blkparse.c

   1 /*
   2  * block queue tracing parse application
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  *  This program is free software; you can redistribute it and/or modify
   8  *  it under the terms of the GNU General Public License as published by
   9  *  the Free Software Foundation; either version 2 of the License, or
  10  *  (at your option) any later version.
  11  *
  12  *  This program is distributed in the hope that it will be useful,
  13  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *  GNU General Public License for more details.
  16  *
  17  *  You should have received a copy of the GNU General Public License
  18  *  along with this program; if not, write to the Free Software
  19  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  20  *
  21  */
  22 #include <sys/types.h>
  23 #include <sys/stat.h>
  24 #include <unistd.h>
  25 #include <stdio.h>
  26 #include <fcntl.h>
  27 #include <stdlib.h>
  28 #include <string.h>
  29 #include <getopt.h>
  30 #include <errno.h>
  31 #include <signal.h>
  32 #include <locale.h>
  33 #include <libgen.h>
  34
  35 #include "blktrace.h"
  36 #include "rbtree.h"
  37 #include "jhash.h"
  38
  39 static char blkparse_version[] = "0.99.3";
  40
  41 struct skip_info {
  42         unsigned long start, end;
  43         struct skip_info *prev, *next;
  44 };
  45
  46 struct per_dev_info {
  47         dev_t dev;
  48         char *name;
  49
  50         int backwards;
  51         unsigned long long events;
  52         unsigned long long first_reported_time;
  53         unsigned long long last_reported_time;
  54         unsigned long long last_read_time;
  55         struct io_stats io_stats;
  56         unsigned long skips;
  57         unsigned long long seq_skips;
  58         unsigned int max_depth[2];
  59         unsigned int cur_depth[2];
  60
  61         struct rb_root rb_track;
  62
  63         int nfiles;
  64         int ncpus;
  65
  66         unsigned long *cpu_map;
  67         unsigned int cpu_map_max;
  68
  69         struct per_cpu_info *cpus;
  70 };
  71
  72 /*
  73  * some duplicated effort here, we can unify this hash and the ppi hash later
  74  */
  75 struct process_pid_map {
  76         pid_t pid;
  77         char comm[16];
  78         struct process_pid_map *hash_next, *list_next;
  79 };
  80
  81 #define PPM_HASH_SHIFT  (8)
  82 #define PPM_HASH_SIZE   (1 << PPM_HASH_SHIFT)
  83 #define PPM_HASH_MASK   (PPM_HASH_SIZE - 1)
  84 static struct process_pid_map *ppm_hash_table[PPM_HASH_SIZE];
  85
  86 struct per_process_info {
  87         struct process_pid_map *ppm;
  88         struct io_stats io_stats;
  89         struct per_process_info *hash_next, *list_next;
  90         int more_than_one;
  91
  92         /*
  93          * individual io stats
  94          */
  95         unsigned long long longest_allocation_wait[2];
  96         unsigned long long longest_dispatch_wait[2];
  97         unsigned long long longest_completion_wait[2];
  98 };
  99
 100 #define PPI_HASH_SHIFT  (8)
 101 #define PPI_HASH_SIZE   (1 << PPI_HASH_SHIFT)
 102 #define PPI_HASH_MASK   (PPI_HASH_SIZE - 1)
 103 static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
 104 static struct per_process_info *ppi_list;
 105 static int ppi_list_entries;
 106
 107 #define S_OPTS  "a:A:b:D:d:f:F:hi:o:Oqstw:vV"
 108 static struct option l_opts[] = {
 109         {
 110                 .name = "act-mask",
 111                 .has_arg = required_argument,
 112                 .flag = NULL,
 113                 .val = 'a'
 114         },
 115         {
 116                 .name = "set-mask",
 117                 .has_arg = required_argument,
 118                 .flag = NULL,
 119                 .val = 'A'
 120         },
 121         {
 122                 .name = "batch",
 123                 .has_arg = required_argument,
 124                 .flag = NULL,
 125                 .val = 'b'
 126         },
 127         {
 128                 .name = "input-directory",
 129                 .has_arg = required_argument,
 130                 .flag = NULL,
 131                 .val = 'D'
 132         },
 133         {
 134                 .name = "dump-binary",
 135                 .has_arg = required_argument,
 136                 .flag = NULL,
 137                 .val = 'd'
 138         },
 139         {
 140                 .name = "format",
 141                 .has_arg = required_argument,
 142                 .flag = NULL,
 143                 .val = 'f'
 144         },
 145         {
 146                 .name = "format-spec",
 147                 .has_arg = required_argument,
 148                 .flag = NULL,
 149                 .val = 'F'
 150         },
 151         {
 152                 .name = "hash-by-name",
 153                 .has_arg = no_argument,
 154                 .flag = NULL,
 155                 .val = 'h'
 156         },
 157         {
 158                 .name = "input",
 159                 .has_arg = required_argument,
 160                 .flag = NULL,
 161                 .val = 'i'
 162         },
 163         {
 164                 .name = "output",
 165                 .has_arg = required_argument,
 166                 .flag = NULL,
 167                 .val = 'o'
 168         },
 169         {
 170                 .name = "no-text-output",
 171                 .has_arg = no_argument,
 172                 .flag = NULL,
 173                 .val = 'O'
 174         },
 175         {
 176                 .name = "quiet",
 177                 .has_arg = no_argument,
 178                 .flag = NULL,
 179                 .val = 'q'
 180         },
 181         {
 182                 .name = "per-program-stats",
 183                 .has_arg = no_argument,
 184                 .flag = NULL,
 185                 .val = 's'
 186         },
 187         {
 188                 .name = "track-ios",
 189                 .has_arg = no_argument,
 190                 .flag = NULL,
 191                 .val = 't'
 192         },
 193         {
 194                 .name = "stopwatch",
 195                 .has_arg = required_argument,
 196                 .flag = NULL,
 197                 .val = 'w'
 198         },
 199         {
 200                 .name = "verbose",
 201                 .has_arg = no_argument,
 202                 .flag = NULL,
 203                 .val = 'v'
 204         },
 205         {
 206                 .name = "version",
 207                 .has_arg = no_argument,
 208                 .flag = NULL,
 209                 .val = 'V'
 210         },
 211         {
 212                 .name = NULL,
 213         }
 214 };
 215
 216 /*
 217  * for sorting the displayed output
 218  */
 219 struct trace {
 220         struct blk_io_trace *bit;
 221         struct rb_node rb_node;
 222         struct trace *next;
 223         unsigned long read_sequence;
 224 };
 225
 226 static struct rb_root rb_sort_root;
 227 static unsigned long rb_sort_entries;
 228
 229 static struct trace *trace_list;
 230
 231 /*
 232  * allocation cache
 233  */
 234 static struct blk_io_trace *bit_alloc_list;
 235 static struct trace *t_alloc_list;
 236
 237 /*
 238  * for tracking individual ios
 239  */
 240 struct io_track {
 241         struct rb_node rb_node;
 242
 243         struct process_pid_map *ppm;
 244         __u64 sector;
 245         unsigned long long allocation_time;
 246         unsigned long long queue_time;
 247         unsigned long long dispatch_time;
 248         unsigned long long completion_time;
 249 };
 250
 251 static int ndevices;
 252 static struct per_dev_info *devices;
 253 static char *get_dev_name(struct per_dev_info *, char *, int);
 254 static int trace_rb_insert_last(struct per_dev_info *, struct trace *);
 255
 256 FILE *ofp = NULL;
 257 static char *output_name;
 258 static char *input_dir;
 259
 260 static unsigned long long genesis_time;
 261 static unsigned long long last_allowed_time;
 262 static unsigned long long stopwatch_start;      /* start from zero by default */
 263 static unsigned long long stopwatch_end = -1ULL;        /* "infinity" */
 264 static unsigned long read_sequence;
 265
 266 static int per_process_stats;
 267 static int per_device_and_cpu_stats = 1;
 268 static int track_ios;
 269 static int ppi_hash_by_pid = 1;
 270 static int verbose;
 271 static unsigned int act_mask = -1U;
 272 static int stats_printed;
 273 int data_is_native = -1;
 274
 275 static FILE *dump_fp;
 276 static char *dump_binary;
 277
 278 static unsigned int t_alloc_cache;
 279 static unsigned int bit_alloc_cache;
 280
 281 #define RB_BATCH_DEFAULT        (512)
 282 static unsigned int rb_batch = RB_BATCH_DEFAULT;
 283
 284 static int pipeline;
 285 static char *pipename;
 286
 287 static int text_output = 1;
 288
 289 #define is_done()       (*(volatile int *)(&done))
 290 static volatile int done;
 291
 292 struct timespec         abs_start_time;
 293 static unsigned long long start_timestamp;
 294
 295 #define JHASH_RANDOM    (0x3af5f2ee)
 296
 297 #define CPUS_PER_LONG   (8 * sizeof(unsigned long))
 298 #define CPU_IDX(cpu)    ((cpu) / CPUS_PER_LONG)
 299 #define CPU_BIT(cpu)    ((cpu) & (CPUS_PER_LONG - 1))
 300
 301 static void output_binary(void *buf, int len)
 302 {
 303         if (dump_binary) {
 304                 size_t n = fwrite(buf, len, 1, dump_fp);
 305                 if (n != 1) {
 306                         perror(dump_binary);
 307                         fclose(dump_fp);
 308                         dump_binary = NULL;
 309                 }
 310         }
 311 }
 312
 313 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
 314 {
 315         struct per_cpu_info *cpus = pdi->cpus;
 316         int ncpus = pdi->ncpus;
 317         int new_count = cpu + 1;
 318         int new_space, size;
 319         char *new_start;
 320
 321         size = new_count * sizeof(struct per_cpu_info);
 322         cpus = realloc(cpus, size);
 323         if (!cpus) {
 324                 char name[20];
 325                 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
 326                         get_dev_name(pdi, name, sizeof(name)), size);
 327                 exit(1);
 328         }
 329
 330         new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
 331         new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
 332         memset(new_start, 0, new_space);
 333
 334         pdi->ncpus = new_count;
 335         pdi->cpus = cpus;
 336
 337         for (new_count = 0; new_count < pdi->ncpus; new_count++) {
 338                 struct per_cpu_info *pci = &pdi->cpus[new_count];
 339
 340                 if (!pci->fd) {
 341                         pci->fd = -1;
 342                         memset(&pci->rb_last, 0, sizeof(pci->rb_last));
 343                         pci->rb_last_entries = 0;
 344                         pci->last_sequence = -1;
 345                 }
 346         }
 347 }
 348
 349 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
 350 {
 351         struct per_cpu_info *pci;
 352
 353         if (cpu >= pdi->ncpus)
 354                 resize_cpu_info(pdi, cpu);
 355
 356         pci = &pdi->cpus[cpu];
 357         pci->cpu = cpu;
 358         return pci;
 359 }
 360
 361
 362 static int resize_devices(char *name)
 363 {
 364         int size = (ndevices + 1) * sizeof(struct per_dev_info);
 365
 366         devices = realloc(devices, size);
 367         if (!devices) {
 368                 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
 369                 return 1;
 370         }
 371         memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
 372         devices[ndevices].name = name;
 373         ndevices++;
 374         return 0;
 375 }
 376
 377 static struct per_dev_info *get_dev_info(dev_t dev)
 378 {
 379         struct per_dev_info *pdi;
 380         int i;
 381
 382         for (i = 0; i < ndevices; i++) {
 383                 if (!devices[i].dev)
 384                         devices[i].dev = dev;
 385                 if (devices[i].dev == dev)
 386                         return &devices[i];
 387         }
 388
 389         if (resize_devices(NULL))
 390                 return NULL;
 391
 392         pdi = &devices[ndevices - 1];
 393         pdi->dev = dev;
 394         pdi->first_reported_time = 0;
 395         pdi->last_read_time = 0;
 396
 397         return pdi;
 398 }
 399
 400 static void insert_skip(struct per_cpu_info *pci, unsigned long start,
 401                         unsigned long end)
 402 {
 403         struct skip_info *sip;
 404
 405         for (sip = pci->skips_tail; sip != NULL; sip = sip->prev) {
 406                 if (end == (sip->start - 1)) {
 407                         sip->start = start;
 408                         return;
 409                 } else if (start == (sip->end + 1)) {
 410                         sip->end = end;
 411                         return;
 412                 }
 413         }
 414
 415         sip = malloc(sizeof(struct skip_info));
 416         sip->start = start;
 417         sip->end = end;
 418         sip->prev = sip->next = NULL;
 419         if (pci->skips_tail == NULL)
 420                 pci->skips_head = pci->skips_tail = sip;
 421         else {
 422                 sip->prev = pci->skips_tail;
 423                 pci->skips_tail->next = sip;
 424                 pci->skips_tail = sip;
 425         }
 426 }
 427
 428 static void remove_sip(struct per_cpu_info *pci, struct skip_info *sip)
 429 {
 430         if (sip->prev == NULL) {
 431                 if (sip->next == NULL)
 432                         pci->skips_head = pci->skips_tail = NULL;
 433                 else {
 434                         pci->skips_head = sip->next;
 435                         sip->next->prev = NULL;
 436                 }
 437         } else if (sip->next == NULL) {
 438                 pci->skips_tail = sip->prev;
 439                 sip->prev->next = NULL;
 440         } else {
 441                 sip->prev->next = sip->next;
 442                 sip->next->prev = sip->prev;
 443         }
 444
 445         sip->prev = sip->next = NULL;
 446         free(sip);
 447 }
 448
 449 #define IN_SKIP(sip,seq) (((sip)->start <= (seq)) && ((seq) <= sip->end))
 450 static int check_current_skips(struct per_cpu_info *pci, unsigned long seq)
 451 {
 452         struct skip_info *sip;
 453
 454         for (sip = pci->skips_tail; sip != NULL; sip = sip->prev) {
 455                 if (IN_SKIP(sip, seq)) {
 456                         if (sip->start == seq) {
 457                                 if (sip->end == seq)
 458                                         remove_sip(pci, sip);
 459                                 else
 460                                         sip->start += 1;
 461                         } else if (sip->end == seq)
 462                                 sip->end -= 1;
 463                         else {
 464                                 sip->end = seq - 1;
 465                                 insert_skip(pci, seq + 1, sip->end);
 466                         }
 467                         return 1;
 468                 }
 469         }
 470
 471         return 0;
 472 }
 473
 474 static void collect_pdi_skips(struct per_dev_info *pdi)
 475 {
 476         struct skip_info *sip;
 477         int cpu;
 478
 479         pdi->skips = 0;
 480         pdi->seq_skips = 0;
 481
 482         for (cpu = 0; cpu < pdi->ncpus; cpu++) {
 483                 struct per_cpu_info *pci = &pdi->cpus[cpu];
 484
 485                 for (sip = pci->skips_head; sip != NULL; sip = sip->next) {
 486                         pdi->skips++;
 487                         pdi->seq_skips += (sip->end - sip->start + 1);
 488                         if (verbose)
 489                                 fprintf(stderr,"(%d,%d): skipping %lu -> %lu\n",
 490                                         MAJOR(pdi->dev), MINOR(pdi->dev),
 491                                         sip->start, sip->end);
 492                 }
 493         }
 494 }
 495
 496 static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu)
 497 {
 498         if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) {
 499                 int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1);
 500                 unsigned long *map = malloc(new_max / sizeof(long));
 501
 502                 memset(map, 0, new_max / sizeof(long));
 503
 504                 if (pdi->cpu_map) {
 505                         memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long));
 506                         free(pdi->cpu_map);
 507                 }
 508
 509                 pdi->cpu_map = map;
 510                 pdi->cpu_map_max = new_max;
 511         }
 512
 513         pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
 514 }
 515
 516 static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu)
 517 {
 518         pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu));
 519 }
 520
 521 static inline int cpu_is_online(struct per_dev_info *pdi, int cpu)
 522 {
 523         return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0;
 524 }
 525
 526 static inline int ppm_hash_pid(pid_t pid)
 527 {
 528         return jhash_1word(pid, JHASH_RANDOM) & PPM_HASH_MASK;
 529 }
 530
 531 static struct process_pid_map *find_ppm(pid_t pid)
 532 {
 533         const int hash_idx = ppm_hash_pid(pid);
 534         struct process_pid_map *ppm;
 535
 536         ppm = ppm_hash_table[hash_idx];
 537         while (ppm) {
 538                 if (ppm->pid == pid)
 539                         return ppm;
 540
 541                 ppm = ppm->hash_next;
 542         }
 543
 544         return NULL;
 545 }
 546
 547 static struct process_pid_map *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         return ppm;
 563 }
 564
 565 static void handle_notify(struct blk_io_trace *bit)
 566 {
 567         void    *payload = (caddr_t) bit + sizeof(*bit);
 568         __u32   two32[2];
 569
 570         switch (bit->action) {
 571         case BLK_TN_PROCESS:
 572                 add_ppm_hash(bit->pid, payload);
 573                 break;
 574
 575         case BLK_TN_TIMESTAMP:
 576                 if (bit->pdu_len != sizeof(two32))
 577                         return;
 578                 memcpy(two32, payload, sizeof(two32));
 579                 if (!data_is_native) {
 580                         two32[0] = be32_to_cpu(two32[0]);
 581                         two32[1] = be32_to_cpu(two32[1]);
 582                 }
 583                 start_timestamp = bit->time;
 584                 abs_start_time.tv_sec  = two32[0];
 585                 abs_start_time.tv_nsec = two32[1];
 586                 if (abs_start_time.tv_nsec < 0) {
 587                         abs_start_time.tv_sec--;
 588                         abs_start_time.tv_nsec += 1000000000;
 589                 }
 590
 591                 break;
 592
 593         default:
 594                 /* Ignore unknown notify events */
 595                 ;
 596         }
 597 }
 598
 599 char *find_process_name(pid_t pid)
 600 {
 601         struct process_pid_map *ppm = find_ppm(pid);
 602
 603         if (ppm)
 604                 return ppm->comm;
 605
 606         return NULL;
 607 }
 608
 609 static inline int ppi_hash_pid(pid_t pid)
 610 {
 611         return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
 612 }
 613
 614 static inline int ppi_hash_name(const char *name)
 615 {
 616         return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
 617 }
 618
 619 static inline int ppi_hash(struct per_process_info *ppi)
 620 {
 621         struct process_pid_map *ppm = ppi->ppm;
 622
 623         if (ppi_hash_by_pid)
 624                 return ppi_hash_pid(ppm->pid);
 625
 626         return ppi_hash_name(ppm->comm);
 627 }
 628
 629 static inline void add_ppi_to_hash(struct per_process_info *ppi)
 630 {
 631         const int hash_idx = ppi_hash(ppi);
 632
 633         ppi->hash_next = ppi_hash_table[hash_idx];
 634         ppi_hash_table[hash_idx] = ppi;
 635 }
 636
 637 static inline void add_ppi_to_list(struct per_process_info *ppi)
 638 {
 639         ppi->list_next = ppi_list;
 640         ppi_list = ppi;
 641         ppi_list_entries++;
 642 }
 643
 644 static struct per_process_info *find_ppi_by_name(char *name)
 645 {
 646         const int hash_idx = ppi_hash_name(name);
 647         struct per_process_info *ppi;
 648
 649         ppi = ppi_hash_table[hash_idx];
 650         while (ppi) {
 651                 struct process_pid_map *ppm = ppi->ppm;
 652
 653                 if (!strcmp(ppm->comm, name))
 654                         return ppi;
 655
 656                 ppi = ppi->hash_next;
 657         }
 658
 659         return NULL;
 660 }
 661
 662 static struct per_process_info *find_ppi_by_pid(pid_t pid)
 663 {
 664         const int hash_idx = ppi_hash_pid(pid);
 665         struct per_process_info *ppi;
 666
 667         ppi = ppi_hash_table[hash_idx];
 668         while (ppi) {
 669                 struct process_pid_map *ppm = ppi->ppm;
 670
 671                 if (ppm->pid == pid)
 672                         return ppi;
 673
 674                 ppi = ppi->hash_next;
 675         }
 676
 677         return NULL;
 678 }
 679
 680 static struct per_process_info *find_ppi(pid_t pid)
 681 {
 682         struct per_process_info *ppi;
 683         char *name;
 684
 685         if (ppi_hash_by_pid)
 686                 return find_ppi_by_pid(pid);
 687
 688         name = find_process_name(pid);
 689         if (!name)
 690                 return NULL;
 691
 692         ppi = find_ppi_by_name(name);
 693         if (ppi && ppi->ppm->pid != pid)
 694                 ppi->more_than_one = 1;
 695
 696         return ppi;
 697 }
 698
 699 /*
 700  * struct trace and blktrace allocation cache, we do potentially
 701  * millions of mallocs for these structures while only using at most
 702  * a few thousand at the time
 703  */
 704 static inline void t_free(struct trace *t)
 705 {
 706         if (t_alloc_cache < 1024) {
 707                 t->next = t_alloc_list;
 708                 t_alloc_list = t;
 709                 t_alloc_cache++;
 710         } else
 711                 free(t);
 712 }
 713
 714 static inline struct trace *t_alloc(void)
 715 {
 716         struct trace *t = t_alloc_list;
 717
 718         if (t) {
 719                 t_alloc_list = t->next;
 720                 t_alloc_cache--;
 721                 return t;
 722         }
 723
 724         return malloc(sizeof(*t));
 725 }
 726
 727 static inline void bit_free(struct blk_io_trace *bit)
 728 {
 729         if (bit_alloc_cache < 1024 && !bit->pdu_len) {
 730                 /*
 731                  * abuse a 64-bit field for a next pointer for the free item
 732                  */
 733                 bit->time = (__u64) (unsigned long) bit_alloc_list;
 734                 bit_alloc_list = (struct blk_io_trace *) bit;
 735                 bit_alloc_cache++;
 736         } else
 737                 free(bit);
 738 }
 739
 740 static inline struct blk_io_trace *bit_alloc(void)
 741 {
 742         struct blk_io_trace *bit = bit_alloc_list;
 743
 744         if (bit) {
 745                 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
 746                                  bit->time;
 747                 bit_alloc_cache--;
 748                 return bit;
 749         }
 750
 751         return malloc(sizeof(*bit));
 752 }
 753
 754 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
 755 {
 756         struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
 757
 758         rb_erase(&t->rb_node, &pci->rb_last);
 759         pci->rb_last_entries--;
 760
 761         bit_free(t->bit);
 762         t_free(t);
 763 }
 764
 765 static void put_trace(struct per_dev_info *pdi, struct trace *t)
 766 {
 767         rb_erase(&t->rb_node, &rb_sort_root);
 768         rb_sort_entries--;
 769
 770         trace_rb_insert_last(pdi, t);
 771 }
 772
 773 static inline int trace_rb_insert(struct trace *t, struct rb_root *root)
 774 {
 775         struct rb_node **p = &root->rb_node;
 776         struct rb_node *parent = NULL;
 777         struct trace *__t;
 778
 779         while (*p) {
 780                 parent = *p;
 781
 782                 __t = rb_entry(parent, struct trace, rb_node);
 783
 784                 if (t->bit->time < __t->bit->time)
 785                         p = &(*p)->rb_left;
 786                 else if (t->bit->time > __t->bit->time)
 787                         p = &(*p)->rb_right;
 788                 else if (t->bit->device < __t->bit->device)
 789                         p = &(*p)->rb_left;
 790                 else if (t->bit->device > __t->bit->device)
 791                         p = &(*p)->rb_right;
 792                 else if (t->bit->sequence < __t->bit->sequence)
 793                         p = &(*p)->rb_left;
 794                 else    /* >= sequence */
 795                         p = &(*p)->rb_right;
 796         }
 797
 798         rb_link_node(&t->rb_node, parent, p);
 799         rb_insert_color(&t->rb_node, root);
 800         return 0;
 801 }
 802
 803 static inline int trace_rb_insert_sort(struct trace *t)
 804 {
 805         if (!trace_rb_insert(t, &rb_sort_root)) {
 806                 rb_sort_entries++;
 807                 return 0;
 808         }
 809
 810         return 1;
 811 }
 812
 813 static int trace_rb_insert_last(struct per_dev_info *pdi, struct trace *t)
 814 {
 815         struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
 816
 817         if (trace_rb_insert(t, &pci->rb_last))
 818                 return 1;
 819
 820         pci->rb_last_entries++;
 821
 822         if (pci->rb_last_entries > rb_batch * pdi->nfiles) {
 823                 struct rb_node *n = rb_first(&pci->rb_last);
 824
 825                 t = rb_entry(n, struct trace, rb_node);
 826                 __put_trace_last(pdi, t);
 827         }
 828
 829         return 0;
 830 }
 831
 832 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
 833                                    struct rb_root *root, int order)
 834 {
 835         struct rb_node *n = root->rb_node;
 836         struct rb_node *prev = NULL;
 837         struct trace *__t;
 838
 839         while (n) {
 840                 __t = rb_entry(n, struct trace, rb_node);
 841                 prev = n;
 842
 843                 if (device < __t->bit->device)
 844                         n = n->rb_left;
 845                 else if (device > __t->bit->device)
 846                         n = n->rb_right;
 847                 else if (sequence < __t->bit->sequence)
 848                         n = n->rb_left;
 849                 else if (sequence > __t->bit->sequence)
 850                         n = n->rb_right;
 851                 else
 852                         return __t;
 853         }
 854
 855         /*
 856          * hack - the list may not be sequence ordered because some
 857          * events don't have sequence and time matched. so we end up
 858          * being a little off in the rb lookup here, because we don't
 859          * know the time we are looking for. compensate by browsing
 860          * a little ahead from the last entry to find the match
 861          */
 862         if (order && prev) {
 863                 int max = 5;
 864
 865                 while (((n = rb_next(prev)) != NULL) && max--) {
 866                         __t = rb_entry(n, struct trace, rb_node);
 867
 868                         if (__t->bit->device == device &&
 869                             __t->bit->sequence == sequence)
 870                                 return __t;
 871
 872                         prev = n;
 873                 }
 874         }
 875
 876         return NULL;
 877 }
 878
 879 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
 880                                                struct per_cpu_info *pci,
 881                                                unsigned long seq)
 882 {
 883         return trace_rb_find(pdi->dev, seq, &pci->rb_last, 0);
 884 }
 885
 886 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
 887 {
 888         struct rb_node **p = &pdi->rb_track.rb_node;
 889         struct rb_node *parent = NULL;
 890         struct io_track *__iot;
 891
 892         while (*p) {
 893                 parent = *p;
 894                 __iot = rb_entry(parent, struct io_track, rb_node);
 895
 896                 if (iot->sector < __iot->sector)
 897                         p = &(*p)->rb_left;
 898                 else if (iot->sector > __iot->sector)
 899                         p = &(*p)->rb_right;
 900                 else {
 901                         fprintf(stderr,
 902                                 "sector alias (%Lu) on device %d,%d!\n",
 903                                 (unsigned long long) iot->sector,
 904                                 MAJOR(pdi->dev), MINOR(pdi->dev));
 905                         return 1;
 906                 }
 907         }
 908
 909         rb_link_node(&iot->rb_node, parent, p);
 910         rb_insert_color(&iot->rb_node, &pdi->rb_track);
 911         return 0;
 912 }
 913
 914 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
 915 {
 916         struct rb_node *n = pdi->rb_track.rb_node;
 917         struct io_track *__iot;
 918
 919         while (n) {
 920                 __iot = rb_entry(n, struct io_track, rb_node);
 921
 922                 if (sector < __iot->sector)
 923                         n = n->rb_left;
 924                 else if (sector > __iot->sector)
 925                         n = n->rb_right;
 926                 else
 927                         return __iot;
 928         }
 929
 930         return NULL;
 931 }
 932
 933 static struct io_track *find_track(struct per_dev_info *pdi, pid_t pid,
 934                                    __u64 sector)
 935 {
 936         struct io_track *iot;
 937
 938         iot = __find_track(pdi, sector);
 939         if (!iot) {
 940                 iot = malloc(sizeof(*iot));
 941                 iot->ppm = find_ppm(pid);
 942                 if (!iot->ppm)
 943                         iot->ppm = add_ppm_hash(pid, "unknown");
 944                 iot->sector = sector;
 945                 track_rb_insert(pdi, iot);
 946         }
 947
 948         return iot;
 949 }
 950
 951 static void log_track_frontmerge(struct per_dev_info *pdi,
 952                                  struct blk_io_trace *t)
 953 {
 954         struct io_track *iot;
 955
 956         if (!track_ios)
 957                 return;
 958
 959         iot = __find_track(pdi, t->sector + t_sec(t));
 960         if (!iot) {
 961                 if (verbose)
 962                         fprintf(stderr, "merge not found for (%d,%d): %llu\n",
 963                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 964                                 (unsigned long long) t->sector + t_sec(t));
 965                 return;
 966         }
 967
 968         rb_erase(&iot->rb_node, &pdi->rb_track);
 969         iot->sector -= t_sec(t);
 970         track_rb_insert(pdi, iot);
 971 }
 972
 973 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
 974 {
 975         struct io_track *iot;
 976
 977         if (!track_ios)
 978                 return;
 979
 980         iot = find_track(pdi, t->pid, t->sector);
 981         iot->allocation_time = t->time;
 982 }
 983
 984 static inline int is_remapper(struct per_dev_info *pdi)
 985 {
 986         int major = MAJOR(pdi->dev);
 987
 988         return (major == 253 || major == 9);
 989 }
 990
 991 /*
 992  * for md/dm setups, the interesting cycle is Q -> C. So track queueing
 993  * time here, as dispatch time
 994  */
 995 static void log_track_queue(struct per_dev_info *pdi, struct blk_io_trace *t)
 996 {
 997         struct io_track *iot;
 998
 999         if (!track_ios)
1000                 return;
1001         if (!is_remapper(pdi))
1002                 return;
1003
1004         iot = find_track(pdi, t->pid, t->sector);
1005         iot->dispatch_time = t->time;
1006 }
1007
1008 /*
1009  * return time between rq allocation and insertion
1010  */
1011 static unsigned long long log_track_insert(struct per_dev_info *pdi,
1012                                            struct blk_io_trace *t)
1013 {
1014         unsigned long long elapsed;
1015         struct io_track *iot;
1016
1017         if (!track_ios)
1018                 return -1;
1019
1020         iot = find_track(pdi, t->pid, t->sector);
1021         iot->queue_time = t->time;
1022
1023         if (!iot->allocation_time)
1024                 return -1;
1025
1026         elapsed = iot->queue_time - iot->allocation_time;
1027
1028         if (per_process_stats) {
1029                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1030                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1031
1032                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
1033                         ppi->longest_allocation_wait[w] = elapsed;
1034         }
1035
1036         return elapsed;
1037 }
1038
1039 /*
1040  * return time between queue and issue
1041  */
1042 static unsigned long long log_track_issue(struct per_dev_info *pdi,
1043                                           struct blk_io_trace *t)
1044 {
1045         unsigned long long elapsed;
1046         struct io_track *iot;
1047
1048         if (!track_ios)
1049                 return -1;
1050         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
1051                 return -1;
1052
1053         iot = __find_track(pdi, t->sector);
1054         if (!iot) {
1055                 if (verbose)
1056                         fprintf(stderr, "issue not found for (%d,%d): %llu\n",
1057                                 MAJOR(pdi->dev), MINOR(pdi->dev),
1058                                 (unsigned long long) t->sector);
1059                 return -1;
1060         }
1061
1062         iot->dispatch_time = t->time;
1063         elapsed = iot->dispatch_time - iot->queue_time;
1064
1065         if (per_process_stats) {
1066                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1067                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1068
1069                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
1070                         ppi->longest_dispatch_wait[w] = elapsed;
1071         }
1072
1073         return elapsed;
1074 }
1075
1076 /*
1077  * return time between dispatch and complete
1078  */
1079 static unsigned long long log_track_complete(struct per_dev_info *pdi,
1080                                              struct blk_io_trace *t)
1081 {
1082         unsigned long long elapsed;
1083         struct io_track *iot;
1084
1085         if (!track_ios)
1086                 return -1;
1087
1088         iot = __find_track(pdi, t->sector);
1089         if (!iot) {
1090                 if (verbose)
1091                         fprintf(stderr,"complete not found for (%d,%d): %llu\n",
1092                                 MAJOR(pdi->dev), MINOR(pdi->dev),
1093                                 (unsigned long long) t->sector);
1094                 return -1;
1095         }
1096
1097         iot->completion_time = t->time;
1098         elapsed = iot->completion_time - iot->dispatch_time;
1099
1100         if (per_process_stats) {
1101                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1102                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1103
1104                 if (ppi && elapsed > ppi->longest_completion_wait[w])
1105                         ppi->longest_completion_wait[w] = elapsed;
1106         }
1107
1108         /*
1109          * kill the trace, we don't need it after completion
1110          */
1111         rb_erase(&iot->rb_node, &pdi->rb_track);
1112         free(iot);
1113
1114         return elapsed;
1115 }
1116
1117
1118 static struct io_stats *find_process_io_stats(pid_t pid)
1119 {
1120         struct per_process_info *ppi = find_ppi(pid);
1121
1122         if (!ppi) {
1123                 ppi = malloc(sizeof(*ppi));
1124                 memset(ppi, 0, sizeof(*ppi));
1125                 ppi->ppm = find_ppm(pid);
1126                 if (!ppi->ppm)
1127                         ppi->ppm = add_ppm_hash(pid, "unknown");
1128                 add_ppi_to_hash(ppi);
1129                 add_ppi_to_list(ppi);
1130         }
1131
1132         return &ppi->io_stats;
1133 }
1134
1135 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
1136 {
1137         if (pdi->name)
1138                 snprintf(buffer, size, "%s", pdi->name);
1139         else
1140                 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
1141         return buffer;
1142 }
1143
1144 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
1145 {
1146         unsigned long long this = bit->time;
1147         unsigned long long last = pdi->last_reported_time;
1148
1149         pdi->backwards = (this < last) ? 'B' : ' ';
1150         pdi->last_reported_time = this;
1151 }
1152
1153 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
1154                                int rw)
1155 {
1156         if (rw) {
1157                 ios->mwrites++;
1158                 ios->mwrite_kb += t_kb(t);
1159         } else {
1160                 ios->mreads++;
1161                 ios->mread_kb += t_kb(t);
1162         }
1163 }
1164
1165 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
1166                              int rw)
1167 {
1168         __account_m(&pci->io_stats, t, rw);
1169
1170         if (per_process_stats) {
1171                 struct io_stats *ios = find_process_io_stats(t->pid);
1172
1173                 __account_m(ios, t, rw);
1174         }
1175 }
1176
1177 static inline void __account_pc_queue(struct io_stats *ios,
1178                                       struct blk_io_trace *t, int rw)
1179 {
1180         if (rw) {
1181                 ios->qwrites_pc++;
1182                 ios->qwrite_kb_pc += t_kb(t);
1183         } else {
1184                 ios->qreads_pc++;
1185                 ios->qread_kb += t_kb(t);
1186         }
1187 }
1188
1189 static inline void account_pc_queue(struct blk_io_trace *t,
1190                                     struct per_cpu_info *pci, int rw)
1191 {
1192         __account_pc_queue(&pci->io_stats, t, rw);
1193
1194         if (per_process_stats) {
1195                 struct io_stats *ios = find_process_io_stats(t->pid);
1196
1197                 __account_pc_queue(ios, t, rw);
1198         }
1199 }
1200
1201 static inline void __account_pc_issue(struct io_stats *ios, int rw,
1202                                       unsigned int bytes)
1203 {
1204         if (rw) {
1205                 ios->iwrites_pc++;
1206                 ios->iwrite_kb_pc += bytes >> 10;
1207         } else {
1208                 ios->ireads_pc++;
1209                 ios->iread_kb_pc += bytes >> 10;
1210         }
1211 }
1212
1213 static inline void account_pc_issue(struct blk_io_trace *t,
1214                                     struct per_cpu_info *pci, int rw)
1215 {
1216         __account_pc_issue(&pci->io_stats, rw, t->bytes);
1217
1218         if (per_process_stats) {
1219                 struct io_stats *ios = find_process_io_stats(t->pid);
1220
1221                 __account_pc_issue(ios, rw, t->bytes);
1222         }
1223 }
1224
1225 static inline void __account_pc_requeue(struct io_stats *ios,
1226                                         struct blk_io_trace *t, int rw)
1227 {
1228         if (rw) {
1229                 ios->wrqueue_pc++;
1230                 ios->iwrite_kb_pc -= t_kb(t);
1231         } else {
1232                 ios->rrqueue_pc++;
1233                 ios->iread_kb_pc -= t_kb(t);
1234         }
1235 }
1236
1237 static inline void account_pc_requeue(struct blk_io_trace *t,
1238                                       struct per_cpu_info *pci, int rw)
1239 {
1240         __account_pc_requeue(&pci->io_stats, t, rw);
1241
1242         if (per_process_stats) {
1243                 struct io_stats *ios = find_process_io_stats(t->pid);
1244
1245                 __account_pc_requeue(ios, t, rw);
1246         }
1247 }
1248
1249 static inline void __account_pc_c(struct io_stats *ios, int rw)
1250 {
1251         if (rw)
1252                 ios->cwrites_pc++;
1253         else
1254                 ios->creads_pc++;
1255 }
1256
1257 static inline void account_pc_c(struct blk_io_trace *t,
1258                                 struct per_cpu_info *pci, int rw)
1259 {
1260         __account_pc_c(&pci->io_stats, rw);
1261
1262         if (per_process_stats) {
1263                 struct io_stats *ios = find_process_io_stats(t->pid);
1264
1265                 __account_pc_c(ios, rw);
1266         }
1267 }
1268
1269 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
1270                                    int rw)
1271 {
1272         if (rw) {
1273                 ios->qwrites++;
1274                 ios->qwrite_kb += t_kb(t);
1275         } else {
1276                 ios->qreads++;
1277                 ios->qread_kb += t_kb(t);
1278         }
1279 }
1280
1281 static inline void account_queue(struct blk_io_trace *t,
1282                                  struct per_cpu_info *pci, int rw)
1283 {
1284         __account_queue(&pci->io_stats, t, rw);
1285
1286         if (per_process_stats) {
1287                 struct io_stats *ios = find_process_io_stats(t->pid);
1288
1289                 __account_queue(ios, t, rw);
1290         }
1291 }
1292
1293 static inline void __account_c(struct io_stats *ios, int rw, int bytes)
1294 {
1295         if (rw) {
1296                 ios->cwrites++;
1297                 ios->cwrite_kb += bytes >> 10;
1298         } else {
1299                 ios->creads++;
1300                 ios->cread_kb += bytes >> 10;
1301         }
1302 }
1303
1304 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
1305                              int rw, int bytes)
1306 {
1307         __account_c(&pci->io_stats, rw, bytes);
1308
1309         if (per_process_stats) {
1310                 struct io_stats *ios = find_process_io_stats(t->pid);
1311
1312                 __account_c(ios, rw, bytes);
1313         }
1314 }
1315
1316 static inline void __account_issue(struct io_stats *ios, int rw,
1317                                    unsigned int bytes)
1318 {
1319         if (rw) {
1320                 ios->iwrites++;
1321                 ios->iwrite_kb += bytes >> 10;
1322         } else {
1323                 ios->ireads++;
1324                 ios->iread_kb += bytes >> 10;
1325         }
1326 }
1327
1328 static inline void account_issue(struct blk_io_trace *t,
1329                                  struct per_cpu_info *pci, int rw)
1330 {
1331         __account_issue(&pci->io_stats, rw, t->bytes);
1332
1333         if (per_process_stats) {
1334                 struct io_stats *ios = find_process_io_stats(t->pid);
1335
1336                 __account_issue(ios, rw, t->bytes);
1337         }
1338 }
1339
1340 static inline void __account_unplug(struct io_stats *ios, int timer)
1341 {
1342         if (timer)
1343                 ios->timer_unplugs++;
1344         else
1345                 ios->io_unplugs++;
1346 }
1347
1348 static inline void account_unplug(struct blk_io_trace *t,
1349                                   struct per_cpu_info *pci, int timer)
1350 {
1351         __account_unplug(&pci->io_stats, timer);
1352
1353         if (per_process_stats) {
1354                 struct io_stats *ios = find_process_io_stats(t->pid);
1355
1356                 __account_unplug(ios, timer);
1357         }
1358 }
1359
1360 static inline void __account_requeue(struct io_stats *ios,
1361                                      struct blk_io_trace *t, int rw)
1362 {
1363         if (rw) {
1364                 ios->wrqueue++;
1365                 ios->iwrite_kb -= t_kb(t);
1366         } else {
1367                 ios->rrqueue++;
1368                 ios->iread_kb -= t_kb(t);
1369         }
1370 }
1371
1372 static inline void account_requeue(struct blk_io_trace *t,
1373                                    struct per_cpu_info *pci, int rw)
1374 {
1375         __account_requeue(&pci->io_stats, t, rw);
1376
1377         if (per_process_stats) {
1378                 struct io_stats *ios = find_process_io_stats(t->pid);
1379
1380                 __account_requeue(ios, t, rw);
1381         }
1382 }
1383
1384 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
1385                          struct blk_io_trace *t, char *act)
1386 {
1387         process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
1388 }
1389
1390 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
1391                        struct blk_io_trace *t, char *act)
1392 {
1393         process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
1394 }
1395
1396 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
1397                       char *act)
1398 {
1399         process_fmt(act, pci, t, -1, 0, NULL);
1400 }
1401
1402 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
1403                       struct blk_io_trace *t, char *act)
1404 {
1405         process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
1406 }
1407
1408 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
1409                       struct blk_io_trace *t, char *act)
1410 {
1411         if (act[0] == 'F')
1412                 log_track_frontmerge(pdi, t);
1413
1414         process_fmt(act, pci, t, -1ULL, 0, NULL);
1415 }
1416
1417 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
1418                         char *act)
1419 {
1420         process_fmt(act, pci, t, -1ULL, 0, NULL);
1421 }
1422
1423 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
1424                         char *act)
1425 {
1426         process_fmt(act, pci, t, -1ULL, 0, NULL);
1427 }
1428
1429 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
1430                       char *act)
1431 {
1432         process_fmt(act, pci, t, -1ULL, 0, NULL);
1433 }
1434
1435 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
1436                       char *act)
1437 {
1438         process_fmt(act, pci, t, -1ULL, 0, NULL);
1439 }
1440
1441 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1442 {
1443         unsigned char *buf = (unsigned char *) t + sizeof(*t);
1444
1445         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1446 }
1447
1448 static void dump_trace_pc(struct blk_io_trace *t, struct per_dev_info *pdi,
1449                           struct per_cpu_info *pci)
1450 {
1451         int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1452         int act = t->action & 0xffff;
1453
1454         switch (act) {
1455                 case __BLK_TA_QUEUE:
1456                         log_generic(pci, t, "Q");
1457                         account_pc_queue(t, pci, w);
1458                         break;
1459                 case __BLK_TA_GETRQ:
1460                         log_generic(pci, t, "G");
1461                         break;
1462                 case __BLK_TA_SLEEPRQ:
1463                         log_generic(pci, t, "S");
1464                         break;
1465                 case __BLK_TA_REQUEUE:
1466                         /*
1467                          * can happen if we miss traces, don't let it go
1468                          * below zero
1469                          */
1470                         if (pdi->cur_depth[w])
1471                                 pdi->cur_depth[w]--;
1472                         account_pc_requeue(t, pci, w);
1473                         log_generic(pci, t, "R");
1474                         break;
1475                 case __BLK_TA_ISSUE:
1476                         account_pc_issue(t, pci, w);
1477                         pdi->cur_depth[w]++;
1478                         if (pdi->cur_depth[w] > pdi->max_depth[w])
1479                                 pdi->max_depth[w] = pdi->cur_depth[w];
1480                         log_pc(pci, t, "D");
1481                         break;
1482                 case __BLK_TA_COMPLETE:
1483                         if (pdi->cur_depth[w])
1484                                 pdi->cur_depth[w]--;
1485                         log_pc(pci, t, "C");
1486                         account_pc_c(t, pci, w);
1487                         break;
1488                 case __BLK_TA_INSERT:
1489                         log_pc(pci, t, "I");
1490                         break;
1491                 default:
1492                         fprintf(stderr, "Bad pc action %x\n", act);
1493                         break;
1494         }
1495 }
1496
1497 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1498                           struct per_cpu_info *pci)
1499 {
1500         int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1501         int act = t->action & 0xffff;
1502
1503         switch (act) {
1504                 case __BLK_TA_QUEUE:
1505                         log_track_queue(pdi, t);
1506                         account_queue(t, pci, w);
1507                         log_queue(pci, t, "Q");
1508                         break;
1509                 case __BLK_TA_INSERT:
1510                         log_insert(pdi, pci, t, "I");
1511                         break;
1512                 case __BLK_TA_BACKMERGE:
1513                         account_m(t, pci, w);
1514                         log_merge(pdi, pci, t, "M");
1515                         break;
1516                 case __BLK_TA_FRONTMERGE:
1517                         account_m(t, pci, w);
1518                         log_merge(pdi, pci, t, "F");
1519                         break;
1520                 case __BLK_TA_GETRQ:
1521                         log_track_getrq(pdi, t);
1522                         log_generic(pci, t, "G");
1523                         break;
1524                 case __BLK_TA_SLEEPRQ:
1525                         log_generic(pci, t, "S");
1526                         break;
1527                 case __BLK_TA_REQUEUE:
1528                         /*
1529                          * can happen if we miss traces, don't let it go
1530                          * below zero
1531                          */
1532                         if (pdi->cur_depth[w])
1533                                 pdi->cur_depth[w]--;
1534                         account_requeue(t, pci, w);
1535                         log_queue(pci, t, "R");
1536                         break;
1537                 case __BLK_TA_ISSUE:
1538                         account_issue(t, pci, w);
1539                         pdi->cur_depth[w]++;
1540                         if (pdi->cur_depth[w] > pdi->max_depth[w])
1541                                 pdi->max_depth[w] = pdi->cur_depth[w];
1542                         log_issue(pdi, pci, t, "D");
1543                         break;
1544                 case __BLK_TA_COMPLETE:
1545                         if (pdi->cur_depth[w])
1546                                 pdi->cur_depth[w]--;
1547                         account_c(t, pci, w, t->bytes);
1548                         log_complete(pdi, pci, t, "C");
1549                         break;
1550                 case __BLK_TA_PLUG:
1551                         log_action(pci, t, "P");
1552                         break;
1553                 case __BLK_TA_UNPLUG_IO:
1554                         account_unplug(t, pci, 0);
1555                         log_unplug(pci, t, "U");
1556                         break;
1557                 case __BLK_TA_UNPLUG_TIMER:
1558                         account_unplug(t, pci, 1);
1559                         log_unplug(pci, t, "UT");
1560                         break;
1561                 case __BLK_TA_SPLIT:
1562                         log_split(pci, t, "X");
1563                         break;
1564                 case __BLK_TA_BOUNCE:
1565                         log_generic(pci, t, "B");
1566                         break;
1567                 case __BLK_TA_REMAP:
1568                         log_generic(pci, t, "A");
1569                         break;
1570                 default:
1571                         fprintf(stderr, "Bad fs action %x\n", t->action);
1572                         break;
1573         }
1574 }
1575
1576 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1577                        struct per_dev_info *pdi)
1578 {
1579         if (text_output) {
1580                 if (t->action & BLK_TC_ACT(BLK_TC_PC))
1581                         dump_trace_pc(t, pdi, pci);
1582                 else
1583                         dump_trace_fs(t, pdi, pci);
1584         }
1585
1586         if (!pdi->events)
1587                 pdi->first_reported_time = t->time;
1588
1589         pdi->events++;
1590
1591         output_binary(t, sizeof(*t) + t->pdu_len);
1592 }
1593
1594 /*
1595  * print in a proper way, not too small and not too big. if more than
1596  * 1000,000K, turn into M and so on
1597  */
1598 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1599 {
1600         char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1601         unsigned int i = 0;
1602
1603         if (in_kb)
1604                 i++;
1605
1606         while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1607                 i++;
1608                 num /= 1000;
1609         }
1610
1611         sprintf(dst, "%'8Lu%c", num, suff[i]);
1612         return dst;
1613 }
1614
1615 static void dump_io_stats(struct per_dev_info *pdi, struct io_stats *ios,
1616                           char *msg)
1617 {
1618         static char x[256], y[256];
1619
1620         fprintf(ofp, "%s\n", msg);
1621
1622         fprintf(ofp, " Reads Queued:    %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1623         fprintf(ofp, " Writes Queued:    %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1624         fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1625         fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1626         fprintf(ofp, " Reads Requeued:  %s\t\t", size_cnv(x, ios->rrqueue, 0));
1627         fprintf(ofp, " Writes Requeued:  %s\n", size_cnv(x, ios->wrqueue, 0));
1628         fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1629         fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1630         fprintf(ofp, " Read Merges:     %s, %siB\t", size_cnv(x, ios->mreads, 0), size_cnv(y, ios->mread_kb, 1));
1631         fprintf(ofp, " Write Merges:     %s, %siB\n", size_cnv(x, ios->mwrites, 0), size_cnv(y, ios->mwrite_kb, 1));
1632         if (pdi) {
1633                 fprintf(ofp, " Read depth:      %'8u%8c\t", pdi->max_depth[0], ' ');
1634                 fprintf(ofp, " Write depth:      %'8u\n", pdi->max_depth[1]);
1635         }
1636         if (ios->qreads_pc || ios->qwrites_pc || ios->ireads_pc || ios->iwrites_pc ||
1637             ios->rrqueue_pc || ios->wrqueue_pc || ios->creads_pc || ios->cwrites_pc) {
1638                 fprintf(ofp, " PC Reads Queued: %s, %siB\t", size_cnv(x, ios->qreads_pc, 0), size_cnv(y, ios->qread_kb_pc, 1));
1639                 fprintf(ofp, " PC Writes Queued: %s, %siB\n", size_cnv(x, ios->qwrites_pc, 0), size_cnv(y, ios->qwrite_kb_pc, 1));
1640                 fprintf(ofp, " PC Read Disp.:   %s, %siB\t", size_cnv(x, ios->ireads_pc, 0), size_cnv(y, ios->iread_kb_pc, 1));
1641                 fprintf(ofp, " PC Write Disp.:   %s, %siB\n", size_cnv(x, ios->iwrites_pc, 0), size_cnv(y, ios->iwrite_kb_pc, 1));
1642                 fprintf(ofp, " PC Reads Req.:   %s\t\t", size_cnv(x, ios->rrqueue_pc, 0));
1643                 fprintf(ofp, " PC Writes Req.:   %s\n", size_cnv(x, ios->wrqueue_pc, 0));
1644                 fprintf(ofp, " PC Reads Compl.: %s\t\t", size_cnv(x, ios->creads_pc, 0));
1645                 fprintf(ofp, " PC Writes Compl.: %s\n", size_cnv(x, ios->cwrites, 0));
1646         }
1647         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
1648         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
1649 }
1650
1651 static void dump_wait_stats(struct per_process_info *ppi)
1652 {
1653         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1654         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1655         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1656         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1657         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1658         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1659
1660         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1661         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
1662         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
1663         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
1664         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1665         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
1666 }
1667
1668 static int ppi_name_compare(const void *p1, const void *p2)
1669 {
1670         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1671         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1672         int res;
1673
1674         res = strverscmp(ppi1->ppm->comm, ppi2->ppm->comm);
1675         if (!res)
1676                 res = ppi1->ppm->pid > ppi2->ppm->pid;
1677
1678         return res;
1679 }
1680
1681 static void sort_process_list(void)
1682 {
1683         struct per_process_info **ppis;
1684         struct per_process_info *ppi;
1685         int i = 0;
1686
1687         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1688
1689         ppi = ppi_list;
1690         while (ppi) {
1691                 ppis[i++] = ppi;
1692                 ppi = ppi->list_next;
1693         }
1694
1695         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1696
1697         i = ppi_list_entries - 1;
1698         ppi_list = NULL;
1699         while (i >= 0) {
1700                 ppi = ppis[i];
1701
1702                 ppi->list_next = ppi_list;
1703                 ppi_list = ppi;
1704                 i--;
1705         }
1706
1707         free(ppis);
1708 }
1709
1710 static void show_process_stats(void)
1711 {
1712         struct per_process_info *ppi;
1713
1714         sort_process_list();
1715
1716         ppi = ppi_list;
1717         while (ppi) {
1718                 struct process_pid_map *ppm = ppi->ppm;
1719                 char name[64];
1720
1721                 if (ppi->more_than_one)
1722                         sprintf(name, "%s (%u, ...)", ppm->comm, ppm->pid);
1723                 else
1724                         sprintf(name, "%s (%u)", ppm->comm, ppm->pid);
1725
1726                 dump_io_stats(NULL, &ppi->io_stats, name);
1727                 dump_wait_stats(ppi);
1728                 ppi = ppi->list_next;
1729         }
1730
1731         fprintf(ofp, "\n");
1732 }
1733
1734 static void show_device_and_cpu_stats(void)
1735 {
1736         struct per_dev_info *pdi;
1737         struct per_cpu_info *pci;
1738         struct io_stats total, *ios;
1739         unsigned long long rrate, wrate, msec;
1740         int i, j, pci_events;
1741         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1742         char name[32];
1743         double ratio;
1744
1745         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1746
1747                 memset(&total, 0, sizeof(total));
1748                 pci_events = 0;
1749
1750                 if (i > 0)
1751                         fprintf(ofp, "\n");
1752
1753                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1754                         if (!pci->nelems)
1755                                 continue;
1756
1757                         ios = &pci->io_stats;
1758                         total.qreads += ios->qreads;
1759                         total.qwrites += ios->qwrites;
1760                         total.creads += ios->creads;
1761                         total.cwrites += ios->cwrites;
1762                         total.mreads += ios->mreads;
1763                         total.mwrites += ios->mwrites;
1764                         total.ireads += ios->ireads;
1765                         total.iwrites += ios->iwrites;
1766                         total.rrqueue += ios->rrqueue;
1767                         total.wrqueue += ios->wrqueue;
1768                         total.qread_kb += ios->qread_kb;
1769                         total.qwrite_kb += ios->qwrite_kb;
1770                         total.cread_kb += ios->cread_kb;
1771                         total.cwrite_kb += ios->cwrite_kb;
1772                         total.iread_kb += ios->iread_kb;
1773                         total.iwrite_kb += ios->iwrite_kb;
1774                         total.mread_kb += ios->mread_kb;
1775                         total.mwrite_kb += ios->mwrite_kb;
1776
1777                         total.qreads_pc += ios->qreads_pc;
1778                         total.qwrites_pc += ios->qwrites_pc;
1779                         total.creads_pc += ios->creads_pc;
1780                         total.cwrites_pc += ios->cwrites_pc;
1781                         total.ireads_pc += ios->ireads_pc;
1782                         total.iwrites_pc += ios->iwrites_pc;
1783                         total.rrqueue_pc += ios->rrqueue_pc;
1784                         total.wrqueue_pc += ios->wrqueue_pc;
1785                         total.qread_kb_pc += ios->qread_kb_pc;
1786                         total.qwrite_kb_pc += ios->qwrite_kb_pc;
1787                         total.iread_kb_pc += ios->iread_kb_pc;
1788                         total.iwrite_kb_pc += ios->iwrite_kb_pc;
1789
1790                         total.timer_unplugs += ios->timer_unplugs;
1791                         total.io_unplugs += ios->io_unplugs;
1792
1793                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1794                                  j, get_dev_name(pdi, name, sizeof(name)));
1795                         dump_io_stats(pdi, ios, line);
1796                         pci_events++;
1797                 }
1798
1799                 if (pci_events > 1) {
1800                         fprintf(ofp, "\n");
1801                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1802                                  get_dev_name(pdi, name, sizeof(name)));
1803                         dump_io_stats(NULL, &total, line);
1804                 }
1805
1806                 wrate = rrate = 0;
1807                 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1808                 if (msec) {
1809                         rrate = 1000 * total.cread_kb / msec;
1810                         wrate = 1000 * total.cwrite_kb / msec;
1811                 }
1812
1813                 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n",
1814                         rrate, wrate);
1815                 fprintf(ofp, "Events (%s): %'Lu entries\n",
1816                         get_dev_name(pdi, line, sizeof(line)), pdi->events);
1817
1818                 collect_pdi_skips(pdi);
1819                 if (!pdi->skips && !pdi->events)
1820                         ratio = 0.0;
1821                 else
1822                         ratio = 100.0 * ((double)pdi->seq_skips /
1823                                         (double)(pdi->events + pdi->seq_skips));
1824                 fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%)\n",
1825                         pdi->skips, pdi->seq_skips, ratio);
1826         }
1827 }
1828
1829 static void find_genesis(void)
1830 {
1831         struct trace *t = trace_list;
1832
1833         genesis_time = -1ULL;
1834         while (t != NULL) {
1835                 if (t->bit->time < genesis_time)
1836                         genesis_time = t->bit->time;
1837
1838                 t = t->next;
1839         }
1840
1841         /* The time stamp record will usually be the first
1842          * record in the trace, but not always.
1843          */
1844         if (start_timestamp
1845          && start_timestamp != genesis_time) {
1846                 long delta = genesis_time - start_timestamp;
1847
1848                 abs_start_time.tv_sec  += SECONDS(delta);
1849                 abs_start_time.tv_nsec += NANO_SECONDS(delta);
1850                 if (abs_start_time.tv_nsec < 0) {
1851                         abs_start_time.tv_nsec += 1000000000;
1852                         abs_start_time.tv_sec -= 1;
1853                 } else
1854                 if (abs_start_time.tv_nsec > 1000000000) {
1855                         abs_start_time.tv_nsec -= 1000000000;
1856                         abs_start_time.tv_sec += 1;
1857                 }
1858         }
1859 }
1860
1861 static inline int check_stopwatch(struct blk_io_trace *bit)
1862 {
1863         if (bit->time < stopwatch_end &&
1864             bit->time >= stopwatch_start)
1865                 return 0;
1866
1867         return 1;
1868 }
1869
1870 /*
1871  * return youngest entry read
1872  */
1873 static int sort_entries(unsigned long long *youngest)
1874 {
1875         struct per_dev_info *pdi = NULL;
1876         struct per_cpu_info *pci = NULL;
1877         struct trace *t;
1878
1879         if (!genesis_time)
1880                 find_genesis();
1881
1882         *youngest = 0;
1883         while ((t = trace_list) != NULL) {
1884                 struct blk_io_trace *bit = t->bit;
1885
1886                 trace_list = t->next;
1887
1888                 bit->time -= genesis_time;
1889
1890                 if (bit->time < *youngest || !*youngest)
1891                         *youngest = bit->time;
1892
1893                 if (!pdi || pdi->dev != bit->device) {
1894                         pdi = get_dev_info(bit->device);
1895                         pci = NULL;
1896                 }
1897
1898                 if (!pci || pci->cpu != bit->cpu)
1899                         pci = get_cpu_info(pdi, bit->cpu);
1900
1901                 if (bit->sequence < pci->smallest_seq_read)
1902                         pci->smallest_seq_read = bit->sequence;
1903
1904                 if (check_stopwatch(bit)) {
1905                         bit_free(bit);
1906                         t_free(t);
1907                         continue;
1908                 }
1909
1910                 if (trace_rb_insert_sort(t))
1911                         return -1;
1912         }
1913
1914         return 0;
1915 }
1916
1917 /*
1918  * to continue, we must have traces from all online cpus in the tree
1919  */
1920 static int check_cpu_map(struct per_dev_info *pdi)
1921 {
1922         unsigned long *cpu_map;
1923         struct rb_node *n;
1924         struct trace *__t;
1925         unsigned int i;
1926         int ret, cpu;
1927
1928         /*
1929          * create a map of the cpus we have traces for
1930          */
1931         cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1932         n = rb_first(&rb_sort_root);
1933         while (n) {
1934                 __t = rb_entry(n, struct trace, rb_node);
1935                 cpu = __t->bit->cpu;
1936
1937                 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1938                 n = rb_next(n);
1939         }
1940
1941         /*
1942          * we can't continue if pdi->cpu_map has entries set that we don't
1943          * have in the sort rbtree. the opposite is not a problem, though
1944          */
1945         ret = 0;
1946         for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1947                 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1948                         ret = 1;
1949                         break;
1950                 }
1951         }
1952
1953         free(cpu_map);
1954         return ret;
1955 }
1956
1957 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1958 {
1959         struct blk_io_trace *bit = t->bit;
1960         unsigned long expected_sequence;
1961         struct per_cpu_info *pci;
1962         struct trace *__t;
1963
1964         pci = get_cpu_info(pdi, bit->cpu);
1965         expected_sequence = pci->last_sequence + 1;
1966
1967         if (!expected_sequence) {
1968                 /*
1969                  * 1 should be the first entry, just allow it
1970                  */
1971                 if (bit->sequence == 1)
1972                         return 0;
1973                 if (bit->sequence == pci->smallest_seq_read)
1974                         return 0;
1975
1976                 return check_cpu_map(pdi);
1977         }
1978
1979         if (bit->sequence == expected_sequence)
1980                 return 0;
1981
1982         /*
1983          * we may not have seen that sequence yet. if we are not doing
1984          * the final run, break and wait for more entries.
1985          */
1986         if (expected_sequence < pci->smallest_seq_read) {
1987                 __t = trace_rb_find_last(pdi, pci, expected_sequence);
1988                 if (!__t)
1989                         goto skip;
1990
1991                 __put_trace_last(pdi, __t);
1992                 return 0;
1993         } else if (!force) {
1994                 return 1;
1995         } else {
1996 skip:
1997                 if (check_current_skips(pci, bit->sequence))
1998                         return 0;
1999
2000                 if (expected_sequence < bit->sequence)
2001                         insert_skip(pci, expected_sequence, bit->sequence - 1);
2002                 return 0;
2003         }
2004 }
2005
2006 static void show_entries_rb(int force)
2007 {
2008         struct per_dev_info *pdi = NULL;
2009         struct per_cpu_info *pci = NULL;
2010         struct blk_io_trace *bit;
2011         struct rb_node *n;
2012         struct trace *t;
2013
2014         while ((n = rb_first(&rb_sort_root)) != NULL) {
2015                 if (is_done() && !force && !pipeline)
2016                         break;
2017
2018                 t = rb_entry(n, struct trace, rb_node);
2019                 bit = t->bit;
2020
2021                 if (read_sequence - t->read_sequence < 1 && !force)
2022                         break;
2023
2024                 if (!pdi || pdi->dev != bit->device) {
2025                         pdi = get_dev_info(bit->device);
2026                         pci = NULL;
2027                 }
2028
2029                 if (!pdi) {
2030                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
2031                                 MAJOR(bit->device), MINOR(bit->device));
2032                         break;
2033                 }
2034
2035                 if (check_sequence(pdi, t, force))
2036                         break;
2037
2038                 if (!force && bit->time > last_allowed_time)
2039                         break;
2040
2041                 check_time(pdi, bit);
2042
2043                 if (!pci || pci->cpu != bit->cpu)
2044                         pci = get_cpu_info(pdi, bit->cpu);
2045
2046                 pci->last_sequence = bit->sequence;
2047
2048                 pci->nelems++;
2049
2050                 if (bit->action & (act_mask << BLK_TC_SHIFT))
2051                         dump_trace(bit, pci, pdi);
2052
2053                 put_trace(pdi, t);
2054         }
2055 }
2056
2057 static int read_data(int fd, void *buffer, int bytes, int block, int *fdblock)
2058 {
2059         int ret, bytes_left, fl;
2060         void *p;
2061
2062         if (block != *fdblock) {
2063                 fl = fcntl(fd, F_GETFL);
2064
2065                 if (!block) {
2066                         *fdblock = 0;
2067                         fcntl(fd, F_SETFL, fl | O_NONBLOCK);
2068                 } else {
2069                         *fdblock = 1;
2070                         fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
2071                 }
2072         }
2073
2074         bytes_left = bytes;
2075         p = buffer;
2076         while (bytes_left > 0) {
2077                 ret = read(fd, p, bytes_left);
2078                 if (!ret)
2079                         return 1;
2080                 else if (ret < 0) {
2081                         if (errno != EAGAIN) {
2082                                 perror("read");
2083                                 return -1;
2084                         }
2085
2086                         /*
2087                          * never do partial reads. we can return if we
2088                          * didn't read anything and we should not block,
2089                          * otherwise wait for data
2090                          */
2091                         if ((bytes_left == bytes) && !block)
2092                                 return 1;
2093
2094                         usleep(10);
2095                         continue;
2096                 } else {
2097                         p += ret;
2098                         bytes_left -= ret;
2099                 }
2100         }
2101
2102         return 0;
2103 }
2104
2105 static inline __u16 get_pdulen(struct blk_io_trace *bit)
2106 {
2107         if (data_is_native)
2108                 return bit->pdu_len;
2109
2110         return __bswap_16(bit->pdu_len);
2111 }
2112
2113 static inline __u32 get_magic(struct blk_io_trace *bit)
2114 {
2115         if (data_is_native)
2116                 return bit->magic;
2117
2118         return __bswap_32(bit->magic);
2119 }
2120
2121 static int read_events(int fd, int always_block, int *fdblock)
2122 {
2123         struct per_dev_info *pdi = NULL;
2124         unsigned int events = 0;
2125
2126         while (!is_done() && events < rb_batch) {
2127                 struct blk_io_trace *bit;
2128                 struct trace *t;
2129                 int pdu_len, should_block, ret;
2130                 __u32 magic;
2131
2132                 bit = bit_alloc();
2133
2134                 should_block = !events || always_block;
2135
2136                 ret = read_data(fd, bit, sizeof(*bit), should_block, fdblock);
2137                 if (ret) {
2138                         bit_free(bit);
2139                         if (!events && ret < 0)
2140                                 events = ret;
2141                         break;
2142                 }
2143
2144                 /*
2145                  * look at first trace to check whether we need to convert
2146                  * data in the future
2147                  */
2148                 if (data_is_native == -1 && check_data_endianness(bit->magic))
2149                         break;
2150
2151                 magic = get_magic(bit);
2152                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2153                         fprintf(stderr, "Bad magic %x\n", magic);
2154                         break;
2155                 }
2156
2157                 pdu_len = get_pdulen(bit);
2158                 if (pdu_len) {
2159                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2160
2161                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1, fdblock)) {
2162                                 bit_free(ptr);
2163                                 break;
2164                         }
2165
2166                         bit = ptr;
2167                 }
2168
2169                 trace_to_cpu(bit);
2170
2171                 if (verify_trace(bit)) {
2172                         bit_free(bit);
2173                         continue;
2174                 }
2175
2176                 /*
2177                  * not a real trace, so grab and handle it here
2178                  */
2179                 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY)) {
2180                         handle_notify(bit);
2181                         output_binary(bit, sizeof(*bit) + bit->pdu_len);
2182                         continue;
2183                 }
2184
2185                 t = t_alloc();
2186                 memset(t, 0, sizeof(*t));
2187                 t->bit = bit;
2188                 t->read_sequence = read_sequence;
2189
2190                 t->next = trace_list;
2191                 trace_list = t;
2192
2193                 if (!pdi || pdi->dev != bit->device)
2194                         pdi = get_dev_info(bit->device);
2195
2196                 if (bit->time > pdi->last_read_time)
2197                         pdi->last_read_time = bit->time;
2198
2199                 events++;
2200         }
2201
2202         return events;
2203 }
2204
2205 /*
2206  * Managing input streams
2207  */
2208
2209 struct ms_stream {
2210         struct ms_stream *next;
2211         struct trace *first, *last;
2212         struct per_dev_info *pdi;
2213         unsigned int cpu;
2214 };
2215
2216 #define MS_HASH(d, c) ((MAJOR(d) & 0xff) ^ (MINOR(d) & 0xff) ^ (cpu & 0xff))
2217
2218 struct ms_stream *ms_head;
2219 struct ms_stream *ms_hash[256];
2220
2221 static void ms_sort(struct ms_stream *msp);
2222 static int ms_prime(struct ms_stream *msp);
2223
2224 static inline struct trace *ms_peek(struct ms_stream *msp)
2225 {
2226         return (msp == NULL) ? NULL : msp->first;
2227 }
2228
2229 static inline __u64 ms_peek_time(struct ms_stream *msp)
2230 {
2231         return ms_peek(msp)->bit->time;
2232 }
2233
2234 static inline void ms_resort(struct ms_stream *msp)
2235 {
2236         if (msp->next && ms_peek_time(msp) > ms_peek_time(msp->next)) {
2237                 ms_head = msp->next;
2238                 msp->next = NULL;
2239                 ms_sort(msp);
2240         }
2241 }
2242
2243 static inline void ms_deq(struct ms_stream *msp)
2244 {
2245         msp->first = msp->first->next;
2246         if (!msp->first) {
2247                 msp->last = NULL;
2248                 if (!ms_prime(msp)) {
2249                         ms_head = msp->next;
2250                         msp->next = NULL;
2251                         return;
2252                 }
2253         }
2254
2255         ms_resort(msp);
2256 }
2257
2258 static void ms_sort(struct ms_stream *msp)
2259 {
2260         __u64 msp_t = ms_peek_time(msp);
2261         struct ms_stream *this_msp = ms_head;
2262
2263         if (this_msp == NULL)
2264                 ms_head = msp;
2265         else if (msp_t < ms_peek_time(this_msp)) {
2266                 msp->next = this_msp;
2267                 ms_head = msp;
2268         }
2269         else {
2270                 while (this_msp->next && ms_peek_time(this_msp->next) < msp_t)
2271                         this_msp = this_msp->next;
2272
2273                 msp->next = this_msp->next;
2274                 this_msp->next = msp;
2275         }
2276 }
2277
2278 static int ms_prime(struct ms_stream *msp)
2279 {
2280         __u32 magic;
2281         unsigned int i;
2282         struct trace *t;
2283         struct per_dev_info *pdi = msp->pdi;
2284         struct per_cpu_info *pci = get_cpu_info(pdi, msp->cpu);
2285         struct blk_io_trace *bit = NULL;
2286         int ret, pdu_len, ndone = 0;
2287
2288         for (i = 0; !is_done() && pci->fd >= 0 && i < rb_batch; i++) {
2289                 bit = bit_alloc();
2290                 ret = read_data(pci->fd, bit, sizeof(*bit), 1, &pci->fdblock);
2291                 if (ret)
2292                         goto err;
2293
2294                 if (data_is_native == -1 && check_data_endianness(bit->magic))
2295                         goto err;
2296
2297                 magic = get_magic(bit);
2298                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2299                         fprintf(stderr, "Bad magic %x\n", magic);
2300                         goto err;
2301
2302                 }
2303
2304                 pdu_len = get_pdulen(bit);
2305                 if (pdu_len) {
2306                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2307                         ret = read_data(pci->fd, ptr + sizeof(*bit), pdu_len,
2308                                                              1, &pci->fdblock);
2309                         if (ret) {
2310                                 free(ptr);
2311                                 bit = NULL;
2312                                 goto err;
2313                         }
2314
2315                         bit = ptr;
2316                 }
2317
2318                 trace_to_cpu(bit);
2319                 if (verify_trace(bit))
2320                         goto err;
2321
2322                 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY)) {
2323                         handle_notify(bit);
2324                         output_binary(bit, sizeof(*bit) + bit->pdu_len);
2325                         bit_free(bit);
2326
2327                         i -= 1;
2328                         continue;
2329                 }
2330
2331                 if (bit->time > pdi->last_read_time)
2332                         pdi->last_read_time = bit->time;
2333
2334                 t = t_alloc();
2335                 memset(t, 0, sizeof(*t));
2336                 t->bit = bit;
2337
2338                 if (msp->first == NULL)
2339                         msp->first = msp->last = t;
2340                 else {
2341                         msp->last->next = t;
2342                         msp->last = t;
2343                 }
2344
2345                 ndone++;
2346         }
2347
2348         return ndone;
2349
2350 err:
2351         if (bit) bit_free(bit);
2352
2353         cpu_mark_offline(pdi, pci->cpu);
2354         close(pci->fd);
2355         pci->fd = -1;
2356
2357         return ndone;
2358 }
2359
2360 static struct ms_stream *ms_alloc(struct per_dev_info *pdi, int cpu)
2361 {
2362         struct ms_stream *msp = malloc(sizeof(*msp));
2363
2364         msp->next = NULL;
2365         msp->first = msp->last = NULL;
2366         msp->pdi = pdi;
2367         msp->cpu = cpu;
2368
2369         if (ms_prime(msp))
2370                 ms_sort(msp);
2371
2372         return msp;
2373 }
2374
2375 static int setup_file(struct per_dev_info *pdi, int cpu)
2376 {
2377         int len = 0;
2378         struct stat st;
2379         char *p, *dname;
2380         struct per_cpu_info *pci = get_cpu_info(pdi, cpu);
2381
2382         pci->cpu = cpu;
2383         pci->fdblock = -1;
2384
2385         p = strdup(pdi->name);
2386         dname = dirname(p);
2387         if (strcmp(dname, ".")) {
2388                 input_dir = dname;
2389                 p = strdup(pdi->name);
2390                 strcpy(pdi->name, basename(p));
2391         }
2392         free(p);
2393
2394         if (input_dir)
2395                 len = sprintf(pci->fname, "%s/", input_dir);
2396
2397         snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
2398                  "%s.blktrace.%d", pdi->name, pci->cpu);
2399         if (stat(pci->fname, &st) < 0)
2400                 return 0;
2401         if (!st.st_size)
2402                 return 1;
2403
2404         pci->fd = open(pci->fname, O_RDONLY);
2405         if (pci->fd < 0) {
2406                 perror(pci->fname);
2407                 return 0;
2408         }
2409
2410         printf("Input file %s added\n", pci->fname);
2411         cpu_mark_online(pdi, pci->cpu);
2412
2413         pdi->nfiles++;
2414         ms_alloc(pdi, pci->cpu);
2415
2416         return 1;
2417 }
2418
2419 static int handle(struct ms_stream *msp)
2420 {
2421         struct trace *t;
2422         struct per_dev_info *pdi;
2423         struct per_cpu_info *pci;
2424         struct blk_io_trace *bit;
2425
2426         t = ms_peek(msp);
2427
2428         bit = t->bit;
2429         pdi = msp->pdi;
2430         pci = get_cpu_info(pdi, msp->cpu);
2431         pci->nelems++;
2432         bit->time -= genesis_time;
2433
2434         if (t->bit->time > stopwatch_end)
2435                 return 0;
2436
2437         pdi->last_reported_time = bit->time;
2438         if ((bit->action & (act_mask << BLK_TC_SHIFT))&&
2439             t->bit->time >= stopwatch_start)
2440                 dump_trace(bit, pci, pdi);
2441
2442         ms_deq(msp);
2443
2444         if (text_output)
2445                 trace_rb_insert_last(pdi, t);
2446         else {
2447                 bit_free(t->bit);
2448                 t_free(t);
2449         }
2450
2451         return 1;
2452 }
2453
2454 /*
2455  * Check if we need to sanitize the name. We allow 'foo', or if foo.blktrace.X
2456  * is given, then strip back down to 'foo' to avoid missing files.
2457  */
2458 static int name_fixup(char *name)
2459 {
2460         char *b;
2461
2462         if (!name)
2463                 return 1;
2464
2465         b = strstr(name, ".blktrace.");
2466         if (b)
2467                 *b = '\0';
2468
2469         return 0;
2470 }
2471
2472 static int do_file(void)
2473 {
2474         int i, cpu, ret;
2475         struct per_dev_info *pdi;
2476
2477         /*
2478          * first prepare all files for reading
2479          */
2480         for (i = 0; i < ndevices; i++) {
2481                 pdi = &devices[i];
2482                 ret = name_fixup(pdi->name);
2483                 if (ret)
2484                         return ret;
2485
2486                 for (cpu = 0; setup_file(pdi, cpu); cpu++)
2487                         ;
2488         }
2489
2490         /*
2491          * Get the initial time stamp
2492          */
2493         if (ms_head)
2494                 genesis_time = ms_peek_time(ms_head);
2495
2496         /*
2497          * Keep processing traces while any are left
2498          */
2499         while (!is_done() && ms_head && handle(ms_head))
2500                 ;
2501
2502         return 0;
2503 }
2504
2505 static void do_pipe(int fd)
2506 {
2507         unsigned long long youngest;
2508         int events, fdblock;
2509
2510         last_allowed_time = -1ULL;
2511         fdblock = -1;
2512         while ((events = read_events(fd, 0, &fdblock)) > 0) {
2513                 read_sequence++;
2514
2515 #if 0
2516                 smallest_seq_read = -1U;
2517 #endif
2518
2519                 if (sort_entries(&youngest))
2520                         break;
2521
2522                 if (youngest > stopwatch_end)
2523                         break;
2524
2525                 show_entries_rb(0);
2526         }
2527
2528         if (rb_sort_entries)
2529                 show_entries_rb(1);
2530 }
2531
2532 static int do_fifo(void)
2533 {
2534         int fd;
2535
2536         if (!strcmp(pipename, "-"))
2537                 fd = dup(STDIN_FILENO);
2538         else
2539                 fd = open(pipename, O_RDONLY);
2540
2541         if (fd == -1) {
2542                 perror("dup stdin");
2543                 return -1;
2544         }
2545
2546         do_pipe(fd);
2547         close(fd);
2548         return 0;
2549 }
2550
2551 static void show_stats(void)
2552 {
2553         if (!ofp)
2554                 return;
2555         if (stats_printed)
2556                 return;
2557
2558         stats_printed = 1;
2559
2560         if (per_process_stats)
2561                 show_process_stats();
2562
2563         if (per_device_and_cpu_stats)
2564                 show_device_and_cpu_stats();
2565
2566         fflush(ofp);
2567 }
2568
2569 static void handle_sigint(__attribute__((__unused__)) int sig)
2570 {
2571         done = 1;
2572 }
2573
2574 /*
2575  * Extract start and duration times from a string, allowing
2576  * us to specify a time interval of interest within a trace.
2577  * Format: "duration" (start is zero) or "start:duration".
2578  */
2579 static int find_stopwatch_interval(char *string)
2580 {
2581         double value;
2582         char *sp;
2583
2584         value = strtod(string, &sp);
2585         if (sp == string) {
2586                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
2587                 return 1;
2588         }
2589         if (*sp == ':') {
2590                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
2591                 string = sp + 1;
2592                 value = strtod(string, &sp);
2593                 if (sp == string || *sp != '\0') {
2594                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
2595                                 string);
2596                         return 1;
2597                 }
2598         } else if (*sp != '\0') {
2599                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
2600                 return 1;
2601         }
2602         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
2603         if (stopwatch_end <= stopwatch_start) {
2604                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
2605                         stopwatch_start, stopwatch_end);
2606                 return 1;
2607         }
2608
2609         return 0;
2610 }
2611
2612 static int is_pipe(const char *str)
2613 {
2614         struct stat st;
2615
2616         if (!strcmp(str, "-"))
2617                 return 1;
2618         if (!stat(str, &st) && S_ISFIFO(st.st_mode))
2619                 return 1;
2620
2621         return 0;
2622 }
2623
2624 #define S_OPTS  "a:A:b:D:d:f:F:hi:o:Oqstw:vV"
2625 static char usage_str[] =    "\n\n" \
2626         "-i <file>           | --input=<file>\n" \
2627         "[ -a <action field> | --act-mask=<action field> ]\n" \
2628         "[ -A <action mask>  | --set-mask=<action mask> ]\n" \
2629         "[ -b <traces>       | --batch=<traces> ]\n" \
2630         "[ -d <file>         | --dump-binary=<file> ]\n" \
2631         "[ -D <dir>          | --input-directory=<dir> ]\n" \
2632         "[ -f <format>       | --format=<format> ]\n" \
2633         "[ -F <spec>         | --format-spec=<spec> ]\n" \
2634         "[ -h                | --hash-by-name ]\n" \
2635         "[ -o <file>         | --output=<file> ]\n" \
2636         "[ -O                | --no-text-output ]\n" \
2637         "[ -q                | --quiet ]\n" \
2638         "[ -s                | --per-program-stats ]\n" \
2639         "[ -t                | --track-ios ]\n" \
2640         "[ -w <time>         | --stopwatch=<time> ]\n" \
2641         "[ -v                | --verbose ]\n" \
2642         "[ -V                | --version ]\n\n" \
2643         "\t-b stdin read batching\n" \
2644         "\t-d Output file. If specified, binary data is written to file\n" \
2645         "\t-D Directory to prepend to input file names\n" \
2646         "\t-f Output format. Customize the output format. The format field\n" \
2647         "\t   identifies can be found in the documentation\n" \
2648         "\t-F Format specification. Can be found in the documentation\n" \
2649         "\t-h Hash processes by name, not pid\n" \
2650         "\t-i Input file containing trace data, or '-' for stdin\n" \
2651         "\t-o Output file. If not given, output is stdout\n" \
2652         "\t-O Do NOT output text data\n" \
2653         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
2654         "\t-s Show per-program io statistics\n" \
2655         "\t-t Track individual ios. Will tell you the time a request took\n" \
2656         "\t   to get queued, to get dispatched, and to get completed\n" \
2657         "\t-w Only parse data between the given time interval in seconds.\n" \
2658         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
2659         "\t-v More verbose for marginal errors\n" \
2660         "\t-V Print program version info\n\n";
2661
2662 static void usage(char *prog)
2663 {
2664         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
2665 }
2666
2667 int main(int argc, char *argv[])
2668 {
2669         int i, c, ret, mode;
2670         int act_mask_tmp = 0;
2671         char *ofp_buffer = NULL;
2672         char *bin_ofp_buffer = NULL;
2673
2674         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
2675                 switch (c) {
2676                 case 'a':
2677                         i = find_mask_map(optarg);
2678                         if (i < 0) {
2679                                 fprintf(stderr,"Invalid action mask %s\n",
2680                                         optarg);
2681                                 return 1;
2682                         }
2683                         act_mask_tmp |= i;
2684                         break;
2685
2686                 case 'A':
2687                         if ((sscanf(optarg, "%x", &i) != 1) ||
2688                                                         !valid_act_opt(i)) {
2689                                 fprintf(stderr,
2690                                         "Invalid set action mask %s/0x%x\n",
2691                                         optarg, i);
2692                                 return 1;
2693                         }
2694                         act_mask_tmp = i;
2695                         break;
2696                 case 'i':
2697                         if (is_pipe(optarg) && !pipeline) {
2698                                 pipeline = 1;
2699                                 pipename = strdup(optarg);
2700                         } else if (resize_devices(optarg) != 0)
2701                                 return 1;
2702                         break;
2703                 case 'D':
2704                         input_dir = optarg;
2705                         break;
2706                 case 'o':
2707                         output_name = optarg;
2708                         break;
2709                 case 'O':
2710                         text_output = 0;
2711                         break;
2712                 case 'b':
2713                         rb_batch = atoi(optarg);
2714                         if (rb_batch <= 0)
2715                                 rb_batch = RB_BATCH_DEFAULT;
2716                         break;
2717                 case 's':
2718                         per_process_stats = 1;
2719                         break;
2720                 case 't':
2721                         track_ios = 1;
2722                         break;
2723                 case 'q':
2724                         per_device_and_cpu_stats = 0;
2725                         break;
2726                 case 'w':
2727                         if (find_stopwatch_interval(optarg) != 0)
2728                                 return 1;
2729                         break;
2730                 case 'f':
2731                         set_all_format_specs(optarg);
2732                         break;
2733                 case 'F':
2734                         if (add_format_spec(optarg) != 0)
2735                                 return 1;
2736                         break;
2737                 case 'h':
2738                         ppi_hash_by_pid = 0;
2739                         break;
2740                 case 'v':
2741                         verbose++;
2742                         break;
2743                 case 'V':
2744                         printf("%s version %s\n", argv[0], blkparse_version);
2745                         return 0;
2746                 case 'd':
2747                         dump_binary = optarg;
2748                         break;
2749                 default:
2750                         usage(argv[0]);
2751                         return 1;
2752                 }
2753         }
2754
2755         while (optind < argc) {
2756                 if (is_pipe(argv[optind]) && !pipeline) {
2757                         pipeline = 1;
2758                         pipename = strdup(argv[optind]);
2759                 } else if (resize_devices(argv[optind]) != 0)
2760                         return 1;
2761                 optind++;
2762         }
2763
2764         if (!pipeline && !ndevices) {
2765                 usage(argv[0]);
2766                 return 1;
2767         }
2768
2769         if (act_mask_tmp != 0)
2770                 act_mask = act_mask_tmp;
2771
2772         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
2773
2774         signal(SIGINT, handle_sigint);
2775         signal(SIGHUP, handle_sigint);
2776         signal(SIGTERM, handle_sigint);
2777
2778         setlocale(LC_NUMERIC, "en_US");
2779
2780         if (text_output) {
2781                 if (!output_name) {
2782                         ofp = fdopen(STDOUT_FILENO, "w");
2783                         mode = _IOLBF;
2784                 } else {
2785                         char ofname[128];
2786
2787                         snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
2788                         ofp = fopen(ofname, "w");
2789                         mode = _IOFBF;
2790                 }
2791
2792                 if (!ofp) {
2793                         perror("fopen");
2794                         return 1;
2795                 }
2796
2797                 ofp_buffer = malloc(4096);
2798                 if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
2799                         perror("setvbuf");
2800                         return 1;
2801                 }
2802         }
2803
2804         if (dump_binary) {
2805                 dump_fp = fopen(dump_binary, "w");
2806                 if (!dump_fp) {
2807                         perror(dump_binary);
2808                         dump_binary = NULL;
2809                         return 1;
2810                 }
2811                 bin_ofp_buffer = malloc(128 * 1024);
2812                 if (setvbuf(dump_fp, bin_ofp_buffer, _IOFBF, 128 * 1024)) {
2813                         perror("setvbuf binary");
2814                         return 1;
2815                 }
2816         }
2817
2818         if (pipeline)
2819                 ret = do_fifo();
2820         else
2821                 ret = do_file();
2822
2823         if (!ret)
2824                 show_stats();
2825
2826         if (ofp_buffer) {
2827                 fflush(ofp);
2828                 free(ofp_buffer);
2829         }
2830         if (bin_ofp_buffer) {
2831                 fflush(dump_fp);
2832                 free(bin_ofp_buffer);
2833         }
2834         return ret;
2835 }