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         case BLK_TN_MESSAGE:
 594                 if (bit->pdu_len > 0) {
 595                         char msg[bit->pdu_len+1];
 596
 597                         memcpy(msg, (char *)payload, bit->pdu_len);
 598                         msg[bit->pdu_len] = '\0';
 599
 600                         fprintf(ofp,
 601                                 "%3d,%-3d %2d %8s %5d.%09lu %5u %2s %3s %s\n",
 602                                 MAJOR(bit->device), MINOR(bit->device),
 603                                 bit->cpu, "0", (int) SECONDS(bit->time),
 604                                 (unsigned long) NANO_SECONDS(bit->time),
 605                                 0, "m", "N", msg);
 606                 }
 607                 break;
 608
 609         default:
 610                 /* Ignore unknown notify events */
 611                 ;
 612         }
 613 }
 614
 615 char *find_process_name(pid_t pid)
 616 {
 617         struct process_pid_map *ppm = find_ppm(pid);
 618
 619         if (ppm)
 620                 return ppm->comm;
 621
 622         return NULL;
 623 }
 624
 625 static inline int ppi_hash_pid(pid_t pid)
 626 {
 627         return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
 628 }
 629
 630 static inline int ppi_hash_name(const char *name)
 631 {
 632         return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
 633 }
 634
 635 static inline int ppi_hash(struct per_process_info *ppi)
 636 {
 637         struct process_pid_map *ppm = ppi->ppm;
 638
 639         if (ppi_hash_by_pid)
 640                 return ppi_hash_pid(ppm->pid);
 641
 642         return ppi_hash_name(ppm->comm);
 643 }
 644
 645 static inline void add_ppi_to_hash(struct per_process_info *ppi)
 646 {
 647         const int hash_idx = ppi_hash(ppi);
 648
 649         ppi->hash_next = ppi_hash_table[hash_idx];
 650         ppi_hash_table[hash_idx] = ppi;
 651 }
 652
 653 static inline void add_ppi_to_list(struct per_process_info *ppi)
 654 {
 655         ppi->list_next = ppi_list;
 656         ppi_list = ppi;
 657         ppi_list_entries++;
 658 }
 659
 660 static struct per_process_info *find_ppi_by_name(char *name)
 661 {
 662         const int hash_idx = ppi_hash_name(name);
 663         struct per_process_info *ppi;
 664
 665         ppi = ppi_hash_table[hash_idx];
 666         while (ppi) {
 667                 struct process_pid_map *ppm = ppi->ppm;
 668
 669                 if (!strcmp(ppm->comm, name))
 670                         return ppi;
 671
 672                 ppi = ppi->hash_next;
 673         }
 674
 675         return NULL;
 676 }
 677
 678 static struct per_process_info *find_ppi_by_pid(pid_t pid)
 679 {
 680         const int hash_idx = ppi_hash_pid(pid);
 681         struct per_process_info *ppi;
 682
 683         ppi = ppi_hash_table[hash_idx];
 684         while (ppi) {
 685                 struct process_pid_map *ppm = ppi->ppm;
 686
 687                 if (ppm->pid == pid)
 688                         return ppi;
 689
 690                 ppi = ppi->hash_next;
 691         }
 692
 693         return NULL;
 694 }
 695
 696 static struct per_process_info *find_ppi(pid_t pid)
 697 {
 698         struct per_process_info *ppi;
 699         char *name;
 700
 701         if (ppi_hash_by_pid)
 702                 return find_ppi_by_pid(pid);
 703
 704         name = find_process_name(pid);
 705         if (!name)
 706                 return NULL;
 707
 708         ppi = find_ppi_by_name(name);
 709         if (ppi && ppi->ppm->pid != pid)
 710                 ppi->more_than_one = 1;
 711
 712         return ppi;
 713 }
 714
 715 /*
 716  * struct trace and blktrace allocation cache, we do potentially
 717  * millions of mallocs for these structures while only using at most
 718  * a few thousand at the time
 719  */
 720 static inline void t_free(struct trace *t)
 721 {
 722         if (t_alloc_cache < 1024) {
 723                 t->next = t_alloc_list;
 724                 t_alloc_list = t;
 725                 t_alloc_cache++;
 726         } else
 727                 free(t);
 728 }
 729
 730 static inline struct trace *t_alloc(void)
 731 {
 732         struct trace *t = t_alloc_list;
 733
 734         if (t) {
 735                 t_alloc_list = t->next;
 736                 t_alloc_cache--;
 737                 return t;
 738         }
 739
 740         return malloc(sizeof(*t));
 741 }
 742
 743 static inline void bit_free(struct blk_io_trace *bit)
 744 {
 745         if (bit_alloc_cache < 1024 && !bit->pdu_len) {
 746                 /*
 747                  * abuse a 64-bit field for a next pointer for the free item
 748                  */
 749                 bit->time = (__u64) (unsigned long) bit_alloc_list;
 750                 bit_alloc_list = (struct blk_io_trace *) bit;
 751                 bit_alloc_cache++;
 752         } else
 753                 free(bit);
 754 }
 755
 756 static inline struct blk_io_trace *bit_alloc(void)
 757 {
 758         struct blk_io_trace *bit = bit_alloc_list;
 759
 760         if (bit) {
 761                 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
 762                                  bit->time;
 763                 bit_alloc_cache--;
 764                 return bit;
 765         }
 766
 767         return malloc(sizeof(*bit));
 768 }
 769
 770 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
 771 {
 772         struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
 773
 774         rb_erase(&t->rb_node, &pci->rb_last);
 775         pci->rb_last_entries--;
 776
 777         bit_free(t->bit);
 778         t_free(t);
 779 }
 780
 781 static void put_trace(struct per_dev_info *pdi, struct trace *t)
 782 {
 783         rb_erase(&t->rb_node, &rb_sort_root);
 784         rb_sort_entries--;
 785
 786         trace_rb_insert_last(pdi, t);
 787 }
 788
 789 static inline int trace_rb_insert(struct trace *t, struct rb_root *root)
 790 {
 791         struct rb_node **p = &root->rb_node;
 792         struct rb_node *parent = NULL;
 793         struct trace *__t;
 794
 795         while (*p) {
 796                 parent = *p;
 797
 798                 __t = rb_entry(parent, struct trace, rb_node);
 799
 800                 if (t->bit->time < __t->bit->time)
 801                         p = &(*p)->rb_left;
 802                 else if (t->bit->time > __t->bit->time)
 803                         p = &(*p)->rb_right;
 804                 else if (t->bit->device < __t->bit->device)
 805                         p = &(*p)->rb_left;
 806                 else if (t->bit->device > __t->bit->device)
 807                         p = &(*p)->rb_right;
 808                 else if (t->bit->sequence < __t->bit->sequence)
 809                         p = &(*p)->rb_left;
 810                 else    /* >= sequence */
 811                         p = &(*p)->rb_right;
 812         }
 813
 814         rb_link_node(&t->rb_node, parent, p);
 815         rb_insert_color(&t->rb_node, root);
 816         return 0;
 817 }
 818
 819 static inline int trace_rb_insert_sort(struct trace *t)
 820 {
 821         if (!trace_rb_insert(t, &rb_sort_root)) {
 822                 rb_sort_entries++;
 823                 return 0;
 824         }
 825
 826         return 1;
 827 }
 828
 829 static int trace_rb_insert_last(struct per_dev_info *pdi, struct trace *t)
 830 {
 831         struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
 832
 833         if (trace_rb_insert(t, &pci->rb_last))
 834                 return 1;
 835
 836         pci->rb_last_entries++;
 837
 838         if (pci->rb_last_entries > rb_batch * pdi->nfiles) {
 839                 struct rb_node *n = rb_first(&pci->rb_last);
 840
 841                 t = rb_entry(n, struct trace, rb_node);
 842                 __put_trace_last(pdi, t);
 843         }
 844
 845         return 0;
 846 }
 847
 848 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
 849                                    struct rb_root *root, int order)
 850 {
 851         struct rb_node *n = root->rb_node;
 852         struct rb_node *prev = NULL;
 853         struct trace *__t;
 854
 855         while (n) {
 856                 __t = rb_entry(n, struct trace, rb_node);
 857                 prev = n;
 858
 859                 if (device < __t->bit->device)
 860                         n = n->rb_left;
 861                 else if (device > __t->bit->device)
 862                         n = n->rb_right;
 863                 else if (sequence < __t->bit->sequence)
 864                         n = n->rb_left;
 865                 else if (sequence > __t->bit->sequence)
 866                         n = n->rb_right;
 867                 else
 868                         return __t;
 869         }
 870
 871         /*
 872          * hack - the list may not be sequence ordered because some
 873          * events don't have sequence and time matched. so we end up
 874          * being a little off in the rb lookup here, because we don't
 875          * know the time we are looking for. compensate by browsing
 876          * a little ahead from the last entry to find the match
 877          */
 878         if (order && prev) {
 879                 int max = 5;
 880
 881                 while (((n = rb_next(prev)) != NULL) && max--) {
 882                         __t = rb_entry(n, struct trace, rb_node);
 883
 884                         if (__t->bit->device == device &&
 885                             __t->bit->sequence == sequence)
 886                                 return __t;
 887
 888                         prev = n;
 889                 }
 890         }
 891
 892         return NULL;
 893 }
 894
 895 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
 896                                                struct per_cpu_info *pci,
 897                                                unsigned long seq)
 898 {
 899         return trace_rb_find(pdi->dev, seq, &pci->rb_last, 0);
 900 }
 901
 902 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
 903 {
 904         struct rb_node **p = &pdi->rb_track.rb_node;
 905         struct rb_node *parent = NULL;
 906         struct io_track *__iot;
 907
 908         while (*p) {
 909                 parent = *p;
 910                 __iot = rb_entry(parent, struct io_track, rb_node);
 911
 912                 if (iot->sector < __iot->sector)
 913                         p = &(*p)->rb_left;
 914                 else if (iot->sector > __iot->sector)
 915                         p = &(*p)->rb_right;
 916                 else {
 917                         fprintf(stderr,
 918                                 "sector alias (%Lu) on device %d,%d!\n",
 919                                 (unsigned long long) iot->sector,
 920                                 MAJOR(pdi->dev), MINOR(pdi->dev));
 921                         return 1;
 922                 }
 923         }
 924
 925         rb_link_node(&iot->rb_node, parent, p);
 926         rb_insert_color(&iot->rb_node, &pdi->rb_track);
 927         return 0;
 928 }
 929
 930 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
 931 {
 932         struct rb_node *n = pdi->rb_track.rb_node;
 933         struct io_track *__iot;
 934
 935         while (n) {
 936                 __iot = rb_entry(n, struct io_track, rb_node);
 937
 938                 if (sector < __iot->sector)
 939                         n = n->rb_left;
 940                 else if (sector > __iot->sector)
 941                         n = n->rb_right;
 942                 else
 943                         return __iot;
 944         }
 945
 946         return NULL;
 947 }
 948
 949 static struct io_track *find_track(struct per_dev_info *pdi, pid_t pid,
 950                                    __u64 sector)
 951 {
 952         struct io_track *iot;
 953
 954         iot = __find_track(pdi, sector);
 955         if (!iot) {
 956                 iot = malloc(sizeof(*iot));
 957                 iot->ppm = find_ppm(pid);
 958                 if (!iot->ppm)
 959                         iot->ppm = add_ppm_hash(pid, "unknown");
 960                 iot->sector = sector;
 961                 track_rb_insert(pdi, iot);
 962         }
 963
 964         return iot;
 965 }
 966
 967 static void log_track_frontmerge(struct per_dev_info *pdi,
 968                                  struct blk_io_trace *t)
 969 {
 970         struct io_track *iot;
 971
 972         if (!track_ios)
 973                 return;
 974
 975         iot = __find_track(pdi, t->sector + t_sec(t));
 976         if (!iot) {
 977                 if (verbose)
 978                         fprintf(stderr, "merge not found for (%d,%d): %llu\n",
 979                                 MAJOR(pdi->dev), MINOR(pdi->dev),
 980                                 (unsigned long long) t->sector + t_sec(t));
 981                 return;
 982         }
 983
 984         rb_erase(&iot->rb_node, &pdi->rb_track);
 985         iot->sector -= t_sec(t);
 986         track_rb_insert(pdi, iot);
 987 }
 988
 989 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
 990 {
 991         struct io_track *iot;
 992
 993         if (!track_ios)
 994                 return;
 995
 996         iot = find_track(pdi, t->pid, t->sector);
 997         iot->allocation_time = t->time;
 998 }
 999
1000 static inline int is_remapper(struct per_dev_info *pdi)
1001 {
1002         int major = MAJOR(pdi->dev);
1003
1004         return (major == 253 || major == 9);
1005 }
1006
1007 /*
1008  * for md/dm setups, the interesting cycle is Q -> C. So track queueing
1009  * time here, as dispatch time
1010  */
1011 static void log_track_queue(struct per_dev_info *pdi, struct blk_io_trace *t)
1012 {
1013         struct io_track *iot;
1014
1015         if (!track_ios)
1016                 return;
1017         if (!is_remapper(pdi))
1018                 return;
1019
1020         iot = find_track(pdi, t->pid, t->sector);
1021         iot->dispatch_time = t->time;
1022 }
1023
1024 /*
1025  * return time between rq allocation and insertion
1026  */
1027 static unsigned long long log_track_insert(struct per_dev_info *pdi,
1028                                            struct blk_io_trace *t)
1029 {
1030         unsigned long long elapsed;
1031         struct io_track *iot;
1032
1033         if (!track_ios)
1034                 return -1;
1035
1036         iot = find_track(pdi, t->pid, t->sector);
1037         iot->queue_time = t->time;
1038
1039         if (!iot->allocation_time)
1040                 return -1;
1041
1042         elapsed = iot->queue_time - iot->allocation_time;
1043
1044         if (per_process_stats) {
1045                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1046                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1047
1048                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
1049                         ppi->longest_allocation_wait[w] = elapsed;
1050         }
1051
1052         return elapsed;
1053 }
1054
1055 /*
1056  * return time between queue and issue
1057  */
1058 static unsigned long long log_track_issue(struct per_dev_info *pdi,
1059                                           struct blk_io_trace *t)
1060 {
1061         unsigned long long elapsed;
1062         struct io_track *iot;
1063
1064         if (!track_ios)
1065                 return -1;
1066         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
1067                 return -1;
1068
1069         iot = __find_track(pdi, t->sector);
1070         if (!iot) {
1071                 if (verbose)
1072                         fprintf(stderr, "issue not found for (%d,%d): %llu\n",
1073                                 MAJOR(pdi->dev), MINOR(pdi->dev),
1074                                 (unsigned long long) t->sector);
1075                 return -1;
1076         }
1077
1078         iot->dispatch_time = t->time;
1079         elapsed = iot->dispatch_time - iot->queue_time;
1080
1081         if (per_process_stats) {
1082                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1083                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1084
1085                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
1086                         ppi->longest_dispatch_wait[w] = elapsed;
1087         }
1088
1089         return elapsed;
1090 }
1091
1092 /*
1093  * return time between dispatch and complete
1094  */
1095 static unsigned long long log_track_complete(struct per_dev_info *pdi,
1096                                              struct blk_io_trace *t)
1097 {
1098         unsigned long long elapsed;
1099         struct io_track *iot;
1100
1101         if (!track_ios)
1102                 return -1;
1103
1104         iot = __find_track(pdi, t->sector);
1105         if (!iot) {
1106                 if (verbose)
1107                         fprintf(stderr,"complete not found for (%d,%d): %llu\n",
1108                                 MAJOR(pdi->dev), MINOR(pdi->dev),
1109                                 (unsigned long long) t->sector);
1110                 return -1;
1111         }
1112
1113         iot->completion_time = t->time;
1114         elapsed = iot->completion_time - iot->dispatch_time;
1115
1116         if (per_process_stats) {
1117                 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1118                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1119
1120                 if (ppi && elapsed > ppi->longest_completion_wait[w])
1121                         ppi->longest_completion_wait[w] = elapsed;
1122         }
1123
1124         /*
1125          * kill the trace, we don't need it after completion
1126          */
1127         rb_erase(&iot->rb_node, &pdi->rb_track);
1128         free(iot);
1129
1130         return elapsed;
1131 }
1132
1133
1134 static struct io_stats *find_process_io_stats(pid_t pid)
1135 {
1136         struct per_process_info *ppi = find_ppi(pid);
1137
1138         if (!ppi) {
1139                 ppi = malloc(sizeof(*ppi));
1140                 memset(ppi, 0, sizeof(*ppi));
1141                 ppi->ppm = find_ppm(pid);
1142                 if (!ppi->ppm)
1143                         ppi->ppm = add_ppm_hash(pid, "unknown");
1144                 add_ppi_to_hash(ppi);
1145                 add_ppi_to_list(ppi);
1146         }
1147
1148         return &ppi->io_stats;
1149 }
1150
1151 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
1152 {
1153         if (pdi->name)
1154                 snprintf(buffer, size, "%s", pdi->name);
1155         else
1156                 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
1157         return buffer;
1158 }
1159
1160 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
1161 {
1162         unsigned long long this = bit->time;
1163         unsigned long long last = pdi->last_reported_time;
1164
1165         pdi->backwards = (this < last) ? 'B' : ' ';
1166         pdi->last_reported_time = this;
1167 }
1168
1169 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
1170                                int rw)
1171 {
1172         if (rw) {
1173                 ios->mwrites++;
1174                 ios->mwrite_kb += t_kb(t);
1175         } else {
1176                 ios->mreads++;
1177                 ios->mread_kb += t_kb(t);
1178         }
1179 }
1180
1181 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
1182                              int rw)
1183 {
1184         __account_m(&pci->io_stats, t, rw);
1185
1186         if (per_process_stats) {
1187                 struct io_stats *ios = find_process_io_stats(t->pid);
1188
1189                 __account_m(ios, t, rw);
1190         }
1191 }
1192
1193 static inline void __account_pc_queue(struct io_stats *ios,
1194                                       struct blk_io_trace *t, int rw)
1195 {
1196         if (rw) {
1197                 ios->qwrites_pc++;
1198                 ios->qwrite_kb_pc += t_kb(t);
1199         } else {
1200                 ios->qreads_pc++;
1201                 ios->qread_kb += t_kb(t);
1202         }
1203 }
1204
1205 static inline void account_pc_queue(struct blk_io_trace *t,
1206                                     struct per_cpu_info *pci, int rw)
1207 {
1208         __account_pc_queue(&pci->io_stats, t, rw);
1209
1210         if (per_process_stats) {
1211                 struct io_stats *ios = find_process_io_stats(t->pid);
1212
1213                 __account_pc_queue(ios, t, rw);
1214         }
1215 }
1216
1217 static inline void __account_pc_issue(struct io_stats *ios, int rw,
1218                                       unsigned int bytes)
1219 {
1220         if (rw) {
1221                 ios->iwrites_pc++;
1222                 ios->iwrite_kb_pc += bytes >> 10;
1223         } else {
1224                 ios->ireads_pc++;
1225                 ios->iread_kb_pc += bytes >> 10;
1226         }
1227 }
1228
1229 static inline void account_pc_issue(struct blk_io_trace *t,
1230                                     struct per_cpu_info *pci, int rw)
1231 {
1232         __account_pc_issue(&pci->io_stats, rw, t->bytes);
1233
1234         if (per_process_stats) {
1235                 struct io_stats *ios = find_process_io_stats(t->pid);
1236
1237                 __account_pc_issue(ios, rw, t->bytes);
1238         }
1239 }
1240
1241 static inline void __account_pc_requeue(struct io_stats *ios,
1242                                         struct blk_io_trace *t, int rw)
1243 {
1244         if (rw) {
1245                 ios->wrqueue_pc++;
1246                 ios->iwrite_kb_pc -= t_kb(t);
1247         } else {
1248                 ios->rrqueue_pc++;
1249                 ios->iread_kb_pc -= t_kb(t);
1250         }
1251 }
1252
1253 static inline void account_pc_requeue(struct blk_io_trace *t,
1254                                       struct per_cpu_info *pci, int rw)
1255 {
1256         __account_pc_requeue(&pci->io_stats, t, rw);
1257
1258         if (per_process_stats) {
1259                 struct io_stats *ios = find_process_io_stats(t->pid);
1260
1261                 __account_pc_requeue(ios, t, rw);
1262         }
1263 }
1264
1265 static inline void __account_pc_c(struct io_stats *ios, int rw)
1266 {
1267         if (rw)
1268                 ios->cwrites_pc++;
1269         else
1270                 ios->creads_pc++;
1271 }
1272
1273 static inline void account_pc_c(struct blk_io_trace *t,
1274                                 struct per_cpu_info *pci, int rw)
1275 {
1276         __account_pc_c(&pci->io_stats, rw);
1277
1278         if (per_process_stats) {
1279                 struct io_stats *ios = find_process_io_stats(t->pid);
1280
1281                 __account_pc_c(ios, rw);
1282         }
1283 }
1284
1285 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
1286                                    int rw)
1287 {
1288         if (rw) {
1289                 ios->qwrites++;
1290                 ios->qwrite_kb += t_kb(t);
1291         } else {
1292                 ios->qreads++;
1293                 ios->qread_kb += t_kb(t);
1294         }
1295 }
1296
1297 static inline void account_queue(struct blk_io_trace *t,
1298                                  struct per_cpu_info *pci, int rw)
1299 {
1300         __account_queue(&pci->io_stats, t, rw);
1301
1302         if (per_process_stats) {
1303                 struct io_stats *ios = find_process_io_stats(t->pid);
1304
1305                 __account_queue(ios, t, rw);
1306         }
1307 }
1308
1309 static inline void __account_c(struct io_stats *ios, int rw, int bytes)
1310 {
1311         if (rw) {
1312                 ios->cwrites++;
1313                 ios->cwrite_kb += bytes >> 10;
1314         } else {
1315                 ios->creads++;
1316                 ios->cread_kb += bytes >> 10;
1317         }
1318 }
1319
1320 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
1321                              int rw, int bytes)
1322 {
1323         __account_c(&pci->io_stats, rw, bytes);
1324
1325         if (per_process_stats) {
1326                 struct io_stats *ios = find_process_io_stats(t->pid);
1327
1328                 __account_c(ios, rw, bytes);
1329         }
1330 }
1331
1332 static inline void __account_issue(struct io_stats *ios, int rw,
1333                                    unsigned int bytes)
1334 {
1335         if (rw) {
1336                 ios->iwrites++;
1337                 ios->iwrite_kb += bytes >> 10;
1338         } else {
1339                 ios->ireads++;
1340                 ios->iread_kb += bytes >> 10;
1341         }
1342 }
1343
1344 static inline void account_issue(struct blk_io_trace *t,
1345                                  struct per_cpu_info *pci, int rw)
1346 {
1347         __account_issue(&pci->io_stats, rw, t->bytes);
1348
1349         if (per_process_stats) {
1350                 struct io_stats *ios = find_process_io_stats(t->pid);
1351
1352                 __account_issue(ios, rw, t->bytes);
1353         }
1354 }
1355
1356 static inline void __account_unplug(struct io_stats *ios, int timer)
1357 {
1358         if (timer)
1359                 ios->timer_unplugs++;
1360         else
1361                 ios->io_unplugs++;
1362 }
1363
1364 static inline void account_unplug(struct blk_io_trace *t,
1365                                   struct per_cpu_info *pci, int timer)
1366 {
1367         __account_unplug(&pci->io_stats, timer);
1368
1369         if (per_process_stats) {
1370                 struct io_stats *ios = find_process_io_stats(t->pid);
1371
1372                 __account_unplug(ios, timer);
1373         }
1374 }
1375
1376 static inline void __account_requeue(struct io_stats *ios,
1377                                      struct blk_io_trace *t, int rw)
1378 {
1379         if (rw) {
1380                 ios->wrqueue++;
1381                 ios->iwrite_kb -= t_kb(t);
1382         } else {
1383                 ios->rrqueue++;
1384                 ios->iread_kb -= t_kb(t);
1385         }
1386 }
1387
1388 static inline void account_requeue(struct blk_io_trace *t,
1389                                    struct per_cpu_info *pci, int rw)
1390 {
1391         __account_requeue(&pci->io_stats, t, rw);
1392
1393         if (per_process_stats) {
1394                 struct io_stats *ios = find_process_io_stats(t->pid);
1395
1396                 __account_requeue(ios, t, rw);
1397         }
1398 }
1399
1400 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
1401                          struct blk_io_trace *t, char *act)
1402 {
1403         process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
1404 }
1405
1406 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
1407                        struct blk_io_trace *t, char *act)
1408 {
1409         process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
1410 }
1411
1412 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
1413                       char *act)
1414 {
1415         process_fmt(act, pci, t, -1, 0, NULL);
1416 }
1417
1418 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
1419                       struct blk_io_trace *t, char *act)
1420 {
1421         process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
1422 }
1423
1424 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
1425                       struct blk_io_trace *t, char *act)
1426 {
1427         if (act[0] == 'F')
1428                 log_track_frontmerge(pdi, t);
1429
1430         process_fmt(act, pci, t, -1ULL, 0, NULL);
1431 }
1432
1433 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
1434                         char *act)
1435 {
1436         process_fmt(act, pci, t, -1ULL, 0, NULL);
1437 }
1438
1439 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
1440                         char *act)
1441 {
1442         process_fmt(act, pci, t, -1ULL, 0, NULL);
1443 }
1444
1445 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
1446                       char *act)
1447 {
1448         process_fmt(act, pci, t, -1ULL, 0, NULL);
1449 }
1450
1451 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
1452                       char *act)
1453 {
1454         process_fmt(act, pci, t, -1ULL, 0, NULL);
1455 }
1456
1457 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1458 {
1459         unsigned char *buf = (unsigned char *) t + sizeof(*t);
1460
1461         process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1462 }
1463
1464 static void dump_trace_pc(struct blk_io_trace *t, struct per_dev_info *pdi,
1465                           struct per_cpu_info *pci)
1466 {
1467         int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1468         int act = t->action & 0xffff;
1469
1470         switch (act) {
1471                 case __BLK_TA_QUEUE:
1472                         log_generic(pci, t, "Q");
1473                         account_pc_queue(t, pci, w);
1474                         break;
1475                 case __BLK_TA_GETRQ:
1476                         log_generic(pci, t, "G");
1477                         break;
1478                 case __BLK_TA_SLEEPRQ:
1479                         log_generic(pci, t, "S");
1480                         break;
1481                 case __BLK_TA_REQUEUE:
1482                         /*
1483                          * can happen if we miss traces, don't let it go
1484                          * below zero
1485                          */
1486                         if (pdi->cur_depth[w])
1487                                 pdi->cur_depth[w]--;
1488                         account_pc_requeue(t, pci, w);
1489                         log_generic(pci, t, "R");
1490                         break;
1491                 case __BLK_TA_ISSUE:
1492                         account_pc_issue(t, pci, w);
1493                         pdi->cur_depth[w]++;
1494                         if (pdi->cur_depth[w] > pdi->max_depth[w])
1495                                 pdi->max_depth[w] = pdi->cur_depth[w];
1496                         log_pc(pci, t, "D");
1497                         break;
1498                 case __BLK_TA_COMPLETE:
1499                         if (pdi->cur_depth[w])
1500                                 pdi->cur_depth[w]--;
1501                         log_pc(pci, t, "C");
1502                         account_pc_c(t, pci, w);
1503                         break;
1504                 case __BLK_TA_INSERT:
1505                         log_pc(pci, t, "I");
1506                         break;
1507                 default:
1508                         fprintf(stderr, "Bad pc action %x\n", act);
1509                         break;
1510         }
1511 }
1512
1513 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1514                           struct per_cpu_info *pci)
1515 {
1516         int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1517         int act = t->action & 0xffff;
1518
1519         switch (act) {
1520                 case __BLK_TA_QUEUE:
1521                         log_track_queue(pdi, t);
1522                         account_queue(t, pci, w);
1523                         log_queue(pci, t, "Q");
1524                         break;
1525                 case __BLK_TA_INSERT:
1526                         log_insert(pdi, pci, t, "I");
1527                         break;
1528                 case __BLK_TA_BACKMERGE:
1529                         account_m(t, pci, w);
1530                         log_merge(pdi, pci, t, "M");
1531                         break;
1532                 case __BLK_TA_FRONTMERGE:
1533                         account_m(t, pci, w);
1534                         log_merge(pdi, pci, t, "F");
1535                         break;
1536                 case __BLK_TA_GETRQ:
1537                         log_track_getrq(pdi, t);
1538                         log_generic(pci, t, "G");
1539                         break;
1540                 case __BLK_TA_SLEEPRQ:
1541                         log_generic(pci, t, "S");
1542                         break;
1543                 case __BLK_TA_REQUEUE:
1544                         /*
1545                          * can happen if we miss traces, don't let it go
1546                          * below zero
1547                          */
1548                         if (pdi->cur_depth[w])
1549                                 pdi->cur_depth[w]--;
1550                         account_requeue(t, pci, w);
1551                         log_queue(pci, t, "R");
1552                         break;
1553                 case __BLK_TA_ISSUE:
1554                         account_issue(t, pci, w);
1555                         pdi->cur_depth[w]++;
1556                         if (pdi->cur_depth[w] > pdi->max_depth[w])
1557                                 pdi->max_depth[w] = pdi->cur_depth[w];
1558                         log_issue(pdi, pci, t, "D");
1559                         break;
1560                 case __BLK_TA_COMPLETE:
1561                         if (pdi->cur_depth[w])
1562                                 pdi->cur_depth[w]--;
1563                         account_c(t, pci, w, t->bytes);
1564                         log_complete(pdi, pci, t, "C");
1565                         break;
1566                 case __BLK_TA_PLUG:
1567                         log_action(pci, t, "P");
1568                         break;
1569                 case __BLK_TA_UNPLUG_IO:
1570                         account_unplug(t, pci, 0);
1571                         log_unplug(pci, t, "U");
1572                         break;
1573                 case __BLK_TA_UNPLUG_TIMER:
1574                         account_unplug(t, pci, 1);
1575                         log_unplug(pci, t, "UT");
1576                         break;
1577                 case __BLK_TA_SPLIT:
1578                         log_split(pci, t, "X");
1579                         break;
1580                 case __BLK_TA_BOUNCE:
1581                         log_generic(pci, t, "B");
1582                         break;
1583                 case __BLK_TA_REMAP:
1584                         log_generic(pci, t, "A");
1585                         break;
1586                 default:
1587                         fprintf(stderr, "Bad fs action %x\n", t->action);
1588                         break;
1589         }
1590 }
1591
1592 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1593                        struct per_dev_info *pdi)
1594 {
1595         if (text_output) {
1596                 if (t->action == BLK_TN_MESSAGE)
1597                         handle_notify(t);
1598                 else if (t->action & BLK_TC_ACT(BLK_TC_PC))
1599                         dump_trace_pc(t, pdi, pci);
1600                 else
1601                         dump_trace_fs(t, pdi, pci);
1602         }
1603
1604         if (!pdi->events)
1605                 pdi->first_reported_time = t->time;
1606
1607         pdi->events++;
1608
1609         output_binary(t, sizeof(*t) + t->pdu_len);
1610 }
1611
1612 /*
1613  * print in a proper way, not too small and not too big. if more than
1614  * 1000,000K, turn into M and so on
1615  */
1616 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1617 {
1618         char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1619         unsigned int i = 0;
1620
1621         if (in_kb)
1622                 i++;
1623
1624         while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1625                 i++;
1626                 num /= 1000;
1627         }
1628
1629         sprintf(dst, "%'8Lu%c", num, suff[i]);
1630         return dst;
1631 }
1632
1633 static void dump_io_stats(struct per_dev_info *pdi, struct io_stats *ios,
1634                           char *msg)
1635 {
1636         static char x[256], y[256];
1637
1638         fprintf(ofp, "%s\n", msg);
1639
1640         fprintf(ofp, " Reads Queued:    %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1641         fprintf(ofp, " Writes Queued:    %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1642         fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1643         fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1644         fprintf(ofp, " Reads Requeued:  %s\t\t", size_cnv(x, ios->rrqueue, 0));
1645         fprintf(ofp, " Writes Requeued:  %s\n", size_cnv(x, ios->wrqueue, 0));
1646         fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1647         fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1648         fprintf(ofp, " Read Merges:     %s, %siB\t", size_cnv(x, ios->mreads, 0), size_cnv(y, ios->mread_kb, 1));
1649         fprintf(ofp, " Write Merges:     %s, %siB\n", size_cnv(x, ios->mwrites, 0), size_cnv(y, ios->mwrite_kb, 1));
1650         if (pdi) {
1651                 fprintf(ofp, " Read depth:      %'8u%8c\t", pdi->max_depth[0], ' ');
1652                 fprintf(ofp, " Write depth:      %'8u\n", pdi->max_depth[1]);
1653         }
1654         if (ios->qreads_pc || ios->qwrites_pc || ios->ireads_pc || ios->iwrites_pc ||
1655             ios->rrqueue_pc || ios->wrqueue_pc || ios->creads_pc || ios->cwrites_pc) {
1656                 fprintf(ofp, " PC Reads Queued: %s, %siB\t", size_cnv(x, ios->qreads_pc, 0), size_cnv(y, ios->qread_kb_pc, 1));
1657                 fprintf(ofp, " PC Writes Queued: %s, %siB\n", size_cnv(x, ios->qwrites_pc, 0), size_cnv(y, ios->qwrite_kb_pc, 1));
1658                 fprintf(ofp, " PC Read Disp.:   %s, %siB\t", size_cnv(x, ios->ireads_pc, 0), size_cnv(y, ios->iread_kb_pc, 1));
1659                 fprintf(ofp, " PC Write Disp.:   %s, %siB\n", size_cnv(x, ios->iwrites_pc, 0), size_cnv(y, ios->iwrite_kb_pc, 1));
1660                 fprintf(ofp, " PC Reads Req.:   %s\t\t", size_cnv(x, ios->rrqueue_pc, 0));
1661                 fprintf(ofp, " PC Writes Req.:   %s\n", size_cnv(x, ios->wrqueue_pc, 0));
1662                 fprintf(ofp, " PC Reads Compl.: %s\t\t", size_cnv(x, ios->creads_pc, 0));
1663                 fprintf(ofp, " PC Writes Compl.: %s\n", size_cnv(x, ios->cwrites, 0));
1664         }
1665         fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
1666         fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
1667 }
1668
1669 static void dump_wait_stats(struct per_process_info *ppi)
1670 {
1671         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1672         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1673         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1674         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1675         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1676         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1677
1678         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1679         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
1680         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
1681         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
1682         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1683         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
1684 }
1685
1686 static int ppi_name_compare(const void *p1, const void *p2)
1687 {
1688         struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1689         struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1690         int res;
1691
1692         res = strverscmp(ppi1->ppm->comm, ppi2->ppm->comm);
1693         if (!res)
1694                 res = ppi1->ppm->pid > ppi2->ppm->pid;
1695
1696         return res;
1697 }
1698
1699 static void sort_process_list(void)
1700 {
1701         struct per_process_info **ppis;
1702         struct per_process_info *ppi;
1703         int i = 0;
1704
1705         ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1706
1707         ppi = ppi_list;
1708         while (ppi) {
1709                 ppis[i++] = ppi;
1710                 ppi = ppi->list_next;
1711         }
1712
1713         qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1714
1715         i = ppi_list_entries - 1;
1716         ppi_list = NULL;
1717         while (i >= 0) {
1718                 ppi = ppis[i];
1719
1720                 ppi->list_next = ppi_list;
1721                 ppi_list = ppi;
1722                 i--;
1723         }
1724
1725         free(ppis);
1726 }
1727
1728 static void show_process_stats(void)
1729 {
1730         struct per_process_info *ppi;
1731
1732         sort_process_list();
1733
1734         ppi = ppi_list;
1735         while (ppi) {
1736                 struct process_pid_map *ppm = ppi->ppm;
1737                 char name[64];
1738
1739                 if (ppi->more_than_one)
1740                         sprintf(name, "%s (%u, ...)", ppm->comm, ppm->pid);
1741                 else
1742                         sprintf(name, "%s (%u)", ppm->comm, ppm->pid);
1743
1744                 dump_io_stats(NULL, &ppi->io_stats, name);
1745                 dump_wait_stats(ppi);
1746                 ppi = ppi->list_next;
1747         }
1748
1749         fprintf(ofp, "\n");
1750 }
1751
1752 static void show_device_and_cpu_stats(void)
1753 {
1754         struct per_dev_info *pdi;
1755         struct per_cpu_info *pci;
1756         struct io_stats total, *ios;
1757         unsigned long long rrate, wrate, msec;
1758         int i, j, pci_events;
1759         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1760         char name[32];
1761         double ratio;
1762
1763         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1764
1765                 memset(&total, 0, sizeof(total));
1766                 pci_events = 0;
1767
1768                 if (i > 0)
1769                         fprintf(ofp, "\n");
1770
1771                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1772                         if (!pci->nelems)
1773                                 continue;
1774
1775                         ios = &pci->io_stats;
1776                         total.qreads += ios->qreads;
1777                         total.qwrites += ios->qwrites;
1778                         total.creads += ios->creads;
1779                         total.cwrites += ios->cwrites;
1780                         total.mreads += ios->mreads;
1781                         total.mwrites += ios->mwrites;
1782                         total.ireads += ios->ireads;
1783                         total.iwrites += ios->iwrites;
1784                         total.rrqueue += ios->rrqueue;
1785                         total.wrqueue += ios->wrqueue;
1786                         total.qread_kb += ios->qread_kb;
1787                         total.qwrite_kb += ios->qwrite_kb;
1788                         total.cread_kb += ios->cread_kb;
1789                         total.cwrite_kb += ios->cwrite_kb;
1790                         total.iread_kb += ios->iread_kb;
1791                         total.iwrite_kb += ios->iwrite_kb;
1792                         total.mread_kb += ios->mread_kb;
1793                         total.mwrite_kb += ios->mwrite_kb;
1794
1795                         total.qreads_pc += ios->qreads_pc;
1796                         total.qwrites_pc += ios->qwrites_pc;
1797                         total.creads_pc += ios->creads_pc;
1798                         total.cwrites_pc += ios->cwrites_pc;
1799                         total.ireads_pc += ios->ireads_pc;
1800                         total.iwrites_pc += ios->iwrites_pc;
1801                         total.rrqueue_pc += ios->rrqueue_pc;
1802                         total.wrqueue_pc += ios->wrqueue_pc;
1803                         total.qread_kb_pc += ios->qread_kb_pc;
1804                         total.qwrite_kb_pc += ios->qwrite_kb_pc;
1805                         total.iread_kb_pc += ios->iread_kb_pc;
1806                         total.iwrite_kb_pc += ios->iwrite_kb_pc;
1807
1808                         total.timer_unplugs += ios->timer_unplugs;
1809                         total.io_unplugs += ios->io_unplugs;
1810
1811                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1812                                  j, get_dev_name(pdi, name, sizeof(name)));
1813                         dump_io_stats(pdi, ios, line);
1814                         pci_events++;
1815                 }
1816
1817                 if (pci_events > 1) {
1818                         fprintf(ofp, "\n");
1819                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1820                                  get_dev_name(pdi, name, sizeof(name)));
1821                         dump_io_stats(NULL, &total, line);
1822                 }
1823
1824                 wrate = rrate = 0;
1825                 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1826                 if (msec) {
1827                         rrate = 1000 * total.cread_kb / msec;
1828                         wrate = 1000 * total.cwrite_kb / msec;
1829                 }
1830
1831                 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n",
1832                         rrate, wrate);
1833                 fprintf(ofp, "Events (%s): %'Lu entries\n",
1834                         get_dev_name(pdi, line, sizeof(line)), pdi->events);
1835
1836                 collect_pdi_skips(pdi);
1837                 if (!pdi->skips && !pdi->events)
1838                         ratio = 0.0;
1839                 else
1840                         ratio = 100.0 * ((double)pdi->seq_skips /
1841                                         (double)(pdi->events + pdi->seq_skips));
1842                 fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%)\n",
1843                         pdi->skips, pdi->seq_skips, ratio);
1844         }
1845 }
1846
1847 static void find_genesis(void)
1848 {
1849         struct trace *t = trace_list;
1850
1851         genesis_time = -1ULL;
1852         while (t != NULL) {
1853                 if (t->bit->time < genesis_time)
1854                         genesis_time = t->bit->time;
1855
1856                 t = t->next;
1857         }
1858
1859         /* The time stamp record will usually be the first
1860          * record in the trace, but not always.
1861          */
1862         if (start_timestamp
1863          && start_timestamp != genesis_time) {
1864                 long delta = genesis_time - start_timestamp;
1865
1866                 abs_start_time.tv_sec  += SECONDS(delta);
1867                 abs_start_time.tv_nsec += NANO_SECONDS(delta);
1868                 if (abs_start_time.tv_nsec < 0) {
1869                         abs_start_time.tv_nsec += 1000000000;
1870                         abs_start_time.tv_sec -= 1;
1871                 } else
1872                 if (abs_start_time.tv_nsec > 1000000000) {
1873                         abs_start_time.tv_nsec -= 1000000000;
1874                         abs_start_time.tv_sec += 1;
1875                 }
1876         }
1877 }
1878
1879 static inline int check_stopwatch(struct blk_io_trace *bit)
1880 {
1881         if (bit->time < stopwatch_end &&
1882             bit->time >= stopwatch_start)
1883                 return 0;
1884
1885         return 1;
1886 }
1887
1888 /*
1889  * return youngest entry read
1890  */
1891 static int sort_entries(unsigned long long *youngest)
1892 {
1893         struct per_dev_info *pdi = NULL;
1894         struct per_cpu_info *pci = NULL;
1895         struct trace *t;
1896
1897         if (!genesis_time)
1898                 find_genesis();
1899
1900         *youngest = 0;
1901         while ((t = trace_list) != NULL) {
1902                 struct blk_io_trace *bit = t->bit;
1903
1904                 trace_list = t->next;
1905
1906                 bit->time -= genesis_time;
1907
1908                 if (bit->time < *youngest || !*youngest)
1909                         *youngest = bit->time;
1910
1911                 if (!pdi || pdi->dev != bit->device) {
1912                         pdi = get_dev_info(bit->device);
1913                         pci = NULL;
1914                 }
1915
1916                 if (!pci || pci->cpu != bit->cpu)
1917                         pci = get_cpu_info(pdi, bit->cpu);
1918
1919                 if (bit->sequence < pci->smallest_seq_read)
1920                         pci->smallest_seq_read = bit->sequence;
1921
1922                 if (check_stopwatch(bit)) {
1923                         bit_free(bit);
1924                         t_free(t);
1925                         continue;
1926                 }
1927
1928                 if (trace_rb_insert_sort(t))
1929                         return -1;
1930         }
1931
1932         return 0;
1933 }
1934
1935 /*
1936  * to continue, we must have traces from all online cpus in the tree
1937  */
1938 static int check_cpu_map(struct per_dev_info *pdi)
1939 {
1940         unsigned long *cpu_map;
1941         struct rb_node *n;
1942         struct trace *__t;
1943         unsigned int i;
1944         int ret, cpu;
1945
1946         /*
1947          * create a map of the cpus we have traces for
1948          */
1949         cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1950         n = rb_first(&rb_sort_root);
1951         while (n) {
1952                 __t = rb_entry(n, struct trace, rb_node);
1953                 cpu = __t->bit->cpu;
1954
1955                 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1956                 n = rb_next(n);
1957         }
1958
1959         /*
1960          * we can't continue if pdi->cpu_map has entries set that we don't
1961          * have in the sort rbtree. the opposite is not a problem, though
1962          */
1963         ret = 0;
1964         for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1965                 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1966                         ret = 1;
1967                         break;
1968                 }
1969         }
1970
1971         free(cpu_map);
1972         return ret;
1973 }
1974
1975 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1976 {
1977         struct blk_io_trace *bit = t->bit;
1978         unsigned long expected_sequence;
1979         struct per_cpu_info *pci;
1980         struct trace *__t;
1981
1982         pci = get_cpu_info(pdi, bit->cpu);
1983         expected_sequence = pci->last_sequence + 1;
1984
1985         if (!expected_sequence) {
1986                 /*
1987                  * 1 should be the first entry, just allow it
1988                  */
1989                 if (bit->sequence == 1)
1990                         return 0;
1991                 if (bit->sequence == pci->smallest_seq_read)
1992                         return 0;
1993
1994                 return check_cpu_map(pdi);
1995         }
1996
1997         if (bit->sequence == expected_sequence)
1998                 return 0;
1999
2000         /*
2001          * we may not have seen that sequence yet. if we are not doing
2002          * the final run, break and wait for more entries.
2003          */
2004         if (expected_sequence < pci->smallest_seq_read) {
2005                 __t = trace_rb_find_last(pdi, pci, expected_sequence);
2006                 if (!__t)
2007                         goto skip;
2008
2009                 __put_trace_last(pdi, __t);
2010                 return 0;
2011         } else if (!force) {
2012                 return 1;
2013         } else {
2014 skip:
2015                 if (check_current_skips(pci, bit->sequence))
2016                         return 0;
2017
2018                 if (expected_sequence < bit->sequence)
2019                         insert_skip(pci, expected_sequence, bit->sequence - 1);
2020                 return 0;
2021         }
2022 }
2023
2024 static void show_entries_rb(int force)
2025 {
2026         struct per_dev_info *pdi = NULL;
2027         struct per_cpu_info *pci = NULL;
2028         struct blk_io_trace *bit;
2029         struct rb_node *n;
2030         struct trace *t;
2031
2032         while ((n = rb_first(&rb_sort_root)) != NULL) {
2033                 if (is_done() && !force && !pipeline)
2034                         break;
2035
2036                 t = rb_entry(n, struct trace, rb_node);
2037                 bit = t->bit;
2038
2039                 if (read_sequence - t->read_sequence < 1 && !force)
2040                         break;
2041
2042                 if (!pdi || pdi->dev != bit->device) {
2043                         pdi = get_dev_info(bit->device);
2044                         pci = NULL;
2045                 }
2046
2047                 if (!pdi) {
2048                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
2049                                 MAJOR(bit->device), MINOR(bit->device));
2050                         break;
2051                 }
2052
2053                 if (check_sequence(pdi, t, force))
2054                         break;
2055
2056                 if (!force && bit->time > last_allowed_time)
2057                         break;
2058
2059                 check_time(pdi, bit);
2060
2061                 if (!pci || pci->cpu != bit->cpu)
2062                         pci = get_cpu_info(pdi, bit->cpu);
2063
2064                 pci->last_sequence = bit->sequence;
2065
2066                 pci->nelems++;
2067
2068                 if (bit->action & (act_mask << BLK_TC_SHIFT))
2069                         dump_trace(bit, pci, pdi);
2070
2071                 put_trace(pdi, t);
2072         }
2073 }
2074
2075 static int read_data(int fd, void *buffer, int bytes, int block, int *fdblock)
2076 {
2077         int ret, bytes_left, fl;
2078         void *p;
2079
2080         if (block != *fdblock) {
2081                 fl = fcntl(fd, F_GETFL);
2082
2083                 if (!block) {
2084                         *fdblock = 0;
2085                         fcntl(fd, F_SETFL, fl | O_NONBLOCK);
2086                 } else {
2087                         *fdblock = 1;
2088                         fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
2089                 }
2090         }
2091
2092         bytes_left = bytes;
2093         p = buffer;
2094         while (bytes_left > 0) {
2095                 ret = read(fd, p, bytes_left);
2096                 if (!ret)
2097                         return 1;
2098                 else if (ret < 0) {
2099                         if (errno != EAGAIN) {
2100                                 perror("read");
2101                                 return -1;
2102                         }
2103
2104                         /*
2105                          * never do partial reads. we can return if we
2106                          * didn't read anything and we should not block,
2107                          * otherwise wait for data
2108                          */
2109                         if ((bytes_left == bytes) && !block)
2110                                 return 1;
2111
2112                         usleep(10);
2113                         continue;
2114                 } else {
2115                         p += ret;
2116                         bytes_left -= ret;
2117                 }
2118         }
2119
2120         return 0;
2121 }
2122
2123 static inline __u16 get_pdulen(struct blk_io_trace *bit)
2124 {
2125         if (data_is_native)
2126                 return bit->pdu_len;
2127
2128         return __bswap_16(bit->pdu_len);
2129 }
2130
2131 static inline __u32 get_magic(struct blk_io_trace *bit)
2132 {
2133         if (data_is_native)
2134                 return bit->magic;
2135
2136         return __bswap_32(bit->magic);
2137 }
2138
2139 static int read_events(int fd, int always_block, int *fdblock)
2140 {
2141         struct per_dev_info *pdi = NULL;
2142         unsigned int events = 0;
2143
2144         while (!is_done() && events < rb_batch) {
2145                 struct blk_io_trace *bit;
2146                 struct trace *t;
2147                 int pdu_len, should_block, ret;
2148                 __u32 magic;
2149
2150                 bit = bit_alloc();
2151
2152                 should_block = !events || always_block;
2153
2154                 ret = read_data(fd, bit, sizeof(*bit), should_block, fdblock);
2155                 if (ret) {
2156                         bit_free(bit);
2157                         if (!events && ret < 0)
2158                                 events = ret;
2159                         break;
2160                 }
2161
2162                 /*
2163                  * look at first trace to check whether we need to convert
2164                  * data in the future
2165                  */
2166                 if (data_is_native == -1 && check_data_endianness(bit->magic))
2167                         break;
2168
2169                 magic = get_magic(bit);
2170                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2171                         fprintf(stderr, "Bad magic %x\n", magic);
2172                         break;
2173                 }
2174
2175                 pdu_len = get_pdulen(bit);
2176                 if (pdu_len) {
2177                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2178
2179                         if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1, fdblock)) {
2180                                 bit_free(ptr);
2181                                 break;
2182                         }
2183
2184                         bit = ptr;
2185                 }
2186
2187                 trace_to_cpu(bit);
2188
2189                 if (verify_trace(bit)) {
2190                         bit_free(bit);
2191                         continue;
2192                 }
2193
2194                 /*
2195                  * not a real trace, so grab and handle it here
2196                  */
2197                 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY) && bit->action != BLK_TN_MESSAGE) {
2198                         handle_notify(bit);
2199                         output_binary(bit, sizeof(*bit) + bit->pdu_len);
2200                         continue;
2201                 }
2202
2203                 t = t_alloc();
2204                 memset(t, 0, sizeof(*t));
2205                 t->bit = bit;
2206                 t->read_sequence = read_sequence;
2207
2208                 t->next = trace_list;
2209                 trace_list = t;
2210
2211                 if (!pdi || pdi->dev != bit->device)
2212                         pdi = get_dev_info(bit->device);
2213
2214                 if (bit->time > pdi->last_read_time)
2215                         pdi->last_read_time = bit->time;
2216
2217                 events++;
2218         }
2219
2220         return events;
2221 }
2222
2223 /*
2224  * Managing input streams
2225  */
2226
2227 struct ms_stream {
2228         struct ms_stream *next;
2229         struct trace *first, *last;
2230         struct per_dev_info *pdi;
2231         unsigned int cpu;
2232 };
2233
2234 #define MS_HASH(d, c) ((MAJOR(d) & 0xff) ^ (MINOR(d) & 0xff) ^ (cpu & 0xff))
2235
2236 struct ms_stream *ms_head;
2237 struct ms_stream *ms_hash[256];
2238
2239 static void ms_sort(struct ms_stream *msp);
2240 static int ms_prime(struct ms_stream *msp);
2241
2242 static inline struct trace *ms_peek(struct ms_stream *msp)
2243 {
2244         return (msp == NULL) ? NULL : msp->first;
2245 }
2246
2247 static inline __u64 ms_peek_time(struct ms_stream *msp)
2248 {
2249         return ms_peek(msp)->bit->time;
2250 }
2251
2252 static inline void ms_resort(struct ms_stream *msp)
2253 {
2254         if (msp->next && ms_peek_time(msp) > ms_peek_time(msp->next)) {
2255                 ms_head = msp->next;
2256                 msp->next = NULL;
2257                 ms_sort(msp);
2258         }
2259 }
2260
2261 static inline void ms_deq(struct ms_stream *msp)
2262 {
2263         msp->first = msp->first->next;
2264         if (!msp->first) {
2265                 msp->last = NULL;
2266                 if (!ms_prime(msp)) {
2267                         ms_head = msp->next;
2268                         msp->next = NULL;
2269                         return;
2270                 }
2271         }
2272
2273         ms_resort(msp);
2274 }
2275
2276 static void ms_sort(struct ms_stream *msp)
2277 {
2278         __u64 msp_t = ms_peek_time(msp);
2279         struct ms_stream *this_msp = ms_head;
2280
2281         if (this_msp == NULL)
2282                 ms_head = msp;
2283         else if (msp_t < ms_peek_time(this_msp)) {
2284                 msp->next = this_msp;
2285                 ms_head = msp;
2286         }
2287         else {
2288                 while (this_msp->next && ms_peek_time(this_msp->next) < msp_t)
2289                         this_msp = this_msp->next;
2290
2291                 msp->next = this_msp->next;
2292                 this_msp->next = msp;
2293         }
2294 }
2295
2296 static int ms_prime(struct ms_stream *msp)
2297 {
2298         __u32 magic;
2299         unsigned int i;
2300         struct trace *t;
2301         struct per_dev_info *pdi = msp->pdi;
2302         struct per_cpu_info *pci = get_cpu_info(pdi, msp->cpu);
2303         struct blk_io_trace *bit = NULL;
2304         int ret, pdu_len, ndone = 0;
2305
2306         for (i = 0; !is_done() && pci->fd >= 0 && i < rb_batch; i++) {
2307                 bit = bit_alloc();
2308                 ret = read_data(pci->fd, bit, sizeof(*bit), 1, &pci->fdblock);
2309                 if (ret)
2310                         goto err;
2311
2312                 if (data_is_native == -1 && check_data_endianness(bit->magic))
2313                         goto err;
2314
2315                 magic = get_magic(bit);
2316                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2317                         fprintf(stderr, "Bad magic %x\n", magic);
2318                         goto err;
2319
2320                 }
2321
2322                 pdu_len = get_pdulen(bit);
2323                 if (pdu_len) {
2324                         void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2325                         ret = read_data(pci->fd, ptr + sizeof(*bit), pdu_len,
2326                                                              1, &pci->fdblock);
2327                         if (ret) {
2328                                 free(ptr);
2329                                 bit = NULL;
2330                                 goto err;
2331                         }
2332
2333                         bit = ptr;
2334                 }
2335
2336                 trace_to_cpu(bit);
2337                 if (verify_trace(bit))
2338                         goto err;
2339
2340                 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY) && bit->action != BLK_TN_MESSAGE) {
2341                         handle_notify(bit);
2342                         output_binary(bit, sizeof(*bit) + bit->pdu_len);
2343                         bit_free(bit);
2344
2345                         i -= 1;
2346                         continue;
2347                 }
2348
2349                 if (bit->time > pdi->last_read_time)
2350                         pdi->last_read_time = bit->time;
2351
2352                 t = t_alloc();
2353                 memset(t, 0, sizeof(*t));
2354                 t->bit = bit;
2355
2356                 if (msp->first == NULL)
2357                         msp->first = msp->last = t;
2358                 else {
2359                         msp->last->next = t;
2360                         msp->last = t;
2361                 }
2362
2363                 ndone++;
2364         }
2365
2366         return ndone;
2367
2368 err:
2369         if (bit) bit_free(bit);
2370
2371         cpu_mark_offline(pdi, pci->cpu);
2372         close(pci->fd);
2373         pci->fd = -1;
2374
2375         return ndone;
2376 }
2377
2378 static struct ms_stream *ms_alloc(struct per_dev_info *pdi, int cpu)
2379 {
2380         struct ms_stream *msp = malloc(sizeof(*msp));
2381
2382         msp->next = NULL;
2383         msp->first = msp->last = NULL;
2384         msp->pdi = pdi;
2385         msp->cpu = cpu;
2386
2387         if (ms_prime(msp))
2388                 ms_sort(msp);
2389
2390         return msp;
2391 }
2392
2393 static int setup_file(struct per_dev_info *pdi, int cpu)
2394 {
2395         int len = 0;
2396         struct stat st;
2397         char *p, *dname;
2398         struct per_cpu_info *pci = get_cpu_info(pdi, cpu);
2399
2400         pci->cpu = cpu;
2401         pci->fdblock = -1;
2402
2403         p = strdup(pdi->name);
2404         dname = dirname(p);
2405         if (strcmp(dname, ".")) {
2406                 input_dir = dname;
2407                 p = strdup(pdi->name);
2408                 strcpy(pdi->name, basename(p));
2409         }
2410         free(p);
2411
2412         if (input_dir)
2413                 len = sprintf(pci->fname, "%s/", input_dir);
2414
2415         snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
2416                  "%s.blktrace.%d", pdi->name, pci->cpu);
2417         if (stat(pci->fname, &st) < 0)
2418                 return 0;
2419         if (!st.st_size)
2420                 return 1;
2421
2422         pci->fd = open(pci->fname, O_RDONLY);
2423         if (pci->fd < 0) {
2424                 perror(pci->fname);
2425                 return 0;
2426         }
2427
2428         printf("Input file %s added\n", pci->fname);
2429         cpu_mark_online(pdi, pci->cpu);
2430
2431         pdi->nfiles++;
2432         ms_alloc(pdi, pci->cpu);
2433
2434         return 1;
2435 }
2436
2437 static int handle(struct ms_stream *msp)
2438 {
2439         struct trace *t;
2440         struct per_dev_info *pdi;
2441         struct per_cpu_info *pci;
2442         struct blk_io_trace *bit;
2443
2444         t = ms_peek(msp);
2445
2446         bit = t->bit;
2447         pdi = msp->pdi;
2448         pci = get_cpu_info(pdi, msp->cpu);
2449         pci->nelems++;
2450         bit->time -= genesis_time;
2451
2452         if (t->bit->time > stopwatch_end)
2453                 return 0;
2454
2455         pdi->last_reported_time = bit->time;
2456         if ((bit->action & (act_mask << BLK_TC_SHIFT))&&
2457             t->bit->time >= stopwatch_start)
2458                 dump_trace(bit, pci, pdi);
2459
2460         ms_deq(msp);
2461
2462         if (text_output)
2463                 trace_rb_insert_last(pdi, t);
2464         else {
2465                 bit_free(t->bit);
2466                 t_free(t);
2467         }
2468
2469         return 1;
2470 }
2471
2472 /*
2473  * Check if we need to sanitize the name. We allow 'foo', or if foo.blktrace.X
2474  * is given, then strip back down to 'foo' to avoid missing files.
2475  */
2476 static int name_fixup(char *name)
2477 {
2478         char *b;
2479
2480         if (!name)
2481                 return 1;
2482
2483         b = strstr(name, ".blktrace.");
2484         if (b)
2485                 *b = '\0';
2486
2487         return 0;
2488 }
2489
2490 static int do_file(void)
2491 {
2492         int i, cpu, ret;
2493         struct per_dev_info *pdi;
2494
2495         /*
2496          * first prepare all files for reading
2497          */
2498         for (i = 0; i < ndevices; i++) {
2499                 pdi = &devices[i];
2500                 ret = name_fixup(pdi->name);
2501                 if (ret)
2502                         return ret;
2503
2504                 for (cpu = 0; setup_file(pdi, cpu); cpu++)
2505                         ;
2506         }
2507
2508         /*
2509          * Get the initial time stamp
2510          */
2511         if (ms_head)
2512                 genesis_time = ms_peek_time(ms_head);
2513
2514         /*
2515          * Keep processing traces while any are left
2516          */
2517         while (!is_done() && ms_head && handle(ms_head))
2518                 ;
2519
2520         return 0;
2521 }
2522
2523 static void do_pipe(int fd)
2524 {
2525         unsigned long long youngest;
2526         int events, fdblock;
2527
2528         last_allowed_time = -1ULL;
2529         fdblock = -1;
2530         while ((events = read_events(fd, 0, &fdblock)) > 0) {
2531                 read_sequence++;
2532
2533 #if 0
2534                 smallest_seq_read = -1U;
2535 #endif
2536
2537                 if (sort_entries(&youngest))
2538                         break;
2539
2540                 if (youngest > stopwatch_end)
2541                         break;
2542
2543                 show_entries_rb(0);
2544         }
2545
2546         if (rb_sort_entries)
2547                 show_entries_rb(1);
2548 }
2549
2550 static int do_fifo(void)
2551 {
2552         int fd;
2553
2554         if (!strcmp(pipename, "-"))
2555                 fd = dup(STDIN_FILENO);
2556         else
2557                 fd = open(pipename, O_RDONLY);
2558
2559         if (fd == -1) {
2560                 perror("dup stdin");
2561                 return -1;
2562         }
2563
2564         do_pipe(fd);
2565         close(fd);
2566         return 0;
2567 }
2568
2569 static void show_stats(void)
2570 {
2571         if (!ofp)
2572                 return;
2573         if (stats_printed)
2574                 return;
2575
2576         stats_printed = 1;
2577
2578         if (per_process_stats)
2579                 show_process_stats();
2580
2581         if (per_device_and_cpu_stats)
2582                 show_device_and_cpu_stats();
2583
2584         fflush(ofp);
2585 }
2586
2587 static void handle_sigint(__attribute__((__unused__)) int sig)
2588 {
2589         done = 1;
2590 }
2591
2592 /*
2593  * Extract start and duration times from a string, allowing
2594  * us to specify a time interval of interest within a trace.
2595  * Format: "duration" (start is zero) or "start:duration".
2596  */
2597 static int find_stopwatch_interval(char *string)
2598 {
2599         double value;
2600         char *sp;
2601
2602         value = strtod(string, &sp);
2603         if (sp == string) {
2604                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
2605                 return 1;
2606         }
2607         if (*sp == ':') {
2608                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
2609                 string = sp + 1;
2610                 value = strtod(string, &sp);
2611                 if (sp == string || *sp != '\0') {
2612                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
2613                                 string);
2614                         return 1;
2615                 }
2616         } else if (*sp != '\0') {
2617                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
2618                 return 1;
2619         }
2620         stopwatch_end = DOUBLE_TO_NANO_ULL(value);
2621         if (stopwatch_end <= stopwatch_start) {
2622                 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
2623                         stopwatch_start, stopwatch_end);
2624                 return 1;
2625         }
2626
2627         return 0;
2628 }
2629
2630 static int is_pipe(const char *str)
2631 {
2632         struct stat st;
2633
2634         if (!strcmp(str, "-"))
2635                 return 1;
2636         if (!stat(str, &st) && S_ISFIFO(st.st_mode))
2637                 return 1;
2638
2639         return 0;
2640 }
2641
2642 #define S_OPTS  "a:A:b:D:d:f:F:hi:o:Oqstw:vV"
2643 static char usage_str[] =    "\n\n" \
2644         "-i <file>           | --input=<file>\n" \
2645         "[ -a <action field> | --act-mask=<action field> ]\n" \
2646         "[ -A <action mask>  | --set-mask=<action mask> ]\n" \
2647         "[ -b <traces>       | --batch=<traces> ]\n" \
2648         "[ -d <file>         | --dump-binary=<file> ]\n" \
2649         "[ -D <dir>          | --input-directory=<dir> ]\n" \
2650         "[ -f <format>       | --format=<format> ]\n" \
2651         "[ -F <spec>         | --format-spec=<spec> ]\n" \
2652         "[ -h                | --hash-by-name ]\n" \
2653         "[ -o <file>         | --output=<file> ]\n" \
2654         "[ -O                | --no-text-output ]\n" \
2655         "[ -q                | --quiet ]\n" \
2656         "[ -s                | --per-program-stats ]\n" \
2657         "[ -t                | --track-ios ]\n" \
2658         "[ -w <time>         | --stopwatch=<time> ]\n" \
2659         "[ -v                | --verbose ]\n" \
2660         "[ -V                | --version ]\n\n" \
2661         "\t-b stdin read batching\n" \
2662         "\t-d Output file. If specified, binary data is written to file\n" \
2663         "\t-D Directory to prepend to input file names\n" \
2664         "\t-f Output format. Customize the output format. The format field\n" \
2665         "\t   identifies can be found in the documentation\n" \
2666         "\t-F Format specification. Can be found in the documentation\n" \
2667         "\t-h Hash processes by name, not pid\n" \
2668         "\t-i Input file containing trace data, or '-' for stdin\n" \
2669         "\t-o Output file. If not given, output is stdout\n" \
2670         "\t-O Do NOT output text data\n" \
2671         "\t-q Quiet. Don't display any stats at the end of the trace\n" \
2672         "\t-s Show per-program io statistics\n" \
2673         "\t-t Track individual ios. Will tell you the time a request took\n" \
2674         "\t   to get queued, to get dispatched, and to get completed\n" \
2675         "\t-w Only parse data between the given time interval in seconds.\n" \
2676         "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
2677         "\t-v More verbose for marginal errors\n" \
2678         "\t-V Print program version info\n\n";
2679
2680 static void usage(char *prog)
2681 {
2682         fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
2683 }
2684
2685 int main(int argc, char *argv[])
2686 {
2687         int i, c, ret, mode;
2688         int act_mask_tmp = 0;
2689         char *ofp_buffer = NULL;
2690         char *bin_ofp_buffer = NULL;
2691
2692         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
2693                 switch (c) {
2694                 case 'a':
2695                         i = find_mask_map(optarg);
2696                         if (i < 0) {
2697                                 fprintf(stderr,"Invalid action mask %s\n",
2698                                         optarg);
2699                                 return 1;
2700                         }
2701                         act_mask_tmp |= i;
2702                         break;
2703
2704                 case 'A':
2705                         if ((sscanf(optarg, "%x", &i) != 1) ||
2706                                                         !valid_act_opt(i)) {
2707                                 fprintf(stderr,
2708                                         "Invalid set action mask %s/0x%x\n",
2709                                         optarg, i);
2710                                 return 1;
2711                         }
2712                         act_mask_tmp = i;
2713                         break;
2714                 case 'i':
2715                         if (is_pipe(optarg) && !pipeline) {
2716                                 pipeline = 1;
2717                                 pipename = strdup(optarg);
2718                         } else if (resize_devices(optarg) != 0)
2719                                 return 1;
2720                         break;
2721                 case 'D':
2722                         input_dir = optarg;
2723                         break;
2724                 case 'o':
2725                         output_name = optarg;
2726                         break;
2727                 case 'O':
2728                         text_output = 0;
2729                         break;
2730                 case 'b':
2731                         rb_batch = atoi(optarg);
2732                         if (rb_batch <= 0)
2733                                 rb_batch = RB_BATCH_DEFAULT;
2734                         break;
2735                 case 's':
2736                         per_process_stats = 1;
2737                         break;
2738                 case 't':
2739                         track_ios = 1;
2740                         break;
2741                 case 'q':
2742                         per_device_and_cpu_stats = 0;
2743                         break;
2744                 case 'w':
2745                         if (find_stopwatch_interval(optarg) != 0)
2746                                 return 1;
2747                         break;
2748                 case 'f':
2749                         set_all_format_specs(optarg);
2750                         break;
2751                 case 'F':
2752                         if (add_format_spec(optarg) != 0)
2753                                 return 1;
2754                         break;
2755                 case 'h':
2756                         ppi_hash_by_pid = 0;
2757                         break;
2758                 case 'v':
2759                         verbose++;
2760                         break;
2761                 case 'V':
2762                         printf("%s version %s\n", argv[0], blkparse_version);
2763                         return 0;
2764                 case 'd':
2765                         dump_binary = optarg;
2766                         break;
2767                 default:
2768                         usage(argv[0]);
2769                         return 1;
2770                 }
2771         }
2772
2773         while (optind < argc) {
2774                 if (is_pipe(argv[optind]) && !pipeline) {
2775                         pipeline = 1;
2776                         pipename = strdup(argv[optind]);
2777                 } else if (resize_devices(argv[optind]) != 0)
2778                         return 1;
2779                 optind++;
2780         }
2781
2782         if (!pipeline && !ndevices) {
2783                 usage(argv[0]);
2784                 return 1;
2785         }
2786
2787         if (act_mask_tmp != 0)
2788                 act_mask = act_mask_tmp;
2789
2790         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
2791
2792         signal(SIGINT, handle_sigint);
2793         signal(SIGHUP, handle_sigint);
2794         signal(SIGTERM, handle_sigint);
2795
2796         setlocale(LC_NUMERIC, "en_US");
2797
2798         if (text_output) {
2799                 if (!output_name) {
2800                         ofp = fdopen(STDOUT_FILENO, "w");
2801                         mode = _IOLBF;
2802                 } else {
2803                         char ofname[128];
2804
2805                         snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
2806                         ofp = fopen(ofname, "w");
2807                         mode = _IOFBF;
2808                 }
2809
2810                 if (!ofp) {
2811                         perror("fopen");
2812                         return 1;
2813                 }
2814
2815                 ofp_buffer = malloc(4096);
2816                 if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
2817                         perror("setvbuf");
2818                         return 1;
2819                 }
2820         }
2821
2822         if (dump_binary) {
2823                 dump_fp = fopen(dump_binary, "w");
2824                 if (!dump_fp) {
2825                         perror(dump_binary);
2826                         dump_binary = NULL;
2827                         return 1;
2828                 }
2829                 bin_ofp_buffer = malloc(128 * 1024);
2830                 if (setvbuf(dump_fp, bin_ofp_buffer, _IOFBF, 128 * 1024)) {
2831                         perror("setvbuf binary");
2832                         return 1;
2833                 }
2834         }
2835
2836         if (pipeline)
2837                 ret = do_fifo();
2838         else
2839                 ret = do_file();
2840
2841         if (!ret)
2842                 show_stats();
2843
2844         if (ofp_buffer) {
2845                 fflush(ofp);
2846                 free(ofp_buffer);
2847         }
2848         if (bin_ofp_buffer) {
2849                 fflush(dump_fp);
2850                 free(bin_ofp_buffer);
2851         }
2852         return ret;
2853 }