t/btrace2fio.c

   1 #include <stdio.h>
   2 #include <stdio.h>
   3 #include <unistd.h>
   4 #include <inttypes.h>
   5 #include <math.h>
   6 #include <assert.h>
   7
   8 #include "../io_ddir.h"
   9 #include "../flist.h"
  10 #include "../hash.h"
  11 #include "../fifo.h"
  12 #include "../blktrace_api.h"
  13 #include "../os/os.h"
  14 #include "../log.h"
  15 #include "../lib/linux-dev-lookup.h"
  16
  17 #define TRACE_FIFO_SIZE 8192
  18
  19 static unsigned int rt_threshold = 1000000;
  20 static unsigned int ios_threshold = 10;
  21 static unsigned int rate_threshold;
  22 static unsigned int set_rate;
  23 static unsigned int max_depth = 256;
  24 static int output_ascii = 1;
  25 static char *filename;
  26
  27 /*
  28  * Collapse defaults
  29  */
  30 static unsigned int collapse_entries = 0;
  31 static unsigned int depth_diff = 1;
  32 static unsigned int random_diff = 5;
  33
  34 struct bs {
  35         unsigned int bs;
  36         unsigned int nr;
  37         int merges;
  38 };
  39
  40 struct trace_file {
  41         char *name;
  42         int major, minor;
  43 };
  44
  45 struct btrace_out {
  46         unsigned long ios[DDIR_RWDIR_CNT];
  47         unsigned long merges[DDIR_RWDIR_CNT];
  48
  49         uint64_t last_end[DDIR_RWDIR_CNT];
  50         uint64_t seq[DDIR_RWDIR_CNT];
  51
  52         struct bs *bs[DDIR_RWDIR_CNT];
  53         unsigned int nr_bs[DDIR_RWDIR_CNT];
  54
  55         int inflight;
  56         unsigned int depth;
  57         int depth_disabled;
  58         int complete_seen;
  59
  60         uint64_t first_ttime[DDIR_RWDIR_CNT];
  61         uint64_t last_ttime[DDIR_RWDIR_CNT];
  62         uint64_t kb[DDIR_RWDIR_CNT];
  63
  64         uint64_t start_delay;
  65 };
  66
  67 struct btrace_pid {
  68         struct flist_head hash_list;
  69         struct flist_head pid_list;
  70         pid_t pid;
  71
  72         pid_t *merge_pids;
  73         unsigned int nr_merge_pids;
  74
  75         struct trace_file *files;
  76         int nr_files;
  77         unsigned int last_major, last_minor;
  78         int numjobs;
  79         int ignore;
  80
  81         struct btrace_out o;
  82 };
  83
  84 struct inflight {
  85         struct flist_head list;
  86         struct btrace_pid *p;
  87         uint64_t end_sector;
  88 };
  89
  90 #define PID_HASH_BITS   10
  91 #define PID_HASH_SIZE   (1U << PID_HASH_BITS)
  92
  93 static struct flist_head pid_hash[PID_HASH_SIZE];
  94 static FLIST_HEAD(pid_list);
  95
  96 #define INFLIGHT_HASH_BITS      8
  97 #define INFLIGHT_HASH_SIZE      (1U << INFLIGHT_HASH_BITS)
  98 static struct flist_head inflight_hash[INFLIGHT_HASH_SIZE];
  99
 100 static uint64_t first_ttime = -1ULL;
 101
 102 static struct inflight *inflight_find(uint64_t sector)
 103 {
 104         struct flist_head *inflight_list;
 105         struct flist_head *e;
 106
 107         inflight_list = &inflight_hash[hash_long(sector, INFLIGHT_HASH_BITS)];
 108
 109         flist_for_each(e, inflight_list) {
 110                 struct inflight *i = flist_entry(e, struct inflight, list);
 111
 112                 if (i->end_sector == sector)
 113                         return i;
 114         }
 115
 116         return NULL;
 117 }
 118
 119 static void inflight_remove(struct inflight *i)
 120 {
 121         struct btrace_out *o = &i->p->o;
 122
 123         o->inflight--;
 124         assert(o->inflight >= 0);
 125         flist_del(&i->list);
 126         free(i);
 127 }
 128
 129 static void __inflight_add(struct inflight *i)
 130 {
 131         struct flist_head *list;
 132
 133         list = &inflight_hash[hash_long(i->end_sector, INFLIGHT_HASH_BITS)];
 134         flist_add_tail(&i->list, list);
 135 }
 136
 137 static void inflight_add(struct btrace_pid *p, uint64_t sector, uint32_t len)
 138 {
 139         struct btrace_out *o = &p->o;
 140         struct inflight *i;
 141
 142         i = calloc(1, sizeof(*i));
 143         i->p = p;
 144         o->inflight++;
 145         if (!o->depth_disabled) {
 146                 o->depth = max((int) o->depth, o->inflight);
 147                 if (o->depth >= max_depth && !o->complete_seen) {
 148                         o->depth_disabled = 1;
 149                         o->depth = max_depth;
 150                 }
 151         }
 152         i->end_sector = sector + (len >> 9);
 153         __inflight_add(i);
 154 }
 155
 156 static void inflight_merge(struct inflight *i, int rw, unsigned int size)
 157 {
 158         i->p->o.merges[rw]++;
 159         if (size) {
 160                 i->end_sector += (size >> 9);
 161                 flist_del(&i->list);
 162                 __inflight_add(i);
 163         }
 164 }
 165
 166 /*
 167  * fifo refill frontend, to avoid reading data in trace sized bites
 168  */
 169 static int refill_fifo(struct fifo *fifo, int fd)
 170 {
 171         char buf[TRACE_FIFO_SIZE];
 172         unsigned int total;
 173         int ret;
 174
 175         total = sizeof(buf);
 176         if (total > fifo_room(fifo))
 177                 total = fifo_room(fifo);
 178
 179         ret = read(fd, buf, total);
 180         if (ret < 0) {
 181                 perror("read refill");
 182                 return -1;
 183         }
 184
 185         if (ret > 0)
 186                 ret = fifo_put(fifo, buf, ret);
 187
 188         return ret;
 189 }
 190
 191 /*
 192  * Retrieve 'len' bytes from the fifo, refilling if necessary.
 193  */
 194 static int trace_fifo_get(struct fifo *fifo, int fd, void *buf,
 195                           unsigned int len)
 196 {
 197         if (fifo_len(fifo) < len) {
 198                 int ret = refill_fifo(fifo, fd);
 199
 200                 if (ret < 0)
 201                         return ret;
 202         }
 203
 204         return fifo_get(fifo, buf, len);
 205 }
 206
 207 /*
 208  * Just discard the pdu by seeking past it.
 209  */
 210 static int discard_pdu(struct fifo *fifo, int fd, struct blk_io_trace *t)
 211 {
 212         if (t->pdu_len == 0)
 213                 return 0;
 214
 215         return trace_fifo_get(fifo, fd, NULL, t->pdu_len);
 216 }
 217
 218 static int handle_trace_notify(struct blk_io_trace *t)
 219 {
 220         switch (t->action) {
 221         case BLK_TN_PROCESS:
 222                 //printf("got process notify: %x, %d\n", t->action, t->pid);
 223                 break;
 224         case BLK_TN_TIMESTAMP:
 225                 //printf("got timestamp notify: %x, %d\n", t->action, t->pid);
 226                 break;
 227         case BLK_TN_MESSAGE:
 228                 break;
 229         default:
 230                 log_err("unknown trace act %x\n", t->action);
 231                 return 1;
 232         }
 233
 234         return 0;
 235 }
 236
 237 static void __add_bs(struct btrace_out *o, unsigned int len, int rw)
 238 {
 239         o->bs[rw] = realloc(o->bs[rw], (o->nr_bs[rw] + 1) * sizeof(struct bs));
 240         o->bs[rw][o->nr_bs[rw]].bs = len;
 241         o->bs[rw][o->nr_bs[rw]].nr = 1;
 242         o->nr_bs[rw]++;
 243 }
 244
 245 static void add_bs(struct btrace_out *o, unsigned int len, int rw)
 246 {
 247         struct bs *bs = o->bs[rw];
 248         int i;
 249
 250         if (!o->nr_bs[rw]) {
 251                 __add_bs(o, len, rw);
 252                 return;
 253         }
 254
 255         for (i = 0; i < o->nr_bs[rw]; i++) {
 256                 if (bs[i].bs == len) {
 257                         bs[i].nr++;
 258                         return;
 259                 }
 260         }
 261
 262         __add_bs(o, len, rw);
 263 }
 264
 265 #define FMINORBITS      20
 266 #define FMINORMASK      ((1U << FMINORBITS) - 1)
 267 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
 268 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
 269
 270 static int btrace_add_file(struct btrace_pid *p, uint32_t devno)
 271 {
 272         unsigned int maj = FMAJOR(devno);
 273         unsigned int min = FMINOR(devno);
 274         struct trace_file *f;
 275         unsigned int i;
 276         char dev[256];
 277
 278         if (filename)
 279                 return 0;
 280         if (p->last_major == maj && p->last_minor == min)
 281                 return 0;
 282
 283         p->last_major = maj;
 284         p->last_minor = min;
 285
 286         /*
 287          * check for this file in our list
 288          */
 289         for (i = 0; i < p->nr_files; i++) {
 290                 f = &p->files[i];
 291
 292                 if (f->major == maj && f->minor == min)
 293                         return 0;
 294         }
 295
 296         strcpy(dev, "/dev");
 297         if (!blktrace_lookup_device(NULL, dev, maj, min)) {
 298                 log_err("fio: failed to find device %u/%u\n", maj, min);
 299                 if (!output_ascii) {
 300                         log_err("fio: use -d to specify device\n");
 301                         return 1;
 302                 }
 303                 return 0;
 304         }
 305
 306         p->files = realloc(p->files, (p->nr_files + 1) * sizeof(*f));
 307         f = &p->files[p->nr_files];
 308         f->name = strdup(dev);
 309         f->major = maj;
 310         f->minor = min;
 311         p->nr_files++;
 312         return 0;
 313 }
 314
 315 static int t_to_rwdir(struct blk_io_trace *t)
 316 {
 317         if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 318                 return DDIR_TRIM;
 319
 320         return (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 321 }
 322
 323 static int handle_trace_discard(struct blk_io_trace *t, struct btrace_pid *p)
 324 {
 325         struct btrace_out *o = &p->o;
 326
 327         if (btrace_add_file(p, t->device))
 328                 return 1;
 329
 330         if (o->first_ttime[2] == -1ULL)
 331                 o->first_ttime[2] = t->time;
 332
 333         o->ios[DDIR_TRIM]++;
 334         add_bs(o, t->bytes, DDIR_TRIM);
 335         return 0;
 336 }
 337
 338 static int handle_trace_fs(struct blk_io_trace *t, struct btrace_pid *p)
 339 {
 340         struct btrace_out *o = &p->o;
 341         int rw;
 342
 343         if (btrace_add_file(p, t->device))
 344                 return 1;
 345
 346         first_ttime = min(first_ttime, (uint64_t) t->time);
 347
 348         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 349
 350         if (o->first_ttime[rw] == -1ULL)
 351                 o->first_ttime[rw] = t->time;
 352
 353         add_bs(o, t->bytes, rw);
 354         o->ios[rw]++;
 355
 356         if (t->sector == o->last_end[rw] || o->last_end[rw] == -1ULL)
 357                 o->seq[rw]++;
 358
 359         o->last_end[rw] = t->sector + (t->bytes >> 9);
 360         return 0;
 361 }
 362
 363 static int handle_queue_trace(struct blk_io_trace *t, struct btrace_pid *p)
 364 {
 365         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 366                 return handle_trace_notify(t);
 367         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 368                 return handle_trace_discard(t, p);
 369         else
 370                 return handle_trace_fs(t, p);
 371 }
 372
 373 static int handle_trace(struct blk_io_trace *t, struct btrace_pid *p)
 374 {
 375         unsigned int act = t->action & 0xffff;
 376         int ret = 0;
 377
 378         if (act == __BLK_TA_QUEUE) {
 379                 inflight_add(p, t->sector, t->bytes);
 380                 ret = handle_queue_trace(t, p);
 381         } else if (act == __BLK_TA_BACKMERGE) {
 382                 struct inflight *i;
 383
 384                 i = inflight_find(t->sector + (t->bytes >> 9));
 385                 if (i)
 386                         inflight_remove(i);
 387
 388                 i = inflight_find(t->sector);
 389                 if (i)
 390                         inflight_merge(i, t_to_rwdir(t), t->bytes);
 391         } else if (act == __BLK_TA_FRONTMERGE) {
 392                 struct inflight *i;
 393
 394                 i = inflight_find(t->sector + (t->bytes >> 9));
 395                 if (i)
 396                         inflight_remove(i);
 397
 398                 i = inflight_find(t->sector);
 399                 if (i)
 400                         inflight_merge(i, t_to_rwdir(t), 0);
 401         } else if (act == __BLK_TA_COMPLETE) {
 402                 struct inflight *i;
 403
 404                 i = inflight_find(t->sector + (t->bytes >> 9));
 405                 if (i) {
 406                         i->p->o.kb[t_to_rwdir(t)] += (t->bytes >> 10);
 407                         i->p->o.complete_seen = 1;
 408                         inflight_remove(i);
 409                 }
 410         }
 411
 412         return ret;
 413 }
 414
 415 static void byteswap_trace(struct blk_io_trace *t)
 416 {
 417         t->magic = fio_swap32(t->magic);
 418         t->sequence = fio_swap32(t->sequence);
 419         t->time = fio_swap64(t->time);
 420         t->sector = fio_swap64(t->sector);
 421         t->bytes = fio_swap32(t->bytes);
 422         t->action = fio_swap32(t->action);
 423         t->pid = fio_swap32(t->pid);
 424         t->device = fio_swap32(t->device);
 425         t->cpu = fio_swap32(t->cpu);
 426         t->error = fio_swap16(t->error);
 427         t->pdu_len = fio_swap16(t->pdu_len);
 428 }
 429
 430 static struct btrace_pid *pid_hash_find(pid_t pid, struct flist_head *list)
 431 {
 432         struct flist_head *e;
 433         struct btrace_pid *p;
 434
 435         flist_for_each(e, list) {
 436                 p = flist_entry(e, struct btrace_pid, hash_list);
 437                 if (p->pid == pid)
 438                         return p;
 439         }
 440
 441         return NULL;
 442 }
 443
 444 static struct btrace_pid *pid_hash_get(pid_t pid)
 445 {
 446         struct flist_head *hash_list;
 447         struct btrace_pid *p;
 448
 449         hash_list = &pid_hash[hash_long(pid, PID_HASH_BITS)];
 450
 451         p = pid_hash_find(pid, hash_list);
 452         if (!p) {
 453                 int i;
 454
 455                 p = calloc(1, sizeof(*p));
 456
 457                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 458                         p->o.first_ttime[i] = -1ULL;
 459                         p->o.last_ttime[i] = -1ULL;
 460                         p->o.last_end[i] = -1ULL;
 461                 }
 462
 463                 p->pid = pid;
 464                 p->numjobs = 1;
 465                 flist_add_tail(&p->hash_list, hash_list);
 466                 flist_add_tail(&p->pid_list, &pid_list);
 467         }
 468
 469         return p;
 470 }
 471
 472 /*
 473  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 474  * them as io_pieces like the fio text version would do.
 475  */
 476 static int load_blktrace(const char *fname, int need_swap)
 477 {
 478         struct btrace_pid *p;
 479         unsigned long traces;
 480         struct blk_io_trace t;
 481         struct fifo *fifo;
 482         int fd, ret = 0;
 483
 484         fd = open(fname, O_RDONLY);
 485         if (fd < 0) {
 486                 perror("open trace file\n");
 487                 return 1;
 488         }
 489
 490         fifo = fifo_alloc(TRACE_FIFO_SIZE);
 491
 492         traces = 0;
 493         do {
 494                 ret = trace_fifo_get(fifo, fd, &t, sizeof(t));
 495                 if (ret < 0)
 496                         goto err;
 497                 else if (!ret)
 498                         break;
 499                 else if (ret < (int) sizeof(t)) {
 500                         log_err("fio: short fifo get\n");
 501                         break;
 502                 }
 503
 504                 if (need_swap)
 505                         byteswap_trace(&t);
 506
 507                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 508                         log_err("fio: bad magic in blktrace data: %x\n", t.magic);
 509                         goto err;
 510                 }
 511                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 512                         log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
 513                         goto err;
 514                 }
 515                 ret = discard_pdu(fifo, fd, &t);
 516                 if (ret < 0) {
 517                         log_err("blktrace lseek\n");
 518                         goto err;
 519                 } else if (t.pdu_len != ret) {
 520                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
 521                         goto err;
 522                 }
 523
 524                 p = pid_hash_get(t.pid);
 525                 ret = handle_trace(&t, p);
 526                 if (ret)
 527                         break;
 528                 p->o.last_ttime[t_to_rwdir(&t)] = t.time;
 529                 traces++;
 530         } while (1);
 531
 532         fifo_free(fifo);
 533         close(fd);
 534
 535         if (ret)
 536                 return ret;
 537
 538         if (output_ascii)
 539                 printf("Traces loaded: %lu\n", traces);
 540
 541         return 0;
 542 err:
 543         close(fd);
 544         fifo_free(fifo);
 545         return 1;
 546 }
 547
 548 static int bs_cmp(const void *ba, const void *bb)
 549 {
 550         const struct bs *bsa = ba;
 551         const struct bs *bsb = bb;
 552
 553         return bsb->nr - bsa->nr;
 554 }
 555
 556 static unsigned long o_to_kb_rate(struct btrace_out *o, int rw)
 557 {
 558         uint64_t usec = (o->last_ttime[rw] - o->first_ttime[rw]) / 1000ULL;
 559         uint64_t val;
 560
 561         if (!usec)
 562                 return 0;
 563
 564         usec /= 1000;
 565         if (!usec)
 566                 return 0;
 567
 568         val = o->kb[rw] * 1000ULL;
 569         return val / usec;
 570 }
 571
 572 static uint64_t o_first_ttime(struct btrace_out *o)
 573 {
 574         uint64_t first;
 575
 576         first = min(o->first_ttime[0], o->first_ttime[1]);
 577         return min(first, o->first_ttime[2]);
 578 }
 579
 580 static uint64_t o_longest_ttime(struct btrace_out *o)
 581 {
 582         uint64_t ret = 0;
 583         int i;
 584
 585         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 586                 uint64_t diff;
 587
 588                 diff = o->last_ttime[i] - o->first_ttime[i];
 589                 ret = max(diff, ret);
 590         }
 591
 592         return ret;
 593 }
 594
 595 static void __output_p_ascii(struct btrace_pid *p, unsigned long *ios)
 596 {
 597         const char *msg[] = { "reads", "writes", "trims" };
 598         struct btrace_out *o = &p->o;
 599         unsigned long total, usec;
 600         int i, j;
 601
 602         printf("[pid:\t%u", p->pid);
 603         if (p->nr_merge_pids)
 604                 for (i = 0; i < p->nr_merge_pids; i++)
 605                         printf(", %u", p->merge_pids[i]);
 606         printf("]\n");
 607
 608         total = ddir_rw_sum(o->ios);
 609         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 610                 float perc;
 611
 612                 if (!o->ios[i])
 613                         continue;
 614
 615                 ios[i] += o->ios[i] + o->merges[i];
 616                 printf("%s\n", msg[i]);
 617                 perc = ((float) o->ios[i] * 100.0) / (float) total;
 618                 printf("\tios:    %lu (perc=%3.2f%%)\n", o->ios[i], perc);
 619                 perc = ((float) o->merges[i] * 100.0) / (float) total;
 620                 printf("\tmerges: %lu (perc=%3.2f%%)\n", o->merges[i], perc);
 621                 perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
 622                 printf("\tseq:    %lu (perc=%3.2f%%)\n", (unsigned long) o->seq[i], perc);
 623                 printf("\trate:   %lu KB/sec\n", o_to_kb_rate(o, i));
 624
 625                 for (j = 0; j < o->nr_bs[i]; j++) {
 626                         struct bs *bs = &o->bs[i][j];
 627
 628                         perc = (((float) bs->nr * 100.0) / (float) o->ios[i]);
 629                         printf("\tbs=%u, perc=%3.2f%%\n", bs->bs, perc);
 630                 }
 631         }
 632
 633         printf("depth:\t%u\n", o->depth);
 634         usec = o_longest_ttime(o) / 1000ULL;
 635         printf("usec:\t%lu (delay=%llu)\n", usec, (unsigned long long) o->start_delay);
 636
 637         printf("files:\t");
 638         for (i = 0; i < p->nr_files; i++)
 639                 printf("%s,", p->files[i].name);
 640         printf("\n");
 641
 642         printf("\n");
 643 }
 644
 645 static int __output_p_fio(struct btrace_pid *p, unsigned long *ios)
 646 {
 647         struct btrace_out *o = &p->o;
 648         unsigned long total;
 649         unsigned long long time;
 650         float perc;
 651         int i, j;
 652
 653         if ((o->ios[0] + o->ios[1]) && o->ios[2]) {
 654                 log_err("fio: trace has both read/write and trim\n");
 655                 return 1;
 656         }
 657         if (!p->nr_files) {
 658                 log_err("fio: no devices found\n");
 659                 return 1;
 660         }
 661
 662         printf("[pid%u", p->pid);
 663         if (p->nr_merge_pids)
 664                 for (i = 0; i < p->nr_merge_pids; i++)
 665                         printf(",pid%u", p->merge_pids[i]);
 666         printf("]\n");
 667
 668         printf("numjobs=%u\n", p->numjobs);
 669         printf("direct=1\n");
 670         if (o->depth == 1)
 671                 printf("ioengine=sync\n");
 672         else
 673                 printf("ioengine=libaio\niodepth=%u\n", o->depth);
 674
 675         if (o->ios[0] && !o->ios[1])
 676                 printf("rw=randread\n");
 677         else if (!o->ios[0] && o->ios[1])
 678                 printf("rw=randwrite\n");
 679         else if (o->ios[2])
 680                 printf("rw=randtrim\n");
 681         else {
 682                 printf("rw=randrw\n");
 683                 total = ddir_rw_sum(o->ios);
 684                 perc = ((float) o->ios[0] * 100.0) / (float) total;
 685                 printf("rwmixread=%u\n", (int) floor(perc + 0.50));
 686         }
 687
 688         printf("percentage_random=");
 689         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 690                 if (o->seq[i] && o->ios[i]) {
 691                         perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
 692                         if (perc >= 99.0)
 693                                 perc = 100.0;
 694                 } else
 695                         perc = 100.0;
 696
 697                 if (i)
 698                         printf(",");
 699                 perc = 100.0 - perc;
 700                 printf("%u", (int) floor(perc + 0.5));
 701         }
 702         printf("\n");
 703
 704         printf("filename=");
 705         for (i = 0; i < p->nr_files; i++) {
 706                 if (i)
 707                         printf(":");
 708                 printf("%s", p->files[i].name);
 709         }
 710         printf("\n");
 711
 712         if (o->start_delay / 1000000ULL)
 713                 printf("startdelay=%llus\n", o->start_delay / 1000000ULL);
 714
 715         time = o_longest_ttime(o);
 716         time = (time + 1000000000ULL - 1) / 1000000000ULL;
 717         printf("runtime=%llus\n", time);
 718
 719         printf("bssplit=");
 720         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 721
 722                 if (i && o->nr_bs[i - 1] && o->nr_bs[i])
 723                         printf(",");
 724
 725                 for (j = 0; j < o->nr_bs[i]; j++) {
 726                         struct bs *bs = &o->bs[i][j];
 727
 728                         perc = (((float) bs->nr * 100.0) / (float) o->ios[i]);
 729                         if (perc < 1.00)
 730                                 continue;
 731                         if (j)
 732                                 printf(":");
 733                         if (j + 1 == o->nr_bs[i])
 734                                 printf("%u/", bs->bs);
 735                         else
 736                                 printf("%u/%u", bs->bs, (int) floor(perc + 0.5));
 737                 }
 738         }
 739         printf("\n");
 740
 741         if (set_rate) {
 742                 printf("rate=");
 743                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 744                         unsigned long rate;
 745
 746                         rate = o_to_kb_rate(o, i);
 747                         if (i)
 748                                 printf(",");
 749                         if (rate)
 750                                 printf("%luk", rate);
 751                 }
 752                 printf("\n");
 753         }
 754
 755         printf("\n");
 756         return 0;
 757 }
 758
 759 static int __output_p(struct btrace_pid *p, unsigned long *ios)
 760 {
 761         struct btrace_out *o = &p->o;
 762         int i, ret = 0;
 763
 764         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 765                 if (o->nr_bs[i] <= 1)
 766                         continue;
 767                 qsort(o->bs[i], o->nr_bs[i], sizeof(struct bs), bs_cmp);
 768         }
 769
 770         if (filename) {
 771                 p->files = malloc(sizeof(struct trace_file));
 772                 p->nr_files++;
 773                 p->files[0].name = filename;
 774         }
 775
 776         if (output_ascii)
 777                 __output_p_ascii(p, ios);
 778         else
 779                 ret = __output_p_fio(p, ios);
 780
 781         return ret;
 782 }
 783
 784 static void remove_ddir(struct btrace_out *o, int rw)
 785 {
 786         o->ios[rw] = 0;
 787 }
 788
 789 static int prune_entry(struct btrace_out *o)
 790 {
 791         unsigned long rate;
 792         uint64_t time;
 793         int i;
 794
 795         if (ddir_rw_sum(o->ios) < ios_threshold)
 796                 return 1;
 797
 798         time = o_longest_ttime(o) / 1000ULL;
 799         if (time < rt_threshold)
 800                 return 1;
 801
 802         rate = 0;
 803         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 804                 unsigned long this_rate;
 805
 806                 this_rate = o_to_kb_rate(o, i);
 807                 if (this_rate < rate_threshold) {
 808                         remove_ddir(o, i);
 809                         this_rate = 0;
 810                 }
 811                 rate += this_rate;
 812         }
 813
 814         if (rate < rate_threshold)
 815                 return 1;
 816
 817         return 0;
 818 }
 819
 820 static int entry_cmp(void *priv, struct flist_head *a, struct flist_head *b)
 821 {
 822         struct btrace_pid *pa = flist_entry(a, struct btrace_pid, pid_list);
 823         struct btrace_pid *pb = flist_entry(b, struct btrace_pid, pid_list);
 824
 825         return ddir_rw_sum(pb->o.ios) - ddir_rw_sum(pa->o.ios);
 826 }
 827
 828 static void free_p(struct btrace_pid *p)
 829 {
 830         struct btrace_out *o = &p->o;
 831         int i;
 832
 833         for (i = 0; i < p->nr_files; i++) {
 834                 if (p->files[i].name && p->files[i].name != filename)
 835                         free(p->files[i].name);
 836         }
 837
 838         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 839                 free(o->bs[i]);
 840
 841         free(p->files);
 842         flist_del(&p->pid_list);
 843         flist_del(&p->hash_list);
 844         free(p);
 845 }
 846
 847 static int entries_close(struct btrace_pid *pida, struct btrace_pid *pidb)
 848 {
 849         float perca, percb, fdiff;
 850         int i, idiff;
 851
 852         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 853                 if ((pida->o.ios[i] && !pidb->o.ios[i]) ||
 854                     (pidb->o.ios[i] && !pida->o.ios[i]))
 855                         return 0;
 856                 if (pida->o.ios[i] && pidb->o.ios[i]) {
 857                         perca = ((float) pida->o.seq[i] * 100.0) / (float) pida->o.ios[i];
 858                         percb = ((float) pidb->o.seq[i] * 100.0) / (float) pidb->o.ios[i];
 859                         fdiff = perca - percb;
 860                         if (fabs(fdiff) > random_diff)
 861                                 return 0;
 862                 }
 863
 864                 idiff = pida->o.depth - pidb->o.depth;
 865                 if (abs(idiff) > depth_diff)
 866                         return 0;
 867         }
 868
 869         return 1;
 870 }
 871
 872 static void merge_bs(struct bs **bsap, unsigned int *nr_bsap,
 873                      struct bs *bsb, unsigned int nr_bsb)
 874 {
 875         struct bs *bsa = *bsap;
 876         unsigned int nr_bsa = *nr_bsap;
 877         int a, b;
 878
 879         for (b = 0; b < nr_bsb; b++) {
 880                 int next, found = 0;
 881
 882                 for (a = 0; a < nr_bsa; a++) {
 883                         if (bsb[b].bs != bsa[a].bs)
 884                                 continue;
 885
 886                         bsa[a].nr += bsb[b].nr;
 887                         bsa[a].merges += bsb[b].merges;
 888                         found = 1;
 889                         break;
 890                 }
 891
 892                 if (found)
 893                         continue;
 894
 895                 next = *nr_bsap;
 896                 bsa = realloc(bsa, (next + 1) * sizeof(struct bs));
 897                 bsa[next].bs = bsb[b].bs;
 898                 bsa[next].nr = bsb[b].nr;
 899                 (*nr_bsap)++;
 900                 *bsap = bsa;
 901         }
 902 }
 903
 904 static int merge_entries(struct btrace_pid *pida, struct btrace_pid *pidb)
 905 {
 906         int i;
 907
 908         if (!entries_close(pida, pidb))
 909                 return 0;
 910
 911         pida->nr_merge_pids++;
 912         pida->merge_pids = realloc(pida->merge_pids, pida->nr_merge_pids * sizeof(pid_t));
 913         pida->merge_pids[pida->nr_merge_pids - 1] = pidb->pid;
 914
 915         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 916                 struct btrace_out *oa = &pida->o;
 917                 struct btrace_out *ob = &pidb->o;
 918
 919                 oa->ios[i] += ob->ios[i];
 920                 oa->merges[i] += ob->merges[i];
 921                 oa->seq[i] += ob->seq[i];
 922                 oa->kb[i] += ob->kb[i];
 923                 oa->first_ttime[i] = min(oa->first_ttime[i], ob->first_ttime[i]);
 924                 oa->last_ttime[i] = max(oa->last_ttime[i], ob->last_ttime[i]);
 925                 merge_bs(&oa->bs[i], &oa->nr_bs[i], ob->bs[i], ob->nr_bs[i]);
 926         }
 927
 928         pida->o.start_delay = min(pida->o.start_delay, pidb->o.start_delay);
 929         pida->o.depth = (pida->o.depth + pidb->o.depth) / 2;
 930         return 1;
 931 }
 932
 933 static void check_merges(struct btrace_pid *p, struct flist_head *pid_list)
 934 {
 935         struct flist_head *e, *tmp;
 936
 937         if (p->ignore)
 938                 return;
 939
 940         flist_for_each_safe(e, tmp, pid_list) {
 941                 struct btrace_pid *pidb;
 942
 943                 pidb = flist_entry(e, struct btrace_pid, pid_list);
 944                 if (pidb == p)
 945                         continue;
 946
 947                 if (merge_entries(p, pidb)) {
 948                         pidb->ignore = 1;
 949                         p->numjobs++;
 950                 }
 951         }
 952 }
 953
 954 static int output_p(void)
 955 {
 956         unsigned long ios[DDIR_RWDIR_CNT];
 957         struct flist_head *e, *tmp;
 958         int depth_disabled = 0;
 959         int ret = 0;
 960
 961         flist_for_each_safe(e, tmp, &pid_list) {
 962                 struct btrace_pid *p;
 963
 964                 p = flist_entry(e, struct btrace_pid, pid_list);
 965                 if (prune_entry(&p->o)) {
 966                         free_p(p);
 967                         continue;
 968                 }
 969                 p->o.start_delay = (o_first_ttime(&p->o) / 1000ULL) - first_ttime;
 970                 depth_disabled += p->o.depth_disabled;
 971         }
 972
 973         if (collapse_entries) {
 974                 struct btrace_pid *p;
 975
 976                 flist_for_each_safe(e, tmp, &pid_list) {
 977                         p = flist_entry(e, struct btrace_pid, pid_list);
 978                         check_merges(p, &pid_list);
 979                 }
 980
 981                 flist_for_each_safe(e, tmp, &pid_list) {
 982                         p = flist_entry(e, struct btrace_pid, pid_list);
 983                         if (p->ignore)
 984                                 free_p(p);
 985                 }
 986         }
 987
 988         if (depth_disabled)
 989                 log_err("fio: missing completion traces, depths capped at %u\n", max_depth);
 990
 991         memset(ios, 0, sizeof(ios));
 992
 993         flist_sort(NULL, &pid_list, entry_cmp);
 994
 995         flist_for_each(e, &pid_list) {
 996                 struct btrace_pid *p;
 997
 998                 p = flist_entry(e, struct btrace_pid, pid_list);
 999                 ret |= __output_p(p, ios);
1000                 if (ret && !output_ascii)
1001                         break;
1002         }
1003
1004         if (output_ascii)
1005                 printf("Total: reads=%lu, writes=%lu\n", ios[0], ios[1]);
1006
1007         return ret;
1008 }
1009
1010 static int usage(char *argv[])
1011 {
1012         log_err("%s: [options] <blktrace bin file>\n", argv[0]);
1013         log_err("\t-t\tUsec threshold to ignore task\n");
1014         log_err("\t-n\tNumber IOS threshold to ignore task\n");
1015         log_err("\t-f\tFio job file output\n");
1016         log_err("\t-d\tUse this file/device for replay\n");
1017         log_err("\t-r\tIgnore jobs with less than this KB/sec rate\n");
1018         log_err("\t-R\tSet rate in fio job (def=%u)\n", set_rate);
1019         log_err("\t-D\tCap queue depth at this value (def=%u)\n", max_depth);
1020         log_err("\t-c\tCollapse \"identical\" jobs (def=%u)\n", collapse_entries);
1021         log_err("\t-u\tDepth difference for collapse (def=%u)\n", depth_diff);
1022         log_err("\t-x\tRandom difference for collapse (def=%u)\n", random_diff);
1023         return 1;
1024 }
1025
1026 static int trace_needs_swap(const char *trace_file, int *swap)
1027 {
1028         struct blk_io_trace t;
1029         int fd, ret;
1030
1031         *swap = -1;
1032
1033         fd = open(trace_file, O_RDONLY);
1034         if (fd < 0) {
1035                 perror("open");
1036                 return 1;
1037         }
1038
1039         ret = read(fd, &t, sizeof(t));
1040         if (ret < 0) {
1041                 close(fd);
1042                 perror("read");
1043                 return 1;
1044         } else if (ret != sizeof(t)) {
1045                 close(fd);
1046                 log_err("fio: short read on trace file\n");
1047                 return 1;
1048         }
1049
1050         close(fd);
1051
1052         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
1053                 *swap = 0;
1054         else {
1055                 /*
1056                  * Maybe it needs to be endian swapped...
1057                  */
1058                 t.magic = fio_swap32(t.magic);
1059                 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
1060                         *swap = 1;
1061         }
1062
1063         if (*swap == -1) {
1064                 log_err("fio: blktrace appears corrupt\n");
1065                 return 1;
1066         }
1067
1068         return 0;
1069 }
1070
1071 int main(int argc, char *argv[])
1072 {
1073         int need_swap, i, c;
1074
1075         if (argc < 2)
1076                 return usage(argv);
1077
1078         while ((c = getopt(argc, argv, "t:n:fd:r:RD:c:u:x:")) != -1) {
1079                 switch (c) {
1080                 case 'R':
1081                         set_rate = 1;
1082                         break;
1083                 case 'r':
1084                         rate_threshold = atoi(optarg);
1085                         break;
1086                 case 't':
1087                         rt_threshold = atoi(optarg);
1088                         break;
1089                 case 'n':
1090                         ios_threshold = atoi(optarg);
1091                         break;
1092                 case 'f':
1093                         output_ascii = 0;
1094                         break;
1095                 case 'd':
1096                         filename = strdup(optarg);
1097                         break;
1098                 case 'D':
1099                         max_depth = atoi(optarg);
1100                         break;
1101                 case 'c':
1102                         collapse_entries = atoi(optarg);
1103                         break;
1104                 case 'u':
1105                         depth_diff = atoi(optarg);
1106                         break;
1107                 case 'x':
1108                         random_diff = atoi(optarg);
1109                         break;
1110                 case '?':
1111                 default:
1112                         return usage(argv);
1113                 }
1114         }
1115
1116         if (argc == optind)
1117                 return usage(argv);
1118
1119         if (trace_needs_swap(argv[optind], &need_swap))
1120                 return 1;
1121
1122         for (i = 0; i < PID_HASH_SIZE; i++)
1123                 INIT_FLIST_HEAD(&pid_hash[i]);
1124         for (i = 0; i < INFLIGHT_HASH_SIZE; i++)
1125                 INIT_FLIST_HEAD(&inflight_hash[i]);
1126
1127         load_blktrace(argv[optind], need_swap);
1128         first_ttime /= 1000ULL;
1129
1130         return output_p();
1131 }