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->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 || l->log_gz_store) {
 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         while (!flist_empty(&log->io_logs)) {
 634                 struct io_logs *cur_log;
 635
 636                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 637                 flist_del_init(&cur_log->list);
 638                 free(cur_log->log);
 639         }
 640
 641         free(log->filename);
 642         sfree(log);
 643 }
 644
 645 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 646 {
 647         struct io_sample *s;
 648         int log_offset;
 649         uint64_t i, nr_samples;
 650
 651         if (!sample_size)
 652                 return;
 653
 654         s = __get_sample(samples, 0, 0);
 655         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 656
 657         nr_samples = sample_size / __log_entry_sz(log_offset);
 658
 659         for (i = 0; i < nr_samples; i++) {
 660                 s = __get_sample(samples, log_offset, i);
 661
 662                 if (!log_offset) {
 663                         fprintf(f, "%lu, %lu, %u, %u\n",
 664                                         (unsigned long) s->time,
 665                                         (unsigned long) s->val,
 666                                         io_sample_ddir(s), s->bs);
 667                 } else {
 668                         struct io_sample_offset *so = (void *) s;
 669
 670                         fprintf(f, "%lu, %lu, %u, %u, %llu\n",
 671                                         (unsigned long) s->time,
 672                                         (unsigned long) s->val,
 673                                         io_sample_ddir(s), s->bs,
 674                                         (unsigned long long) so->offset);
 675                 }
 676         }
 677 }
 678
 679 #ifdef CONFIG_ZLIB
 680
 681 struct iolog_flush_data {
 682         struct workqueue_work work;
 683         struct io_log *log;
 684         void *samples;
 685         uint32_t nr_samples;
 686         bool free;
 687 };
 688
 689 #define GZ_CHUNK        131072
 690
 691 static struct iolog_compress *get_new_chunk(unsigned int seq)
 692 {
 693         struct iolog_compress *c;
 694
 695         c = malloc(sizeof(*c));
 696         INIT_FLIST_HEAD(&c->list);
 697         c->buf = malloc(GZ_CHUNK);
 698         c->len = 0;
 699         c->seq = seq;
 700         return c;
 701 }
 702
 703 static void free_chunk(struct iolog_compress *ic)
 704 {
 705         free(ic->buf);
 706         free(ic);
 707 }
 708
 709 static int z_stream_init(z_stream *stream, int gz_hdr)
 710 {
 711         int wbits = 15;
 712
 713         memset(stream, 0, sizeof(*stream));
 714         stream->zalloc = Z_NULL;
 715         stream->zfree = Z_NULL;
 716         stream->opaque = Z_NULL;
 717         stream->next_in = Z_NULL;
 718
 719         /*
 720          * zlib magic - add 32 for auto-detection of gz header or not,
 721          * if we decide to store files in a gzip friendly format.
 722          */
 723         if (gz_hdr)
 724                 wbits += 32;
 725
 726         if (inflateInit2(stream, wbits) != Z_OK)
 727                 return 1;
 728
 729         return 0;
 730 }
 731
 732 struct inflate_chunk_iter {
 733         unsigned int seq;
 734         int err;
 735         void *buf;
 736         size_t buf_size;
 737         size_t buf_used;
 738         size_t chunk_sz;
 739 };
 740
 741 static void finish_chunk(z_stream *stream, FILE *f,
 742                          struct inflate_chunk_iter *iter)
 743 {
 744         int ret;
 745
 746         ret = inflateEnd(stream);
 747         if (ret != Z_OK)
 748                 log_err("fio: failed to end log inflation (%d)\n", ret);
 749
 750         flush_samples(f, iter->buf, iter->buf_used);
 751         free(iter->buf);
 752         iter->buf = NULL;
 753         iter->buf_size = iter->buf_used = 0;
 754 }
 755
 756 /*
 757  * Iterative chunk inflation. Handles cases where we cross into a new
 758  * sequence, doing flush finish of previous chunk if needed.
 759  */
 760 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 761                             z_stream *stream, struct inflate_chunk_iter *iter)
 762 {
 763         size_t ret;
 764
 765         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
 766                                 (unsigned long) ic->len, ic->seq);
 767
 768         if (ic->seq != iter->seq) {
 769                 if (iter->seq)
 770                         finish_chunk(stream, f, iter);
 771
 772                 z_stream_init(stream, gz_hdr);
 773                 iter->seq = ic->seq;
 774         }
 775
 776         stream->avail_in = ic->len;
 777         stream->next_in = ic->buf;
 778
 779         if (!iter->buf_size) {
 780                 iter->buf_size = iter->chunk_sz;
 781                 iter->buf = malloc(iter->buf_size);
 782         }
 783
 784         while (stream->avail_in) {
 785                 size_t this_out = iter->buf_size - iter->buf_used;
 786                 int err;
 787
 788                 stream->avail_out = this_out;
 789                 stream->next_out = iter->buf + iter->buf_used;
 790
 791                 err = inflate(stream, Z_NO_FLUSH);
 792                 if (err < 0) {
 793                         log_err("fio: failed inflating log: %d\n", err);
 794                         iter->err = err;
 795                         break;
 796                 }
 797
 798                 iter->buf_used += this_out - stream->avail_out;
 799
 800                 if (!stream->avail_out) {
 801                         iter->buf_size += iter->chunk_sz;
 802                         iter->buf = realloc(iter->buf, iter->buf_size);
 803                         continue;
 804                 }
 805
 806                 if (err == Z_STREAM_END)
 807                         break;
 808         }
 809
 810         ret = (void *) stream->next_in - ic->buf;
 811
 812         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) ret);
 813
 814         return ret;
 815 }
 816
 817 /*
 818  * Inflate stored compressed chunks, or write them directly to the log
 819  * file if so instructed.
 820  */
 821 static int inflate_gz_chunks(struct io_log *log, FILE *f)
 822 {
 823         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
 824         z_stream stream;
 825
 826         while (!flist_empty(&log->chunk_list)) {
 827                 struct iolog_compress *ic;
 828
 829                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
 830                 flist_del(&ic->list);
 831
 832                 if (log->log_gz_store) {
 833                         size_t ret;
 834
 835                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
 836                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
 837
 838                         ret = fwrite(ic->buf, ic->len, 1, f);
 839                         if (ret != 1 || ferror(f)) {
 840                                 iter.err = errno;
 841                                 log_err("fio: error writing compressed log\n");
 842                         }
 843                 } else
 844                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
 845
 846                 free_chunk(ic);
 847         }
 848
 849         if (iter.seq) {
 850                 finish_chunk(&stream, f, &iter);
 851                 free(iter.buf);
 852         }
 853
 854         return iter.err;
 855 }
 856
 857 /*
 858  * Open compressed log file and decompress the stored chunks and
 859  * write them to stdout. The chunks are stored sequentially in the
 860  * file, so we iterate over them and do them one-by-one.
 861  */
 862 int iolog_file_inflate(const char *file)
 863 {
 864         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
 865         struct iolog_compress ic;
 866         z_stream stream;
 867         struct stat sb;
 868         ssize_t ret;
 869         size_t total;
 870         void *buf;
 871         FILE *f;
 872
 873         f = fopen(file, "r");
 874         if (!f) {
 875                 perror("fopen");
 876                 return 1;
 877         }
 878
 879         if (stat(file, &sb) < 0) {
 880                 fclose(f);
 881                 perror("stat");
 882                 return 1;
 883         }
 884
 885         ic.buf = buf = malloc(sb.st_size);
 886         ic.len = sb.st_size;
 887         ic.seq = 1;
 888
 889         ret = fread(ic.buf, ic.len, 1, f);
 890         if (ret < 0) {
 891                 perror("fread");
 892                 fclose(f);
 893                 free(buf);
 894                 return 1;
 895         } else if (ret != 1) {
 896                 log_err("fio: short read on reading log\n");
 897                 fclose(f);
 898                 free(buf);
 899                 return 1;
 900         }
 901
 902         fclose(f);
 903
 904         /*
 905          * Each chunk will return Z_STREAM_END. We don't know how many
 906          * chunks are in the file, so we just keep looping and incrementing
 907          * the sequence number until we have consumed the whole compressed
 908          * file.
 909          */
 910         total = ic.len;
 911         do {
 912                 size_t iret;
 913
 914                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
 915                 total -= iret;
 916                 if (!total)
 917                         break;
 918                 if (iter.err)
 919                         break;
 920
 921                 ic.seq++;
 922                 ic.len -= iret;
 923                 ic.buf += iret;
 924         } while (1);
 925
 926         if (iter.seq) {
 927                 finish_chunk(&stream, stdout, &iter);
 928                 free(iter.buf);
 929         }
 930
 931         free(buf);
 932         return iter.err;
 933 }
 934
 935 #else
 936
 937 static int inflate_gz_chunks(struct io_log *log, FILE *f)
 938 {
 939         return 0;
 940 }
 941
 942 int iolog_file_inflate(const char *file)
 943 {
 944         log_err("fio: log inflation not possible without zlib\n");
 945         return 1;
 946 }
 947
 948 #endif
 949
 950 void flush_log(struct io_log *log, int do_append)
 951 {
 952         void *buf;
 953         FILE *f;
 954
 955         if (!do_append)
 956                 f = fopen(log->filename, "w");
 957         else
 958                 f = fopen(log->filename, "a");
 959         if (!f) {
 960                 perror("fopen log");
 961                 return;
 962         }
 963
 964         buf = set_file_buffer(f);
 965
 966         inflate_gz_chunks(log, f);
 967
 968         while (!flist_empty(&log->io_logs)) {
 969                 struct io_logs *cur_log;
 970
 971                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 972                 flist_del_init(&cur_log->list);
 973                 flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
 974         }
 975
 976         fclose(f);
 977         clear_file_buffer(buf);
 978 }
 979
 980 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
 981 {
 982         if (td->flags & TD_F_COMPRESS_LOG)
 983                 iolog_flush(log);
 984
 985         if (trylock) {
 986                 if (fio_trylock_file(log->filename))
 987                         return 1;
 988         } else
 989                 fio_lock_file(log->filename);
 990
 991         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
 992                 fio_send_iolog(td, log, log->filename);
 993         else
 994                 flush_log(log, !td->o.per_job_logs);
 995
 996         fio_unlock_file(log->filename);
 997         free_log(log);
 998         return 0;
 999 }
1000
1001 size_t log_chunk_sizes(struct io_log *log)
1002 {
1003         struct flist_head *entry;
1004         size_t ret;
1005
1006         if (flist_empty(&log->chunk_list))
1007                 return 0;
1008
1009         ret = 0;
1010         pthread_mutex_lock(&log->chunk_lock);
1011         flist_for_each(entry, &log->chunk_list) {
1012                 struct iolog_compress *c;
1013
1014                 c = flist_entry(entry, struct iolog_compress, list);
1015                 ret += c->len;
1016         }
1017         pthread_mutex_unlock(&log->chunk_lock);
1018         return ret;
1019 }
1020
1021 #ifdef CONFIG_ZLIB
1022
1023 static int gz_work(struct iolog_flush_data *data)
1024 {
1025         struct iolog_compress *c;
1026         struct flist_head list;
1027         unsigned int seq;
1028         z_stream stream;
1029         size_t total = 0;
1030         int ret;
1031
1032         INIT_FLIST_HEAD(&list);
1033
1034         memset(&stream, 0, sizeof(stream));
1035         stream.zalloc = Z_NULL;
1036         stream.zfree = Z_NULL;
1037         stream.opaque = Z_NULL;
1038
1039         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1040         if (ret != Z_OK) {
1041                 log_err("fio: failed to init gz stream\n");
1042                 goto err;
1043         }
1044
1045         seq = ++data->log->chunk_seq;
1046
1047         stream.next_in = (void *) data->samples;
1048         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1049
1050         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u\n",
1051                                 (unsigned long) stream.avail_in, seq);
1052         do {
1053                 c = get_new_chunk(seq);
1054                 stream.avail_out = GZ_CHUNK;
1055                 stream.next_out = c->buf;
1056                 ret = deflate(&stream, Z_NO_FLUSH);
1057                 if (ret < 0) {
1058                         log_err("fio: deflate log (%d)\n", ret);
1059                         free_chunk(c);
1060                         goto err;
1061                 }
1062
1063                 c->len = GZ_CHUNK - stream.avail_out;
1064                 flist_add_tail(&c->list, &list);
1065                 total += c->len;
1066         } while (stream.avail_in);
1067
1068         stream.next_out = c->buf + c->len;
1069         stream.avail_out = GZ_CHUNK - c->len;
1070
1071         ret = deflate(&stream, Z_FINISH);
1072         if (ret == Z_STREAM_END)
1073                 c->len = GZ_CHUNK - stream.avail_out;
1074         else {
1075                 do {
1076                         c = get_new_chunk(seq);
1077                         stream.avail_out = GZ_CHUNK;
1078                         stream.next_out = c->buf;
1079                         ret = deflate(&stream, Z_FINISH);
1080                         c->len = GZ_CHUNK - stream.avail_out;
1081                         total += c->len;
1082                         flist_add_tail(&c->list, &list);
1083                 } while (ret != Z_STREAM_END);
1084         }
1085
1086         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1087
1088         ret = deflateEnd(&stream);
1089         if (ret != Z_OK)
1090                 log_err("fio: deflateEnd %d\n", ret);
1091
1092         free(data->samples);
1093
1094         if (!flist_empty(&list)) {
1095                 pthread_mutex_lock(&data->log->chunk_lock);
1096                 flist_splice_tail(&list, &data->log->chunk_list);
1097                 pthread_mutex_unlock(&data->log->chunk_lock);
1098         }
1099
1100         ret = 0;
1101 done:
1102         if (data->free)
1103                 free(data);
1104         return ret;
1105 err:
1106         while (!flist_empty(&list)) {
1107                 c = flist_first_entry(list.next, struct iolog_compress, list);
1108                 flist_del(&c->list);
1109                 free_chunk(c);
1110         }
1111         ret = 1;
1112         goto done;
1113 }
1114
1115 /*
1116  * Invoked from our compress helper thread, when logging would have exceeded
1117  * the specified memory limitation. Compresses the previously stored
1118  * entries.
1119  */
1120 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1121 {
1122         return gz_work(container_of(work, struct iolog_flush_data, work));
1123 }
1124
1125 static int gz_init_worker(struct submit_worker *sw)
1126 {
1127         struct thread_data *td = sw->wq->td;
1128
1129         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1130                 return 0;
1131
1132         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1133                 log_err("gz: failed to set CPU affinity\n");
1134                 return 1;
1135         }
1136
1137         return 0;
1138 }
1139
1140 static struct workqueue_ops log_compress_wq_ops = {
1141         .fn             = gz_work_async,
1142         .init_worker_fn = gz_init_worker,
1143         .nice           = 1,
1144 };
1145
1146 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1147 {
1148         if (!(td->flags & TD_F_COMPRESS_LOG))
1149                 return 0;
1150
1151         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1152         return 0;
1153 }
1154
1155 void iolog_compress_exit(struct thread_data *td)
1156 {
1157         if (!(td->flags & TD_F_COMPRESS_LOG))
1158                 return;
1159
1160         workqueue_exit(&td->log_compress_wq);
1161 }
1162
1163 /*
1164  * Queue work item to compress the existing log entries. We reset the
1165  * current log to a small size, and reference the existing log in the
1166  * data that we queue for compression. Once compression has been done,
1167  * this old log is freed. If called with finish == true, will not return
1168  * until the log compression has completed, and will flush all previous
1169  * logs too
1170  */
1171 static int iolog_flush(struct io_log *log)
1172 {
1173         struct iolog_flush_data *data;
1174
1175         data = malloc(sizeof(*data));
1176         if (!data)
1177                 return 1;
1178
1179         data->log = log;
1180         data->free = false;
1181
1182         while (!flist_empty(&log->io_logs)) {
1183                 struct io_logs *cur_log;
1184
1185                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1186                 flist_del_init(&cur_log->list);
1187
1188                 data->samples = cur_log->log;
1189                 data->nr_samples = cur_log->nr_samples;
1190
1191                 cur_log->nr_samples = 0;
1192                 cur_log->max_samples = 0;
1193                 cur_log->log = NULL;
1194
1195                 gz_work(data);
1196         }
1197
1198         free(data);
1199         return 0;
1200 }
1201
1202 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1203 {
1204         struct iolog_flush_data *data;
1205
1206         data = malloc(sizeof(*data));
1207         if (!data)
1208                 return 1;
1209
1210         data->log = log;
1211
1212         data->samples = cur_log->log;
1213         data->nr_samples = cur_log->nr_samples;
1214         data->free = true;
1215
1216         cur_log->nr_samples = cur_log->max_samples = 0;
1217         cur_log->log = NULL;
1218
1219         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1220         return 0;
1221 }
1222 #else
1223
1224 static int iolog_flush(struct io_log *log)
1225 {
1226         return 1;
1227 }
1228
1229 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1230 {
1231         return 1;
1232 }
1233
1234 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1235 {
1236         return 0;
1237 }
1238
1239 void iolog_compress_exit(struct thread_data *td)
1240 {
1241 }
1242
1243 #endif
1244
1245 struct io_logs *iolog_cur_log(struct io_log *log)
1246 {
1247         if (flist_empty(&log->io_logs))
1248                 return NULL;
1249
1250         return flist_last_entry(&log->io_logs, struct io_logs, list);
1251 }
1252
1253 uint64_t iolog_nr_samples(struct io_log *iolog)
1254 {
1255         struct flist_head *entry;
1256         uint64_t ret = 0;
1257
1258         flist_for_each(entry, &iolog->io_logs) {
1259                 struct io_logs *cur_log;
1260
1261                 cur_log = flist_entry(entry, struct io_logs, list);
1262                 ret += cur_log->nr_samples;
1263         }
1264
1265         return ret;
1266 }
1267
1268 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1269 {
1270         if (log)
1271                 return finish_log(td, log, try);
1272
1273         return 0;
1274 }
1275
1276 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1277 {
1278         int ret;
1279
1280         if (per_unit_log(td->iops_log) != unit_log)
1281                 return 0;
1282
1283         ret = __write_log(td, td->iops_log, try);
1284         if (!ret)
1285                 td->iops_log = NULL;
1286
1287         return ret;
1288 }
1289
1290 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1291 {
1292         int ret;
1293
1294         if (!unit_log)
1295                 return 0;
1296
1297         ret = __write_log(td, td->slat_log, try);
1298         if (!ret)
1299                 td->slat_log = NULL;
1300
1301         return ret;
1302 }
1303
1304 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1305 {
1306         int ret;
1307
1308         if (!unit_log)
1309                 return 0;
1310
1311         ret = __write_log(td, td->clat_log, try);
1312         if (!ret)
1313                 td->clat_log = NULL;
1314
1315         return ret;
1316 }
1317
1318 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1319 {
1320         int ret;
1321
1322         if (!unit_log)
1323                 return 0;
1324
1325         ret = __write_log(td, td->lat_log, try);
1326         if (!ret)
1327                 td->lat_log = NULL;
1328
1329         return ret;
1330 }
1331
1332 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1333 {
1334         int ret;
1335
1336         if (per_unit_log(td->bw_log) != unit_log)
1337                 return 0;
1338
1339         ret = __write_log(td, td->bw_log, try);
1340         if (!ret)
1341                 td->bw_log = NULL;
1342
1343         return ret;
1344 }
1345
1346 enum {
1347         BW_LOG_MASK     = 1,
1348         LAT_LOG_MASK    = 2,
1349         SLAT_LOG_MASK   = 4,
1350         CLAT_LOG_MASK   = 8,
1351         IOPS_LOG_MASK   = 16,
1352
1353         ALL_LOG_NR      = 5,
1354 };
1355
1356 struct log_type {
1357         unsigned int mask;
1358         int (*fn)(struct thread_data *, int, bool);
1359 };
1360
1361 static struct log_type log_types[] = {
1362         {
1363                 .mask   = BW_LOG_MASK,
1364                 .fn     = write_bandw_log,
1365         },
1366         {
1367                 .mask   = LAT_LOG_MASK,
1368                 .fn     = write_lat_log,
1369         },
1370         {
1371                 .mask   = SLAT_LOG_MASK,
1372                 .fn     = write_slat_log,
1373         },
1374         {
1375                 .mask   = CLAT_LOG_MASK,
1376                 .fn     = write_clat_log,
1377         },
1378         {
1379                 .mask   = IOPS_LOG_MASK,
1380                 .fn     = write_iops_log,
1381         },
1382 };
1383
1384 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1385 {
1386         unsigned int log_mask = 0;
1387         unsigned int log_left = ALL_LOG_NR;
1388         int old_state, i;
1389
1390         old_state = td_bump_runstate(td, TD_FINISHING);
1391
1392         finalize_logs(td, unit_logs);
1393
1394         while (log_left) {
1395                 int prev_log_left = log_left;
1396
1397                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1398                         struct log_type *lt = &log_types[i];
1399                         int ret;
1400
1401                         if (!(log_mask & lt->mask)) {
1402                                 ret = lt->fn(td, log_left != 1, unit_logs);
1403                                 if (!ret) {
1404                                         log_left--;
1405                                         log_mask |= lt->mask;
1406                                 }
1407                         }
1408                 }
1409
1410                 if (prev_log_left == log_left)
1411                         usleep(5000);
1412         }
1413
1414         td_restore_runstate(td, old_state);
1415 }
1416
1417 void fio_writeout_logs(bool unit_logs)
1418 {
1419         struct thread_data *td;
1420         int i;
1421
1422         for_each_td(td, i)
1423                 td_writeout_logs(td, unit_logs);
1424 }