrqueue_pc++;
1227                 ios->iread_kb_pc -= t_kb(t);
1228         }
1229 }
1230
1231 static inline void account_pc_requeue(struct blk_io_trace *t,
1232                                       struct per_cpu_info *pci, int rw)
1233 {
1234         __account_pc_requeue(&pci->io_stats, t, rw);
1235
1236         if (per_process_stats) {
1237                 struct io_stats *ios = find_process_io_stats(t->pid);
1238
1239                 __account_pc_requeue(ios, t, rw);
1240         }
1241 }
1242
1243 static inline void __account_pc_c(struct io_stats *ios, int rw)
1244 {
1245         if (rw)
1246                 ios->cwrites_pc++;
1247         else
1248                 ios->creads_pc++;
1249 }
1250
1251 static inline void account_pc_c(struct blk_io_trace *t,
1252                                 struct per_cpu_info *pci, int rw)
1253 {
1254         __account_pc_c(&pci->io_stats, rw);
1255
1256         if (per_process_stats) {
1257                 struct io_stats *ios = find_process_io_stats(t->pid);
1258
1259                 __account_pc_c(ios, rw);
1260         }
1261 }
1262
1263 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
1264                                    int rw)
1265 {
1266         if (rw) {
1267                 ios->qwrites++;
1268                 ios->qwrite_kb += t_kb(t);
1269         } else {
1270                 ios->qreads++;
1271                 ios->qread_kb += t_kb(t);
1272         }
1273 }
1274
1275 static inline void account_queue(struct blk_io_trace *t,
1276                                  struct per_cpu_info *pci, int rw)
1277 {
1278         __account_queue(&pci->io_stats, t, rw);
1279
1280         if (per_process_stats) {
1281                 struct io_stats *ios = find_process_io_stats(t->pid);
1282
1283                 __account_queue(ios, t, rw);
1284         }
1285 }
1286
1287 static inline void __account_c(struct io_stats *ios, int rw, int bytes)
1288 {
1289         if (rw) {
1290                 ios->cwrites++;
1291                 ios->cwrite_kb += bytes >> 10;
1292         } else {
1293                 ios->creads++;
1294                 ios->cread_kb += bytes >> 10;
1295         }
1296 }
1297
1298 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
1299                              int rw, int bytes)
1300 {
1301         __account_c(&pci->io_stats, rw, bytes);
1302
1303         if (per_process_stats) {
1304                 struct io_stats *ios = find_process_io_stats(t->pid);
1305
1306                 __account_c(ios, rw, bytes);
1307         }
1308 }
1309
1310 static inline void __account_issue(struct io_stats *ios, int rw,
1311                                    unsigned int bytes)
1312 {
1313         if (rw) {
1314                 ios->iwrites++;
1315                 ios->iwrite_kb += bytes >> 10;
1316         } else {
1317                 ios->ireads++;
1318                 ios->iread_kb += bytes >> 10;
1319         }
1320 }
1321
1322 static inline void account_issue(struct blk_io_trace *t,
1323                                  struct per_cpu_info *pci, int rw)
1324 {
1325         __account_issue(&pci->io_stats, rw, t->bytes);
1326
1327         if (per_process_stats) {
1328                 struct io_stats *ios = find_process_io_stats(t->pid);
1329
1330                 __account_issue(ios, rw, t->bytes);
1331         }
1332 }
1333
1334 static inline void __account_unplug(struct io_stats *ios, int timer)
1335 {
1336         if (timer)
1337                 ios->timer_unplugs++;
1338         else
1339                 ios->io_unplugs++;
1340 }
1341
1342 static inline void account_unplug(struct blk_io_trace *t,
1343                                   struct per_cpu_info *pci, int timer)
1344 {
1345         __account_unplug(&pci->io_stats, timer);
1346
1347         if (per_process_stats) {
1348                 struct io_stats *ios = find_process_io_stats(t->pid);
1349
1350                 __account_unplug(ios, timer);
1351         }
1352 }
1353
1354 static inline void __account_requeue(struct io_stats *ios,
1355                                      struct blk_io_trace *t, int rw)
1356 {
1357         if (rw) {
1358                 ios->wrqueue++;
1359                 ios->iwrite_kb -= t_kb(t);
1360         } else {
1361                 ios->rrqueue++;
1362                 ios->iread_kb -= t_kb(t);
1363         }
1364 }
1365
1366 static inline void account_requeue(struct blk_io_trace *t,
1367                                    struct per_cpu_info *pci, int rw)
1368 {
1369         __account_requeue(&pci->io_stats, t, rw);
1370
1371         if (per_process_stats) {
1372                 struct io_stats *ios = find_process_io_stats(t->pid);
1373
1374                 __account_requeue(ios, t, rw);
1375         }
1376 }
1377
1378 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
1379                          struct blk_io_trace *t, char *act)
1380 {
1381         process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
1382 }
1383
1384 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
1385                        struct blk_io_trace *t, char *act)
1386 {
1387         process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
1388 }
1389
1390 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
1391                       char *act)
1392 {
1393         process_fmt(act, pci, t, -1, 0, NULL);
1394 }
1395
1396 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
1397                       struct blk_io_trace *t, char *act)
1398 {
1399         process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
1400 }
1401
1402 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
1403                       struct blk_io_trace *t, char *act)
1404 {
1405         if (act[0] == 'F')
1406                 log_track_frontmerge(pdi, t);
1407
1408         process_fmt(act, pci, t, -1ULL, 0, NULL);
1409 }
1410
1411 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
1412                         char *act)
1413 {
1414         process_fmt(act, pci, t, -1ULL, 0, NULL);
1415 }
1416
1417 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
1418                         char *act)
1419 {
1420         process_fmt(act, pci, t, -1ULL, 0, NULL);
1421 }
1422
1423 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
1424                       char *act)
1425 {
1426         process_fmt(act, pci, t, -1ULL, 0, NULL);
1427 }
1428
1429 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
1430                       char *act)
1431 {
1432         process_fmt(act, pci, t, -1ULL, 0, NULL);
1433 }
1434
1435 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1436 {
1437         unsigned char *buf = (unsigned char *) t + sizeof(*t);
1438
1439         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1440 }
1441
1442 static void dump_trace_pc(struct blk_io_trace *t, struct per_dev_info *pdi,
1443                           struct per_cpu_info *pci)
1444 {
1445         int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1446         int act = t->action & 0xffff;
1447
1448         switch (act) {
1449                 case __BLK_TA_QUEUE:
1450                         log_generic(pci, t, "Q");
1451                         account_pc_queue(t, pci, w);
1452                         break;
1453                 case __BLK_TA_GETRQ:
1454                         log_generic(pci, t, "G");
1455                         break;
1456                 case __BLK_TA_SLEEPRQ:
1457                         log_generic(pci, t, "S");
1458                         break;
1459                 case __BLK_TA_REQUEUE:
1460                         /*
1461                          * can happen if we miss traces, don't let it go
1462                          * below zero
1463                          */
1464                         if (pdi->cur_depth[w])
1465                                 pdi->cur_depth[w]--;
1466                         account_pc_requeue(t, pci, w);
1467                         log_generic(pci, t, "R");
1468                         break;
1469                 case __BLK_TA_ISSUE:
1470                         account_pc_issue(t, pci, w);
1471                         pdi->cur_depth[w]++;
1472                         if (pdi->cur_depth[w] > pdi->max_depth[w])
1473                                 pdi->max_depth[w] = pdi->cur_depth[w];
1474                         log_pc(pci, t, "D");
1475                         break;
1476                 case __BLK_TA_COMPLETE:
1477                         if (pdi->cur_depth[w])
1478                                 pdi->cur_depth[w]--;
1479                         log_pc(pci, t, "C");
1480                         account_pc_c(t, pci, w);
1481                         break;
1482                 case __BLK_TA_INSERT:
1483                         log_pc(pci, t, "I");
1484                         break;
1485                 default:
1486                         fprintf(stderr, "Bad pc action %x\n", act);
1487                         break;
1488         }
1489 }
1490
1491 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1492                           struct per_cpu_info *pci)
1493 {
1494         int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1495         int act = t->action & 0xffff;
1496
1497         switch (act) {
1498                 case __BLK_TA_QUEUE:
1499                         log_track_queue(pdi, t);
1500                         account_queue(t, pci, w);
1501                         log_queue(pci, t, "Q");
1502                         break;
1503                 case __BLK_TA_INSERT:
1504                         log_insert(pdi, pci, t, "I");
1505                         break;
1506                 case __BLK_TA_BACKMERGE:
1507                         account_m(t, pci, w);
1508                         log_merge(pdi, pci, t, "M");
1509                         break;
1510                 case __BLK_TA_FRONTMERGE:
1511                         account_m(t, pci, w);
1512                         log_merge(pdi, pci, t, "F");
1513                         break;
1514                 case __BLK_TA_GETRQ:
1515                         log_track_getrq(pdi, t);
1516                         log_generic(pci, t, "G");
1517                         break;
1518                 case __BLK_TA_SLEEPRQ:
1519                         log_generic(pci, t, "S");
1520                         break;
1521                 case __BLK_TA_REQUEUE:
1522                         /*
1523                          * can happen if we miss traces, don't let it go
1524                          * below zero
1525                          */
1526                         if (pdi->cur_depth[w])
1527                                 pdi->cur_depth[w]--;
1528                         account_requeue(t, pci, w);
1529                         log_queue(pci, t, "R");
1530                         break;
1531                 case __BLK_TA_ISSUE:
1532                         account_issue(t, pci, w);
1533                         pdi->cur_depth[w]++;
1534                         if (pdi->cur_depth[w] > pdi->max_depth[w])
1535                                 pdi->max_depth[w] = pdi->cur_depth[w];
1536                         log_issue(pdi, pci, t, "D");
1537                         break;
1538                 case __BLK_TA_COMPLETE:
1539                         if (pdi->cur_depth[w])
1540                                 pdi->cur_depth[w]--;
1541                         account_c(t, pci, w, t->bytes);
1542                         log_complete(pdi, pci, t, "C");
1543                         break;
1544                 case __BLK_TA_PLUG:
1545                         log_action(pci, t, "P");
1546                         break;
1547                 case __BLK_TA_UNPLUG_IO:
1548                         account_unplug(t, pci, 0);
1549                         log_unplug(pci, t, "U");
1550                         break;
1551                 case __BLK_TA_UNPLUG_TIMER:
1552                         account_unplug(t, pci, 1);
1553                         log_unplug(pci, t, "UT");
1554                         break;
1555                 case __BLK_TA_SPLIT:
1556                         log_split(pci, t, "X");
1557                         break;
1558                 case __BLK_TA_BOUNCE:
1559                         log_generic(pci, t, "B");
1560                         break;
1561                 case __BLK_TA_REMAP:
1562                         log_generic(pci, t, "A");
1563                         break;
1564                 default:
1565                         fprintf(stderr, "Bad fs action %x\n", t->action);
1566                         break;
1567         }
1568 }
1569
1570 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1571                        struct per_dev_info *pdi)
1572 {
1573         if (text_output) {
1574                 if (t->action & BLK_TC_ACT(BLK_TC_PC))
1575                         dump_trace_pc(t, pdi, pci);
1576                 else
1577                         dump_trace_fs(t, pdi, pci);
1578         }
1579
1580         if (!pdi->events)
1581                 pdi->first_reported_time = t->time;
1582
1583         pdi->events++;
1584
1585         output_binary(t, sizeof(*t) + t->pdu_len);
1586 }
1587
1588 /*
1589  * print in a proper way, not too small and not too big. if more than
1590  * 1000,000K, turn into M and so on
1591  */
1592 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1593 {
1594         char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1595         unsigned int i = 0;
1596
1597         if (in_kb)
1598                 i++;
1599
1600         while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1601                 i++;
1602                 num /= 1000;
1603         }
1604
1605         sprintf(dst, "%'8Lu%c", num, suff[i]);
1606         return dst;
1607 }
1608
1609 static void dump_io_stats(struct per_dev_info *pdi, struct io_stats *ios,
1610                           char *msg)
1611 {
1612         static char x[256], y[256];
1613
1614         fprintf(ofp, "%s\n", msg);
1615
1616         fprintf(ofp, " Reads Queued:    %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1617         fprintf(ofp, " Writes Queued:    %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1618         fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1619         fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1620         fprintf(ofp, " Reads Requeued:  %s\t\t", size_cnv(x, ios->rrqueue, 0));
1621         fprintf(ofp, " Writes Requeued:  %s\n", size_cnv(x, ios->wrqueue, 0));
1622         fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1623         fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1624         fprintf(ofp, " Read Merges:     %s, %siB\t", size_cnv(x, ios->mreads, 0), size_cnv(y, ios->mread_kb, 1));
1625         fprintf(ofp, " Write Merges:     %s, %siB\n", size_cnv(x, ios->mwrites, 0), size_cnv(y, ios->mwrite_kb, 1));
1626         if (pdi) {
1627                 fprintf(ofp, " Read depth:      %'8u%8c\t", pdi->max_depth[0], ' ');
1628                 fprintf(ofp, " Write depth:      %'8u\n", pdi->max_depth[1]);
1629         }
1630         if (ios->qreads_pc || ios->qwrites_pc || ios->ireads_pc || ios->iwrites_pc ||
1631             ios->rrqueue_pc || ios->wrqueue_pc || ios->creads_pc || ios->cwrites_pc) {
1632                 fprintf(ofp, " PC Reads Queued: %s, %siB\t", size_cnv(x, ios->qreads_pc, 0), size_cnv(y, ios->qread_kb_pc, 1));
1633                 fprintf(ofp, " PC Writes Queued: %s, %siB\n", size_cnv(x, ios->qwrites_pc, 0), size_cnv(y, ios->qwrite_kb_pc, 1));
1634                 fprintf(ofp, " PC Read Disp.:   %s, %siB\t", size_cnv(x, ios->ireads_pc, 0), size_cnv(y, ios->iread_kb_pc, 1));
1635                 fprintf(ofp, " PC Write Disp.:   %s, %siB\n", size_cnv(x, ios->iwrites_pc, 0), size_cnv(y, ios->iwrite_kb_pc, 1));
1636                 fprintf(ofp, " PC Reads Req.:   %s\t\t", size_cnv(x, ios->rrqueue_pc, 0));
1637                 fprintf(ofp, " PC Writes Req.:   %s\n", size_cnv(x, ios->wrqueue_pc, 0));
1638                 fprintf(ofp, " PC Reads Compl.: %s\t\t", size_cnv(x, ios->creads_pc, 0));
1639                 fprintf(ofp, " PC Writes Compl.: %s\n", size_cnv(x, ios->cwrites, 0));
1640         }
1641         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
1642         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
1643 }
1644
1645 static void dump_wait_stats(struct per_process_info *ppi)
1646 {
1647         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1648         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1649         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1650         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1651         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1652         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1653
1654         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1655         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
1656         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
1657         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
1658         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1659         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
1660 }
1661
1662 static int ppi_name_compare(const void *p1, const void *p2)
1663 {
1664         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1665         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1666         int res;
1667
1668         res = strverscmp(ppi1->ppm->comm, ppi2->ppm->comm);
1669         if (!res)
1670                 res = ppi1->ppm->pid > ppi2->ppm->pid;
1671
1672         return res;
1673 }
1674
1675 static void sort_process_list(void)
1676 {
1677         struct per_process_info **ppis;
1678         struct per_process_info *ppi;
1679         int i = 0;
1680
1681         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1682
1683         ppi = ppi_list;
1684         while (ppi) {
1685                 ppis[i++] = ppi;
1686                 ppi = ppi->list_next;
1687         }
1688
1689         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1690
1691         i = ppi_list_entries - 1;
1692         ppi_list = NULL;
1693         while (i >= 0) {
1694                 ppi = ppis[i];
1695
1696                 ppi->list_next = ppi_list;
1697                 ppi_list = ppi;
1698                 i--;
1699         }
1700
1701         free(ppis);
1702 }
1703
1704 static void show_process_stats(void)
1705 {
1706         struct per_process_info *ppi;
1707
1708         sort_process_list();
1709
1710         ppi = ppi_list;
1711         while (ppi) {
1712                 struct process_pid_map *ppm = ppi->ppm;
1713                 char name[64];
1714
1715                 if (ppi->more_than_one)
1716                         sprintf(name, "%s (%u, ...)", ppm->comm, ppm->pid);
1717                 else
1718                         sprintf(name, "%s (%u)", ppm->comm, ppm->pid);
1719
1720                 dump_io_stats(NULL, &ppi->io_stats, name);
1721                 dump_wait_stats(ppi);
1722                 ppi = ppi->list_next;
1723         }
1724
1725         fprintf(ofp, "\n");
1726 }
1727
1728 static void show_device_and_cpu_stats(void)
1729 {
1730         struct per_dev_info *pdi;
1731         struct per_cpu_info *pci;
1732         struct io_stats total, *ios;
1733         unsigned long long rrate, wrate, msec;
1734         int i, j, pci_events;
1735         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1736         char name[32];
1737         double ratio;
1738
1739         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1740
1741                 memset(&total, 0, sizeof(total));
1742                 pci_events = 0;
1743
1744                 if (i > 0)
1745                         fprintf(ofp, "\n");
1746
1747                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1748                         if (!pci->nelems)
1749                                 continue;
1750
1751                         ios = &pci->io_stats;
1752                         total.qreads += ios->qreads;
1753                         total.qwrites += ios->qwrites;
1754                         total.creads += ios->creads;
1755                         total.cwrites += ios->cwrites;
1756                         total.mreads += ios->mreads;
1757                         total.mwrites += ios->mwrites;
1758                         total.ireads += ios->ireads;
1759                         total.iwrites += ios->iwrites;
1760                         total.rrqueue += ios->rrqueue;
1761                         total.wrqueue += ios->wrqueue;
1762                         total.qread_kb += ios->qread_kb;
1763                         total.qwrite_kb += ios->qwrite_kb;
1764                         total.cread_kb += ios->cread_kb;
1765                         total.cwrite_kb += ios->cwrite_kb;
1766                         total.iread_kb += ios->iread_kb;
1767                         total.iwrite_kb += ios->iwrite_kb;
1768                         total.mread_kb += ios->mread_kb;
1769                         total.mwrite_kb += ios->mwrite_kb;
1770
1771                         total.qreads_pc += ios->qreads_pc;
1772                         total.qwrites_pc += ios->qwrites_pc;
1773                         total.creads_pc += ios->creads_pc;
1774                         total.cwrites_pc += ios->cwrites_pc;
1775                         total.ireads_pc += ios->ireads_pc;
1776                         total.iwrites_pc += ios->iwrites_pc;
1777                         total.rrqueue_pc += ios->rrqueue_pc;
1778                         total.wrqueue_pc += ios->wrqueue_pc;
1779                         total.qread_kb_pc += ios->qread_kb_pc;
1780                         total.qwrite_kb_pc += ios->qwrite_kb_pc;
1781                         total.iread_kb_pc += ios->iread_kb_pc;
1782                         total.iwrite_kb_pc += ios->iwrite_kb_pc;
1783
1784                         total.timer_unplugs += ios->timer_unplugs;
1785                         total.io_unplugs += ios->io_unplugs;
1786
1787                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1788                                  j, get_dev_name(pdi, name, sizeof(name)));
1789                         dump_io_stats(pdi, ios, line);
1790                         pci_events++;
1791                 }
1792
1793                 if (pci_events > 1) {
1794                         fprintf(ofp, "\n");
1795                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1796                                  get_dev_name(pdi, name, sizeof(name)));
1797                         dump_io_stats(NULL, &total, line);
1798                 }
1799
1800                 wrate = rrate = 0;
1801                 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1802                 if (msec) {
1803                         rrate = 1000 * total.cread_kb / msec;
1804                         wrate = 1000 * total.cwrite_kb / msec;
1805                 }
1806
1807                 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n",
1808                         rrate, wrate);
1809                 fprintf(ofp, "Events (%s): %'Lu entries\n",
1810                         get_dev_name(pdi, line, sizeof(line)), pdi->events);
1811
1812                 collect_pdi_skips(pdi);
1813                 if (!pdi->skips && !pdi->events)
1814                         ratio = 0.0;
1815                 else
1816                         ratio = 100.0 * ((double)pdi->seq_skips /
1817                                         (double)(pdi->events + pdi->seq_skips));
1818                 fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%)\n",
1819                         pdi->skips, pdi->seq_skips, ratio);
1820         }
1821 }
1822
1823 static void find_genesis(void)
1824 {
1825         struct trace *t = trace_list;
1826
1827         genesis_time = -1ULL;
1828         while (t != NULL) {
1829                 if (t->bit->time < genesis_time)
1830                         genesis_time = t->bit->time;
1831
1832                 t = t->next;
1833         }
1834
1835         /* The time stamp record will usually be the first
1836          * record in the trace, but not always.
1837          */
1838         if (start_timestamp
1839          && start_timestamp != genesis_time) {
1840                 long delta = genesis_time - start_timestamp;
1841
1842                 abs_start_time.tv_sec  += SECONDS(delta);
1843                 abs_start_time.tv_nsec += NANO_SECONDS(delta);
1844                 if (abs_start_time.tv_nsec < 0) {
1845                         abs_start_time.tv_nsec += 1000000000;
1846                         abs_start_time.tv_sec -= 1;
1847                 } else
1848                 if (abs_start_time.tv_nsec > 1000000000) {
1849                         abs_start_time.tv_nsec -= 1000000000;
1850                         abs_start_time.tv_sec += 1;
1851                 }
1852         }
1853 }
1854
1855 static inline int check_stopwatch(struct blk_io_trace *bit)
1856 {
1857         if (bit->time < stopwatch_end &&
1858             bit->time >= stopwatch_start)
1859                 return 0;
1860
1861         return 1;
1862 }
1863
1864 /*
1865  * return youngest entry read
1866  */
1867 static int sort_entries(unsigned long long *youngest)
1868 {
1869         struct per_dev_info *pdi = NULL;
1870         struct per_cpu_info *pci = NULL;
1871         struct trace *t;
1872
1873         if (!genesis_time)
1874                 find_genesis();
1875
1876         *youngest = 0;
1877         while ((t = trace_list) != NULL) {
1878                 struct blk_io_trace *bit = t->bit;
1879
1880                 trace_list = t->next;
1881
1882                 bit->time -= genesis_time;
1883
1884                 if (bit->time < *youngest || !*youngest)
1885                         *youngest = bit->time;
1886
1887                 if (!pdi || pdi->dev != bit->device) {
1888                         pdi = get_dev_info(bit->device);
1889                         pci = NULL;
1890                 }
1891
1892                 if (!pci || pci->cpu != bit->cpu)
1893                         pci = get_cpu_info(pdi, bit->cpu);
1894
1895                 if (bit->sequence < pci->smallest_seq_read)
1896                         pci->smallest_seq_read = bit->sequence;
1897
1898                 if (check_stopwatch(bit)) {
1899                         bit_free(bit);
1900                         t_free(t);
1901                         continue;
1902                 }
1903
1904                 if (trace_rb_insert_sort(t))
1905                         return -1;
1906         }
1907
1908         return 0;
1909 }
1910
1911 /*
1912  * to continue, we must have traces from all online cpus in the tree
1913  */
1914 static int check_cpu_map(struct per_dev_info *pdi)
1915 {
1916         unsigned long *cpu_map;
1917         struct rb_node *n;
1918         struct trace *__t;
1919         unsigned int i;
1920         int ret, cpu;
1921
1922         /*
1923          * create a map of the cpus we have traces for
1924          */
1925         cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1926         n = rb_first(&rb_sort_root);
1927         while (n) {
1928                 __t = rb_entry(n, struct trace, rb_node);
1929                 cpu = __t->bit->cpu;
1930
1931                 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1932                 n = rb_next(n);
1933         }
1934
1935         /*
1936          * we can't continue if pdi->cpu_map has entries set that we don't
1937          * have in the sort rbtree. the opposite is not a problem, though
1938          */
1939         ret = 0;
1940         for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1941                 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1942                         ret = 1;
1943                         break;
1944                 }
1945         }
1946
1947         free(cpu_map);
1948         return ret;
1949 }
1950
1951 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1952 {
1953         struct blk_io_trace *bit = t->bit;
1954         unsigned long expected_sequence;
1955         struct per_cpu_info *pci;
1956         struct trace *__t;
1957
1958         pci = get_cpu_info(pdi, bit->cpu);
1959         expected_sequence = pci->last_sequence + 1;
1960
1961         if (!expected_sequence) {
1962                 /*
1963                  * 1 should be the first entry, just allow it
1964                  */
1965                 if (bit->sequence == 1)
1966                         return 0;
1967                 if (bit->sequence == pci->smallest_seq_read)
1968                         return 0;
1969
1970                 return check_cpu_map(pdi);
1971         }
1972
1973         if (bit->sequence == expected_sequence)
1974                 return 0;
1975
1976         /*
1977          * we may not have seen that sequence yet. if we are not doing
1978          * the final run, break and wait for more entries.
1979          */
1980         if (expected_sequence < pci->smallest_seq_read) {
1981                 __t = trace_rb_find_last(pdi, pci, expected_sequence);
1982                 if (!__t)
1983                         goto skip;
1984
1985                 __put_trace_last(pdi, __t);
1986                 return 0;
1987         } else if (!force) {
1988                 return 1;
1989         } else {
1990 skip:
1991                 if (check_current_skips(pci, bit->sequence))
1992                         return 0;
1993
1994                 if (expected_sequence < bit->sequence)
1995                         insert_skip(pci, expected_sequence, bit->sequence - 1);
1996                 return 0;
1997         }
1998 }
1999
2000 static void show_entries_rb(int force)
2001 {
2002         struct per_dev_info *pdi = NULL;
2003         struct per_cpu_info *pci = NULL;
2004         struct blk_io_trace *bit;
2005         struct rb_node *n;
2006         struct trace *t;
2007
2008         while ((n = rb_first(&rb_sort_root)) != NULL) {
2009                 if (is_done() && !force && !pipeline)
2010                         break;
2011
2012                 t = rb_entry(n, struct trace, rb_node);
2013                 bit = t->bit;
2014
2015                 if (read_sequence - t->read_sequence < 1 && !force)
2016                         break;
2017
2018                 if (!pdi || pdi->dev != bit->device) {
2019                         pdi = get_dev_info(bit->device);
2020                         pci = NULL;
2021                 }
2022
2023                 if (!pdi) {
2024                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
2025                                 MAJOR(bit->device), MINOR(bit->device));
2026                         break;
2027                 }
2028
2029                 if (check_sequence(pdi, t, force))
2030                         break;
2031
2032                 if (!force && bit->time > last_allowed_time)
2033                         break;
2034
2035                 check_time(pdi, bit);
2036
2037                 if (!pci || pci->cpu != bit->cpu)
2038                         pci = get_cpu_info(pdi, bit->cpu);
2039
2040                 pci->last_sequence = bit->sequence;
2041
2042                 pci->nelems++;
2043
2044                 if (bit->action & (act_mask << BLK_TC_SHIFT))
2045                         dump_trace(bit, pci, pdi);
2046
2047                 put_trace(pdi, t);
2048         }
2049 }
2050
2051 static int read_data(int fd, void *buffer, int bytes, int block, int *fdblock)
2052 {
2053         int ret, bytes_left, fl;
2054         void *p;
2055
2056         if (block != *fdblock) {
2057                 fl = fcntl(fd, F_GETFL);
2058
2059                 if (!block) {
2060                         *fdblock = 0;
2061                         fcntl(fd, F_SETFL, fl | O_NONBLOCK);
2062                 } else {
2063                         *fdblock = 1;
2064                         fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
2065                 }
2066         }
2067
2068         bytes_left = bytes;
2069         p = buffer;
2070         while (bytes_left > 0) {
2071                 ret = read(fd, p, bytes_left);
2072                 if (!ret)
2073                         return 1;
2074                 else if (ret < 0) {
2075                         if (errno != EAGAIN) {
2076                                 perror("read");
2077                                 return -1;
2078                         }
2079
2080                         /*
2081                          * never do partial reads. we can return if we
2082                          * didn't read anything and we should not block,
2083                          * otherwise wait for data
2084                          */
2085                         if ((bytes_left == bytes) && !block)
2086                                 return 1;
2087
2088                         usleep(10);
2089                         continue;
2090                 } else {
2091                         p += ret;
2092                         bytes_left -= ret;
2093                 }
2094         }
2095
2096         return 0;
2097 }
2098
2099 static inline __u16 get_pdulen(struct blk_io_trace *bit)
2100 {
2101         if (data_is_native)
2102                 return bit->pdu_len;
2103
2104         return __bswap_16(bit->pdu_len);
2105 }
2106
2107 static inline __u32 get_magic(struct blk_io_trace *bit)
2108 {
2109         if (data_is_native)
2110                 return bit->magic;
2111
2112         return __bswap_32(bit->magic);
2113 }
2114
2115 static int read_events(int fd, int always_block, int *fdblock)
2116 {
2117         struct per_dev_info *pdi = NULL;
2118         unsigned int events = 0;
2119
2120         while (!is_done() && events < rb_batch) {
2121                 struct blk_io_trace *bit;
2122                 struct trace *t;
2123                 int pdu_len, should_block, ret;
2124                 __u32 magic;
2125
2126                 bit = bit_alloc();
2127
2128                 should_block = !events || always_block;
2129
2130                 ret = read_data(fd, bit, sizeof(*bit), should_block, fdblock);
2131                 if (ret) {
2132                         bit_free(bit);
2133                         if (!events && ret < 0)
2134                                 events = ret;
2135                         break;
2136                 }
2137
2138                 /*
2139                  * look at first trace to check whether we need to convert
2140                  * data in the future
2141                  */
2142                 if (data_is_native == -1 && check_data_endianness(bit->magic))
2143                         break;
2144
2145                 magic = get_magic(bit);
2146                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2147                         fprintf(stderr, "Bad magic %x\n", magic);
2148                         break;
2149                 }
2150
2151                 pdu_len = get_pdulen(bit);
2152                 if (pdu_len) {
2153                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2154
2155                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1, fdblock)) {
2156                                 bit_free(ptr);
2157                                 break;
2158                         }
2159
2160                         bit = ptr;
2161                 }
2162
2163                 trace_to_cpu(bit);
2164
2165                 if (verify_trace(bit)) {
2166                         bit_free(bit);
2167                         continue;
2168                 }
2169
2170                 /*
2171                  * not a real trace, so grab and handle it here
2172                  */
2173                 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY)) {
2174                         handle_notify(bit);
2175                         output_binary(bit, sizeof(*bit) + bit->pdu_len);
2176                         continue;
2177                 }
2178
2179                 t = t_alloc();
2180                 memset(t, 0, sizeof(*t));
2181                 t->bit = bit;
2182                 t->read_sequence = read_sequence;
2183
2184                 t->next = trace_list;
2185                 trace_list = t;
2186
2187                 if (!pdi || pdi->dev != bit->device)
2188                         pdi = get_dev_info(bit->device);
2189
2190                 if (bit->time > pdi->last_read_time)
2191                         pdi->last_read_time = bit->time;
2192
2193                 events++;
2194         }
2195
2196         return events;
2197 }
2198
2199 /*
2200  * Managing input streams
2201  */
2202
2203 struct ms_stream {
2204         struct ms_stream *next;
2205         struct trace *first, *last;
2206         struct per_dev_info *pdi;
2207         unsigned int cpu;
2208 };
2209
2210 #define MS_HASH(d, c) ((MAJOR(d) & 0xff) ^ (MINOR(d) & 0xff) ^ (cpu & 0xff))
2211
2212 struct ms_stream *ms_head;
2213 struct ms_stream *ms_hash[256];
2214
2215 static void ms_sort(struct ms_stream *msp);
2216 static int ms_prime(struct ms_stream *msp);
2217
2218 static inline struct trace *ms_peek(struct ms_stream *msp)
2219 {
2220         return (msp == NULL) ? NULL : msp->first;
2221 }
2222
2223 static inline __u64 ms_peek_time(struct ms_stream *msp)
2224 {
2225         return ms_peek(msp)->bit->time;
2226 }
2227
2228 static inline void ms_resort(struct ms_stream *msp)
2229 {
2230         if (msp->next && ms_peek_time(msp) > ms_peek_time(msp->next)) {
2231                 ms_head = msp->next;
2232                 msp->next = NULL;
2233                 ms_sort(msp);
2234         }
2235 }
2236
2237 static inline void ms_deq(struct ms_stream *msp)
2238 {
2239         msp->first = msp->first->next;
2240         if (!msp->first) {
2241                 msp->last = NULL;
2242                 if (!ms_prime(msp)) {
2243                         ms_head = msp->next;
2244                         msp->next = NULL;
2245                         return;
2246                 }
2247         }
2248
2249         ms_resort(msp);
2250 }
2251
2252 static void ms_sort(struct ms_stream *msp)
2253 {
2254         __u64 msp_t = ms_peek_time(msp);
2255         struct ms_stream *this_msp = ms_head;
2256
2257         if (this_msp == NULL)
2258                 ms_head = msp;
2259         else if (msp_t < ms_peek_time(this_msp)) {
2260                 msp->next = this_msp;
2261                 ms_head = msp;
2262         }
2263         else {
2264                 while (this_msp->next && ms_peek_time(this_msp->next) < msp_t)
2265                         this_msp = this_msp->next;
2266
2267                 msp->next = this_msp->next;
2268                 this_msp->next = msp;
2269         }
2270 }
2271
2272 static int ms_prime(struct ms_stream *msp)
2273 {
2274         __u32 magic;
2275         unsigned int i;
2276         struct trace *t;
2277         struct per_dev_info *pdi = msp->pdi;
2278         struct per_cpu_info *pci = get_cpu_info(pdi, msp->cpu);
2279         struct blk_io_trace *bit = NULL;
2280         int ret, pdu_len, ndone = 0;
2281
2282         for (i = 0; !is_done() && pci->fd >= 0 && i < rb_batch; i++) {
2283                 bit = bit_alloc();
2284                 ret = read_data(pci->fd, bit, sizeof(*bit), 1, &pci->fdblock);
2285                 if (ret)
2286                         goto err;
2287
2288                 if (data_is_native == -1 && check_data_endianness(bit->magic))
2289                         goto err;
2290
2291                 magic = get_magic(bit);
2292                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2293                         fprintf(stderr, "Bad magic %x\n", magic);
2294                         goto err;
2295
2296                 }
2297
2298                 pdu_len = get_pdulen(bit);
2299                 if (pdu_len) {
2300                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2301                         ret = read_data(pci->fd, ptr + sizeof(*bit), pdu_len,
2302                                                              1, &pci->fdblock);
2303                         if (ret) {
2304                                 free(ptr);
2305                                 bit = NULL;
2306                                 goto err;
2307                         }
2308
2309                         bit = ptr;
2310                 }
2311
2312                 trace_to_cpu(bit);
2313                 if (verify_trace(bit))
2314                         goto err;
2315
2316                 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY)) {
2317                         handle_notify(bit);
2318                         output_binary(bit, sizeof(*bit) + bit->pdu_len);
2319                         bit_free(bit);
2320
2321                         i -= 1;
2322                         continue;
2323                 }
2324
2325                 if (bit->time > pdi->last_read_time)
2326                         pdi->last_read_time = bit->time;
2327
2328                 t = t_alloc();
2329                 memset(t, 0, sizeof(*t));
2330                 t->bit = bit;
2331
2332                 if (msp->first == NULL)
2333                         msp->first = msp->last = t;
2334                 else {
2335                         msp->last->next = t;
2336                         msp->last = t;
2337                 }
2338
2339                 ndone++;
2340         }
2341
2342         return ndone;
2343
2344 err:
2345         if (bit) bit_free(bit);
2346
2347         cpu_mark_offline(pdi, pci->cpu);
2348         close(pci->fd);
2349         pci->fd = -1;
2350
2351         return ndone;
2352 }
2353
2354 static struct ms_stream *ms_alloc(struct per_dev_info *pdi, int cpu)
2355 {
2356         struct ms_stream *msp = malloc(sizeof(*msp));
2357
2358         msp->next = NULL;
2359         msp->first = msp->last = NULL;
2360         msp->pdi = pdi;
2361         msp->cpu = cpu;
2362
2363         if (ms_prime(msp))
2364                 ms_sort(msp);
2365
2366         return msp;
2367 }
2368
2369 static int setup_file(struct per_dev_info *pdi, int cpu)
2370 {
2371         int len = 0;
2372         struct stat st;
2373         char *p, *dname;
2374         struct per_cpu_info *pci = get_cpu_info(pdi, cpu);
2375
2376         pci->cpu = cpu;
2377         pci->fdblock = -1;
2378
2379         p = strdup(pdi->name);
2380         dname = dirname(p);
2381         if (strcmp(dname, ".")) {
2382                 input_dir = dname;
2383                 p = strdup(pdi->name);
2384                 strcpy(pdi->name, basename(p));
2385         }
2386         free(p);
2387
2388         if (input_dir)
2389                 len = sprintf(pci->fname, "%s/", input_dir);
2390
2391         snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
2392                  "%s.blktrace.%d", pdi->name, pci->cpu);
2393         if (stat(pci->fname, &st) < 0)
2394                 return 0;
2395         if (!st.st_size)
2396                 return 1;
2397
2398         pci->fd = open(pci->fname, O_RDONLY);
2399         if (pci->fd < 0) {
2400                 perror(pci->fname);
2401                 return 0;
2402         }
2403
2404         printf("Input file %s added\n", pci->fname);
2405         cpu_mark_online(pdi, pci->cpu);
2406
2407         pdi->nfiles++;
2408         ms_alloc(pdi, pci->cpu);
2409
2410         return 1;
2411 }
2412
2413 static int handle(struct ms_stream *msp)
2414 {
2415         struct trace *t;
2416         struct per_dev_info *pdi;
2417         struct per_cpu_info *pci;
2418         struct blk_io_trace *bit;
2419
2420         t = ms_peek(msp);
2421         if (t->bit->time > stopwatch_end)
2422                 return 0;
2423
2424         bit = t->bit;
2425         pdi = msp->pdi;
2426         pci = get_cpu_info(pdi, msp->cpu);
2427         pci->nelems++;
2428
2429         bit->time -= genesis_time;
2430         pdi->last_reported_time = bit->time;
2431         if (bit->action & (act_mask << BLK_TC_SHIFT))
2432                 dump_trace(bit, pci, pdi);
2433
2434         ms_deq(msp);
2435
2436         if (text_output)
2437                 trace_rb_insert_last(pdi, t);
2438         else {
2439                 bit_free(t->bit);
2440                 t_free(t);
2441         }
2442
2443         return 1;
2444 }
2445
2446 /*
2447  * Check if we need to sanitize the name. We allow 'foo', or if foo.blktrace.X
2448  * is given, then strip back down to 'foo' to avoid missing files.
2449  */
2450 static int name_fixup(char *name)
2451 {
2452         char *b;
2453
2454         if (!name)
2455                 return 1;
2456
2457         b = strstr(name, ".blktrace.");
2458         if (b)
2459                 *b = '\0';
2460
2461         return 0;
2462 }
2463
2464 static int do_file(void)
2465 {
2466         int i, cpu, ret;
2467         struct per_dev_info *pdi;
2468
2469         /*
2470          * first prepare all files for reading
2471          */
2472         for (i = 0; i < ndevices; i++) {
2473                 pdi = &devices[i];
2474                 ret = name_fixup(pdi->name);
2475                 if (ret)
2476                         return ret;
2477
2478                 for (cpu = 0; setup_file(pdi, cpu); cpu++)
2479                         ;
2480         }
2481
2482         /*
2483          * Get the initial time stamp
2484          */
2485         if (ms_head)
2486                 genesis_time = ms_peek_time(ms_head);
2487
2488         /*
2489          * Keep processing traces while any are left
2490          */
2491         while (!is_done() && ms_head && handle(ms_head))
2492                 ;
2493
2494         return 0;
2495 }
2496
2497 static void do_pipe(int fd)
2498 {
2499         unsigned long long youngest;
2500         int events, fdblock;
2501
2502         last_allowed_time = -1ULL;
2503         fdblock = -1;
2504         while ((events = read_events(fd, 0, &fdblock)) > 0) {
2505                 read_sequence++;
2506
2507 #if 0
2508                 smallest_seq_read = -1U;
2509 #endif
2510
2511                 if (sort_entries(&youngest))
2512                         break;
2513
2514                 if (youngest > stopwatch_end)
2515                         break;
2516
2517                 show_entries_rb(0);
2518         }
2519
2520         if (rb_sort_entries)
2521                 show_entries_rb(1);
2522 }
2523
2524 static int do_fifo(void)
2525 {
2526         int fd;
2527
2528         if (!strcmp(pipename, "-"))
2529                 fd = dup(STDIN_FILENO);
2530         else
2531                 fd = open(pipename, O_RDONLY);
2532
2533         if (fd == -1) {
2534                 perror("dup stdin");
2535                 return -1;
2536         }
2537
2538         do_pipe(fd);
2539         close(fd);
2540         return 0;
2541 }
2542
2543 static void show_stats(void)
2544 {
2545         if (!ofp)
2546                 return;
2547         if (stats_printed)
2548                 return;
2549
2550         stats_printed = 1;
2551
2552         if (per_process_stats)
2553                 show_process_stats();
2554
2555         if (per_device_and_cpu_stats)
2556                 show_device_and_cpu_stats();
2557
2558         fflush(ofp);
2559 }
2560
2561 static void handle_sigint(__attribute__((__unused__)) int sig)
2562 {
2563         done = 1;
2564 }
2565
2566 /*
2567  * Extract start and duration times from a string, allowing
2568  * us to specify a time interval of interest within a trace.
2569  * Format: "duration" (start is zero) or "start:duration".
2570  */
2571 static int find_stopwatch_interval(char *string)
2572 {
2573         double value;
2574         char *sp;
2575
2576         value = strtod(string, &sp);
2577         if (sp == string) {
2578                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
2579                 return 1;
2580         }
2581         if (*sp == ':') {
2582                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
2583                 string = sp + 1;
2584                 value = strtod(string, &sp);
2585                 if (sp == string || *sp != '\0') {
2586                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
2587                                 string);
2588                         return 1;
2589                 }
2590         } else if (*sp != '\0') {
2591                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
2592                 return 1;
2593         }
2594         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
2595         if (stopwatch_end <= stopwatch_start) {
2596                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
2597                         stopwatch_start, stopwatch_end);
2598                 return 1;
2599         }
2600
2601         return 0;
2602 }
2603
2604 static int is_pipe(const char *str)
2605 {
2606         struct stat st;
2607
2608         if (!strcmp(str, "-"))
2609                 return 1;
2610         if (!stat(str, &st) && S_ISFIFO(st.st_mode))
2611                 return 1;
2612
2613         return 0;
2614 }
2615
2616 #define S_OPTS  "a:A:b:D:d:f:F:hi:o:Oqstw:vV"
2617 static char usage_str[] =    "\n\n" \
2618         "-i <file>           | --input=<file>\n" \
2619         "[ -a <action field> | --act-mask=<action field> ]\n" \
2620         "[ -A <action mask>  | --set-mask=<action mask> ]\n" \
2621         "[ -b <traces>       | --batch=<traces> ]\n" \
2622         "[ -d <file>         | --dump-binary=<file> ]\n" \
2623         "[ -D <dir>          | --input-directory=<dir> ]\n" \
2624         "[ -f <format>       | --format=<format> ]\n" \
2625         "[ -F <spec>         | --format-spec=<spec> ]\n" \
2626         "[ -h                | --hash-by-name ]\n" \
2627         "[ -o <file>         | --output=<file> ]\n" \
2628         "[ -O                | --no-text-output ]\n" \
2629         "[ -q                | --quiet ]\n" \
2630         "[ -s                | --per-program-stats ]\n" \
2631         "[ -t                | --track-ios ]\n" \
2632         "[ -w <time>         | --stopwatch=<time> ]\n" \
2633         "[ -v                | --verbose ]\n" \
2634         "[ -V                | --version ]\n\n" \
2635         "\t-b stdin read batching\n" \
2636         "\t-d Output file. If specified, binary data is written to file\n" \
2637         "\t-D Directory to prepend to input file names\n" \
2638         "\t-f Output format. Customize the output format. The format field\n" \
2639         "\t   identifies can be found in the documentation\n" \
2640         "\t-F Format specification. Can be found in the documentation\n" \
2641         "\t-h Hash processes by name, not pid\n" \
2642         "\t-i Input file containing trace data, or '-' for stdin\n" \
2643         "\t-o Output file. If not given, output is stdout\n" \
2644         "\t-O Do NOT output text data\n" \
2645         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
2646         "\t-s Show per-program io statistics\n" \
2647         "\t-t Track individual ios. Will tell you the time a request took\n" \
2648         "\t   to get queued, to get dispatched, and to get completed\n" \
2649         "\t-w Only parse data between the given time interval in seconds.\n" \
2650         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
2651         "\t-v More verbose for marginal errors\n" \
2652         "\t-V Print program version info\n\n";
2653
2654 static void usage(char *prog)
2655 {
2656         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
2657 }
2658
2659 int main(int argc, char *argv[])
2660 {
2661         int i, c, ret, mode;
2662         int act_mask_tmp = 0;
2663         char *ofp_buffer = NULL;
2664         char *bin_ofp_buffer = NULL;
2665
2666         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
2667                 switch (c) {
2668                 case 'a':
2669                         i = find_mask_map(optarg);
2670                         if (i < 0) {
2671                                 fprintf(stderr,"Invalid action mask %s\n",
2672                                         optarg);
2673                                 return 1;
2674                         }
2675                         act_mask_tmp |= i;
2676                         break;
2677
2678                 case 'A':
2679                         if ((sscanf(optarg, "%x", &i) != 1) ||
2680                                                         !valid_act_opt(i)) {
2681                                 fprintf(stderr,
2682                                         "Invalid set action mask %s/0x%x\n",
2683                                         optarg, i);
2684                                 return 1;
2685                         }
2686                         act_mask_tmp = i;
2687                         break;
2688                 case 'i':
2689                         if (is_pipe(optarg) && !pipeline) {
2690                                 pipeline = 1;
2691                                 pipename = strdup(optarg);
2692                         } else if (resize_devices(optarg) != 0)
2693                                 return 1;
2694                         break;
2695                 case 'D':
2696                         input_dir = optarg;
2697                         break;
2698                 case 'o':
2699                         output_name = optarg;
2700                         break;
2701                 case 'O':
2702                         text_output = 0;
2703                         break;
2704                 case 'b':
2705                         rb_batch = atoi(optarg);
2706                         if (rb_batch <= 0)
2707                                 rb_batch = RB_BATCH_DEFAULT;
2708                         break;
2709                 case 's':
2710                         per_process_stats = 1;
2711                         break;
2712                 case 't':
2713                         track_ios = 1;
2714                         break;
2715                 case 'q':
2716                         per_device_and_cpu_stats = 0;
2717                         break;
2718                 case 'w':
2719                         if (find_stopwatch_interval(optarg) != 0)
2720                                 return 1;
2721                         break;
2722                 case 'f':
2723                         set_all_format_specs(optarg);
2724                         break;
2725                 case 'F':
2726                         if (add_format_spec(optarg) != 0)
2727                                 return 1;
2728                         break;
2729                 case 'h':
2730                         ppi_hash_by_pid = 0;
2731                         break;
2732                 case 'v':
2733                         verbose++;
2734                         break;
2735                 case 'V':
2736                         printf("%s version %s\n", argv[0], blkparse_version);
2737                         return 0;
2738                 case 'd':
2739                         dump_binary = optarg;
2740                         break;
2741                 default:
2742                         usage(argv[0]);
2743                         return 1;
2744                 }
2745         }
2746
2747         while (optind < argc) {
2748                 if (is_pipe(argv[optind]) && !pipeline) {
2749                         pipeline = 1;
2750                         pipename = strdup(argv[optind]);
2751                 } else if (resize_devices(argv[optind]) != 0)
2752                         return 1;
2753                 optind++;
2754         }
2755
2756         if (!pipeline && !ndevices) {
2757                 usage(argv[0]);
2758                 return 1;
2759         }
2760
2761         if (act_mask_tmp != 0)
2762                 act_mask = act_mask_tmp;
2763
2764         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
2765
2766         signal(SIGINT, handle_sigint);
2767         signal(SIGHUP, handle_sigint);
2768         signal(SIGTERM, handle_sigint);
2769
2770         setlocale(LC_NUMERIC, "en_US");
2771
2772         if (text_output) {
2773                 if (!output_name) {
2774                         ofp = fdopen(STDOUT_FILENO, "w");
2775                         mode = _IOLBF;
2776                 } else {
2777                         char ofname[128];
2778
2779                         snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
2780                         ofp = fopen(ofname, "w");
2781                         mode = _IOFBF;
2782                 }
2783
2784                 if (!ofp) {
2785                         perror("fopen");
2786                         return 1;
2787                 }
2788
2789                 ofp_buffer = malloc(4096);
2790                 if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
2791                         perror("setvbuf");
2792                         return 1;
2793                 }
2794         }
2795
2796         if (dump_binary) {
2797                 dump_fp = fopen(dump_binary, "w");
2798                 if (!dump_fp) {
2799                         perror(dump_binary);
2800                         dump_binary = NULL;
2801                         return 1;
2802                 }
2803                 bin_ofp_buffer = malloc(128 * 1024);
2804                 if (setvbuf(dump_fp, bin_ofp_buffer, _IOFBF, 128 * 1024)) {
2805                         perror("setvbuf binary");
2806                         return 1;
2807                 }
2808         }
2809
2810         if (pipeline)
2811                 ret = do_fifo();
2812         else
2813                 ret = do_file();
2814
2815         if (!ret)
2816                 show_stats();
2817
2818         if (ofp_buffer) {
2819                 fflush(ofp);
2820                 free(ofp_buffer);
2821         }
2822         if (bin_ofp_buffer) {
2823                 fflush(dump_fp);
2824                 free(bin_ofp_buffer);
2825         }
2826         return ret;
2827 }