iolog.c

   1 /*
   2  * Code related to writing an iolog of what a thread is doing, and to
   3  * later read that back and replay
   4  */
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <libgen.h>
   8 #include <assert.h>
   9 #include <sys/types.h>
  10 #include <sys/stat.h>
  11 #include <unistd.h>
  12 #ifdef CONFIG_ZLIB
  13 #include <zlib.h>
  14 #endif
  15
  16 #include "flist.h"
  17 #include "fio.h"
  18 #include "verify.h"
  19 #include "trim.h"
  20 #include "filelock.h"
  21 #include "smalloc.h"
  22
  23 static int iolog_flush(struct io_log *log);
  24
  25 static const char iolog_ver2[] = "fio version 2 iolog";
  26
  27 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  28 {
  29         flist_add_tail(&ipo->list, &td->io_log_list);
  30         td->total_io_size += ipo->len;
  31 }
  32
  33 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  34 {
  35         if (!td->o.write_iolog_file)
  36                 return;
  37
  38         fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
  39                                                 io_ddir_name(io_u->ddir),
  40                                                 io_u->offset, io_u->buflen);
  41 }
  42
  43 void log_file(struct thread_data *td, struct fio_file *f,
  44               enum file_log_act what)
  45 {
  46         const char *act[] = { "add", "open", "close" };
  47
  48         assert(what < 3);
  49
  50         if (!td->o.write_iolog_file)
  51                 return;
  52
  53
  54         /*
  55          * this happens on the pre-open/close done before the job starts
  56          */
  57         if (!td->iolog_f)
  58                 return;
  59
  60         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  61 }
  62
  63 static void iolog_delay(struct thread_data *td, unsigned long delay)
  64 {
  65         uint64_t usec = utime_since_now(&td->last_issue);
  66         uint64_t this_delay;
  67         struct timeval tv;
  68
  69         if (delay < td->time_offset) {
  70                 td->time_offset = 0;
  71                 return;
  72         }
  73
  74         delay -= td->time_offset;
  75         if (delay < usec)
  76                 return;
  77
  78         delay -= usec;
  79
  80         fio_gettime(&tv, NULL);
  81         while (delay && !td->terminate) {
  82                 this_delay = delay;
  83                 if (this_delay > 500000)
  84                         this_delay = 500000;
  85
  86                 usec_sleep(td, this_delay);
  87                 delay -= this_delay;
  88         }
  89
  90         usec = utime_since_now(&tv);
  91         if (usec > delay)
  92                 td->time_offset = usec - delay;
  93         else
  94                 td->time_offset = 0;
  95 }
  96
  97 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
  98 {
  99         struct fio_file *f;
 100         int ret;
 101
 102         /*
 103          * Not a special ipo
 104          */
 105         if (ipo->ddir != DDIR_INVAL)
 106                 return 0;
 107
 108         f = td->files[ipo->fileno];
 109
 110         switch (ipo->file_action) {
 111         case FIO_LOG_OPEN_FILE:
 112                 ret = td_io_open_file(td, f);
 113                 if (!ret)
 114                         break;
 115                 td_verror(td, ret, "iolog open file");
 116                 return -1;
 117         case FIO_LOG_CLOSE_FILE:
 118                 td_io_close_file(td, f);
 119                 break;
 120         case FIO_LOG_UNLINK_FILE:
 121                 td_io_unlink_file(td, f);
 122                 break;
 123         default:
 124                 log_err("fio: bad file action %d\n", ipo->file_action);
 125                 break;
 126         }
 127
 128         return 1;
 129 }
 130
 131 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 132 {
 133         struct io_piece *ipo;
 134         unsigned long elapsed;
 135
 136         while (!flist_empty(&td->io_log_list)) {
 137                 int ret;
 138
 139                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 140                 flist_del(&ipo->list);
 141                 remove_trim_entry(td, ipo);
 142
 143                 ret = ipo_special(td, ipo);
 144                 if (ret < 0) {
 145                         free(ipo);
 146                         break;
 147                 } else if (ret > 0) {
 148                         free(ipo);
 149                         continue;
 150                 }
 151
 152                 io_u->ddir = ipo->ddir;
 153                 if (ipo->ddir != DDIR_WAIT) {
 154                         io_u->offset = ipo->offset;
 155                         io_u->buflen = ipo->len;
 156                         io_u->file = td->files[ipo->fileno];
 157                         get_file(io_u->file);
 158                         dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
 159                                                 io_u->buflen, io_u->file->file_name);
 160                         if (ipo->delay)
 161                                 iolog_delay(td, ipo->delay);
 162                 } else {
 163                         elapsed = mtime_since_genesis();
 164                         if (ipo->delay > elapsed)
 165                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 166                 }
 167
 168                 free(ipo);
 169
 170                 if (io_u->ddir != DDIR_WAIT)
 171                         return 0;
 172         }
 173
 174         td->done = 1;
 175         return 1;
 176 }
 177
 178 void prune_io_piece_log(struct thread_data *td)
 179 {
 180         struct io_piece *ipo;
 181         struct rb_node *n;
 182
 183         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 184                 ipo = rb_entry(n, struct io_piece, rb_node);
 185                 rb_erase(n, &td->io_hist_tree);
 186                 remove_trim_entry(td, ipo);
 187                 td->io_hist_len--;
 188                 free(ipo);
 189         }
 190
 191         while (!flist_empty(&td->io_hist_list)) {
 192                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 193                 flist_del(&ipo->list);
 194                 remove_trim_entry(td, ipo);
 195                 td->io_hist_len--;
 196                 free(ipo);
 197         }
 198 }
 199
 200 /*
 201  * log a successful write, so we can unwind the log for verify
 202  */
 203 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 204 {
 205         struct rb_node **p, *parent;
 206         struct io_piece *ipo, *__ipo;
 207
 208         ipo = malloc(sizeof(struct io_piece));
 209         init_ipo(ipo);
 210         ipo->file = io_u->file;
 211         ipo->offset = io_u->offset;
 212         ipo->len = io_u->buflen;
 213         ipo->numberio = io_u->numberio;
 214         ipo->flags = IP_F_IN_FLIGHT;
 215
 216         io_u->ipo = ipo;
 217
 218         if (io_u_should_trim(td, io_u)) {
 219                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 220                 td->trim_entries++;
 221         }
 222
 223         /*
 224          * We don't need to sort the entries, if:
 225          *
 226          *      Sequential writes, or
 227          *      Random writes that lay out the file as it goes along
 228          *
 229          * For both these cases, just reading back data in the order we
 230          * wrote it out is the fastest.
 231          *
 232          * One exception is if we don't have a random map AND we are doing
 233          * verifies, in that case we need to check for duplicate blocks and
 234          * drop the old one, which we rely on the rb insert/lookup for
 235          * handling.
 236          */
 237         if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
 238               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
 239                 INIT_FLIST_HEAD(&ipo->list);
 240                 flist_add_tail(&ipo->list, &td->io_hist_list);
 241                 ipo->flags |= IP_F_ONLIST;
 242                 td->io_hist_len++;
 243                 return;
 244         }
 245
 246         RB_CLEAR_NODE(&ipo->rb_node);
 247
 248         /*
 249          * Sort the entry into the verification list
 250          */
 251 restart:
 252         p = &td->io_hist_tree.rb_node;
 253         parent = NULL;
 254         while (*p) {
 255                 int overlap = 0;
 256                 parent = *p;
 257
 258                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 259                 if (ipo->file < __ipo->file)
 260                         p = &(*p)->rb_left;
 261                 else if (ipo->file > __ipo->file)
 262                         p = &(*p)->rb_right;
 263                 else if (ipo->offset < __ipo->offset) {
 264                         p = &(*p)->rb_left;
 265                         overlap = ipo->offset + ipo->len > __ipo->offset;
 266                 }
 267                 else if (ipo->offset > __ipo->offset) {
 268                         p = &(*p)->rb_right;
 269                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 270                 }
 271                 else
 272                         overlap = 1;
 273
 274                 if (overlap) {
 275                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
 276                                 __ipo->offset, __ipo->len,
 277                                 ipo->offset, ipo->len);
 278                         td->io_hist_len--;
 279                         rb_erase(parent, &td->io_hist_tree);
 280                         remove_trim_entry(td, __ipo);
 281                         free(__ipo);
 282                         goto restart;
 283                 }
 284         }
 285
 286         rb_link_node(&ipo->rb_node, parent, p);
 287         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 288         ipo->flags |= IP_F_ONRB;
 289         td->io_hist_len++;
 290 }
 291
 292 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 293 {
 294         struct io_piece *ipo = io_u->ipo;
 295
 296         if (td->ts.nr_block_infos) {
 297                 uint32_t *info = io_u_block_info(td, io_u);
 298                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 299                         if (io_u->ddir == DDIR_TRIM)
 300                                 *info = BLOCK_INFO_SET_STATE(*info,
 301                                                 BLOCK_STATE_TRIM_FAILURE);
 302                         else if (io_u->ddir == DDIR_WRITE)
 303                                 *info = BLOCK_INFO_SET_STATE(*info,
 304                                                 BLOCK_STATE_WRITE_FAILURE);
 305                 }
 306         }
 307
 308         if (!ipo)
 309                 return;
 310
 311         if (ipo->flags & IP_F_ONRB)
 312                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 313         else if (ipo->flags & IP_F_ONLIST)
 314                 flist_del(&ipo->list);
 315
 316         free(ipo);
 317         io_u->ipo = NULL;
 318         td->io_hist_len--;
 319 }
 320
 321 void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
 322 {
 323         struct io_piece *ipo = io_u->ipo;
 324
 325         if (!ipo)
 326                 return;
 327
 328         ipo->len = io_u->xfer_buflen - io_u->resid;
 329 }
 330
 331 void write_iolog_close(struct thread_data *td)
 332 {
 333         fflush(td->iolog_f);
 334         fclose(td->iolog_f);
 335         free(td->iolog_buf);
 336         td->iolog_f = NULL;
 337         td->iolog_buf = NULL;
 338 }
 339
 340 /*
 341  * Read version 2 iolog data. It is enhanced to include per-file logging,
 342  * syncs, etc.
 343  */
 344 static int read_iolog2(struct thread_data *td, FILE *f)
 345 {
 346         unsigned long long offset;
 347         unsigned int bytes;
 348         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 349         char *fname, *act;
 350         char *str, *p;
 351         enum fio_ddir rw;
 352
 353         free_release_files(td);
 354
 355         /*
 356          * Read in the read iolog and store it, reuse the infrastructure
 357          * for doing verifications.
 358          */
 359         str = malloc(4096);
 360         fname = malloc(256+16);
 361         act = malloc(256+16);
 362
 363         reads = writes = waits = 0;
 364         while ((p = fgets(str, 4096, f)) != NULL) {
 365                 struct io_piece *ipo;
 366                 int r;
 367
 368                 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
 369                                                                         &bytes);
 370                 if (r == 4) {
 371                         /*
 372                          * Check action first
 373                          */
 374                         if (!strcmp(act, "wait"))
 375                                 rw = DDIR_WAIT;
 376                         else if (!strcmp(act, "read"))
 377                                 rw = DDIR_READ;
 378                         else if (!strcmp(act, "write"))
 379                                 rw = DDIR_WRITE;
 380                         else if (!strcmp(act, "sync"))
 381                                 rw = DDIR_SYNC;
 382                         else if (!strcmp(act, "datasync"))
 383                                 rw = DDIR_DATASYNC;
 384                         else if (!strcmp(act, "trim"))
 385                                 rw = DDIR_TRIM;
 386                         else {
 387                                 log_err("fio: bad iolog file action: %s\n",
 388                                                                         act);
 389                                 continue;
 390                         }
 391                         fileno = get_fileno(td, fname);
 392                 } else if (r == 2) {
 393                         rw = DDIR_INVAL;
 394                         if (!strcmp(act, "add")) {
 395                                 fileno = add_file(td, fname, 0, 1);
 396                                 file_action = FIO_LOG_ADD_FILE;
 397                                 continue;
 398                         } else if (!strcmp(act, "open")) {
 399                                 fileno = get_fileno(td, fname);
 400                                 file_action = FIO_LOG_OPEN_FILE;
 401                         } else if (!strcmp(act, "close")) {
 402                                 fileno = get_fileno(td, fname);
 403                                 file_action = FIO_LOG_CLOSE_FILE;
 404                         } else {
 405                                 log_err("fio: bad iolog file action: %s\n",
 406                                                                         act);
 407                                 continue;
 408                         }
 409                 } else {
 410                         log_err("bad iolog2: %s", p);
 411                         continue;
 412                 }
 413
 414                 if (rw == DDIR_READ)
 415                         reads++;
 416                 else if (rw == DDIR_WRITE) {
 417                         /*
 418                          * Don't add a write for ro mode
 419                          */
 420                         if (read_only)
 421                                 continue;
 422                         writes++;
 423                 } else if (rw == DDIR_WAIT) {
 424                         waits++;
 425                 } else if (rw == DDIR_INVAL) {
 426                 } else if (!ddir_sync(rw)) {
 427                         log_err("bad ddir: %d\n", rw);
 428                         continue;
 429                 }
 430
 431                 /*
 432                  * Make note of file
 433                  */
 434                 ipo = malloc(sizeof(*ipo));
 435                 init_ipo(ipo);
 436                 ipo->ddir = rw;
 437                 if (rw == DDIR_WAIT) {
 438                         ipo->delay = offset;
 439                 } else {
 440                         ipo->offset = offset;
 441                         ipo->len = bytes;
 442                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
 443                                 td->o.max_bs[rw] = bytes;
 444                         ipo->fileno = fileno;
 445                         ipo->file_action = file_action;
 446                         td->o.size += bytes;
 447                 }
 448
 449                 queue_io_piece(td, ipo);
 450         }
 451
 452         free(str);
 453         free(act);
 454         free(fname);
 455
 456         if (writes && read_only) {
 457                 log_err("fio: <%s> skips replay of %d writes due to"
 458                         " read-only\n", td->o.name, writes);
 459                 writes = 0;
 460         }
 461
 462         if (!reads && !writes && !waits)
 463                 return 1;
 464         else if (reads && !writes)
 465                 td->o.td_ddir = TD_DDIR_READ;
 466         else if (!reads && writes)
 467                 td->o.td_ddir = TD_DDIR_WRITE;
 468         else
 469                 td->o.td_ddir = TD_DDIR_RW;
 470
 471         return 0;
 472 }
 473
 474 /*
 475  * open iolog, check version, and call appropriate parser
 476  */
 477 static int init_iolog_read(struct thread_data *td)
 478 {
 479         char buffer[256], *p;
 480         FILE *f;
 481         int ret;
 482
 483         f = fopen(td->o.read_iolog_file, "r");
 484         if (!f) {
 485                 perror("fopen read iolog");
 486                 return 1;
 487         }
 488
 489         p = fgets(buffer, sizeof(buffer), f);
 490         if (!p) {
 491                 td_verror(td, errno, "iolog read");
 492                 log_err("fio: unable to read iolog\n");
 493                 fclose(f);
 494                 return 1;
 495         }
 496
 497         /*
 498          * version 2 of the iolog stores a specific string as the
 499          * first line, check for that
 500          */
 501         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 502                 ret = read_iolog2(td, f);
 503         else {
 504                 log_err("fio: iolog version 1 is no longer supported\n");
 505                 ret = 1;
 506         }
 507
 508         fclose(f);
 509         return ret;
 510 }
 511
 512 /*
 513  * Set up a log for storing io patterns.
 514  */
 515 static int init_iolog_write(struct thread_data *td)
 516 {
 517         struct fio_file *ff;
 518         FILE *f;
 519         unsigned int i;
 520
 521         f = fopen(td->o.write_iolog_file, "a");
 522         if (!f) {
 523                 perror("fopen write iolog");
 524                 return 1;
 525         }
 526
 527         /*
 528          * That's it for writing, setup a log buffer and we're done.
 529           */
 530         td->iolog_f = f;
 531         td->iolog_buf = malloc(8192);
 532         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 533
 534         /*
 535          * write our version line
 536          */
 537         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 538                 perror("iolog init\n");
 539                 return 1;
 540         }
 541
 542         /*
 543          * add all known files
 544          */
 545         for_each_file(td, ff, i)
 546                 log_file(td, ff, FIO_LOG_ADD_FILE);
 547
 548         return 0;
 549 }
 550
 551 int init_iolog(struct thread_data *td)
 552 {
 553         int ret = 0;
 554
 555         if (td->o.read_iolog_file) {
 556                 int need_swap;
 557
 558                 /*
 559                  * Check if it's a blktrace file and load that if possible.
 560                  * Otherwise assume it's a normal log file and load that.
 561                  */
 562                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 563                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 564                 else
 565                         ret = init_iolog_read(td);
 566         } else if (td->o.write_iolog_file)
 567                 ret = init_iolog_write(td);
 568
 569         if (ret)
 570                 td_verror(td, EINVAL, "failed initializing iolog");
 571
 572         return ret;
 573 }
 574
 575 void setup_log(struct io_log **log, struct log_params *p,
 576                const char *filename)
 577 {
 578         struct io_log *l;
 579
 580         l = scalloc(1, sizeof(*l));
 581         INIT_FLIST_HEAD(&l->io_logs);
 582         l->log_type = p->log_type;
 583         l->log_offset = p->log_offset;
 584         l->log_gz = p->log_gz;
 585         l->log_gz_store = p->log_gz_store;
 586         l->avg_msec = p->avg_msec;
 587         l->hist_msec = p->hist_msec;
 588         l->hist_coarseness = p->hist_coarseness;
 589         l->filename = strdup(filename);
 590         l->td = p->td;
 591
 592         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 593                 struct io_logs *p;
 594
 595                 p = calloc(1, sizeof(*l->pending));
 596                 p->max_samples = DEF_LOG_ENTRIES;
 597                 p->log = calloc(p->max_samples, log_entry_sz(l));
 598                 l->pending = p;
 599         }
 600
 601         if (l->log_offset)
 602                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 603
 604         INIT_FLIST_HEAD(&l->chunk_list);
 605
 606         if (l->log_gz && !p->td)
 607                 l->log_gz = 0;
 608         else if (l->log_gz || l->log_gz_store) {
 609                 mutex_init_pshared(&l->chunk_lock);
 610                 p->td->flags |= TD_F_COMPRESS_LOG;
 611         }
 612
 613         *log = l;
 614 }
 615
 616 #ifdef CONFIG_SETVBUF
 617 static void *set_file_buffer(FILE *f)
 618 {
 619         size_t size = 1048576;
 620         void *buf;
 621
 622         buf = malloc(size);
 623         setvbuf(f, buf, _IOFBF, size);
 624         return buf;
 625 }
 626
 627 static void clear_file_buffer(void *buf)
 628 {
 629         free(buf);
 630 }
 631 #else
 632 static void *set_file_buffer(FILE *f)
 633 {
 634         return NULL;
 635 }
 636
 637 static void clear_file_buffer(void *buf)
 638 {
 639 }
 640 #endif
 641
 642 void free_log(struct io_log *log)
 643 {
 644         while (!flist_empty(&log->io_logs)) {
 645                 struct io_logs *cur_log;
 646
 647                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 648                 flist_del_init(&cur_log->list);
 649                 free(cur_log->log);
 650                 sfree(cur_log);
 651         }
 652
 653         if (log->pending) {
 654                 free(log->pending->log);
 655                 free(log->pending);
 656                 log->pending = NULL;
 657         }
 658
 659         free(log->pending);
 660         free(log->filename);
 661         sfree(log);
 662 }
 663
 664 static inline unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat)
 665 {
 666         unsigned long sum;
 667         int k;
 668
 669         for (k = sum = 0; k < stride; k++)
 670                 sum += io_u_plat[j + k];
 671
 672         return sum;
 673 }
 674
 675 void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 676                         uint64_t sample_size)
 677 {
 678         struct io_sample *s;
 679         int log_offset;
 680         uint64_t i, j, nr_samples;
 681         unsigned int *io_u_plat;
 682
 683         int stride = 1 << hist_coarseness;
 684
 685         if (!sample_size)
 686                 return;
 687
 688         s = __get_sample(samples, 0, 0);
 689         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 690
 691         nr_samples = sample_size / __log_entry_sz(log_offset);
 692
 693         for (i = 0; i < nr_samples; i++) {
 694                 s = __get_sample(samples, log_offset, i);
 695                 io_u_plat = (unsigned int *) s->val;
 696                 fprintf(f, "%lu, %u, %u, ", (unsigned long)s->time,
 697                         io_sample_ddir(s), s->bs);
 698                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 699                         fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat));
 700                 }
 701                 fprintf(f, "%lu\n", (unsigned long)
 702                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat));
 703                 free(io_u_plat);
 704         }
 705 }
 706
 707 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 708 {
 709         struct io_sample *s;
 710         int log_offset;
 711         uint64_t i, nr_samples;
 712
 713         if (!sample_size)
 714                 return;
 715
 716         s = __get_sample(samples, 0, 0);
 717         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 718
 719         nr_samples = sample_size / __log_entry_sz(log_offset);
 720
 721         for (i = 0; i < nr_samples; i++) {
 722                 s = __get_sample(samples, log_offset, i);
 723
 724                 if (!log_offset) {
 725                         fprintf(f, "%lu, %lu, %u, %u\n",
 726                                         (unsigned long) s->time,
 727                                         (unsigned long) s->val,
 728                                         io_sample_ddir(s), s->bs);
 729                 } else {
 730                         struct io_sample_offset *so = (void *) s;
 731
 732                         fprintf(f, "%lu, %lu, %u, %u, %llu\n",
 733                                         (unsigned long) s->time,
 734                                         (unsigned long) s->val,
 735                                         io_sample_ddir(s), s->bs,
 736                                         (unsigned long long) so->offset);
 737                 }
 738         }
 739 }
 740
 741 #ifdef CONFIG_ZLIB
 742
 743 struct iolog_flush_data {
 744         struct workqueue_work work;
 745         struct io_log *log;
 746         void *samples;
 747         uint32_t nr_samples;
 748         bool free;
 749 };
 750
 751 #define GZ_CHUNK        131072
 752
 753 static struct iolog_compress *get_new_chunk(unsigned int seq)
 754 {
 755         struct iolog_compress *c;
 756
 757         c = malloc(sizeof(*c));
 758         INIT_FLIST_HEAD(&c->list);
 759         c->buf = malloc(GZ_CHUNK);
 760         c->len = 0;
 761         c->seq = seq;
 762         return c;
 763 }
 764
 765 static void free_chunk(struct iolog_compress *ic)
 766 {
 767         free(ic->buf);
 768         free(ic);
 769 }
 770
 771 static int z_stream_init(z_stream *stream, int gz_hdr)
 772 {
 773         int wbits = 15;
 774
 775         memset(stream, 0, sizeof(*stream));
 776         stream->zalloc = Z_NULL;
 777         stream->zfree = Z_NULL;
 778         stream->opaque = Z_NULL;
 779         stream->next_in = Z_NULL;
 780
 781         /*
 782          * zlib magic - add 32 for auto-detection of gz header or not,
 783          * if we decide to store files in a gzip friendly format.
 784          */
 785         if (gz_hdr)
 786                 wbits += 32;
 787
 788         if (inflateInit2(stream, wbits) != Z_OK)
 789                 return 1;
 790
 791         return 0;
 792 }
 793
 794 struct inflate_chunk_iter {
 795         unsigned int seq;
 796         int err;
 797         void *buf;
 798         size_t buf_size;
 799         size_t buf_used;
 800         size_t chunk_sz;
 801 };
 802
 803 static void finish_chunk(z_stream *stream, FILE *f,
 804                          struct inflate_chunk_iter *iter)
 805 {
 806         int ret;
 807
 808         ret = inflateEnd(stream);
 809         if (ret != Z_OK)
 810                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 811                                 iter->seq, ret);
 812
 813         flush_samples(f, iter->buf, iter->buf_used);
 814         free(iter->buf);
 815         iter->buf = NULL;
 816         iter->buf_size = iter->buf_used = 0;
 817 }
 818
 819 /*
 820  * Iterative chunk inflation. Handles cases where we cross into a new
 821  * sequence, doing flush finish of previous chunk if needed.
 822  */
 823 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 824                             z_stream *stream, struct inflate_chunk_iter *iter)
 825 {
 826         size_t ret;
 827
 828         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
 829                                 (unsigned long) ic->len, ic->seq);
 830
 831         if (ic->seq != iter->seq) {
 832                 if (iter->seq)
 833                         finish_chunk(stream, f, iter);
 834
 835                 z_stream_init(stream, gz_hdr);
 836                 iter->seq = ic->seq;
 837         }
 838
 839         stream->avail_in = ic->len;
 840         stream->next_in = ic->buf;
 841
 842         if (!iter->buf_size) {
 843                 iter->buf_size = iter->chunk_sz;
 844                 iter->buf = malloc(iter->buf_size);
 845         }
 846
 847         while (stream->avail_in) {
 848                 size_t this_out = iter->buf_size - iter->buf_used;
 849                 int err;
 850
 851                 stream->avail_out = this_out;
 852                 stream->next_out = iter->buf + iter->buf_used;
 853
 854                 err = inflate(stream, Z_NO_FLUSH);
 855                 if (err < 0) {
 856                         log_err("fio: failed inflating log: %d\n", err);
 857                         iter->err = err;
 858                         break;
 859                 }
 860
 861                 iter->buf_used += this_out - stream->avail_out;
 862
 863                 if (!stream->avail_out) {
 864                         iter->buf_size += iter->chunk_sz;
 865                         iter->buf = realloc(iter->buf, iter->buf_size);
 866                         continue;
 867                 }
 868
 869                 if (err == Z_STREAM_END)
 870                         break;
 871         }
 872
 873         ret = (void *) stream->next_in - ic->buf;
 874
 875         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
 876
 877         return ret;
 878 }
 879
 880 /*
 881  * Inflate stored compressed chunks, or write them directly to the log
 882  * file if so instructed.
 883  */
 884 static int inflate_gz_chunks(struct io_log *log, FILE *f)
 885 {
 886         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
 887         z_stream stream;
 888
 889         while (!flist_empty(&log->chunk_list)) {
 890                 struct iolog_compress *ic;
 891
 892                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
 893                 flist_del(&ic->list);
 894
 895                 if (log->log_gz_store) {
 896                         size_t ret;
 897
 898                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
 899                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
 900
 901                         ret = fwrite(ic->buf, ic->len, 1, f);
 902                         if (ret != 1 || ferror(f)) {
 903                                 iter.err = errno;
 904                                 log_err("fio: error writing compressed log\n");
 905                         }
 906                 } else
 907                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
 908
 909                 free_chunk(ic);
 910         }
 911
 912         if (iter.seq) {
 913                 finish_chunk(&stream, f, &iter);
 914                 free(iter.buf);
 915         }
 916
 917         return iter.err;
 918 }
 919
 920 /*
 921  * Open compressed log file and decompress the stored chunks and
 922  * write them to stdout. The chunks are stored sequentially in the
 923  * file, so we iterate over them and do them one-by-one.
 924  */
 925 int iolog_file_inflate(const char *file)
 926 {
 927         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
 928         struct iolog_compress ic;
 929         z_stream stream;
 930         struct stat sb;
 931         ssize_t ret;
 932         size_t total;
 933         void *buf;
 934         FILE *f;
 935
 936         f = fopen(file, "r");
 937         if (!f) {
 938                 perror("fopen");
 939                 return 1;
 940         }
 941
 942         if (stat(file, &sb) < 0) {
 943                 fclose(f);
 944                 perror("stat");
 945                 return 1;
 946         }
 947
 948         ic.buf = buf = malloc(sb.st_size);
 949         ic.len = sb.st_size;
 950         ic.seq = 1;
 951
 952         ret = fread(ic.buf, ic.len, 1, f);
 953         if (ret < 0) {
 954                 perror("fread");
 955                 fclose(f);
 956                 free(buf);
 957                 return 1;
 958         } else if (ret != 1) {
 959                 log_err("fio: short read on reading log\n");
 960                 fclose(f);
 961                 free(buf);
 962                 return 1;
 963         }
 964
 965         fclose(f);
 966
 967         /*
 968          * Each chunk will return Z_STREAM_END. We don't know how many
 969          * chunks are in the file, so we just keep looping and incrementing
 970          * the sequence number until we have consumed the whole compressed
 971          * file.
 972          */
 973         total = ic.len;
 974         do {
 975                 size_t iret;
 976
 977                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
 978                 total -= iret;
 979                 if (!total)
 980                         break;
 981                 if (iter.err)
 982                         break;
 983
 984                 ic.seq++;
 985                 ic.len -= iret;
 986                 ic.buf += iret;
 987         } while (1);
 988
 989         if (iter.seq) {
 990                 finish_chunk(&stream, stdout, &iter);
 991                 free(iter.buf);
 992         }
 993
 994         free(buf);
 995         return iter.err;
 996 }
 997
 998 #else
 999
1000 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1001 {
1002         return 0;
1003 }
1004
1005 int iolog_file_inflate(const char *file)
1006 {
1007         log_err("fio: log inflation not possible without zlib\n");
1008         return 1;
1009 }
1010
1011 #endif
1012
1013 void flush_log(struct io_log *log, bool do_append)
1014 {
1015         void *buf;
1016         FILE *f;
1017
1018         if (!do_append)
1019                 f = fopen(log->filename, "w");
1020         else
1021                 f = fopen(log->filename, "a");
1022         if (!f) {
1023                 perror("fopen log");
1024                 return;
1025         }
1026
1027         buf = set_file_buffer(f);
1028
1029         inflate_gz_chunks(log, f);
1030
1031         while (!flist_empty(&log->io_logs)) {
1032                 struct io_logs *cur_log;
1033
1034                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1035                 flist_del_init(&cur_log->list);
1036
1037                 if (log == log->td->clat_hist_log)
1038                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1039                                            cur_log->nr_samples * log_entry_sz(log));
1040                 else
1041                         flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
1042
1043                 sfree(cur_log);
1044         }
1045
1046         fclose(f);
1047         clear_file_buffer(buf);
1048 }
1049
1050 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1051 {
1052         if (td->flags & TD_F_COMPRESS_LOG)
1053                 iolog_flush(log);
1054
1055         if (trylock) {
1056                 if (fio_trylock_file(log->filename))
1057                         return 1;
1058         } else
1059                 fio_lock_file(log->filename);
1060
1061         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1062                 fio_send_iolog(td, log, log->filename);
1063         else
1064                 flush_log(log, !td->o.per_job_logs);
1065
1066         fio_unlock_file(log->filename);
1067         free_log(log);
1068         return 0;
1069 }
1070
1071 size_t log_chunk_sizes(struct io_log *log)
1072 {
1073         struct flist_head *entry;
1074         size_t ret;
1075
1076         if (flist_empty(&log->chunk_list))
1077                 return 0;
1078
1079         ret = 0;
1080         pthread_mutex_lock(&log->chunk_lock);
1081         flist_for_each(entry, &log->chunk_list) {
1082                 struct iolog_compress *c;
1083
1084                 c = flist_entry(entry, struct iolog_compress, list);
1085                 ret += c->len;
1086         }
1087         pthread_mutex_unlock(&log->chunk_lock);
1088         return ret;
1089 }
1090
1091 #ifdef CONFIG_ZLIB
1092
1093 static int gz_work(struct iolog_flush_data *data)
1094 {
1095         struct iolog_compress *c = NULL;
1096         struct flist_head list;
1097         unsigned int seq;
1098         z_stream stream;
1099         size_t total = 0;
1100         int ret;
1101
1102         INIT_FLIST_HEAD(&list);
1103
1104         memset(&stream, 0, sizeof(stream));
1105         stream.zalloc = Z_NULL;
1106         stream.zfree = Z_NULL;
1107         stream.opaque = Z_NULL;
1108
1109         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1110         if (ret != Z_OK) {
1111                 log_err("fio: failed to init gz stream\n");
1112                 goto err;
1113         }
1114
1115         seq = ++data->log->chunk_seq;
1116
1117         stream.next_in = (void *) data->samples;
1118         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1119
1120         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1121                                 (unsigned long) stream.avail_in, seq,
1122                                 data->log->filename);
1123         do {
1124                 if (c)
1125                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1126                 c = get_new_chunk(seq);
1127                 stream.avail_out = GZ_CHUNK;
1128                 stream.next_out = c->buf;
1129                 ret = deflate(&stream, Z_NO_FLUSH);
1130                 if (ret < 0) {
1131                         log_err("fio: deflate log (%d)\n", ret);
1132                         free_chunk(c);
1133                         goto err;
1134                 }
1135
1136                 c->len = GZ_CHUNK - stream.avail_out;
1137                 flist_add_tail(&c->list, &list);
1138                 total += c->len;
1139         } while (stream.avail_in);
1140
1141         stream.next_out = c->buf + c->len;
1142         stream.avail_out = GZ_CHUNK - c->len;
1143
1144         ret = deflate(&stream, Z_FINISH);
1145         if (ret < 0) {
1146                 /*
1147                  * Z_BUF_ERROR is special, it just means we need more
1148                  * output space. We'll handle that below. Treat any other
1149                  * error as fatal.
1150                  */
1151                 if (ret != Z_BUF_ERROR) {
1152                         log_err("fio: deflate log (%d)\n", ret);
1153                         flist_del(&c->list);
1154                         free_chunk(c);
1155                         goto err;
1156                 }
1157         }
1158
1159         total -= c->len;
1160         c->len = GZ_CHUNK - stream.avail_out;
1161         total += c->len;
1162         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1163
1164         if (ret != Z_STREAM_END) {
1165                 do {
1166                         c = get_new_chunk(seq);
1167                         stream.avail_out = GZ_CHUNK;
1168                         stream.next_out = c->buf;
1169                         ret = deflate(&stream, Z_FINISH);
1170                         c->len = GZ_CHUNK - stream.avail_out;
1171                         total += c->len;
1172                         flist_add_tail(&c->list, &list);
1173                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1174                 } while (ret != Z_STREAM_END);
1175         }
1176
1177         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1178
1179         ret = deflateEnd(&stream);
1180         if (ret != Z_OK)
1181                 log_err("fio: deflateEnd %d\n", ret);
1182
1183         free(data->samples);
1184
1185         if (!flist_empty(&list)) {
1186                 pthread_mutex_lock(&data->log->chunk_lock);
1187                 flist_splice_tail(&list, &data->log->chunk_list);
1188                 pthread_mutex_unlock(&data->log->chunk_lock);
1189         }
1190
1191         ret = 0;
1192 done:
1193         if (data->free)
1194                 free(data);
1195         return ret;
1196 err:
1197         while (!flist_empty(&list)) {
1198                 c = flist_first_entry(list.next, struct iolog_compress, list);
1199                 flist_del(&c->list);
1200                 free_chunk(c);
1201         }
1202         ret = 1;
1203         goto done;
1204 }
1205
1206 /*
1207  * Invoked from our compress helper thread, when logging would have exceeded
1208  * the specified memory limitation. Compresses the previously stored
1209  * entries.
1210  */
1211 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1212 {
1213         return gz_work(container_of(work, struct iolog_flush_data, work));
1214 }
1215
1216 static int gz_init_worker(struct submit_worker *sw)
1217 {
1218         struct thread_data *td = sw->wq->td;
1219
1220         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1221                 return 0;
1222
1223         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1224                 log_err("gz: failed to set CPU affinity\n");
1225                 return 1;
1226         }
1227
1228         return 0;
1229 }
1230
1231 static struct workqueue_ops log_compress_wq_ops = {
1232         .fn             = gz_work_async,
1233         .init_worker_fn = gz_init_worker,
1234         .nice           = 1,
1235 };
1236
1237 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1238 {
1239         if (!(td->flags & TD_F_COMPRESS_LOG))
1240                 return 0;
1241
1242         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1243         return 0;
1244 }
1245
1246 void iolog_compress_exit(struct thread_data *td)
1247 {
1248         if (!(td->flags & TD_F_COMPRESS_LOG))
1249                 return;
1250
1251         workqueue_exit(&td->log_compress_wq);
1252 }
1253
1254 /*
1255  * Queue work item to compress the existing log entries. We reset the
1256  * current log to a small size, and reference the existing log in the
1257  * data that we queue for compression. Once compression has been done,
1258  * this old log is freed. If called with finish == true, will not return
1259  * until the log compression has completed, and will flush all previous
1260  * logs too
1261  */
1262 static int iolog_flush(struct io_log *log)
1263 {
1264         struct iolog_flush_data *data;
1265
1266         data = malloc(sizeof(*data));
1267         if (!data)
1268                 return 1;
1269
1270         data->log = log;
1271         data->free = false;
1272
1273         while (!flist_empty(&log->io_logs)) {
1274                 struct io_logs *cur_log;
1275
1276                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1277                 flist_del_init(&cur_log->list);
1278
1279                 data->samples = cur_log->log;
1280                 data->nr_samples = cur_log->nr_samples;
1281
1282                 sfree(cur_log);
1283
1284                 gz_work(data);
1285         }
1286
1287         free(data);
1288         return 0;
1289 }
1290
1291 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1292 {
1293         struct iolog_flush_data *data;
1294
1295         data = malloc(sizeof(*data));
1296         if (!data)
1297                 return 1;
1298
1299         data->log = log;
1300
1301         data->samples = cur_log->log;
1302         data->nr_samples = cur_log->nr_samples;
1303         data->free = true;
1304
1305         cur_log->nr_samples = cur_log->max_samples = 0;
1306         cur_log->log = NULL;
1307
1308         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1309         return 0;
1310 }
1311 #else
1312
1313 static int iolog_flush(struct io_log *log)
1314 {
1315         return 1;
1316 }
1317
1318 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1319 {
1320         return 1;
1321 }
1322
1323 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1324 {
1325         return 0;
1326 }
1327
1328 void iolog_compress_exit(struct thread_data *td)
1329 {
1330 }
1331
1332 #endif
1333
1334 struct io_logs *iolog_cur_log(struct io_log *log)
1335 {
1336         if (flist_empty(&log->io_logs))
1337                 return NULL;
1338
1339         return flist_last_entry(&log->io_logs, struct io_logs, list);
1340 }
1341
1342 uint64_t iolog_nr_samples(struct io_log *iolog)
1343 {
1344         struct flist_head *entry;
1345         uint64_t ret = 0;
1346
1347         flist_for_each(entry, &iolog->io_logs) {
1348                 struct io_logs *cur_log;
1349
1350                 cur_log = flist_entry(entry, struct io_logs, list);
1351                 ret += cur_log->nr_samples;
1352         }
1353
1354         return ret;
1355 }
1356
1357 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1358 {
1359         if (log)
1360                 return finish_log(td, log, try);
1361
1362         return 0;
1363 }
1364
1365 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1366 {
1367         int ret;
1368
1369         if (per_unit_log(td->iops_log) != unit_log)
1370                 return 0;
1371
1372         ret = __write_log(td, td->iops_log, try);
1373         if (!ret)
1374                 td->iops_log = NULL;
1375
1376         return ret;
1377 }
1378
1379 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1380 {
1381         int ret;
1382
1383         if (!unit_log)
1384                 return 0;
1385
1386         ret = __write_log(td, td->slat_log, try);
1387         if (!ret)
1388                 td->slat_log = NULL;
1389
1390         return ret;
1391 }
1392
1393 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1394 {
1395         int ret;
1396
1397         if (!unit_log)
1398                 return 0;
1399
1400         ret = __write_log(td, td->clat_log, try);
1401         if (!ret)
1402                 td->clat_log = NULL;
1403
1404         return ret;
1405 }
1406
1407 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1408 {
1409         int ret;
1410
1411         if (!unit_log)
1412                 return 0;
1413
1414         ret = __write_log(td, td->clat_hist_log, try);
1415         if (!ret)
1416                 td->clat_hist_log = NULL;
1417
1418         return ret;
1419 }
1420
1421 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1422 {
1423         int ret;
1424
1425         if (!unit_log)
1426                 return 0;
1427
1428         ret = __write_log(td, td->lat_log, try);
1429         if (!ret)
1430                 td->lat_log = NULL;
1431
1432         return ret;
1433 }
1434
1435 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1436 {
1437         int ret;
1438
1439         if (per_unit_log(td->bw_log) != unit_log)
1440                 return 0;
1441
1442         ret = __write_log(td, td->bw_log, try);
1443         if (!ret)
1444                 td->bw_log = NULL;
1445
1446         return ret;
1447 }
1448
1449 enum {
1450         BW_LOG_MASK     = 1,
1451         LAT_LOG_MASK    = 2,
1452         SLAT_LOG_MASK   = 4,
1453         CLAT_LOG_MASK   = 8,
1454         IOPS_LOG_MASK   = 16,
1455         CLAT_HIST_LOG_MASK = 32,
1456
1457         ALL_LOG_NR      = 6,
1458 };
1459
1460 struct log_type {
1461         unsigned int mask;
1462         int (*fn)(struct thread_data *, int, bool);
1463 };
1464
1465 static struct log_type log_types[] = {
1466         {
1467                 .mask   = BW_LOG_MASK,
1468                 .fn     = write_bandw_log,
1469         },
1470         {
1471                 .mask   = LAT_LOG_MASK,
1472                 .fn     = write_lat_log,
1473         },
1474         {
1475                 .mask   = SLAT_LOG_MASK,
1476                 .fn     = write_slat_log,
1477         },
1478         {
1479                 .mask   = CLAT_LOG_MASK,
1480                 .fn     = write_clat_log,
1481         },
1482         {
1483                 .mask   = IOPS_LOG_MASK,
1484                 .fn     = write_iops_log,
1485         },
1486         {
1487                 .mask   = CLAT_HIST_LOG_MASK,
1488                 .fn     = write_clat_hist_log,
1489         }
1490 };
1491
1492 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1493 {
1494         unsigned int log_mask = 0;
1495         unsigned int log_left = ALL_LOG_NR;
1496         int old_state, i;
1497
1498         old_state = td_bump_runstate(td, TD_FINISHING);
1499
1500         finalize_logs(td, unit_logs);
1501
1502         while (log_left) {
1503                 int prev_log_left = log_left;
1504
1505                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1506                         struct log_type *lt = &log_types[i];
1507                         int ret;
1508
1509                         if (!(log_mask & lt->mask)) {
1510                                 ret = lt->fn(td, log_left != 1, unit_logs);
1511                                 if (!ret) {
1512                                         log_left--;
1513                                         log_mask |= lt->mask;
1514                                 }
1515                         }
1516                 }
1517
1518                 if (prev_log_left == log_left)
1519                         usleep(5000);
1520         }
1521
1522         td_restore_runstate(td, old_state);
1523 }
1524
1525 void fio_writeout_logs(bool unit_logs)
1526 {
1527         struct thread_data *td;
1528         int i;
1529
1530         for_each_td(td, i)
1531                 td_writeout_logs(td, unit_logs);
1532 }