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