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