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