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