t/btrace2fio.c

   1 #include <stdio.h>
   2 #include <unistd.h>
   3 #include <inttypes.h>
   4 #include <math.h>
   5 #include <assert.h>
   6
   7 #include "../io_ddir.h"
   8 #include "../flist.h"
   9 #include "../hash.h"
  10 #include "../fifo.h"
  11 #include "../blktrace_api.h"
  12 #include "../os/os.h"
  13 #include "../log.h"
  14 #include "../minmax.h"
  15 #include "../oslib/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 static char **add_opts;
  28 static int n_add_opts;
  29
  30 /*
  31  * Collapse defaults
  32  */
  33 static unsigned int collapse_entries = 0;
  34 static unsigned int depth_diff = 1;
  35 static unsigned int random_diff = 5;
  36
  37 struct bs {
  38         unsigned int bs;
  39         unsigned int nr;
  40         int merges;
  41 };
  42
  43 struct trace_file {
  44         char *name;
  45         int major, minor;
  46 };
  47
  48 struct btrace_out {
  49         unsigned long ios[DDIR_RWDIR_CNT];
  50         unsigned long merges[DDIR_RWDIR_CNT];
  51
  52         uint64_t last_end[DDIR_RWDIR_CNT];
  53         uint64_t seq[DDIR_RWDIR_CNT];
  54
  55         struct bs *bs[DDIR_RWDIR_CNT];
  56         unsigned int nr_bs[DDIR_RWDIR_CNT];
  57
  58         int inflight;
  59         unsigned int depth;
  60         int depth_disabled;
  61         int complete_seen;
  62
  63         uint64_t first_ttime[DDIR_RWDIR_CNT];
  64         uint64_t last_ttime[DDIR_RWDIR_CNT];
  65         uint64_t kib[DDIR_RWDIR_CNT];
  66
  67         uint64_t start_delay;
  68 };
  69
  70 struct btrace_pid {
  71         struct flist_head hash_list;
  72         struct flist_head pid_list;
  73         pid_t pid;
  74
  75         pid_t *merge_pids;
  76         unsigned int nr_merge_pids;
  77
  78         struct trace_file *files;
  79         int nr_files;
  80         unsigned int last_major, last_minor;
  81         int numjobs;
  82         int ignore;
  83
  84         struct btrace_out o;
  85 };
  86
  87 struct inflight {
  88         struct flist_head list;
  89         struct btrace_pid *p;
  90         uint64_t end_sector;
  91 };
  92
  93 #define PID_HASH_BITS   10
  94 #define PID_HASH_SIZE   (1U << PID_HASH_BITS)
  95
  96 static struct flist_head pid_hash[PID_HASH_SIZE];
  97 static FLIST_HEAD(pid_list);
  98
  99 #define INFLIGHT_HASH_BITS      8
 100 #define INFLIGHT_HASH_SIZE      (1U << INFLIGHT_HASH_BITS)
 101 static struct flist_head inflight_hash[INFLIGHT_HASH_SIZE];
 102
 103 static uint64_t first_ttime = -1ULL;
 104
 105 static struct inflight *inflight_find(uint64_t sector)
 106 {
 107         struct flist_head *inflight_list;
 108         struct flist_head *e;
 109
 110         inflight_list = &inflight_hash[hash_long(sector, INFLIGHT_HASH_BITS)];
 111
 112         flist_for_each(e, inflight_list) {
 113                 struct inflight *i = flist_entry(e, struct inflight, list);
 114
 115                 if (i->end_sector == sector)
 116                         return i;
 117         }
 118
 119         return NULL;
 120 }
 121
 122 static void inflight_remove(struct inflight *i)
 123 {
 124         struct btrace_out *o = &i->p->o;
 125
 126         o->inflight--;
 127         assert(o->inflight >= 0);
 128         flist_del(&i->list);
 129         free(i);
 130 }
 131
 132 static void __inflight_add(struct inflight *i)
 133 {
 134         struct flist_head *list;
 135
 136         list = &inflight_hash[hash_long(i->end_sector, INFLIGHT_HASH_BITS)];
 137         flist_add_tail(&i->list, list);
 138 }
 139
 140 static void inflight_add(struct btrace_pid *p, uint64_t sector, uint32_t len)
 141 {
 142         struct btrace_out *o = &p->o;
 143         struct inflight *i;
 144
 145         i = calloc(1, sizeof(*i));
 146         i->p = p;
 147         o->inflight++;
 148         if (!o->depth_disabled) {
 149                 o->depth = max((int) o->depth, o->inflight);
 150                 if (o->depth >= max_depth && !o->complete_seen) {
 151                         o->depth_disabled = 1;
 152                         o->depth = max_depth;
 153                 }
 154         }
 155         i->end_sector = sector + (len >> 9);
 156         __inflight_add(i);
 157 }
 158
 159 static void inflight_merge(struct inflight *i, int rw, unsigned int size)
 160 {
 161         i->p->o.merges[rw]++;
 162         if (size) {
 163                 i->end_sector += (size >> 9);
 164                 flist_del(&i->list);
 165                 __inflight_add(i);
 166         }
 167 }
 168
 169 /*
 170  * fifo refill frontend, to avoid reading data in trace sized bites
 171  */
 172 static int refill_fifo(struct fifo *fifo, int fd)
 173 {
 174         char buf[TRACE_FIFO_SIZE];
 175         unsigned int total;
 176         int ret;
 177
 178         total = sizeof(buf);
 179         if (total > fifo_room(fifo))
 180                 total = fifo_room(fifo);
 181
 182         ret = read(fd, buf, total);
 183         if (ret < 0) {
 184                 perror("read refill");
 185                 return -1;
 186         }
 187
 188         if (ret > 0)
 189                 ret = fifo_put(fifo, buf, ret);
 190
 191         return ret;
 192 }
 193
 194 /*
 195  * Retrieve 'len' bytes from the fifo, refilling if necessary.
 196  */
 197 static int trace_fifo_get(struct fifo *fifo, int fd, void *buf,
 198                           unsigned int len)
 199 {
 200         if (fifo_len(fifo) < len) {
 201                 int ret = refill_fifo(fifo, fd);
 202
 203                 if (ret < 0)
 204                         return ret;
 205         }
 206
 207         return fifo_get(fifo, buf, len);
 208 }
 209
 210 /*
 211  * Just discard the pdu by seeking past it.
 212  */
 213 static int discard_pdu(struct fifo *fifo, int fd, struct blk_io_trace *t)
 214 {
 215         if (t->pdu_len == 0)
 216                 return 0;
 217
 218         return trace_fifo_get(fifo, fd, NULL, t->pdu_len);
 219 }
 220
 221 static int handle_trace_notify(struct blk_io_trace *t)
 222 {
 223         switch (t->action) {
 224         case BLK_TN_PROCESS:
 225                 //printf("got process notify: %x, %d\n", t->action, t->pid);
 226                 break;
 227         case BLK_TN_TIMESTAMP:
 228                 //printf("got timestamp notify: %x, %d\n", t->action, t->pid);
 229                 break;
 230         case BLK_TN_MESSAGE:
 231                 break;
 232         default:
 233                 log_err("unknown trace act %x\n", t->action);
 234                 return 1;
 235         }
 236
 237         return 0;
 238 }
 239
 240 static void __add_bs(struct btrace_out *o, unsigned int len, int rw)
 241 {
 242         o->bs[rw] = realloc(o->bs[rw], (o->nr_bs[rw] + 1) * sizeof(struct bs));
 243         o->bs[rw][o->nr_bs[rw]].bs = len;
 244         o->bs[rw][o->nr_bs[rw]].nr = 1;
 245         o->nr_bs[rw]++;
 246 }
 247
 248 static void add_bs(struct btrace_out *o, unsigned int len, int rw)
 249 {
 250         struct bs *bs = o->bs[rw];
 251         int i;
 252
 253         if (!o->nr_bs[rw]) {
 254                 __add_bs(o, len, rw);
 255                 return;
 256         }
 257
 258         for (i = 0; i < o->nr_bs[rw]; i++) {
 259                 if (bs[i].bs == len) {
 260                         bs[i].nr++;
 261                         return;
 262                 }
 263         }
 264
 265         __add_bs(o, len, rw);
 266 }
 267
 268 #define FMINORBITS      20
 269 #define FMINORMASK      ((1U << FMINORBITS) - 1)
 270 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
 271 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
 272
 273 static int btrace_add_file(struct btrace_pid *p, uint32_t devno)
 274 {
 275         unsigned int maj = FMAJOR(devno);
 276         unsigned int min = FMINOR(devno);
 277         struct trace_file *f;
 278         unsigned int i;
 279         char dev[256];
 280
 281         if (filename)
 282                 return 0;
 283         if (p->last_major == maj && p->last_minor == min)
 284                 return 0;
 285
 286         p->last_major = maj;
 287         p->last_minor = min;
 288
 289         /*
 290          * check for this file in our list
 291          */
 292         for (i = 0; i < p->nr_files; i++) {
 293                 f = &p->files[i];
 294
 295                 if (f->major == maj && f->minor == min)
 296                         return 0;
 297         }
 298
 299         strcpy(dev, "/dev");
 300         if (!blktrace_lookup_device(NULL, dev, maj, min)) {
 301                 log_err("fio: failed to find device %u/%u\n", maj, min);
 302                 if (!output_ascii) {
 303                         log_err("fio: use -d to specify device\n");
 304                         return 1;
 305                 }
 306                 return 0;
 307         }
 308
 309         p->files = realloc(p->files, (p->nr_files + 1) * sizeof(*f));
 310         f = &p->files[p->nr_files];
 311         f->name = strdup(dev);
 312         f->major = maj;
 313         f->minor = min;
 314         p->nr_files++;
 315         return 0;
 316 }
 317
 318 static int t_to_rwdir(struct blk_io_trace *t)
 319 {
 320         if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 321                 return DDIR_TRIM;
 322
 323         return (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 324 }
 325
 326 static int handle_trace_discard(struct blk_io_trace *t, struct btrace_pid *p)
 327 {
 328         struct btrace_out *o = &p->o;
 329
 330         if (btrace_add_file(p, t->device))
 331                 return 1;
 332
 333         if (o->first_ttime[2] == -1ULL)
 334                 o->first_ttime[2] = t->time;
 335
 336         o->ios[DDIR_TRIM]++;
 337         add_bs(o, t->bytes, DDIR_TRIM);
 338         return 0;
 339 }
 340
 341 static int handle_trace_fs(struct blk_io_trace *t, struct btrace_pid *p)
 342 {
 343         struct btrace_out *o = &p->o;
 344         int rw;
 345
 346         if (btrace_add_file(p, t->device))
 347                 return 1;
 348
 349         first_ttime = min(first_ttime, (uint64_t) t->time);
 350
 351         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 352
 353         if (o->first_ttime[rw] == -1ULL)
 354                 o->first_ttime[rw] = t->time;
 355
 356         add_bs(o, t->bytes, rw);
 357         o->ios[rw]++;
 358
 359         if (t->sector == o->last_end[rw] || o->last_end[rw] == -1ULL)
 360                 o->seq[rw]++;
 361
 362         o->last_end[rw] = t->sector + (t->bytes >> 9);
 363         return 0;
 364 }
 365
 366 static int handle_queue_trace(struct blk_io_trace *t, struct btrace_pid *p)
 367 {
 368         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 369                 return handle_trace_notify(t);
 370         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 371                 return handle_trace_discard(t, p);
 372         else
 373                 return handle_trace_fs(t, p);
 374 }
 375
 376 static int handle_trace(struct blk_io_trace *t, struct btrace_pid *p)
 377 {
 378         unsigned int act = t->action & 0xffff;
 379         int ret = 0;
 380
 381         if (act == __BLK_TA_QUEUE) {
 382                 inflight_add(p, t->sector, t->bytes);
 383                 ret = handle_queue_trace(t, p);
 384         } else if (act == __BLK_TA_BACKMERGE) {
 385                 struct inflight *i;
 386
 387                 i = inflight_find(t->sector + (t->bytes >> 9));
 388                 if (i)
 389                         inflight_remove(i);
 390
 391                 i = inflight_find(t->sector);
 392                 if (i)
 393                         inflight_merge(i, t_to_rwdir(t), t->bytes);
 394         } else if (act == __BLK_TA_FRONTMERGE) {
 395                 struct inflight *i;
 396
 397                 i = inflight_find(t->sector + (t->bytes >> 9));
 398                 if (i)
 399                         inflight_remove(i);
 400
 401                 i = inflight_find(t->sector);
 402                 if (i)
 403                         inflight_merge(i, t_to_rwdir(t), 0);
 404         } else if (act == __BLK_TA_COMPLETE) {
 405                 struct inflight *i;
 406
 407                 i = inflight_find(t->sector + (t->bytes >> 9));
 408                 if (i) {
 409                         i->p->o.kib[t_to_rwdir(t)] += (t->bytes >> 10);
 410                         i->p->o.complete_seen = 1;
 411                         inflight_remove(i);
 412                 }
 413         }
 414
 415         return ret;
 416 }
 417
 418 static void byteswap_trace(struct blk_io_trace *t)
 419 {
 420         t->magic = fio_swap32(t->magic);
 421         t->sequence = fio_swap32(t->sequence);
 422         t->time = fio_swap64(t->time);
 423         t->sector = fio_swap64(t->sector);
 424         t->bytes = fio_swap32(t->bytes);
 425         t->action = fio_swap32(t->action);
 426         t->pid = fio_swap32(t->pid);
 427         t->device = fio_swap32(t->device);
 428         t->cpu = fio_swap32(t->cpu);
 429         t->error = fio_swap16(t->error);
 430         t->pdu_len = fio_swap16(t->pdu_len);
 431 }
 432
 433 static struct btrace_pid *pid_hash_find(pid_t pid, struct flist_head *list)
 434 {
 435         struct flist_head *e;
 436         struct btrace_pid *p;
 437
 438         flist_for_each(e, list) {
 439                 p = flist_entry(e, struct btrace_pid, hash_list);
 440                 if (p->pid == pid)
 441                         return p;
 442         }
 443
 444         return NULL;
 445 }
 446
 447 static struct btrace_pid *pid_hash_get(pid_t pid)
 448 {
 449         struct flist_head *hash_list;
 450         struct btrace_pid *p;
 451
 452         hash_list = &pid_hash[hash_long(pid, PID_HASH_BITS)];
 453
 454         p = pid_hash_find(pid, hash_list);
 455         if (!p) {
 456                 int i;
 457
 458                 p = calloc(1, sizeof(*p));
 459
 460                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 461                         p->o.first_ttime[i] = -1ULL;
 462                         p->o.last_ttime[i] = -1ULL;
 463                         p->o.last_end[i] = -1ULL;
 464                 }
 465
 466                 p->pid = pid;
 467                 p->numjobs = 1;
 468                 flist_add_tail(&p->hash_list, hash_list);
 469                 flist_add_tail(&p->pid_list, &pid_list);
 470         }
 471
 472         return p;
 473 }
 474
 475 /*
 476  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 477  * them as io_pieces like the fio text version would do.
 478  */
 479 static int load_blktrace(const char *fname, int need_swap)
 480 {
 481         struct btrace_pid *p;
 482         unsigned long traces;
 483         struct blk_io_trace t;
 484         struct fifo *fifo;
 485         int fd, ret = 0;
 486
 487         fd = open(fname, O_RDONLY);
 488         if (fd < 0) {
 489                 perror("open trace file\n");
 490                 return 1;
 491         }
 492
 493         fifo = fifo_alloc(TRACE_FIFO_SIZE);
 494
 495         traces = 0;
 496         do {
 497                 ret = trace_fifo_get(fifo, fd, &t, sizeof(t));
 498                 if (ret < 0)
 499                         goto err;
 500                 else if (!ret)
 501                         break;
 502                 else if (ret < (int) sizeof(t)) {
 503                         log_err("fio: short fifo get\n");
 504                         break;
 505                 }
 506
 507                 if (need_swap)
 508                         byteswap_trace(&t);
 509
 510                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 511                         log_err("fio: bad magic in blktrace data: %x\n", t.magic);
 512                         goto err;
 513                 }
 514                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 515                         log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
 516                         goto err;
 517                 }
 518                 ret = discard_pdu(fifo, fd, &t);
 519                 if (ret < 0) {
 520                         log_err("blktrace lseek\n");
 521                         goto err;
 522                 } else if (t.pdu_len != ret) {
 523                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
 524                         goto err;
 525                 }
 526
 527                 p = pid_hash_get(t.pid);
 528                 ret = handle_trace(&t, p);
 529                 if (ret)
 530                         break;
 531                 p->o.last_ttime[t_to_rwdir(&t)] = t.time;
 532                 traces++;
 533         } while (1);
 534
 535         fifo_free(fifo);
 536         close(fd);
 537
 538         if (ret)
 539                 return ret;
 540
 541         if (output_ascii)
 542                 printf("Traces loaded: %lu\n", traces);
 543
 544         return 0;
 545 err:
 546         close(fd);
 547         fifo_free(fifo);
 548         return 1;
 549 }
 550
 551 static int bs_cmp(const void *ba, const void *bb)
 552 {
 553         const struct bs *bsa = ba;
 554         const struct bs *bsb = bb;
 555
 556         return bsb->nr - bsa->nr;
 557 }
 558
 559 static unsigned long o_to_kib_rate(struct btrace_out *o, int rw)
 560 {
 561         uint64_t usec = (o->last_ttime[rw] - o->first_ttime[rw]) / 1000ULL;
 562         uint64_t val;
 563
 564         if (!usec)
 565                 return 0;
 566
 567         usec /= 1000;
 568         if (!usec)
 569                 return 0;
 570
 571         val = o->kib[rw] * 1000ULL;
 572         return val / usec;
 573 }
 574
 575 static uint64_t o_first_ttime(struct btrace_out *o)
 576 {
 577         uint64_t first;
 578
 579         first = min(o->first_ttime[0], o->first_ttime[1]);
 580         return min(first, o->first_ttime[2]);
 581 }
 582
 583 static uint64_t o_longest_ttime(struct btrace_out *o)
 584 {
 585         uint64_t ret = 0;
 586         int i;
 587
 588         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 589                 uint64_t diff;
 590
 591                 diff = o->last_ttime[i] - o->first_ttime[i];
 592                 ret = max(diff, ret);
 593         }
 594
 595         return ret;
 596 }
 597
 598 static void __output_p_ascii(struct btrace_pid *p, unsigned long *ios)
 599 {
 600         const char *msg[] = { "reads", "writes", "trims" };
 601         struct btrace_out *o = &p->o;
 602         unsigned long total, usec;
 603         int i, j;
 604
 605         printf("[pid:\t%u", p->pid);
 606         if (p->nr_merge_pids)
 607                 for (i = 0; i < p->nr_merge_pids; i++)
 608                         printf(", %u", p->merge_pids[i]);
 609         printf("]\n");
 610
 611         total = ddir_rw_sum(o->ios);
 612         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 613                 float perc;
 614
 615                 if (!o->ios[i])
 616                         continue;
 617
 618                 ios[i] += o->ios[i] + o->merges[i];
 619                 printf("%s\n", msg[i]);
 620                 perc = ((float) o->ios[i] * 100.0) / (float) total;
 621                 printf("\tios:    %lu (perc=%3.2f%%)\n", o->ios[i], perc);
 622                 perc = ((float) o->merges[i] * 100.0) / (float) total;
 623                 printf("\tmerges: %lu (perc=%3.2f%%)\n", o->merges[i], perc);
 624                 perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
 625                 printf("\tseq:    %lu (perc=%3.2f%%)\n", (unsigned long) o->seq[i], perc);
 626                 printf("\trate:   %lu KiB/sec\n", o_to_kib_rate(o, i));
 627
 628                 for (j = 0; j < o->nr_bs[i]; j++) {
 629                         struct bs *bs = &o->bs[i][j];
 630
 631                         perc = (((float) bs->nr * 100.0) / (float) o->ios[i]);
 632                         printf("\tbs=%u, perc=%3.2f%%\n", bs->bs, perc);
 633                 }
 634         }
 635
 636         printf("depth:\t%u\n", o->depth);
 637         usec = o_longest_ttime(o) / 1000ULL;
 638         printf("usec:\t%lu (delay=%llu)\n", usec, (unsigned long long) o->start_delay);
 639
 640         printf("files:\t");
 641         for (i = 0; i < p->nr_files; i++)
 642                 printf("%s,", p->files[i].name);
 643         printf("\n");
 644
 645         printf("\n");
 646 }
 647
 648 static int __output_p_fio(struct btrace_pid *p, unsigned long *ios,
 649                           const char *name_postfix)
 650 {
 651         struct btrace_out *o = &p->o;
 652         unsigned long total;
 653         unsigned long long time;
 654         float perc;
 655         int i, j;
 656
 657         if ((o->ios[0] + o->ios[1]) && o->ios[2]) {
 658                 unsigned long ios_bak[DDIR_RWDIR_CNT];
 659
 660                 memcpy(ios_bak, o->ios, DDIR_RWDIR_CNT * sizeof(unsigned long));
 661
 662                 /* create job for read/write */
 663                 o->ios[2] = 0;
 664                 __output_p_fio(p, ios, "");
 665                 o->ios[2] = ios_bak[2];
 666
 667                 /* create job for trim */
 668                 o->ios[0] = 0;
 669                 o->ios[1] = 0;
 670                 __output_p_fio(p, ios, "_trim");
 671                 o->ios[0] = ios_bak[0];
 672                 o->ios[1] = ios_bak[1];
 673
 674                 return 0;
 675         }
 676         if (!p->nr_files) {
 677                 log_err("fio: no devices found\n");
 678                 return 1;
 679         }
 680
 681         printf("[pid%u%s", p->pid, name_postfix);
 682         if (p->nr_merge_pids)
 683                 for (i = 0; i < p->nr_merge_pids; i++)
 684                         printf(",pid%u", p->merge_pids[i]);
 685         printf("]\n");
 686
 687         printf("numjobs=%u\n", p->numjobs);
 688         printf("direct=1\n");
 689         if (o->depth == 1)
 690                 printf("ioengine=sync\n");
 691         else
 692                 printf("ioengine=libaio\niodepth=%u\n", o->depth);
 693
 694         if (o->ios[0] && !o->ios[1])
 695                 printf("rw=randread\n");
 696         else if (!o->ios[0] && o->ios[1])
 697                 printf("rw=randwrite\n");
 698         else if (o->ios[2])
 699                 printf("rw=randtrim\n");
 700         else {
 701                 printf("rw=randrw\n");
 702                 total = ddir_rw_sum(o->ios);
 703                 perc = ((float) o->ios[0] * 100.0) / (float) total;
 704                 printf("rwmixread=%u\n", (int) floor(perc + 0.50));
 705         }
 706
 707         printf("percentage_random=");
 708         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 709                 if (o->seq[i] && o->ios[i]) {
 710                         perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
 711                         if (perc >= 99.0)
 712                                 perc = 100.0;
 713                 } else
 714                         perc = 100.0;
 715
 716                 if (i)
 717                         printf(",");
 718                 perc = 100.0 - perc;
 719                 printf("%u", (int) floor(perc + 0.5));
 720         }
 721         printf("\n");
 722
 723         printf("filename=");
 724         for (i = 0; i < p->nr_files; i++) {
 725                 if (i)
 726                         printf(":");
 727                 printf("%s", p->files[i].name);
 728         }
 729         printf("\n");
 730
 731         if (o->start_delay / 1000000ULL)
 732                 printf("startdelay=%llus\n", o->start_delay / 1000000ULL);
 733
 734         time = o_longest_ttime(o);
 735         time = (time + 1000000000ULL - 1) / 1000000000ULL;
 736         printf("runtime=%llus\n", time);
 737
 738         printf("bssplit=");
 739         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 740
 741                 if (i && o->nr_bs[i - 1] && o->nr_bs[i])
 742                         printf(",");
 743
 744                 for (j = 0; j < o->nr_bs[i]; j++) {
 745                         struct bs *bs = &o->bs[i][j];
 746
 747                         perc = (((float) bs->nr * 100.0) / (float) o->ios[i]);
 748                         if (perc < 1.00)
 749                                 continue;
 750                         if (j)
 751                                 printf(":");
 752                         if (j + 1 == o->nr_bs[i])
 753                                 printf("%u/", bs->bs);
 754                         else
 755                                 printf("%u/%u", bs->bs, (int) floor(perc + 0.5));
 756                 }
 757         }
 758         printf("\n");
 759
 760         if (set_rate) {
 761                 printf("rate=");
 762                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 763                         unsigned long rate;
 764
 765                         rate = o_to_kib_rate(o, i);
 766                         if (i)
 767                                 printf(",");
 768                         if (rate)
 769                                 printf("%luk", rate);
 770                 }
 771                 printf("\n");
 772         }
 773
 774         if (n_add_opts)
 775                 for (i = 0; i < n_add_opts; i++)
 776                         printf("%s\n", add_opts[i]);
 777
 778         printf("\n");
 779         return 0;
 780 }
 781
 782 static int __output_p(struct btrace_pid *p, unsigned long *ios)
 783 {
 784         struct btrace_out *o = &p->o;
 785         int i, ret = 0;
 786
 787         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 788                 if (o->nr_bs[i] <= 1)
 789                         continue;
 790                 qsort(o->bs[i], o->nr_bs[i], sizeof(struct bs), bs_cmp);
 791         }
 792
 793         if (filename) {
 794                 p->files = malloc(sizeof(struct trace_file));
 795                 p->nr_files++;
 796                 p->files[0].name = filename;
 797         }
 798
 799         if (output_ascii)
 800                 __output_p_ascii(p, ios);
 801         else
 802                 ret = __output_p_fio(p, ios, "");
 803
 804         return ret;
 805 }
 806
 807 static void remove_ddir(struct btrace_out *o, int rw)
 808 {
 809         o->ios[rw] = 0;
 810 }
 811
 812 static int prune_entry(struct btrace_out *o)
 813 {
 814         unsigned long rate;
 815         uint64_t time;
 816         int i;
 817
 818         if (ddir_rw_sum(o->ios) < ios_threshold)
 819                 return 1;
 820
 821         time = o_longest_ttime(o) / 1000ULL;
 822         if (time < rt_threshold)
 823                 return 1;
 824
 825         rate = 0;
 826         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 827                 unsigned long this_rate;
 828
 829                 this_rate = o_to_kib_rate(o, i);
 830                 if (this_rate < rate_threshold) {
 831                         remove_ddir(o, i);
 832                         this_rate = 0;
 833                 }
 834                 rate += this_rate;
 835         }
 836
 837         if (rate < rate_threshold)
 838                 return 1;
 839
 840         return 0;
 841 }
 842
 843 static int entry_cmp(void *priv, struct flist_head *a, struct flist_head *b)
 844 {
 845         struct btrace_pid *pa = flist_entry(a, struct btrace_pid, pid_list);
 846         struct btrace_pid *pb = flist_entry(b, struct btrace_pid, pid_list);
 847
 848         return ddir_rw_sum(pb->o.ios) - ddir_rw_sum(pa->o.ios);
 849 }
 850
 851 static void free_p(struct btrace_pid *p)
 852 {
 853         struct btrace_out *o = &p->o;
 854         int i;
 855
 856         for (i = 0; i < p->nr_files; i++) {
 857                 if (p->files[i].name && p->files[i].name != filename)
 858                         free(p->files[i].name);
 859         }
 860
 861         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 862                 free(o->bs[i]);
 863
 864         free(p->files);
 865         flist_del(&p->pid_list);
 866         flist_del(&p->hash_list);
 867         free(p);
 868 }
 869
 870 static int entries_close(struct btrace_pid *pida, struct btrace_pid *pidb)
 871 {
 872         float perca, percb, fdiff;
 873         int i, idiff;
 874
 875         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 876                 if ((pida->o.ios[i] && !pidb->o.ios[i]) ||
 877                     (pidb->o.ios[i] && !pida->o.ios[i]))
 878                         return 0;
 879                 if (pida->o.ios[i] && pidb->o.ios[i]) {
 880                         perca = ((float) pida->o.seq[i] * 100.0) / (float) pida->o.ios[i];
 881                         percb = ((float) pidb->o.seq[i] * 100.0) / (float) pidb->o.ios[i];
 882                         fdiff = perca - percb;
 883                         if (fabs(fdiff) > random_diff)
 884                                 return 0;
 885                 }
 886
 887                 idiff = pida->o.depth - pidb->o.depth;
 888                 if (abs(idiff) > depth_diff)
 889                         return 0;
 890         }
 891
 892         return 1;
 893 }
 894
 895 static void merge_bs(struct bs **bsap, unsigned int *nr_bsap,
 896                      struct bs *bsb, unsigned int nr_bsb)
 897 {
 898         struct bs *bsa = *bsap;
 899         unsigned int nr_bsa = *nr_bsap;
 900         int a, b;
 901
 902         for (b = 0; b < nr_bsb; b++) {
 903                 int next, found = 0;
 904
 905                 for (a = 0; a < nr_bsa; a++) {
 906                         if (bsb[b].bs != bsa[a].bs)
 907                                 continue;
 908
 909                         bsa[a].nr += bsb[b].nr;
 910                         bsa[a].merges += bsb[b].merges;
 911                         found = 1;
 912                         break;
 913                 }
 914
 915                 if (found)
 916                         continue;
 917
 918                 next = *nr_bsap;
 919                 bsa = realloc(bsa, (next + 1) * sizeof(struct bs));
 920                 bsa[next].bs = bsb[b].bs;
 921                 bsa[next].nr = bsb[b].nr;
 922                 (*nr_bsap)++;
 923                 *bsap = bsa;
 924         }
 925 }
 926
 927 static int merge_entries(struct btrace_pid *pida, struct btrace_pid *pidb)
 928 {
 929         int i;
 930
 931         if (!entries_close(pida, pidb))
 932                 return 0;
 933
 934         pida->nr_merge_pids++;
 935         pida->merge_pids = realloc(pida->merge_pids, pida->nr_merge_pids * sizeof(pid_t));
 936         pida->merge_pids[pida->nr_merge_pids - 1] = pidb->pid;
 937
 938         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 939                 struct btrace_out *oa = &pida->o;
 940                 struct btrace_out *ob = &pidb->o;
 941
 942                 oa->ios[i] += ob->ios[i];
 943                 oa->merges[i] += ob->merges[i];
 944                 oa->seq[i] += ob->seq[i];
 945                 oa->kib[i] += ob->kib[i];
 946                 oa->first_ttime[i] = min(oa->first_ttime[i], ob->first_ttime[i]);
 947                 oa->last_ttime[i] = max(oa->last_ttime[i], ob->last_ttime[i]);
 948                 merge_bs(&oa->bs[i], &oa->nr_bs[i], ob->bs[i], ob->nr_bs[i]);
 949         }
 950
 951         pida->o.start_delay = min(pida->o.start_delay, pidb->o.start_delay);
 952         pida->o.depth = (pida->o.depth + pidb->o.depth) / 2;
 953         return 1;
 954 }
 955
 956 static void check_merges(struct btrace_pid *p, struct flist_head *pidlist)
 957 {
 958         struct flist_head *e, *tmp;
 959
 960         if (p->ignore)
 961                 return;
 962
 963         flist_for_each_safe(e, tmp, pidlist) {
 964                 struct btrace_pid *pidb;
 965
 966                 pidb = flist_entry(e, struct btrace_pid, pid_list);
 967                 if (pidb == p)
 968                         continue;
 969
 970                 if (merge_entries(p, pidb)) {
 971                         pidb->ignore = 1;
 972                         p->numjobs++;
 973                 }
 974         }
 975 }
 976
 977 static int output_p(void)
 978 {
 979         unsigned long ios[DDIR_RWDIR_CNT];
 980         struct flist_head *e, *tmp;
 981         int depth_disabled = 0;
 982         int ret = 0;
 983
 984         flist_for_each_safe(e, tmp, &pid_list) {
 985                 struct btrace_pid *p;
 986
 987                 p = flist_entry(e, struct btrace_pid, pid_list);
 988                 if (prune_entry(&p->o)) {
 989                         free_p(p);
 990                         continue;
 991                 }
 992                 p->o.start_delay = (o_first_ttime(&p->o) / 1000ULL) - first_ttime;
 993                 depth_disabled += p->o.depth_disabled;
 994         }
 995
 996         if (collapse_entries) {
 997                 struct btrace_pid *p;
 998
 999                 flist_for_each_safe(e, tmp, &pid_list) {
1000                         p = flist_entry(e, struct btrace_pid, pid_list);
1001                         check_merges(p, &pid_list);
1002                 }
1003
1004                 flist_for_each_safe(e, tmp, &pid_list) {
1005                         p = flist_entry(e, struct btrace_pid, pid_list);
1006                         if (p->ignore)
1007                                 free_p(p);
1008                 }
1009         }
1010
1011         if (depth_disabled)
1012                 log_err("fio: missing completion traces, depths capped at %u\n", max_depth);
1013
1014         memset(ios, 0, sizeof(ios));
1015
1016         flist_sort(NULL, &pid_list, entry_cmp);
1017
1018         flist_for_each(e, &pid_list) {
1019                 struct btrace_pid *p;
1020
1021                 p = flist_entry(e, struct btrace_pid, pid_list);
1022                 ret |= __output_p(p, ios);
1023                 if (ret && !output_ascii)
1024                         break;
1025         }
1026
1027         if (output_ascii)
1028                 printf("Total: reads=%lu, writes=%lu\n", ios[0], ios[1]);
1029
1030         return ret;
1031 }
1032
1033 static int usage(char *argv[])
1034 {
1035         log_err("%s: [options] <blktrace bin file>\n", argv[0]);
1036         log_err("\t-t\tUsec threshold to ignore task\n");
1037         log_err("\t-n\tNumber IOS threshold to ignore task\n");
1038         log_err("\t-f\tFio job file output\n");
1039         log_err("\t-d\tUse this file/device for replay\n");
1040         log_err("\t-r\tIgnore jobs with less than this KiB/sec rate\n");
1041         log_err("\t-R\tSet rate in fio job (def=%u)\n", set_rate);
1042         log_err("\t-D\tCap queue depth at this value (def=%u)\n", max_depth);
1043         log_err("\t-c\tCollapse \"identical\" jobs (def=%u)\n", collapse_entries);
1044         log_err("\t-u\tDepth difference for collapse (def=%u)\n", depth_diff);
1045         log_err("\t-x\tRandom difference for collapse (def=%u)\n", random_diff);
1046         log_err("\t-a\tAdditional fio option to add to job file\n");
1047         return 1;
1048 }
1049
1050 static int trace_needs_swap(const char *trace_file, int *swap)
1051 {
1052         struct blk_io_trace t;
1053         int fd, ret;
1054
1055         *swap = -1;
1056
1057         fd = open(trace_file, O_RDONLY);
1058         if (fd < 0) {
1059                 perror("open");
1060                 return 1;
1061         }
1062
1063         ret = read(fd, &t, sizeof(t));
1064         if (ret < 0) {
1065                 close(fd);
1066                 perror("read");
1067                 return 1;
1068         } else if (ret != sizeof(t)) {
1069                 close(fd);
1070                 log_err("fio: short read on trace file\n");
1071                 return 1;
1072         }
1073
1074         close(fd);
1075
1076         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
1077                 *swap = 0;
1078         else {
1079                 /*
1080                  * Maybe it needs to be endian swapped...
1081                  */
1082                 t.magic = fio_swap32(t.magic);
1083                 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
1084                         *swap = 1;
1085         }
1086
1087         if (*swap == -1) {
1088                 log_err("fio: blktrace appears corrupt\n");
1089                 return 1;
1090         }
1091
1092         return 0;
1093 }
1094
1095 int main(int argc, char *argv[])
1096 {
1097         int need_swap, i, c;
1098
1099         if (argc < 2)
1100                 return usage(argv);
1101
1102         while ((c = getopt(argc, argv, "t:n:fd:r:RD:c:u:x:a:")) != -1) {
1103                 switch (c) {
1104                 case 'R':
1105                         set_rate = 1;
1106                         break;
1107                 case 'r':
1108                         rate_threshold = atoi(optarg);
1109                         break;
1110                 case 't':
1111                         rt_threshold = atoi(optarg);
1112                         break;
1113                 case 'n':
1114                         ios_threshold = atoi(optarg);
1115                         break;
1116                 case 'f':
1117                         output_ascii = 0;
1118                         break;
1119                 case 'd':
1120                         filename = strdup(optarg);
1121                         break;
1122                 case 'D':
1123                         max_depth = atoi(optarg);
1124                         break;
1125                 case 'c':
1126                         collapse_entries = atoi(optarg);
1127                         break;
1128                 case 'u':
1129                         depth_diff = atoi(optarg);
1130                         break;
1131                 case 'x':
1132                         random_diff = atoi(optarg);
1133                         break;
1134                 case 'a':
1135                         add_opts = realloc(add_opts, (n_add_opts + 1) * sizeof(char *));
1136                         add_opts[n_add_opts] = strdup(optarg);
1137                         n_add_opts++;
1138                         break;
1139                 case '?':
1140                 default:
1141                         return usage(argv);
1142                 }
1143         }
1144
1145         if (argc == optind)
1146                 return usage(argv);
1147
1148         if (trace_needs_swap(argv[optind], &need_swap))
1149                 return 1;
1150
1151         for (i = 0; i < PID_HASH_SIZE; i++)
1152                 INIT_FLIST_HEAD(&pid_hash[i]);
1153         for (i = 0; i < INFLIGHT_HASH_SIZE; i++)
1154                 INIT_FLIST_HEAD(&inflight_hash[i]);
1155
1156         load_blktrace(argv[optind], need_swap);
1157         first_ttime /= 1000ULL;
1158
1159         return output_p();
1160 }