iolog.c

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