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 int hist_sum(int j, int stride, unsigned int *io_u_plat) {
 665         int sum = 0;
 666         for (int k = 0; k < stride; k++) {
 667                 sum += io_u_plat[j + k];
 668         }
 669         return sum;
 670 }
 671
 672 void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 673         uint64_t sample_size)
 674 {
 675         struct io_sample *s;
 676         int log_offset;
 677         uint64_t i, j, nr_samples;
 678         unsigned int *io_u_plat;
 679
 680         int stride = 1 << hist_coarseness;
 681
 682         if (!sample_size)
 683                 return;
 684
 685         s = __get_sample(samples, 0, 0);
 686         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 687
 688         nr_samples = sample_size / __log_entry_sz(log_offset);
 689
 690         for (i = 0; i < nr_samples; i++) {
 691                 s = __get_sample(samples, log_offset, i);
 692                 io_u_plat = (unsigned int *)(s->val);
 693                 fprintf(f, "%lu, %u, %u, ", (unsigned long)s->time,
 694                         io_sample_ddir(s), s->bs);
 695                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 696                         fprintf(f, "%lu, ", (unsigned long) hist_sum(j, stride, io_u_plat));
 697                 }
 698                 fprintf(f, "%lu\n", (unsigned long)
 699                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat));
 700                 free(io_u_plat);
 701         }
 702 }
 703
 704 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 705 {
 706         struct io_sample *s;
 707         int log_offset;
 708         uint64_t i, nr_samples;
 709
 710         if (!sample_size)
 711                 return;
 712
 713         s = __get_sample(samples, 0, 0);
 714         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 715
 716         nr_samples = sample_size / __log_entry_sz(log_offset);
 717
 718         for (i = 0; i < nr_samples; i++) {
 719                 s = __get_sample(samples, log_offset, i);
 720
 721                 if (!log_offset) {
 722                         fprintf(f, "%lu, %lu, %u, %u\n",
 723                                         (unsigned long) s->time,
 724                                         (unsigned long) s->val,
 725                                         io_sample_ddir(s), s->bs);
 726                 } else {
 727                         struct io_sample_offset *so = (void *) s;
 728
 729                         fprintf(f, "%lu, %lu, %u, %u, %llu\n",
 730                                         (unsigned long) s->time,
 731                                         (unsigned long) s->val,
 732                                         io_sample_ddir(s), s->bs,
 733                                         (unsigned long long) so->offset);
 734                 }
 735         }
 736 }
 737
 738 #ifdef CONFIG_ZLIB
 739
 740 struct iolog_flush_data {
 741         struct workqueue_work work;
 742         struct io_log *log;
 743         void *samples;
 744         uint32_t nr_samples;
 745         bool free;
 746 };
 747
 748 #define GZ_CHUNK        131072
 749
 750 static struct iolog_compress *get_new_chunk(unsigned int seq)
 751 {
 752         struct iolog_compress *c;
 753
 754         c = malloc(sizeof(*c));
 755         INIT_FLIST_HEAD(&c->list);
 756         c->buf = malloc(GZ_CHUNK);
 757         c->len = 0;
 758         c->seq = seq;
 759         return c;
 760 }
 761
 762 static void free_chunk(struct iolog_compress *ic)
 763 {
 764         free(ic->buf);
 765         free(ic);
 766 }
 767
 768 static int z_stream_init(z_stream *stream, int gz_hdr)
 769 {
 770         int wbits = 15;
 771
 772         memset(stream, 0, sizeof(*stream));
 773         stream->zalloc = Z_NULL;
 774         stream->zfree = Z_NULL;
 775         stream->opaque = Z_NULL;
 776         stream->next_in = Z_NULL;
 777
 778         /*
 779          * zlib magic - add 32 for auto-detection of gz header or not,
 780          * if we decide to store files in a gzip friendly format.
 781          */
 782         if (gz_hdr)
 783                 wbits += 32;
 784
 785         if (inflateInit2(stream, wbits) != Z_OK)
 786                 return 1;
 787
 788         return 0;
 789 }
 790
 791 struct inflate_chunk_iter {
 792         unsigned int seq;
 793         int err;
 794         void *buf;
 795         size_t buf_size;
 796         size_t buf_used;
 797         size_t chunk_sz;
 798 };
 799
 800 static void finish_chunk(z_stream *stream, FILE *f,
 801                          struct inflate_chunk_iter *iter)
 802 {
 803         int ret;
 804
 805         ret = inflateEnd(stream);
 806         if (ret != Z_OK)
 807                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 808                                 iter->seq, ret);
 809
 810         flush_samples(f, iter->buf, iter->buf_used);
 811         free(iter->buf);
 812         iter->buf = NULL;
 813         iter->buf_size = iter->buf_used = 0;
 814 }
 815
 816 /*
 817  * Iterative chunk inflation. Handles cases where we cross into a new
 818  * sequence, doing flush finish of previous chunk if needed.
 819  */
 820 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 821                             z_stream *stream, struct inflate_chunk_iter *iter)
 822 {
 823         size_t ret;
 824
 825         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
 826                                 (unsigned long) ic->len, ic->seq);
 827
 828         if (ic->seq != iter->seq) {
 829                 if (iter->seq)
 830                         finish_chunk(stream, f, iter);
 831
 832                 z_stream_init(stream, gz_hdr);
 833                 iter->seq = ic->seq;
 834         }
 835
 836         stream->avail_in = ic->len;
 837         stream->next_in = ic->buf;
 838
 839         if (!iter->buf_size) {
 840                 iter->buf_size = iter->chunk_sz;
 841                 iter->buf = malloc(iter->buf_size);
 842         }
 843
 844         while (stream->avail_in) {
 845                 size_t this_out = iter->buf_size - iter->buf_used;
 846                 int err;
 847
 848                 stream->avail_out = this_out;
 849                 stream->next_out = iter->buf + iter->buf_used;
 850
 851                 err = inflate(stream, Z_NO_FLUSH);
 852                 if (err < 0) {
 853                         log_err("fio: failed inflating log: %d\n", err);
 854                         iter->err = err;
 855                         break;
 856                 }
 857
 858                 iter->buf_used += this_out - stream->avail_out;
 859
 860                 if (!stream->avail_out) {
 861                         iter->buf_size += iter->chunk_sz;
 862                         iter->buf = realloc(iter->buf, iter->buf_size);
 863                         continue;
 864                 }
 865
 866                 if (err == Z_STREAM_END)
 867                         break;
 868         }
 869
 870         ret = (void *) stream->next_in - ic->buf;
 871
 872         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
 873
 874         return ret;
 875 }
 876
 877 /*
 878  * Inflate stored compressed chunks, or write them directly to the log
 879  * file if so instructed.
 880  */
 881 static int inflate_gz_chunks(struct io_log *log, FILE *f)
 882 {
 883         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
 884         z_stream stream;
 885
 886         while (!flist_empty(&log->chunk_list)) {
 887                 struct iolog_compress *ic;
 888
 889                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
 890                 flist_del(&ic->list);
 891
 892                 if (log->log_gz_store) {
 893                         size_t ret;
 894
 895                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
 896                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
 897
 898                         ret = fwrite(ic->buf, ic->len, 1, f);
 899                         if (ret != 1 || ferror(f)) {
 900                                 iter.err = errno;
 901                                 log_err("fio: error writing compressed log\n");
 902                         }
 903                 } else
 904                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
 905
 906                 free_chunk(ic);
 907         }
 908
 909         if (iter.seq) {
 910                 finish_chunk(&stream, f, &iter);
 911                 free(iter.buf);
 912         }
 913
 914         return iter.err;
 915 }
 916
 917 /*
 918  * Open compressed log file and decompress the stored chunks and
 919  * write them to stdout. The chunks are stored sequentially in the
 920  * file, so we iterate over them and do them one-by-one.
 921  */
 922 int iolog_file_inflate(const char *file)
 923 {
 924         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
 925         struct iolog_compress ic;
 926         z_stream stream;
 927         struct stat sb;
 928         ssize_t ret;
 929         size_t total;
 930         void *buf;
 931         FILE *f;
 932
 933         f = fopen(file, "r");
 934         if (!f) {
 935                 perror("fopen");
 936                 return 1;
 937         }
 938
 939         if (stat(file, &sb) < 0) {
 940                 fclose(f);
 941                 perror("stat");
 942                 return 1;
 943         }
 944
 945         ic.buf = buf = malloc(sb.st_size);
 946         ic.len = sb.st_size;
 947         ic.seq = 1;
 948
 949         ret = fread(ic.buf, ic.len, 1, f);
 950         if (ret < 0) {
 951                 perror("fread");
 952                 fclose(f);
 953                 free(buf);
 954                 return 1;
 955         } else if (ret != 1) {
 956                 log_err("fio: short read on reading log\n");
 957                 fclose(f);
 958                 free(buf);
 959                 return 1;
 960         }
 961
 962         fclose(f);
 963
 964         /*
 965          * Each chunk will return Z_STREAM_END. We don't know how many
 966          * chunks are in the file, so we just keep looping and incrementing
 967          * the sequence number until we have consumed the whole compressed
 968          * file.
 969          */
 970         total = ic.len;
 971         do {
 972                 size_t iret;
 973
 974                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
 975                 total -= iret;
 976                 if (!total)
 977                         break;
 978                 if (iter.err)
 979                         break;
 980
 981                 ic.seq++;
 982                 ic.len -= iret;
 983                 ic.buf += iret;
 984         } while (1);
 985
 986         if (iter.seq) {
 987                 finish_chunk(&stream, stdout, &iter);
 988                 free(iter.buf);
 989         }
 990
 991         free(buf);
 992         return iter.err;
 993 }
 994
 995 #else
 996
 997 static int inflate_gz_chunks(struct io_log *log, FILE *f)
 998 {
 999         return 0;
1000 }
1001
1002 int iolog_file_inflate(const char *file)
1003 {
1004         log_err("fio: log inflation not possible without zlib\n");
1005         return 1;
1006 }
1007
1008 #endif
1009
1010 void flush_log(struct io_log *log, bool do_append)
1011 {
1012         void *buf;
1013         FILE *f;
1014
1015         if (!do_append)
1016                 f = fopen(log->filename, "w");
1017         else
1018                 f = fopen(log->filename, "a");
1019         if (!f) {
1020                 perror("fopen log");
1021                 return;
1022         }
1023
1024         buf = set_file_buffer(f);
1025
1026         inflate_gz_chunks(log, f);
1027
1028         while (!flist_empty(&log->io_logs)) {
1029                 struct io_logs *cur_log;
1030
1031                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1032                 flist_del_init(&cur_log->list);
1033
1034                 if (log == log->td->clat_hist_log)
1035                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1036                                            cur_log->nr_samples * log_entry_sz(log));
1037                 else
1038                         flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
1039
1040                 sfree(cur_log);
1041         }
1042
1043         fclose(f);
1044         clear_file_buffer(buf);
1045 }
1046
1047 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1048 {
1049         if (td->flags & TD_F_COMPRESS_LOG)
1050                 iolog_flush(log);
1051
1052         if (trylock) {
1053                 if (fio_trylock_file(log->filename))
1054                         return 1;
1055         } else
1056                 fio_lock_file(log->filename);
1057
1058         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1059                 fio_send_iolog(td, log, log->filename);
1060         else
1061                 flush_log(log, !td->o.per_job_logs);
1062
1063         fio_unlock_file(log->filename);
1064         free_log(log);
1065         return 0;
1066 }
1067
1068 size_t log_chunk_sizes(struct io_log *log)
1069 {
1070         struct flist_head *entry;
1071         size_t ret;
1072
1073         if (flist_empty(&log->chunk_list))
1074                 return 0;
1075
1076         ret = 0;
1077         pthread_mutex_lock(&log->chunk_lock);
1078         flist_for_each(entry, &log->chunk_list) {
1079                 struct iolog_compress *c;
1080
1081                 c = flist_entry(entry, struct iolog_compress, list);
1082                 ret += c->len;
1083         }
1084         pthread_mutex_unlock(&log->chunk_lock);
1085         return ret;
1086 }
1087
1088 #ifdef CONFIG_ZLIB
1089
1090 static int gz_work(struct iolog_flush_data *data)
1091 {
1092         struct iolog_compress *c = NULL;
1093         struct flist_head list;
1094         unsigned int seq;
1095         z_stream stream;
1096         size_t total = 0;
1097         int ret;
1098
1099         INIT_FLIST_HEAD(&list);
1100
1101         memset(&stream, 0, sizeof(stream));
1102         stream.zalloc = Z_NULL;
1103         stream.zfree = Z_NULL;
1104         stream.opaque = Z_NULL;
1105
1106         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1107         if (ret != Z_OK) {
1108                 log_err("fio: failed to init gz stream\n");
1109                 goto err;
1110         }
1111
1112         seq = ++data->log->chunk_seq;
1113
1114         stream.next_in = (void *) data->samples;
1115         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1116
1117         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1118                                 (unsigned long) stream.avail_in, seq,
1119                                 data->log->filename);
1120         do {
1121                 if (c)
1122                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1123                 c = get_new_chunk(seq);
1124                 stream.avail_out = GZ_CHUNK;
1125                 stream.next_out = c->buf;
1126                 ret = deflate(&stream, Z_NO_FLUSH);
1127                 if (ret < 0) {
1128                         log_err("fio: deflate log (%d)\n", ret);
1129                         free_chunk(c);
1130                         goto err;
1131                 }
1132
1133                 c->len = GZ_CHUNK - stream.avail_out;
1134                 flist_add_tail(&c->list, &list);
1135                 total += c->len;
1136         } while (stream.avail_in);
1137
1138         stream.next_out = c->buf + c->len;
1139         stream.avail_out = GZ_CHUNK - c->len;
1140
1141         ret = deflate(&stream, Z_FINISH);
1142         if (ret < 0) {
1143                 /*
1144                  * Z_BUF_ERROR is special, it just means we need more
1145                  * output space. We'll handle that below. Treat any other
1146                  * error as fatal.
1147                  */
1148                 if (ret != Z_BUF_ERROR) {
1149                         log_err("fio: deflate log (%d)\n", ret);
1150                         flist_del(&c->list);
1151                         free_chunk(c);
1152                         goto err;
1153                 }
1154         }
1155
1156         total -= c->len;
1157         c->len = GZ_CHUNK - stream.avail_out;
1158         total += c->len;
1159         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1160
1161         if (ret != Z_STREAM_END) {
1162                 do {
1163                         c = get_new_chunk(seq);
1164                         stream.avail_out = GZ_CHUNK;
1165                         stream.next_out = c->buf;
1166                         ret = deflate(&stream, Z_FINISH);
1167                         c->len = GZ_CHUNK - stream.avail_out;
1168                         total += c->len;
1169                         flist_add_tail(&c->list, &list);
1170                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1171                 } while (ret != Z_STREAM_END);
1172         }
1173
1174         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1175
1176         ret = deflateEnd(&stream);
1177         if (ret != Z_OK)
1178                 log_err("fio: deflateEnd %d\n", ret);
1179
1180         free(data->samples);
1181
1182         if (!flist_empty(&list)) {
1183                 pthread_mutex_lock(&data->log->chunk_lock);
1184                 flist_splice_tail(&list, &data->log->chunk_list);
1185                 pthread_mutex_unlock(&data->log->chunk_lock);
1186         }
1187
1188         ret = 0;
1189 done:
1190         if (data->free)
1191                 free(data);
1192         return ret;
1193 err:
1194         while (!flist_empty(&list)) {
1195                 c = flist_first_entry(list.next, struct iolog_compress, list);
1196                 flist_del(&c->list);
1197                 free_chunk(c);
1198         }
1199         ret = 1;
1200         goto done;
1201 }
1202
1203 /*
1204  * Invoked from our compress helper thread, when logging would have exceeded
1205  * the specified memory limitation. Compresses the previously stored
1206  * entries.
1207  */
1208 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1209 {
1210         return gz_work(container_of(work, struct iolog_flush_data, work));
1211 }
1212
1213 static int gz_init_worker(struct submit_worker *sw)
1214 {
1215         struct thread_data *td = sw->wq->td;
1216
1217         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1218                 return 0;
1219
1220         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1221                 log_err("gz: failed to set CPU affinity\n");
1222                 return 1;
1223         }
1224
1225         return 0;
1226 }
1227
1228 static struct workqueue_ops log_compress_wq_ops = {
1229         .fn             = gz_work_async,
1230         .init_worker_fn = gz_init_worker,
1231         .nice           = 1,
1232 };
1233
1234 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1235 {
1236         if (!(td->flags & TD_F_COMPRESS_LOG))
1237                 return 0;
1238
1239         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1240         return 0;
1241 }
1242
1243 void iolog_compress_exit(struct thread_data *td)
1244 {
1245         if (!(td->flags & TD_F_COMPRESS_LOG))
1246                 return;
1247
1248         workqueue_exit(&td->log_compress_wq);
1249 }
1250
1251 /*
1252  * Queue work item to compress the existing log entries. We reset the
1253  * current log to a small size, and reference the existing log in the
1254  * data that we queue for compression. Once compression has been done,
1255  * this old log is freed. If called with finish == true, will not return
1256  * until the log compression has completed, and will flush all previous
1257  * logs too
1258  */
1259 static int iolog_flush(struct io_log *log)
1260 {
1261         struct iolog_flush_data *data;
1262
1263         data = malloc(sizeof(*data));
1264         if (!data)
1265                 return 1;
1266
1267         data->log = log;
1268         data->free = false;
1269
1270         while (!flist_empty(&log->io_logs)) {
1271                 struct io_logs *cur_log;
1272
1273                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1274                 flist_del_init(&cur_log->list);
1275
1276                 data->samples = cur_log->log;
1277                 data->nr_samples = cur_log->nr_samples;
1278
1279                 sfree(cur_log);
1280
1281                 gz_work(data);
1282         }
1283
1284         free(data);
1285         return 0;
1286 }
1287
1288 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1289 {
1290         struct iolog_flush_data *data;
1291
1292         data = malloc(sizeof(*data));
1293         if (!data)
1294                 return 1;
1295
1296         data->log = log;
1297
1298         data->samples = cur_log->log;
1299         data->nr_samples = cur_log->nr_samples;
1300         data->free = true;
1301
1302         cur_log->nr_samples = cur_log->max_samples = 0;
1303         cur_log->log = NULL;
1304
1305         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1306         return 0;
1307 }
1308 #else
1309
1310 static int iolog_flush(struct io_log *log)
1311 {
1312         return 1;
1313 }
1314
1315 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1316 {
1317         return 1;
1318 }
1319
1320 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1321 {
1322         return 0;
1323 }
1324
1325 void iolog_compress_exit(struct thread_data *td)
1326 {
1327 }
1328
1329 #endif
1330
1331 struct io_logs *iolog_cur_log(struct io_log *log)
1332 {
1333         if (flist_empty(&log->io_logs))
1334                 return NULL;
1335
1336         return flist_last_entry(&log->io_logs, struct io_logs, list);
1337 }
1338
1339 uint64_t iolog_nr_samples(struct io_log *iolog)
1340 {
1341         struct flist_head *entry;
1342         uint64_t ret = 0;
1343
1344         flist_for_each(entry, &iolog->io_logs) {
1345                 struct io_logs *cur_log;
1346
1347                 cur_log = flist_entry(entry, struct io_logs, list);
1348                 ret += cur_log->nr_samples;
1349         }
1350
1351         return ret;
1352 }
1353
1354 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1355 {
1356         if (log)
1357                 return finish_log(td, log, try);
1358
1359         return 0;
1360 }
1361
1362 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1363 {
1364         int ret;
1365
1366         if (per_unit_log(td->iops_log) != unit_log)
1367                 return 0;
1368
1369         ret = __write_log(td, td->iops_log, try);
1370         if (!ret)
1371                 td->iops_log = NULL;
1372
1373         return ret;
1374 }
1375
1376 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1377 {
1378         int ret;
1379
1380         if (!unit_log)
1381                 return 0;
1382
1383         ret = __write_log(td, td->slat_log, try);
1384         if (!ret)
1385                 td->slat_log = NULL;
1386
1387         return ret;
1388 }
1389
1390 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1391 {
1392         int ret;
1393
1394         if (!unit_log)
1395                 return 0;
1396
1397         ret = __write_log(td, td->clat_log, try);
1398         if (!ret)
1399                 td->clat_log = NULL;
1400
1401         return ret;
1402 }
1403
1404 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1405 {
1406         int ret;
1407
1408         if (!unit_log)
1409                 return 0;
1410
1411         ret = __write_log(td, td->clat_hist_log, try);
1412         if (!ret)
1413                 td->clat_hist_log = NULL;
1414
1415         return ret;
1416 }
1417
1418 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1419 {
1420         int ret;
1421
1422         if (!unit_log)
1423                 return 0;
1424
1425         ret = __write_log(td, td->lat_log, try);
1426         if (!ret)
1427                 td->lat_log = NULL;
1428
1429         return ret;
1430 }
1431
1432 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1433 {
1434         int ret;
1435
1436         if (per_unit_log(td->bw_log) != unit_log)
1437                 return 0;
1438
1439         ret = __write_log(td, td->bw_log, try);
1440         if (!ret)
1441                 td->bw_log = NULL;
1442
1443         return ret;
1444 }
1445
1446 enum {
1447         BW_LOG_MASK     = 1,
1448         LAT_LOG_MASK    = 2,
1449         SLAT_LOG_MASK   = 4,
1450         CLAT_LOG_MASK   = 8,
1451         IOPS_LOG_MASK   = 16,
1452         CLAT_HIST_LOG_MASK = 32,
1453
1454         ALL_LOG_NR      = 6,
1455 };
1456
1457 struct log_type {
1458         unsigned int mask;
1459         int (*fn)(struct thread_data *, int, bool);
1460 };
1461
1462 static struct log_type log_types[] = {
1463         {
1464                 .mask   = BW_LOG_MASK,
1465                 .fn     = write_bandw_log,
1466         },
1467         {
1468                 .mask   = LAT_LOG_MASK,
1469                 .fn     = write_lat_log,
1470         },
1471         {
1472                 .mask   = SLAT_LOG_MASK,
1473                 .fn     = write_slat_log,
1474         },
1475         {
1476                 .mask   = CLAT_LOG_MASK,
1477                 .fn     = write_clat_log,
1478         },
1479         {
1480                 .mask   = IOPS_LOG_MASK,
1481                 .fn     = write_iops_log,
1482         },
1483         {
1484                 .mask   = CLAT_HIST_LOG_MASK,
1485                 .fn     = write_clat_hist_log,
1486         }
1487 };
1488
1489 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1490 {
1491         unsigned int log_mask = 0;
1492         unsigned int log_left = ALL_LOG_NR;
1493         int old_state, i;
1494
1495         old_state = td_bump_runstate(td, TD_FINISHING);
1496
1497         finalize_logs(td, unit_logs);
1498
1499         while (log_left) {
1500                 int prev_log_left = log_left;
1501
1502                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1503                         struct log_type *lt = &log_types[i];
1504                         int ret;
1505
1506                         if (!(log_mask & lt->mask)) {
1507                                 ret = lt->fn(td, log_left != 1, unit_logs);
1508                                 if (!ret) {
1509                                         log_left--;
1510                                         log_mask |= lt->mask;
1511                                 }
1512                         }
1513                 }
1514
1515                 if (prev_log_left == log_left)
1516                         usleep(5000);
1517         }
1518
1519         td_restore_runstate(td, old_state);
1520 }
1521
1522 void fio_writeout_logs(bool unit_logs)
1523 {
1524         struct thread_data *td;
1525         int i;
1526
1527         for_each_td(td, i)
1528                 td_writeout_logs(td, unit_logs);
1529 }