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[] = "0.99.2";
  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_queue(struct io_stats *ios, struct blk_io_trace *t,
1172                                    int rw)
1173 {
1174         if (rw) {
1175                 ios->qwrites++;
1176                 ios->qwrite_kb += t_kb(t);
1177         } else {
1178                 ios->qreads++;
1179                 ios->qread_kb += t_kb(t);
1180         }
1181 }
1182
1183 static inline void account_queue(struct blk_io_trace *t,
1184                                  struct per_cpu_info *pci, int rw)
1185 {
1186         __account_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_queue(ios, t, rw);
1192         }
1193 }
1194
1195 static inline void __account_c(struct io_stats *ios, int rw, int bytes)
1196 {
1197         if (rw) {
1198                 ios->cwrites++;
1199                 ios->cwrite_kb += bytes >> 10;
1200         } else {
1201                 ios->creads++;
1202                 ios->cread_kb += bytes >> 10;
1203         }
1204 }
1205
1206 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
1207                              int rw, int bytes)
1208 {
1209         __account_c(&pci->io_stats, rw, bytes);
1210
1211         if (per_process_stats) {
1212                 struct io_stats *ios = find_process_io_stats(t->pid);
1213
1214                 __account_c(ios, rw, bytes);
1215         }
1216 }
1217
1218 static inline void __account_issue(struct io_stats *ios, int rw,
1219                                    unsigned int bytes)
1220 {
1221         if (rw) {
1222                 ios->iwrites++;
1223                 ios->iwrite_kb += bytes >> 10;
1224         } else {
1225                 ios->ireads++;
1226                 ios->iread_kb += bytes >> 10;
1227         }
1228 }
1229
1230 static inline void account_issue(struct blk_io_trace *t,
1231                                  struct per_cpu_info *pci, int rw)
1232 {
1233         __account_issue(&pci->io_stats, rw, t->bytes);
1234
1235         if (per_process_stats) {
1236                 struct io_stats *ios = find_process_io_stats(t->pid);
1237
1238                 __account_issue(ios, rw, t->bytes);
1239         }
1240 }
1241
1242 static inline void __account_unplug(struct io_stats *ios, int timer)
1243 {
1244         if (timer)
1245                 ios->timer_unplugs++;
1246         else
1247                 ios->io_unplugs++;
1248 }
1249
1250 static inline void account_unplug(struct blk_io_trace *t,
1251                                   struct per_cpu_info *pci, int timer)
1252 {
1253         __account_unplug(&pci->io_stats, timer);
1254
1255         if (per_process_stats) {
1256                 struct io_stats *ios = find_process_io_stats(t->pid);
1257
1258                 __account_unplug(ios, timer);
1259         }
1260 }
1261
1262 static inline void __account_requeue(struct io_stats *ios,
1263                                      struct blk_io_trace *t, int rw)
1264 {
1265         if (rw) {
1266                 ios->wrqueue++;
1267                 ios->iwrite_kb -= t_kb(t);
1268         } else {
1269                 ios->rrqueue++;
1270                 ios->iread_kb -= t_kb(t);
1271         }
1272 }
1273
1274 static inline void account_requeue(struct blk_io_trace *t,
1275                                    struct per_cpu_info *pci, int rw)
1276 {
1277         __account_requeue(&pci->io_stats, t, rw);
1278
1279         if (per_process_stats) {
1280                 struct io_stats *ios = find_process_io_stats(t->pid);
1281
1282                 __account_requeue(ios, t, rw);
1283         }
1284 }
1285
1286 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
1287                          struct blk_io_trace *t, char *act)
1288 {
1289         process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
1290 }
1291
1292 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
1293                        struct blk_io_trace *t, char *act)
1294 {
1295         process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
1296 }
1297
1298 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
1299                       char *act)
1300 {
1301         process_fmt(act, pci, t, -1, 0, NULL);
1302 }
1303
1304 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
1305                       struct blk_io_trace *t, char *act)
1306 {
1307         process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
1308 }
1309
1310 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
1311                       struct blk_io_trace *t, char *act)
1312 {
1313         if (act[0] == 'F')
1314                 log_track_frontmerge(pdi, t);
1315
1316         process_fmt(act, pci, t, -1ULL, 0, NULL);
1317 }
1318
1319 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
1320                         char *act)
1321 {
1322         process_fmt(act, pci, t, -1ULL, 0, NULL);
1323 }
1324
1325 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
1326                         char *act)
1327 {
1328         process_fmt(act, pci, t, -1ULL, 0, NULL);
1329 }
1330
1331 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
1332                       char *act)
1333 {
1334         process_fmt(act, pci, t, -1ULL, 0, NULL);
1335 }
1336
1337 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
1338                       char *act)
1339 {
1340         process_fmt(act, pci, t, -1ULL, 0, NULL);
1341 }
1342
1343 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1344 {
1345         unsigned char *buf = (unsigned char *) t + sizeof(*t);
1346
1347         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1348 }
1349
1350 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
1351 {
1352         int act = t->action & 0xffff;
1353
1354         switch (act) {
1355                 case __BLK_TA_QUEUE:
1356                         log_generic(pci, t, "Q");
1357                         break;
1358                 case __BLK_TA_GETRQ:
1359                         log_generic(pci, t, "G");
1360                         break;
1361                 case __BLK_TA_SLEEPRQ:
1362                         log_generic(pci, t, "S");
1363                         break;
1364                 case __BLK_TA_REQUEUE:
1365                         log_generic(pci, t, "R");
1366                         break;
1367                 case __BLK_TA_ISSUE:
1368                         log_pc(pci, t, "D");
1369                         break;
1370                 case __BLK_TA_COMPLETE:
1371                         log_pc(pci, t, "C");
1372                         break;
1373                 case __BLK_TA_INSERT:
1374                         log_pc(pci, t, "I");
1375                         break;
1376                 default:
1377                         fprintf(stderr, "Bad pc action %x\n", act);
1378                         break;
1379         }
1380 }
1381
1382 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1383                           struct per_cpu_info *pci)
1384 {
1385         int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1386         int act = t->action & 0xffff;
1387
1388         switch (act) {
1389                 case __BLK_TA_QUEUE:
1390                         log_track_queue(pdi, t);
1391                         account_queue(t, pci, w);
1392                         log_queue(pci, t, "Q");
1393                         break;
1394                 case __BLK_TA_INSERT:
1395                         log_insert(pdi, pci, t, "I");
1396                         break;
1397                 case __BLK_TA_BACKMERGE:
1398                         account_m(t, pci, w);
1399                         log_merge(pdi, pci, t, "M");
1400                         break;
1401                 case __BLK_TA_FRONTMERGE:
1402                         account_m(t, pci, w);
1403                         log_merge(pdi, pci, t, "F");
1404                         break;
1405                 case __BLK_TA_GETRQ:
1406                         log_track_getrq(pdi, t);
1407                         log_generic(pci, t, "G");
1408                         break;
1409                 case __BLK_TA_SLEEPRQ:
1410                         log_generic(pci, t, "S");
1411                         break;
1412                 case __BLK_TA_REQUEUE:
1413                         /*
1414                          * can happen if we miss traces, don't let it go
1415                          * below zero
1416                          */
1417                         if (pdi->cur_depth[w])
1418                                 pdi->cur_depth[w]--;
1419                         account_requeue(t, pci, w);
1420                         log_queue(pci, t, "R");
1421                         break;
1422                 case __BLK_TA_ISSUE:
1423                         account_issue(t, pci, w);
1424                         pdi->cur_depth[w]++;
1425                         if (pdi->cur_depth[w] > pdi->max_depth[w])
1426                                 pdi->max_depth[w] = pdi->cur_depth[w];
1427                         log_issue(pdi, pci, t, "D");
1428                         break;
1429                 case __BLK_TA_COMPLETE:
1430                         if (pdi->cur_depth[w])
1431                                 pdi->cur_depth[w]--;
1432                         account_c(t, pci, w, t->bytes);
1433                         log_complete(pdi, pci, t, "C");
1434                         break;
1435                 case __BLK_TA_PLUG:
1436                         log_action(pci, t, "P");
1437                         break;
1438                 case __BLK_TA_UNPLUG_IO:
1439                         account_unplug(t, pci, 0);
1440                         log_unplug(pci, t, "U");
1441                         break;
1442                 case __BLK_TA_UNPLUG_TIMER:
1443                         account_unplug(t, pci, 1);
1444                         log_unplug(pci, t, "UT");
1445                         break;
1446                 case __BLK_TA_SPLIT:
1447                         log_split(pci, t, "X");
1448                         break;
1449                 case __BLK_TA_BOUNCE:
1450                         log_generic(pci, t, "B");
1451                         break;
1452                 case __BLK_TA_REMAP:
1453                         log_generic(pci, t, "A");
1454                         break;
1455                 default:
1456                         fprintf(stderr, "Bad fs action %x\n", t->action);
1457                         break;
1458         }
1459 }
1460
1461 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1462                        struct per_dev_info *pdi)
1463 {
1464         if (text_output) {
1465                 if (t->action & BLK_TC_ACT(BLK_TC_PC))
1466                         dump_trace_pc(t, pci);
1467                 else
1468                         dump_trace_fs(t, pdi, pci);
1469         }
1470
1471         if (!pdi->events)
1472                 pdi->first_reported_time = t->time;
1473
1474         pdi->events++;
1475
1476         output_binary(t, sizeof(*t) + t->pdu_len);
1477 }
1478
1479 /*
1480  * print in a proper way, not too small and not too big. if more than
1481  * 1000,000K, turn into M and so on
1482  */
1483 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1484 {
1485         char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1486         unsigned int i = 0;
1487
1488         if (in_kb)
1489                 i++;
1490
1491         while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1492                 i++;
1493                 num /= 1000;
1494         }
1495
1496         sprintf(dst, "%'8Lu%c", num, suff[i]);
1497         return dst;
1498 }
1499
1500 static void dump_io_stats(struct per_dev_info *pdi, struct io_stats *ios,
1501                           char *msg)
1502 {
1503         static char x[256], y[256];
1504
1505         fprintf(ofp, "%s\n", msg);
1506
1507         fprintf(ofp, " Reads Queued:    %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1508         fprintf(ofp, " Writes Queued:    %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1509         fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1510         fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1511         fprintf(ofp, " Reads Requeued:  %s\t\t", size_cnv(x, ios->rrqueue, 0));
1512         fprintf(ofp, " Writes Requeued:  %s\n", size_cnv(x, ios->wrqueue, 0));
1513         fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1514         fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1515         fprintf(ofp, " Read Merges:     %s, %siB\t", size_cnv(x, ios->mreads, 0), size_cnv(y, ios->mread_kb, 1));
1516         fprintf(ofp, " Write Merges:     %s, %siB\n", size_cnv(x, ios->mwrites, 0), size_cnv(y, ios->mwrite_kb, 1));
1517         if (pdi) {
1518                 fprintf(ofp, " Read depth:      %'8u%8c\t", pdi->max_depth[0], ' ');
1519                 fprintf(ofp, " Write depth:      %'8u\n", pdi->max_depth[1]);
1520         }
1521         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
1522         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
1523 }
1524
1525 static void dump_wait_stats(struct per_process_info *ppi)
1526 {
1527         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1528         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1529         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1530         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1531         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1532         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1533
1534         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1535         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
1536         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
1537         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
1538         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1539         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
1540 }
1541
1542 static int ppi_name_compare(const void *p1, const void *p2)
1543 {
1544         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1545         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1546         int res;
1547
1548         res = strverscmp(ppi1->ppm->comm, ppi2->ppm->comm);
1549         if (!res)
1550                 res = ppi1->ppm->pid > ppi2->ppm->pid;
1551
1552         return res;
1553 }
1554
1555 static void sort_process_list(void)
1556 {
1557         struct per_process_info **ppis;
1558         struct per_process_info *ppi;
1559         int i = 0;
1560
1561         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1562
1563         ppi = ppi_list;
1564         while (ppi) {
1565                 ppis[i++] = ppi;
1566                 ppi = ppi->list_next;
1567         }
1568
1569         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1570
1571         i = ppi_list_entries - 1;
1572         ppi_list = NULL;
1573         while (i >= 0) {
1574                 ppi = ppis[i];
1575
1576                 ppi->list_next = ppi_list;
1577                 ppi_list = ppi;
1578                 i--;
1579         }
1580
1581         free(ppis);
1582 }
1583
1584 static void show_process_stats(void)
1585 {
1586         struct per_process_info *ppi;
1587
1588         sort_process_list();
1589
1590         ppi = ppi_list;
1591         while (ppi) {
1592                 struct process_pid_map *ppm = ppi->ppm;
1593                 char name[64];
1594
1595                 if (ppi->more_than_one)
1596                         sprintf(name, "%s (%u, ...)", ppm->comm, ppm->pid);
1597                 else
1598                         sprintf(name, "%s (%u)", ppm->comm, ppm->pid);
1599
1600                 dump_io_stats(NULL, &ppi->io_stats, name);
1601                 dump_wait_stats(ppi);
1602                 ppi = ppi->list_next;
1603         }
1604
1605         fprintf(ofp, "\n");
1606 }
1607
1608 static void show_device_and_cpu_stats(void)
1609 {
1610         struct per_dev_info *pdi;
1611         struct per_cpu_info *pci;
1612         struct io_stats total, *ios;
1613         unsigned long long rrate, wrate, msec;
1614         int i, j, pci_events;
1615         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1616         char name[32];
1617         double ratio;
1618
1619         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1620
1621                 memset(&total, 0, sizeof(total));
1622                 pci_events = 0;
1623
1624                 if (i > 0)
1625                         fprintf(ofp, "\n");
1626
1627                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1628                         if (!pci->nelems)
1629                                 continue;
1630
1631                         ios = &pci->io_stats;
1632                         total.qreads += ios->qreads;
1633                         total.qwrites += ios->qwrites;
1634                         total.creads += ios->creads;
1635                         total.cwrites += ios->cwrites;
1636                         total.mreads += ios->mreads;
1637                         total.mwrites += ios->mwrites;
1638                         total.ireads += ios->ireads;
1639                         total.iwrites += ios->iwrites;
1640                         total.rrqueue += ios->rrqueue;
1641                         total.wrqueue += ios->wrqueue;
1642                         total.qread_kb += ios->qread_kb;
1643                         total.qwrite_kb += ios->qwrite_kb;
1644                         total.cread_kb += ios->cread_kb;
1645                         total.cwrite_kb += ios->cwrite_kb;
1646                         total.iread_kb += ios->iread_kb;
1647                         total.iwrite_kb += ios->iwrite_kb;
1648                         total.mread_kb += ios->mread_kb;
1649                         total.mwrite_kb += ios->mwrite_kb;
1650                         total.timer_unplugs += ios->timer_unplugs;
1651                         total.io_unplugs += ios->io_unplugs;
1652
1653                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1654                                  j, get_dev_name(pdi, name, sizeof(name)));
1655                         dump_io_stats(pdi, ios, line);
1656                         pci_events++;
1657                 }
1658
1659                 if (pci_events > 1) {
1660                         fprintf(ofp, "\n");
1661                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1662                                  get_dev_name(pdi, name, sizeof(name)));
1663                         dump_io_stats(NULL, &total, line);
1664                 }
1665
1666                 wrate = rrate = 0;
1667                 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1668                 if (msec) {
1669                         rrate = 1000 * total.cread_kb / msec;
1670                         wrate = 1000 * total.cwrite_kb / msec;
1671                 }
1672
1673                 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n",
1674                         rrate, wrate);
1675                 fprintf(ofp, "Events (%s): %'Lu entries\n",
1676                         get_dev_name(pdi, line, sizeof(line)), pdi->events);
1677
1678                 collect_pdi_skips(pdi);
1679                 if (!pdi->skips && !pdi->events)
1680                         ratio = 0.0;
1681                 else
1682                         ratio = 100.0 * ((double)pdi->seq_skips /
1683                                         (double)(pdi->events + pdi->seq_skips));
1684                 fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%)\n",
1685                         pdi->skips, pdi->seq_skips, ratio);
1686         }
1687 }
1688
1689 static void find_genesis(void)
1690 {
1691         struct trace *t = trace_list;
1692
1693         genesis_time = -1ULL;
1694         while (t != NULL) {
1695                 if (t->bit->time < genesis_time)
1696                         genesis_time = t->bit->time;
1697
1698                 t = t->next;
1699         }
1700
1701         /* The time stamp record will usually be the first
1702          * record in the trace, but not always.
1703          */
1704         if (start_timestamp
1705          && start_timestamp != genesis_time) {
1706                 long delta = genesis_time - start_timestamp;
1707
1708                 abs_start_time.tv_sec  += SECONDS(delta);
1709                 abs_start_time.tv_nsec += NANO_SECONDS(delta);
1710                 if (abs_start_time.tv_nsec < 0) {
1711                         abs_start_time.tv_nsec += 1000000000;
1712                         abs_start_time.tv_sec -= 1;
1713                 } else
1714                 if (abs_start_time.tv_nsec > 1000000000) {
1715                         abs_start_time.tv_nsec -= 1000000000;
1716                         abs_start_time.tv_sec += 1;
1717                 }
1718         }
1719 }
1720
1721 static inline int check_stopwatch(struct blk_io_trace *bit)
1722 {
1723         if (bit->time < stopwatch_end &&
1724             bit->time >= stopwatch_start)
1725                 return 0;
1726
1727         return 1;
1728 }
1729
1730 /*
1731  * return youngest entry read
1732  */
1733 static int sort_entries(unsigned long long *youngest)
1734 {
1735         struct per_dev_info *pdi = NULL;
1736         struct per_cpu_info *pci = NULL;
1737         struct trace *t;
1738
1739         if (!genesis_time)
1740                 find_genesis();
1741
1742         *youngest = 0;
1743         while ((t = trace_list) != NULL) {
1744                 struct blk_io_trace *bit = t->bit;
1745
1746                 trace_list = t->next;
1747
1748                 bit->time -= genesis_time;
1749
1750                 if (bit->time < *youngest || !*youngest)
1751                         *youngest = bit->time;
1752
1753                 if (!pdi || pdi->dev != bit->device) {
1754                         pdi = get_dev_info(bit->device);
1755                         pci = NULL;
1756                 }
1757
1758                 if (!pci || pci->cpu != bit->cpu)
1759                         pci = get_cpu_info(pdi, bit->cpu);
1760
1761                 if (bit->sequence < pci->smallest_seq_read)
1762                         pci->smallest_seq_read = bit->sequence;
1763
1764                 if (check_stopwatch(bit)) {
1765                         bit_free(bit);
1766                         t_free(t);
1767                         continue;
1768                 }
1769
1770                 if (trace_rb_insert_sort(t))
1771                         return -1;
1772         }
1773
1774         return 0;
1775 }
1776
1777 /*
1778  * to continue, we must have traces from all online cpus in the tree
1779  */
1780 static int check_cpu_map(struct per_dev_info *pdi)
1781 {
1782         unsigned long *cpu_map;
1783         struct rb_node *n;
1784         struct trace *__t;
1785         unsigned int i;
1786         int ret, cpu;
1787
1788         /*
1789          * create a map of the cpus we have traces for
1790          */
1791         cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1792         n = rb_first(&rb_sort_root);
1793         while (n) {
1794                 __t = rb_entry(n, struct trace, rb_node);
1795                 cpu = __t->bit->cpu;
1796
1797                 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1798                 n = rb_next(n);
1799         }
1800
1801         /*
1802          * we can't continue if pdi->cpu_map has entries set that we don't
1803          * have in the sort rbtree. the opposite is not a problem, though
1804          */
1805         ret = 0;
1806         for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1807                 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1808                         ret = 1;
1809                         break;
1810                 }
1811         }
1812
1813         free(cpu_map);
1814         return ret;
1815 }
1816
1817 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1818 {
1819         struct blk_io_trace *bit = t->bit;
1820         unsigned long expected_sequence;
1821         struct per_cpu_info *pci;
1822         struct trace *__t;
1823
1824         pci = get_cpu_info(pdi, bit->cpu);
1825         expected_sequence = pci->last_sequence + 1;
1826
1827         if (!expected_sequence) {
1828                 /*
1829                  * 1 should be the first entry, just allow it
1830                  */
1831                 if (bit->sequence == 1)
1832                         return 0;
1833                 if (bit->sequence == pci->smallest_seq_read)
1834                         return 0;
1835
1836                 return check_cpu_map(pdi);
1837         }
1838
1839         if (bit->sequence == expected_sequence)
1840                 return 0;
1841
1842         /*
1843          * we may not have seen that sequence yet. if we are not doing
1844          * the final run, break and wait for more entries.
1845          */
1846         if (expected_sequence < pci->smallest_seq_read) {
1847                 __t = trace_rb_find_last(pdi, pci, expected_sequence);
1848                 if (!__t)
1849                         goto skip;
1850
1851                 __put_trace_last(pdi, __t);
1852                 return 0;
1853         } else if (!force) {
1854                 return 1;
1855         } else {
1856 skip:
1857                 if (check_current_skips(pci, bit->sequence))
1858                         return 0;
1859
1860                 if (expected_sequence < bit->sequence)
1861                         insert_skip(pci, expected_sequence, bit->sequence - 1);
1862                 return 0;
1863         }
1864 }
1865
1866 static void show_entries_rb(int force)
1867 {
1868         struct per_dev_info *pdi = NULL;
1869         struct per_cpu_info *pci = NULL;
1870         struct blk_io_trace *bit;
1871         struct rb_node *n;
1872         struct trace *t;
1873
1874         while ((n = rb_first(&rb_sort_root)) != NULL) {
1875                 if (is_done() && !force && !pipeline)
1876                         break;
1877
1878                 t = rb_entry(n, struct trace, rb_node);
1879                 bit = t->bit;
1880
1881                 if (read_sequence - t->read_sequence < 1 && !force)
1882                         break;
1883
1884                 if (!pdi || pdi->dev != bit->device) {
1885                         pdi = get_dev_info(bit->device);
1886                         pci = NULL;
1887                 }
1888
1889                 if (!pdi) {
1890                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1891                                 MAJOR(bit->device), MINOR(bit->device));
1892                         break;
1893                 }
1894
1895                 if (check_sequence(pdi, t, force))
1896                         break;
1897
1898                 if (!force && bit->time > last_allowed_time)
1899                         break;
1900
1901                 check_time(pdi, bit);
1902
1903                 if (!pci || pci->cpu != bit->cpu)
1904                         pci = get_cpu_info(pdi, bit->cpu);
1905
1906                 pci->last_sequence = bit->sequence;
1907
1908                 pci->nelems++;
1909
1910                 if (bit->action & (act_mask << BLK_TC_SHIFT))
1911                         dump_trace(bit, pci, pdi);
1912
1913                 put_trace(pdi, t);
1914         }
1915 }
1916
1917 static int read_data(int fd, void *buffer, int bytes, int block, int *fdblock)
1918 {
1919         int ret, bytes_left, fl;
1920         void *p;
1921
1922         if (block != *fdblock) {
1923                 fl = fcntl(fd, F_GETFL);
1924
1925                 if (!block) {
1926                         *fdblock = 0;
1927                         fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1928                 } else {
1929                         *fdblock = 1;
1930                         fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1931                 }
1932         }
1933
1934         bytes_left = bytes;
1935         p = buffer;
1936         while (bytes_left > 0) {
1937                 ret = read(fd, p, bytes_left);
1938                 if (!ret)
1939                         return 1;
1940                 else if (ret < 0) {
1941                         if (errno != EAGAIN) {
1942                                 perror("read");
1943                                 return -1;
1944                         }
1945
1946                         /*
1947                          * never do partial reads. we can return if we
1948                          * didn't read anything and we should not block,
1949                          * otherwise wait for data
1950                          */
1951                         if ((bytes_left == bytes) && !block)
1952                                 return 1;
1953
1954                         usleep(10);
1955                         continue;
1956                 } else {
1957                         p += ret;
1958                         bytes_left -= ret;
1959                 }
1960         }
1961
1962         return 0;
1963 }
1964
1965 static inline __u16 get_pdulen(struct blk_io_trace *bit)
1966 {
1967         if (data_is_native)
1968                 return bit->pdu_len;
1969
1970         return __bswap_16(bit->pdu_len);
1971 }
1972
1973 static inline __u32 get_magic(struct blk_io_trace *bit)
1974 {
1975         if (data_is_native)
1976                 return bit->magic;
1977
1978         return __bswap_32(bit->magic);
1979 }
1980
1981 static int read_events(int fd, int always_block, int *fdblock)
1982 {
1983         struct per_dev_info *pdi = NULL;
1984         unsigned int events = 0;
1985
1986         while (!is_done() && events < rb_batch) {
1987                 struct blk_io_trace *bit;
1988                 struct trace *t;
1989                 int pdu_len, should_block, ret;
1990                 __u32 magic;
1991
1992                 bit = bit_alloc();
1993
1994                 should_block = !events || always_block;
1995
1996                 ret = read_data(fd, bit, sizeof(*bit), should_block, fdblock);
1997                 if (ret) {
1998                         bit_free(bit);
1999                         if (!events && ret < 0)
2000                                 events = ret;
2001                         break;
2002                 }
2003
2004                 /*
2005                  * look at first trace to check whether we need to convert
2006                  * data in the future
2007                  */
2008                 if (data_is_native == -1 && check_data_endianness(bit->magic))
2009                         break;
2010
2011                 magic = get_magic(bit);
2012                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2013                         fprintf(stderr, "Bad magic %x\n", magic);
2014                         break;
2015                 }
2016
2017                 pdu_len = get_pdulen(bit);
2018                 if (pdu_len) {
2019                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2020
2021                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1, fdblock)) {
2022                                 bit_free(ptr);
2023                                 break;
2024                         }
2025
2026                         bit = ptr;
2027                 }
2028
2029                 trace_to_cpu(bit);
2030
2031                 if (verify_trace(bit)) {
2032                         bit_free(bit);
2033                         continue;
2034                 }
2035
2036                 /*
2037                  * not a real trace, so grab and handle it here
2038                  */
2039                 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY)) {
2040                         handle_notify(bit);
2041                         output_binary(bit, sizeof(*bit) + bit->pdu_len);
2042                         continue;
2043                 }
2044
2045                 t = t_alloc();
2046                 memset(t, 0, sizeof(*t));
2047                 t->bit = bit;
2048                 t->read_sequence = read_sequence;
2049
2050                 t->next = trace_list;
2051                 trace_list = t;
2052
2053                 if (!pdi || pdi->dev != bit->device)
2054                         pdi = get_dev_info(bit->device);
2055
2056                 if (bit->time > pdi->last_read_time)
2057                         pdi->last_read_time = bit->time;
2058
2059                 events++;
2060         }
2061
2062         return events;
2063 }
2064
2065 /*
2066  * Managing input streams
2067  */
2068
2069 struct ms_stream {
2070         struct ms_stream *next;
2071         struct trace *first, *last;
2072         struct per_dev_info *pdi;
2073         unsigned int cpu;
2074 };
2075
2076 #define MS_HASH(d, c) ((MAJOR(d) & 0xff) ^ (MINOR(d) & 0xff) ^ (cpu & 0xff))
2077
2078 struct ms_stream *ms_head;
2079 struct ms_stream *ms_hash[256];
2080
2081 static void ms_sort(struct ms_stream *msp);
2082 static int ms_prime(struct ms_stream *msp);
2083
2084 static inline struct trace *ms_peek(struct ms_stream *msp)
2085 {
2086         return (msp == NULL) ? NULL : msp->first;
2087 }
2088
2089 static inline __u64 ms_peek_time(struct ms_stream *msp)
2090 {
2091         return ms_peek(msp)->bit->time;
2092 }
2093
2094 static inline void ms_resort(struct ms_stream *msp)
2095 {
2096         if (msp->next && ms_peek_time(msp) > ms_peek_time(msp->next)) {
2097                 ms_head = msp->next;
2098                 msp->next = NULL;
2099                 ms_sort(msp);
2100         }
2101 }
2102
2103 static inline void ms_deq(struct ms_stream *msp)
2104 {
2105         msp->first = msp->first->next;
2106         if (!msp->first) {
2107                 msp->last = NULL;
2108                 if (!ms_prime(msp)) {
2109                         ms_head = msp->next;
2110                         msp->next = NULL;
2111                         return;
2112                 }
2113         }
2114
2115         ms_resort(msp);
2116 }
2117
2118 static void ms_sort(struct ms_stream *msp)
2119 {
2120         __u64 msp_t = ms_peek_time(msp);
2121         struct ms_stream *this_msp = ms_head;
2122
2123         if (this_msp == NULL)
2124                 ms_head = msp;
2125         else if (msp_t < ms_peek_time(this_msp)) {
2126                 msp->next = this_msp;
2127                 ms_head = msp;
2128         }
2129         else {
2130                 while (this_msp->next && ms_peek_time(this_msp->next) < msp_t)
2131                         this_msp = this_msp->next;
2132
2133                 msp->next = this_msp->next;
2134                 this_msp->next = msp;
2135         }
2136 }
2137
2138 static int ms_prime(struct ms_stream *msp)
2139 {
2140         __u32 magic;
2141         unsigned int i;
2142         struct trace *t;
2143         struct per_dev_info *pdi = msp->pdi;
2144         struct per_cpu_info *pci = get_cpu_info(pdi, msp->cpu);
2145         struct blk_io_trace *bit = NULL;
2146         int ret, pdu_len, ndone = 0;
2147
2148         for (i = 0; !is_done() && pci->fd >= 0 && i < rb_batch; i++) {
2149                 bit = bit_alloc();
2150                 ret = read_data(pci->fd, bit, sizeof(*bit), 1, &pci->fdblock);
2151                 if (ret)
2152                         goto err;
2153
2154                 if (data_is_native == -1 && check_data_endianness(bit->magic))
2155                         goto err;
2156
2157                 magic = get_magic(bit);
2158                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2159                         fprintf(stderr, "Bad magic %x\n", magic);
2160                         goto err;
2161
2162                 }
2163
2164                 pdu_len = get_pdulen(bit);
2165                 if (pdu_len) {
2166                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2167                         ret = read_data(pci->fd, ptr + sizeof(*bit), pdu_len,
2168                                                              1, &pci->fdblock);
2169                         if (ret) {
2170                                 free(ptr);
2171                                 bit = NULL;
2172                                 goto err;
2173                         }
2174
2175                         bit = ptr;
2176                 }
2177
2178                 trace_to_cpu(bit);
2179                 if (verify_trace(bit))
2180                         goto err;
2181
2182                 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY)) {
2183                         handle_notify(bit);
2184                         output_binary(bit, sizeof(*bit) + bit->pdu_len);
2185                         bit_free(bit);
2186
2187                         i -= 1;
2188                         continue;
2189                 }
2190
2191                 if (bit->time > pdi->last_read_time)
2192                         pdi->last_read_time = bit->time;
2193
2194                 t = t_alloc();
2195                 memset(t, 0, sizeof(*t));
2196                 t->bit = bit;
2197
2198                 if (msp->first == NULL)
2199                         msp->first = msp->last = t;
2200                 else {
2201                         msp->last->next = t;
2202                         msp->last = t;
2203                 }
2204
2205                 ndone++;
2206         }
2207
2208         return ndone;
2209
2210 err:
2211         if (bit) bit_free(bit);
2212
2213         cpu_mark_offline(pdi, pci->cpu);
2214         close(pci->fd);
2215         pci->fd = -1;
2216
2217         return ndone;
2218 }
2219
2220 static struct ms_stream *ms_alloc(struct per_dev_info *pdi, int cpu)
2221 {
2222         struct ms_stream *msp = malloc(sizeof(*msp));
2223
2224         msp->next = NULL;
2225         msp->first = msp->last = NULL;
2226         msp->pdi = pdi;
2227         msp->cpu = cpu;
2228
2229         if (ms_prime(msp))
2230                 ms_sort(msp);
2231
2232         return msp;
2233 }
2234
2235 static int setup_file(struct per_dev_info *pdi, int cpu)
2236 {
2237         int len = 0;
2238         struct stat st;
2239         char *p, *dname;
2240         struct per_cpu_info *pci = get_cpu_info(pdi, cpu);
2241
2242         pci->cpu = cpu;
2243         pci->fdblock = -1;
2244
2245         p = strdup(pdi->name);
2246         dname = dirname(p);
2247         if (strcmp(dname, ".")) {
2248                 input_dir = dname;
2249                 p = strdup(pdi->name);
2250                 strcpy(pdi->name, basename(p));
2251         }
2252         free(p);
2253
2254         if (input_dir)
2255                 len = sprintf(pci->fname, "%s/", input_dir);
2256
2257         snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
2258                  "%s.blktrace.%d", pdi->name, pci->cpu);
2259         if (stat(pci->fname, &st) < 0)
2260                 return 0;
2261         if (!st.st_size)
2262                 return 1;
2263
2264         pci->fd = open(pci->fname, O_RDONLY);
2265         if (pci->fd < 0) {
2266                 perror(pci->fname);
2267                 return 0;
2268         }
2269
2270         printf("Input file %s added\n", pci->fname);
2271         cpu_mark_online(pdi, pci->cpu);
2272
2273         pdi->nfiles++;
2274         ms_alloc(pdi, pci->cpu);
2275
2276         return 1;
2277 }
2278
2279 static int handle(struct ms_stream *msp)
2280 {
2281         struct trace *t;
2282         struct per_dev_info *pdi;
2283         struct per_cpu_info *pci;
2284         struct blk_io_trace *bit;
2285
2286         t = ms_peek(msp);
2287         if (t->bit->time > stopwatch_end)
2288                 return 0;
2289
2290         bit = t->bit;
2291         pdi = msp->pdi;
2292         pci = get_cpu_info(pdi, msp->cpu);
2293         pci->nelems++;
2294
2295         bit->time -= genesis_time;
2296         pdi->last_reported_time = bit->time;
2297         if (bit->action & (act_mask << BLK_TC_SHIFT))
2298                 dump_trace(bit, pci, pdi);
2299
2300         ms_deq(msp);
2301
2302         if (text_output)
2303                 trace_rb_insert_last(pdi, t);
2304         else {
2305                 bit_free(t->bit);
2306                 t_free(t);
2307         }
2308
2309         return 1;
2310 }
2311
2312 /*
2313  * Check if we need to sanitize the name. We allow 'foo', or if foo.blktrace.X
2314  * is given, then strip back down to 'foo' to avoid missing files.
2315  */
2316 static int name_fixup(char *name)
2317 {
2318         char *b;
2319
2320         if (!name)
2321                 return 1;
2322
2323         b = strstr(name, ".blktrace.");
2324         if (b)
2325                 *b = '\0';
2326
2327         return 0;
2328 }
2329
2330 static int do_file(void)
2331 {
2332         int i, cpu, ret;
2333         struct per_dev_info *pdi;
2334
2335         /*
2336          * first prepare all files for reading
2337          */
2338         for (i = 0; i < ndevices; i++) {
2339                 pdi = &devices[i];
2340                 ret = name_fixup(pdi->name);
2341                 if (ret)
2342                         return ret;
2343
2344                 for (cpu = 0; setup_file(pdi, cpu); cpu++)
2345                         ;
2346         }
2347
2348         /*
2349          * Get the initial time stamp
2350          */
2351         if (ms_head)
2352                 genesis_time = ms_peek_time(ms_head);
2353
2354         /*
2355          * Keep processing traces while any are left
2356          */
2357         while (!is_done() && ms_head && handle(ms_head))
2358                 ;
2359
2360         return 0;
2361 }
2362
2363 static void do_pipe(int fd)
2364 {
2365         unsigned long long youngest;
2366         int events, fdblock;
2367
2368         last_allowed_time = -1ULL;
2369         fdblock = -1;
2370         while ((events = read_events(fd, 0, &fdblock)) > 0) {
2371                 read_sequence++;
2372
2373 #if 0
2374                 smallest_seq_read = -1U;
2375 #endif
2376
2377                 if (sort_entries(&youngest))
2378                         break;
2379
2380                 if (youngest > stopwatch_end)
2381                         break;
2382
2383                 show_entries_rb(0);
2384         }
2385
2386         if (rb_sort_entries)
2387                 show_entries_rb(1);
2388 }
2389
2390 static int do_fifo(void)
2391 {
2392         int fd;
2393
2394         if (!strcmp(pipename, "-"))
2395                 fd = dup(STDIN_FILENO);
2396         else
2397                 fd = open(pipename, O_RDONLY);
2398
2399         if (fd == -1) {
2400                 perror("dup stdin");
2401                 return -1;
2402         }
2403
2404         do_pipe(fd);
2405         close(fd);
2406         return 0;
2407 }
2408
2409 static void show_stats(void)
2410 {
2411         if (!ofp)
2412                 return;
2413         if (stats_printed)
2414                 return;
2415
2416         stats_printed = 1;
2417
2418         if (per_process_stats)
2419                 show_process_stats();
2420
2421         if (per_device_and_cpu_stats)
2422                 show_device_and_cpu_stats();
2423
2424         fflush(ofp);
2425 }
2426
2427 static void handle_sigint(__attribute__((__unused__)) int sig)
2428 {
2429         done = 1;
2430 }
2431
2432 /*
2433  * Extract start and duration times from a string, allowing
2434  * us to specify a time interval of interest within a trace.
2435  * Format: "duration" (start is zero) or "start:duration".
2436  */
2437 static int find_stopwatch_interval(char *string)
2438 {
2439         double value;
2440         char *sp;
2441
2442         value = strtod(string, &sp);
2443         if (sp == string) {
2444                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
2445                 return 1;
2446         }
2447         if (*sp == ':') {
2448                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
2449                 string = sp + 1;
2450                 value = strtod(string, &sp);
2451                 if (sp == string || *sp != '\0') {
2452                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
2453                                 string);
2454                         return 1;
2455                 }
2456         } else if (*sp != '\0') {
2457                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
2458                 return 1;
2459         }
2460         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
2461         if (stopwatch_end <= stopwatch_start) {
2462                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
2463                         stopwatch_start, stopwatch_end);
2464                 return 1;
2465         }
2466
2467         return 0;
2468 }
2469
2470 static int is_pipe(const char *str)
2471 {
2472         struct stat st;
2473
2474         if (!strcmp(str, "-"))
2475                 return 1;
2476         if (!stat(str, &st) && S_ISFIFO(st.st_mode))
2477                 return 1;
2478
2479         return 0;
2480 }
2481
2482 #define S_OPTS  "a:A:b:D:d:f:F:hi:o:Oqstw:vV"
2483 static char usage_str[] =    "\n\n" \
2484         "-i <file>           | --input=<file>\n" \
2485         "[ -a <action field> | --act-mask=<action field> ]\n" \
2486         "[ -A <action mask>  | --set-mask=<action mask> ]\n" \
2487         "[ -b <traces>       | --batch=<traces> ]\n" \
2488         "[ -d <file>         | --dump-binary=<file> ]\n" \
2489         "[ -D <dir>          | --input-directory=<dir> ]\n" \
2490         "[ -f <format>       | --format=<format> ]\n" \
2491         "[ -F <spec>         | --format-spec=<spec> ]\n" \
2492         "[ -h                | --hash-by-name ]\n" \
2493         "[ -o <file>         | --output=<file> ]\n" \
2494         "[ -O                | --no-text-output ]\n" \
2495         "[ -q                | --quiet ]\n" \
2496         "[ -s                | --per-program-stats ]\n" \
2497         "[ -t                | --track-ios ]\n" \
2498         "[ -w <time>         | --stopwatch=<time> ]\n" \
2499         "[ -v                | --verbose ]\n" \
2500         "[ -V                | --version ]\n\n" \
2501         "\t-b stdin read batching\n" \
2502         "\t-d Output file. If specified, binary data is written to file\n" \
2503         "\t-D Directory to prepend to input file names\n" \
2504         "\t-f Output format. Customize the output format. The format field\n" \
2505         "\t   identifies can be found in the documentation\n" \
2506         "\t-F Format specification. Can be found in the documentation\n" \
2507         "\t-h Hash processes by name, not pid\n" \
2508         "\t-i Input file containing trace data, or '-' for stdin\n" \
2509         "\t-o Output file. If not given, output is stdout\n" \
2510         "\t-O Do NOT output text data\n" \
2511         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
2512         "\t-s Show per-program io statistics\n" \
2513         "\t-t Track individual ios. Will tell you the time a request took\n" \
2514         "\t   to get queued, to get dispatched, and to get completed\n" \
2515         "\t-w Only parse data between the given time interval in seconds.\n" \
2516         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
2517         "\t-v More verbose for marginal errors\n" \
2518         "\t-V Print program version info\n\n";
2519
2520 static void usage(char *prog)
2521 {
2522         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
2523 }
2524
2525 int main(int argc, char *argv[])
2526 {
2527         int i, c, ret, mode;
2528         int act_mask_tmp = 0;
2529         char *ofp_buffer = NULL;
2530         char *bin_ofp_buffer = NULL;
2531
2532         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
2533                 switch (c) {
2534                 case 'a':
2535                         i = find_mask_map(optarg);
2536                         if (i < 0) {
2537                                 fprintf(stderr,"Invalid action mask %s\n",
2538                                         optarg);
2539                                 return 1;
2540                         }
2541                         act_mask_tmp |= i;
2542                         break;
2543
2544                 case 'A':
2545                         if ((sscanf(optarg, "%x", &i) != 1) ||
2546                                                         !valid_act_opt(i)) {
2547                                 fprintf(stderr,
2548                                         "Invalid set action mask %s/0x%x\n",
2549                                         optarg, i);
2550                                 return 1;
2551                         }
2552                         act_mask_tmp = i;
2553                         break;
2554                 case 'i':
2555                         if (is_pipe(optarg) && !pipeline) {
2556                                 pipeline = 1;
2557                                 pipename = strdup(optarg);
2558                         } else if (resize_devices(optarg) != 0)
2559                                 return 1;
2560                         break;
2561                 case 'D':
2562                         input_dir = optarg;
2563                         break;
2564                 case 'o':
2565                         output_name = optarg;
2566                         break;
2567                 case 'O':
2568                         text_output = 0;
2569                         break;
2570                 case 'b':
2571                         rb_batch = atoi(optarg);
2572                         if (rb_batch <= 0)
2573                                 rb_batch = RB_BATCH_DEFAULT;
2574                         break;
2575                 case 's':
2576                         per_process_stats = 1;
2577                         break;
2578                 case 't':
2579                         track_ios = 1;
2580                         break;
2581                 case 'q':
2582                         per_device_and_cpu_stats = 0;
2583                         break;
2584                 case 'w':
2585                         if (find_stopwatch_interval(optarg) != 0)
2586                                 return 1;
2587                         break;
2588                 case 'f':
2589                         set_all_format_specs(optarg);
2590                         break;
2591                 case 'F':
2592                         if (add_format_spec(optarg) != 0)
2593                                 return 1;
2594                         break;
2595                 case 'h':
2596                         ppi_hash_by_pid = 0;
2597                         break;
2598                 case 'v':
2599                         verbose++;
2600                         break;
2601                 case 'V':
2602                         printf("%s version %s\n", argv[0], blkparse_version);
2603                         return 0;
2604                 case 'd':
2605                         dump_binary = optarg;
2606                         break;
2607                 default:
2608                         usage(argv[0]);
2609                         return 1;
2610                 }
2611         }
2612
2613         while (optind < argc) {
2614                 if (is_pipe(argv[optind]) && !pipeline) {
2615                         pipeline = 1;
2616                         pipename = strdup(argv[optind]);
2617                 } else if (resize_devices(argv[optind]) != 0)
2618                         return 1;
2619                 optind++;
2620         }
2621
2622         if (!pipeline && !ndevices) {
2623                 usage(argv[0]);
2624                 return 1;
2625         }
2626
2627         if (act_mask_tmp != 0)
2628                 act_mask = act_mask_tmp;
2629
2630         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
2631
2632         signal(SIGINT, handle_sigint);
2633         signal(SIGHUP, handle_sigint);
2634         signal(SIGTERM, handle_sigint);
2635
2636         setlocale(LC_NUMERIC, "en_US");
2637
2638         if (text_output) {
2639                 if (!output_name) {
2640                         ofp = fdopen(STDOUT_FILENO, "w");
2641                         mode = _IOLBF;
2642                 } else {
2643                         char ofname[128];
2644
2645                         snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
2646                         ofp = fopen(ofname, "w");
2647                         mode = _IOFBF;
2648                 }
2649
2650                 if (!ofp) {
2651                         perror("fopen");
2652                         return 1;
2653                 }
2654
2655                 ofp_buffer = malloc(4096);
2656                 if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
2657                         perror("setvbuf");
2658                         return 1;
2659                 }
2660         }
2661
2662         if (dump_binary) {
2663                 dump_fp = fopen(dump_binary, "w");
2664                 if (!dump_fp) {
2665                         perror(dump_binary);
2666                         dump_binary = NULL;
2667                         return 1;
2668                 }
2669                 bin_ofp_buffer = malloc(128 * 1024);
2670                 if (setvbuf(dump_fp, bin_ofp_buffer, _IOFBF, 128 * 1024)) {
2671                         perror("setvbuf binary");
2672                         return 1;
2673                 }
2674         }
2675
2676         if (pipeline)
2677                 ret = do_fifo();
2678         else
2679                 ret = do_file();
2680
2681         if (!ret)
2682                 show_stats();
2683
2684         if (ofp_buffer) {
2685                 fflush(ofp);
2686                 free(ofp_buffer);
2687         }
2688         if (bin_ofp_buffer) {
2689                 fflush(dump_fp);
2690                 free(bin_ofp_buffer);
2691         }
2692         return ret;
2693 }