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 "flist.h"
  10 #include "fio.h"
  11 #include "verify.h"
  12 #include "trim.h"
  13
  14 static const char iolog_ver2[] = "fio version 2 iolog";
  15
  16 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  17 {
  18         flist_add_tail(&ipo->list, &td->io_log_list);
  19         td->total_io_size += ipo->len;
  20 }
  21
  22 void log_io_u(struct thread_data *td, struct io_u *io_u)
  23 {
  24         const char *act[] = { "read", "write", "sync", "datasync",
  25                                 "sync_file_range", "wait", "trim" };
  26
  27         assert(io_u->ddir <= 6);
  28
  29         if (!td->o.write_iolog_file)
  30                 return;
  31
  32         fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
  33                                                 act[io_u->ddir], io_u->offset,
  34                                                 io_u->buflen);
  35 }
  36
  37 void log_file(struct thread_data *td, struct fio_file *f,
  38               enum file_log_act what)
  39 {
  40         const char *act[] = { "add", "open", "close" };
  41
  42         assert(what < 3);
  43
  44         if (!td->o.write_iolog_file)
  45                 return;
  46
  47
  48         /*
  49          * this happens on the pre-open/close done before the job starts
  50          */
  51         if (!td->iolog_f)
  52                 return;
  53
  54         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  55 }
  56
  57 static void iolog_delay(struct thread_data *td, unsigned long delay)
  58 {
  59         unsigned long usec = utime_since_now(&td->last_issue);
  60
  61         if (delay < usec)
  62                 return;
  63
  64         delay -= usec;
  65
  66         /*
  67          * less than 100 usec delay, just regard it as noise
  68          */
  69         if (delay < 100)
  70                 return;
  71
  72         usec_sleep(td, delay);
  73 }
  74
  75 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
  76 {
  77         struct fio_file *f;
  78         int ret;
  79
  80         /*
  81          * Not a special ipo
  82          */
  83         if (ipo->ddir != DDIR_INVAL)
  84                 return 0;
  85
  86         f = td->files[ipo->fileno];
  87
  88         switch (ipo->file_action) {
  89         case FIO_LOG_OPEN_FILE:
  90                 ret = td_io_open_file(td, f);
  91                 if (!ret)
  92                         break;
  93                 td_verror(td, ret, "iolog open file");
  94                 return -1;
  95         case FIO_LOG_CLOSE_FILE:
  96                 td_io_close_file(td, f);
  97                 break;
  98         case FIO_LOG_UNLINK_FILE:
  99                 unlink(f->file_name);
 100                 break;
 101         default:
 102                 log_err("fio: bad file action %d\n", ipo->file_action);
 103                 break;
 104         }
 105
 106         return 1;
 107 }
 108
 109 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 110 {
 111         struct io_piece *ipo;
 112         unsigned long elapsed;
 113
 114         while (!flist_empty(&td->io_log_list)) {
 115                 int ret;
 116
 117                 ipo = flist_entry(td->io_log_list.next, struct io_piece, list);
 118                 flist_del(&ipo->list);
 119                 remove_trim_entry(td, ipo);
 120
 121                 ret = ipo_special(td, ipo);
 122                 if (ret < 0) {
 123                         free(ipo);
 124                         break;
 125                 } else if (ret > 0) {
 126                         free(ipo);
 127                         continue;
 128                 }
 129
 130                 io_u->ddir = ipo->ddir;
 131                 if (ipo->ddir != DDIR_WAIT) {
 132                         io_u->offset = ipo->offset;
 133                         io_u->buflen = ipo->len;
 134                         io_u->file = td->files[ipo->fileno];
 135                         get_file(io_u->file);
 136                         dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
 137                                                 io_u->buflen, io_u->file->file_name);
 138                         if (ipo->delay)
 139                                 iolog_delay(td, ipo->delay);
 140                 } else {
 141                         elapsed = mtime_since_genesis();
 142                         if (ipo->delay > elapsed)
 143                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 144                 }
 145
 146                 free(ipo);
 147
 148                 if (io_u->ddir != DDIR_WAIT)
 149                         return 0;
 150         }
 151
 152         td->done = 1;
 153         return 1;
 154 }
 155
 156 void prune_io_piece_log(struct thread_data *td)
 157 {
 158         struct io_piece *ipo;
 159         struct rb_node *n;
 160
 161         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 162                 ipo = rb_entry(n, struct io_piece, rb_node);
 163                 rb_erase(n, &td->io_hist_tree);
 164                 remove_trim_entry(td, ipo);
 165                 td->io_hist_len--;
 166                 free(ipo);
 167         }
 168
 169         while (!flist_empty(&td->io_hist_list)) {
 170                 ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
 171                 flist_del(&ipo->list);
 172                 remove_trim_entry(td, ipo);
 173                 td->io_hist_len--;
 174                 free(ipo);
 175         }
 176 }
 177
 178 /*
 179  * log a successful write, so we can unwind the log for verify
 180  */
 181 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 182 {
 183         struct rb_node **p, *parent;
 184         struct io_piece *ipo, *__ipo;
 185
 186         ipo = malloc(sizeof(struct io_piece));
 187         init_ipo(ipo);
 188         ipo->file = io_u->file;
 189         ipo->offset = io_u->offset;
 190         ipo->len = io_u->buflen;
 191
 192         if (io_u_should_trim(td, io_u)) {
 193                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 194                 td->trim_entries++;
 195         }
 196
 197         /*
 198          * We don't need to sort the entries, if:
 199          *
 200          *      Sequential writes, or
 201          *      Random writes that lay out the file as it goes along
 202          *
 203          * For both these cases, just reading back data in the order we
 204          * wrote it out is the fastest.
 205          *
 206          * One exception is if we don't have a random map AND we are doing
 207          * verifies, in that case we need to check for duplicate blocks and
 208          * drop the old one, which we rely on the rb insert/lookup for
 209          * handling.
 210          */
 211         if ((!td_random(td) || !td->o.overwrite) &&
 212               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
 213                 INIT_FLIST_HEAD(&ipo->list);
 214                 flist_add_tail(&ipo->list, &td->io_hist_list);
 215                 ipo->flags |= IP_F_ONLIST;
 216                 td->io_hist_len++;
 217                 return;
 218         }
 219
 220         RB_CLEAR_NODE(&ipo->rb_node);
 221
 222         /*
 223          * Sort the entry into the verification list
 224          */
 225 restart:
 226         p = &td->io_hist_tree.rb_node;
 227         parent = NULL;
 228         while (*p) {
 229                 parent = *p;
 230
 231                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 232                 if (ipo->file < __ipo->file)
 233                         p = &(*p)->rb_left;
 234                 else if (ipo->file > __ipo->file)
 235                         p = &(*p)->rb_right;
 236                 else if (ipo->offset < __ipo->offset)
 237                         p = &(*p)->rb_left;
 238                 else if (ipo->offset > __ipo->offset)
 239                         p = &(*p)->rb_right;
 240                 else {
 241                         assert(ipo->len == __ipo->len);
 242                         td->io_hist_len--;
 243                         rb_erase(parent, &td->io_hist_tree);
 244                         remove_trim_entry(td, __ipo);
 245                         free(__ipo);
 246                         goto restart;
 247                 }
 248         }
 249
 250         rb_link_node(&ipo->rb_node, parent, p);
 251         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 252         ipo->flags |= IP_F_ONRB;
 253         td->io_hist_len++;
 254 }
 255
 256 void write_iolog_close(struct thread_data *td)
 257 {
 258         fflush(td->iolog_f);
 259         fclose(td->iolog_f);
 260         free(td->iolog_buf);
 261         td->iolog_f = NULL;
 262         td->iolog_buf = NULL;
 263 }
 264
 265 /*
 266  * Read version 2 iolog data. It is enhanced to include per-file logging,
 267  * syncs, etc.
 268  */
 269 static int read_iolog2(struct thread_data *td, FILE *f)
 270 {
 271         unsigned long long offset;
 272         unsigned int bytes;
 273         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 274         char *fname, *act;
 275         char *str, *p;
 276         enum fio_ddir rw;
 277
 278         free_release_files(td);
 279
 280         /*
 281          * Read in the read iolog and store it, reuse the infrastructure
 282          * for doing verifications.
 283          */
 284         str = malloc(4096);
 285         fname = malloc(256+16);
 286         act = malloc(256+16);
 287
 288         reads = writes = waits = 0;
 289         while ((p = fgets(str, 4096, f)) != NULL) {
 290                 struct io_piece *ipo;
 291                 int r;
 292
 293                 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
 294                                                                         &bytes);
 295                 if (r == 4) {
 296                         /*
 297                          * Check action first
 298                          */
 299                         if (!strcmp(act, "wait"))
 300                                 rw = DDIR_WAIT;
 301                         else if (!strcmp(act, "read"))
 302                                 rw = DDIR_READ;
 303                         else if (!strcmp(act, "write"))
 304                                 rw = DDIR_WRITE;
 305                         else if (!strcmp(act, "sync"))
 306                                 rw = DDIR_SYNC;
 307                         else if (!strcmp(act, "datasync"))
 308                                 rw = DDIR_DATASYNC;
 309                         else if (!strcmp(act, "trim"))
 310                                 rw = DDIR_TRIM;
 311                         else {
 312                                 log_err("fio: bad iolog file action: %s\n",
 313                                                                         act);
 314                                 continue;
 315                         }
 316                 } else if (r == 2) {
 317                         rw = DDIR_INVAL;
 318                         if (!strcmp(act, "add")) {
 319                                 td->o.nr_files++;
 320                                 fileno = add_file(td, fname);
 321                                 file_action = FIO_LOG_ADD_FILE;
 322                                 continue;
 323                         } else if (!strcmp(act, "open")) {
 324                                 fileno = get_fileno(td, fname);
 325                                 file_action = FIO_LOG_OPEN_FILE;
 326                         } else if (!strcmp(act, "close")) {
 327                                 fileno = get_fileno(td, fname);
 328                                 file_action = FIO_LOG_CLOSE_FILE;
 329                         } else {
 330                                 log_err("fio: bad iolog file action: %s\n",
 331                                                                         act);
 332                                 continue;
 333                         }
 334                 } else {
 335                         log_err("bad iolog2: %s", p);
 336                         continue;
 337                 }
 338
 339                 if (rw == DDIR_READ)
 340                         reads++;
 341                 else if (rw == DDIR_WRITE) {
 342                         /*
 343                          * Don't add a write for ro mode
 344                          */
 345                         if (read_only)
 346                                 continue;
 347                         writes++;
 348                 } else if (rw == DDIR_WAIT) {
 349                         waits++;
 350                 } else if (rw == DDIR_INVAL) {
 351                 } else if (!ddir_sync(rw)) {
 352                         log_err("bad ddir: %d\n", rw);
 353                         continue;
 354                 }
 355
 356                 /*
 357                  * Make note of file
 358                  */
 359                 ipo = malloc(sizeof(*ipo));
 360                 init_ipo(ipo);
 361                 ipo->ddir = rw;
 362                 if (rw == DDIR_WAIT) {
 363                         ipo->delay = offset;
 364                 } else {
 365                         ipo->offset = offset;
 366                         ipo->len = bytes;
 367                         if (bytes > td->o.max_bs[rw])
 368                                 td->o.max_bs[rw] = bytes;
 369                         ipo->fileno = fileno;
 370                         ipo->file_action = file_action;
 371                 }
 372
 373                 queue_io_piece(td, ipo);
 374         }
 375
 376         free(str);
 377         free(act);
 378         free(fname);
 379
 380         if (writes && read_only) {
 381                 log_err("fio: <%s> skips replay of %d writes due to"
 382                         " read-only\n", td->o.name, writes);
 383                 writes = 0;
 384         }
 385
 386         if (!reads && !writes && !waits)
 387                 return 1;
 388         else if (reads && !writes)
 389                 td->o.td_ddir = TD_DDIR_READ;
 390         else if (!reads && writes)
 391                 td->o.td_ddir = TD_DDIR_WRITE;
 392         else
 393                 td->o.td_ddir = TD_DDIR_RW;
 394
 395         return 0;
 396 }
 397
 398 /*
 399  * open iolog, check version, and call appropriate parser
 400  */
 401 static int init_iolog_read(struct thread_data *td)
 402 {
 403         char buffer[256], *p;
 404         FILE *f;
 405         int ret;
 406
 407         f = fopen(td->o.read_iolog_file, "r");
 408         if (!f) {
 409                 perror("fopen read iolog");
 410                 return 1;
 411         }
 412
 413         p = fgets(buffer, sizeof(buffer), f);
 414         if (!p) {
 415                 td_verror(td, errno, "iolog read");
 416                 log_err("fio: unable to read iolog\n");
 417                 fclose(f);
 418                 return 1;
 419         }
 420
 421         /*
 422          * version 2 of the iolog stores a specific string as the
 423          * first line, check for that
 424          */
 425         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 426                 ret = read_iolog2(td, f);
 427         else {
 428                 log_err("fio: iolog version 1 is no longer supported\n");
 429                 ret = 1;
 430         }
 431
 432         fclose(f);
 433         return ret;
 434 }
 435
 436 /*
 437  * Set up a log for storing io patterns.
 438  */
 439 static int init_iolog_write(struct thread_data *td)
 440 {
 441         struct fio_file *ff;
 442         FILE *f;
 443         unsigned int i;
 444
 445         f = fopen(td->o.write_iolog_file, "a");
 446         if (!f) {
 447                 perror("fopen write iolog");
 448                 return 1;
 449         }
 450
 451         /*
 452          * That's it for writing, setup a log buffer and we're done.
 453           */
 454         td->iolog_f = f;
 455         td->iolog_buf = malloc(8192);
 456         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 457
 458         /*
 459          * write our version line
 460          */
 461         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 462                 perror("iolog init\n");
 463                 return 1;
 464         }
 465
 466         /*
 467          * add all known files
 468          */
 469         for_each_file(td, ff, i)
 470                 log_file(td, ff, FIO_LOG_ADD_FILE);
 471
 472         return 0;
 473 }
 474
 475 int init_iolog(struct thread_data *td)
 476 {
 477         int ret = 0;
 478
 479         if (td->o.read_iolog_file) {
 480                 /*
 481                  * Check if it's a blktrace file and load that if possible.
 482                  * Otherwise assume it's a normal log file and load that.
 483                  */
 484                 if (is_blktrace(td->o.read_iolog_file))
 485                         ret = load_blktrace(td, td->o.read_iolog_file);
 486                 else
 487                         ret = init_iolog_read(td);
 488         } else if (td->o.write_iolog_file)
 489                 ret = init_iolog_write(td);
 490
 491         return ret;
 492 }
 493
 494 void setup_log(struct io_log **log, unsigned long avg_msec, int log_type)
 495 {
 496         struct io_log *l = malloc(sizeof(*l));
 497
 498         memset(l, 0, sizeof(*l));
 499         l->nr_samples = 0;
 500         l->max_samples = 1024;
 501         l->log_type = log_type;
 502         l->log = malloc(l->max_samples * sizeof(struct io_sample));
 503         l->avg_msec = avg_msec;
 504         *log = l;
 505 }
 506
 507 void __finish_log(struct io_log *log, const char *name)
 508 {
 509         unsigned int i;
 510         FILE *f;
 511
 512         f = fopen(name, "a");
 513         if (!f) {
 514                 perror("fopen log");
 515                 return;
 516         }
 517
 518         for (i = 0; i < log->nr_samples; i++) {
 519                 fprintf(f, "%lu, %lu, %u, %u\n",
 520                                 (unsigned long) log->log[i].time,
 521                                 (unsigned long) log->log[i].val,
 522                                 log->log[i].ddir, log->log[i].bs);
 523         }
 524
 525         fclose(f);
 526         free(log->log);
 527         free(log);
 528 }
 529
 530 void finish_log_named(struct thread_data *td, struct io_log *log,
 531                        const char *prefix, const char *postfix)
 532 {
 533         char file_name[256], *p;
 534
 535         snprintf(file_name, 200, "%s_%s.log", prefix, postfix);
 536         p = basename(file_name);
 537
 538         if (td->client_type == FIO_CLIENT_TYPE_GUI) {
 539                 fio_send_iolog(td, log, p);
 540                 free(log->log);
 541                 free(log);
 542         } else
 543                 __finish_log(log, p);
 544 }
 545
 546 void finish_log(struct thread_data *td, struct io_log *log, const char *name)
 547 {
 548         finish_log_named(td, log, td->o.name, name);
 549 }