mmap engine: allow large files on 32-bit archs
[fio.git] / filesetup.c
1 #include <unistd.h>
2 #include <fcntl.h>
3 #include <string.h>
4 #include <assert.h>
5 #include <dirent.h>
6 #include <sys/stat.h>
7 #include <sys/mman.h>
8 #include <sys/types.h>
9
10 #include "fio.h"
11 #include "smalloc.h"
12 #include "filehash.h"
13
14 static int root_warn;
15
16 /*
17  * Leaves f->fd open on success, caller must close
18  */
19 static int extend_file(struct thread_data *td, struct fio_file *f)
20 {
21         int r, new_layout = 0, unlink_file = 0, flags;
22         unsigned long long left;
23         unsigned int bs;
24         char *b;
25
26         if (read_only) {
27                 log_err("fio: refusing extend of file due to read-only\n");
28                 return 0;
29         }
30
31         /*
32          * check if we need to lay the file out complete again. fio
33          * does that for operations involving reads, or for writes
34          * where overwrite is set
35          */
36         if (td_read(td) || (td_write(td) && td->o.overwrite) ||
37             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
38                 new_layout = 1;
39         if (td_write(td) && !td->o.overwrite)
40                 unlink_file = 1;
41
42         if (unlink_file || new_layout) {
43                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
44                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
45                         td_verror(td, errno, "unlink");
46                         return 1;
47                 }
48         }
49
50         flags = O_WRONLY | O_CREAT;
51         if (new_layout)
52                 flags |= O_TRUNC;
53
54         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
55         f->fd = open(f->file_name, flags, 0644);
56         if (f->fd < 0) {
57                 td_verror(td, errno, "open");
58                 return 1;
59         }
60
61         if (!new_layout)
62                 goto done;
63
64         dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
65                                                         f->real_file_size);
66         if (ftruncate(f->fd, f->real_file_size) == -1) {
67                 td_verror(td, errno, "ftruncate");
68                 goto err;
69         }
70
71 #ifdef FIO_HAVE_FALLOCATE
72         dprint(FD_FILE, "fallocate file %s, size %llu\n", f->file_name,
73                                                         f->real_file_size);
74         r = posix_fallocate(f->fd, 0, f->real_file_size);
75         if (r < 0)
76                 log_err("fio: posix_fallocate fails: %s\n", strerror(-r));
77 #endif
78
79         b = malloc(td->o.max_bs[DDIR_WRITE]);
80         memset(b, 0, td->o.max_bs[DDIR_WRITE]);
81
82         left = f->real_file_size;
83         while (left && !td->terminate) {
84                 bs = td->o.max_bs[DDIR_WRITE];
85                 if (bs > left)
86                         bs = left;
87
88                 r = write(f->fd, b, bs);
89
90                 if (r == (int) bs) {
91                         left -= bs;
92                         continue;
93                 } else {
94                         if (r < 0)
95                                 td_verror(td, errno, "write");
96                         else
97                                 td_verror(td, EIO, "write");
98
99                         break;
100                 }
101         }
102
103         if (td->terminate) {
104                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
105                 unlink(f->file_name);
106         } else if (td->o.create_fsync) {
107                 if (fsync(f->fd) < 0) {
108                         td_verror(td, errno, "fsync");
109                         goto err;
110                 }
111         }
112
113         free(b);
114 done:
115         return 0;
116 err:
117         close(f->fd);
118         f->fd = -1;
119         return 1;
120 }
121
122 static int pre_read_file(struct thread_data *td, struct fio_file *f)
123 {
124         int r, did_open = 0, old_runstate;
125         unsigned long long left;
126         unsigned int bs;
127         char *b;
128
129         if (!(f->flags & FIO_FILE_OPEN)) {
130                 if (td->io_ops->open_file(td, f)) {
131                         log_err("fio: cannot pre-read, failed to open file\n");
132                         return 1;
133                 }
134                 did_open = 1;
135         }
136
137         old_runstate = td->runstate;
138         td_set_runstate(td, TD_PRE_READING);
139
140         bs = td->o.max_bs[DDIR_READ];
141         b = malloc(bs);
142         memset(b, 0, bs);
143
144         lseek(f->fd, f->file_offset, SEEK_SET);
145         left = f->io_size;
146
147         while (left && !td->terminate) {
148                 if (bs > left)
149                         bs = left;
150
151                 r = read(f->fd, b, bs);
152
153                 if (r == (int) bs) {
154                         left -= bs;
155                         continue;
156                 } else {
157                         td_verror(td, EIO, "pre_read");
158                         break;
159                 }
160         }
161
162         td_set_runstate(td, old_runstate);
163
164         if (did_open)
165                 td->io_ops->close_file(td, f);
166         free(b);
167         return 0;
168 }
169
170 static unsigned long long get_rand_file_size(struct thread_data *td)
171 {
172         unsigned long long ret, sized;
173         long r;
174
175         r = os_random_long(&td->file_size_state);
176         sized = td->o.file_size_high - td->o.file_size_low;
177         ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
178         ret += td->o.file_size_low;
179         ret -= (ret % td->o.rw_min_bs);
180         return ret;
181 }
182
183 static int file_size(struct thread_data *td, struct fio_file *f)
184 {
185         struct stat st;
186
187         if (stat(f->file_name, &st) == -1) {
188                 td_verror(td, errno, "fstat");
189                 return 1;
190         }
191
192         f->real_file_size = st.st_size;
193         return 0;
194 }
195
196 static int bdev_size(struct thread_data *td, struct fio_file *f)
197 {
198         unsigned long long bytes;
199         int r;
200
201         if (td->io_ops->open_file(td, f)) {
202                 log_err("fio: failed opening blockdev %s for size check\n",
203                         f->file_name);
204                 return 1;
205         }
206
207         r = blockdev_size(f->fd, &bytes);
208         if (r) {
209                 td_verror(td, r, "blockdev_size");
210                 goto err;
211         }
212
213         if (!bytes) {
214                 log_err("%s: zero sized block device?\n", f->file_name);
215                 goto err;
216         }
217
218         f->real_file_size = bytes;
219         return 0;
220 err:
221         td->io_ops->close_file(td, f);
222         return 1;
223 }
224
225 static int get_file_size(struct thread_data *td, struct fio_file *f)
226 {
227         int ret = 0;
228
229         if (f->flags & FIO_SIZE_KNOWN)
230                 return 0;
231
232         if (f->filetype == FIO_TYPE_FILE)
233                 ret = file_size(td, f);
234         else if (f->filetype == FIO_TYPE_BD)
235                 ret = bdev_size(td, f);
236         else
237                 f->real_file_size = -1;
238
239         if (ret)
240                 return ret;
241
242         if (f->file_offset > f->real_file_size) {
243                 log_err("%s: offset extends end (%Lu > %Lu)\n", td->o.name,
244                                         f->file_offset, f->real_file_size);
245                 return 1;
246         }
247
248         f->flags |= FIO_SIZE_KNOWN;
249         return 0;
250 }
251
252 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
253                                    unsigned long long off,
254                                    unsigned long long len)
255 {
256         int ret = 0;
257
258         if (len == -1ULL) {
259                 len = f->io_size;
260                 if (len == -1ULL && td->o.fill_device)
261                         return 0;
262         }
263         if (off == -1ULL)
264                 off = f->file_offset;
265
266         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
267                                                                 len);
268
269         /*
270          * FIXME: add blockdev flushing too
271          */
272         if (f->mmap_ptr)
273                 ret = madvise(f->mmap_ptr, f->mmap_sz, MADV_DONTNEED);
274         else if (f->filetype == FIO_TYPE_FILE) {
275                 ret = fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
276         } else if (f->filetype == FIO_TYPE_BD) {
277                 ret = blockdev_invalidate_cache(f->fd);
278                 if (ret < 0 && errno == EACCES && geteuid()) {
279                         if (!root_warn) {
280                                 log_err("fio: only root may flush block "
281                                         "devices. Cache flush bypassed!\n");
282                                 root_warn = 1;
283                         }
284                         ret = 0;
285                 }
286         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
287                 ret = 0;
288
289         if (ret < 0) {
290                 td_verror(td, errno, "invalidate_cache");
291                 return 1;
292         } else if (ret > 0) {
293                 td_verror(td, ret, "invalidate_cache");
294                 return 1;
295         }
296
297         return ret;
298
299 }
300
301 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
302 {
303         if (!(f->flags & FIO_FILE_OPEN))
304                 return 0;
305
306         return __file_invalidate_cache(td, f, -1, -1);
307 }
308
309 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
310 {
311         int ret = 0;
312
313         dprint(FD_FILE, "fd close %s\n", f->file_name);
314
315         remove_file_hash(f);
316
317         if (close(f->fd) < 0)
318                 ret = errno;
319
320         f->fd = -1;
321         return ret;
322 }
323
324 static int file_lookup_open(struct fio_file *f, int flags)
325 {
326         struct fio_file *__f;
327         int from_hash;
328
329         __f = lookup_file_hash(f->file_name);
330         if (__f) {
331                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
332                 /*
333                  * racy, need the __f->lock locked
334                  */
335                 f->lock = __f->lock;
336                 f->lock_owner = __f->lock_owner;
337                 f->lock_batch = __f->lock_batch;
338                 f->lock_ddir = __f->lock_ddir;
339                 from_hash = 1;
340         } else {
341                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
342                 from_hash = 0;
343         }
344
345         f->fd = open(f->file_name, flags, 0600);
346         return from_hash;
347 }
348
349 int generic_open_file(struct thread_data *td, struct fio_file *f)
350 {
351         int is_std = 0;
352         int flags = 0;
353         int from_hash = 0;
354
355         dprint(FD_FILE, "fd open %s\n", f->file_name);
356
357         if (!strcmp(f->file_name, "-")) {
358                 if (td_rw(td)) {
359                         log_err("fio: can't read/write to stdin/out\n");
360                         return 1;
361                 }
362                 is_std = 1;
363
364                 /*
365                  * move output logging to stderr, if we are writing to stdout
366                  */
367                 if (td_write(td))
368                         f_out = stderr;
369         }
370
371         if (td->o.odirect)
372                 flags |= OS_O_DIRECT;
373         if (td->o.sync_io)
374                 flags |= O_SYNC;
375         if (f->filetype != FIO_TYPE_FILE)
376                 flags |= FIO_O_NOATIME;
377         if (td->o.create_on_open)
378                 flags |= O_CREAT;
379
380 open_again:
381         if (td_write(td)) {
382                 if (!read_only)
383                         flags |= O_RDWR;
384
385                 if (f->filetype == FIO_TYPE_FILE)
386                         flags |= O_CREAT;
387
388                 if (is_std)
389                         f->fd = dup(STDOUT_FILENO);
390                 else
391                         from_hash = file_lookup_open(f, flags);
392         } else {
393                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
394                         flags |= O_RDWR;
395                 else
396                         flags |= O_RDONLY;
397
398                 if (is_std)
399                         f->fd = dup(STDIN_FILENO);
400                 else
401                         from_hash = file_lookup_open(f, flags);
402         }
403
404         if (f->fd == -1) {
405                 char buf[FIO_VERROR_SIZE];
406                 int __e = errno;
407
408                 if (errno == EPERM && (flags & FIO_O_NOATIME)) {
409                         flags &= ~FIO_O_NOATIME;
410                         goto open_again;
411                 }
412
413                 snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
414
415                 td_verror(td, __e, buf);
416         }
417
418         if (!from_hash && f->fd != -1) {
419                 if (add_file_hash(f)) {
420                         int ret;
421
422                         /*
423                          * OK to ignore, we haven't done anything with it
424                          */
425                         ret = generic_close_file(td, f);
426                         goto open_again;
427                 }
428         }
429
430         return 0;
431 }
432
433 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
434 {
435         return get_file_size(td, f);
436 }
437
438 /*
439  * open/close all files, so that ->real_file_size gets set
440  */
441 static int get_file_sizes(struct thread_data *td)
442 {
443         struct fio_file *f;
444         unsigned int i;
445         int err = 0;
446
447         for_each_file(td, f, i) {
448                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
449                                                                 f->file_name);
450
451                 if (td_io_get_file_size(td, f)) {
452                         if (td->error != ENOENT) {
453                                 log_err("%s\n", td->verror);
454                                 err = 1;
455                         }
456                         clear_error(td);
457                 }
458
459                 if (f->real_file_size == -1ULL && td->o.size)
460                         f->real_file_size = td->o.size / td->o.nr_files;
461         }
462
463         return err;
464 }
465
466 /*
467  * Open the files and setup files sizes, creating files if necessary.
468  */
469 int setup_files(struct thread_data *td)
470 {
471         unsigned long long total_size, extend_size;
472         struct fio_file *f;
473         unsigned int i;
474         int err = 0, need_extend;
475
476         dprint(FD_FILE, "setup files\n");
477
478         if (td->o.read_iolog_file)
479                 return 0;
480
481         /*
482          * if ioengine defines a setup() method, it's responsible for
483          * opening the files and setting f->real_file_size to indicate
484          * the valid range for that file.
485          */
486         if (td->io_ops->setup)
487                 err = td->io_ops->setup(td);
488         else
489                 err = get_file_sizes(td);
490
491         if (err)
492                 return err;
493
494         /*
495          * check sizes. if the files/devices do not exist and the size
496          * isn't passed to fio, abort.
497          */
498         total_size = 0;
499         for_each_file(td, f, i) {
500                 if (f->real_file_size == -1ULL)
501                         total_size = -1ULL;
502                 else
503                         total_size += f->real_file_size;
504         }
505
506         /*
507          * device/file sizes are zero and no size given, punt
508          */
509         if ((!total_size || total_size == -1ULL) && !td->o.size &&
510             !(td->io_ops->flags & FIO_NOIO) && !td->o.fill_device) {
511                 log_err("%s: you need to specify size=\n", td->o.name);
512                 td_verror(td, EINVAL, "total_file_size");
513                 return 1;
514         }
515
516         /*
517          * now file sizes are known, so we can set ->io_size. if size= is
518          * not given, ->io_size is just equal to ->real_file_size. if size
519          * is given, ->io_size is size / nr_files.
520          */
521         extend_size = total_size = 0;
522         need_extend = 0;
523         for_each_file(td, f, i) {
524                 f->file_offset = td->o.start_offset;
525
526                 if (!td->o.file_size_low) {
527                         /*
528                          * no file size range given, file size is equal to
529                          * total size divided by number of files. if that is
530                          * zero, set it to the real file size.
531                          */
532                         f->io_size = td->o.size / td->o.nr_files;
533                         if (!f->io_size)
534                                 f->io_size = f->real_file_size - f->file_offset;
535                 } else if (f->real_file_size < td->o.file_size_low ||
536                            f->real_file_size > td->o.file_size_high) {
537                         if (f->file_offset > td->o.file_size_low)
538                                 goto err_offset;
539                         /*
540                          * file size given. if it's fixed, use that. if it's a
541                          * range, generate a random size in-between.
542                          */
543                         if (td->o.file_size_low == td->o.file_size_high) {
544                                 f->io_size = td->o.file_size_low
545                                                 - f->file_offset;
546                         } else {
547                                 f->io_size = get_rand_file_size(td)
548                                                 - f->file_offset;
549                         }
550                 } else
551                         f->io_size = f->real_file_size - f->file_offset;
552
553                 if (f->io_size == -1ULL)
554                         total_size = -1ULL;
555                 else
556                         total_size += f->io_size;
557
558                 if (f->filetype == FIO_TYPE_FILE &&
559                     (f->io_size + f->file_offset) > f->real_file_size &&
560                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
561                         if (!td->o.create_on_open) {
562                                 need_extend++;
563                                 extend_size += (f->io_size + f->file_offset);
564                         } else
565                                 f->real_file_size = f->io_size + f->file_offset;
566                         f->flags |= FIO_FILE_EXTEND;
567                 }
568         }
569
570         if (!td->o.size || td->o.size > total_size)
571                 td->o.size = total_size;
572
573         /*
574          * See if we need to extend some files
575          */
576         if (need_extend) {
577                 temp_stall_ts = 1;
578                 if (!terse_output)
579                         log_info("%s: Laying out IO file(s) (%u file(s) /"
580                                  " %LuMiB)\n", td->o.name, need_extend,
581                                         extend_size >> 20);
582
583                 for_each_file(td, f, i) {
584                         unsigned long long old_len, extend_len;
585
586                         if (!(f->flags & FIO_FILE_EXTEND))
587                                 continue;
588
589                         assert(f->filetype == FIO_TYPE_FILE);
590                         f->flags &= ~FIO_FILE_EXTEND;
591                         old_len = f->real_file_size;
592                         extend_len = f->io_size + f->file_offset - old_len;
593                         f->real_file_size = (f->io_size + f->file_offset);
594                         err = extend_file(td, f);
595                         if (err)
596                                 break;
597                         
598                         err = __file_invalidate_cache(td, f, old_len,
599                                                                 extend_len);
600                         close(f->fd);
601                         f->fd = -1;
602                         if (err)
603                                 break;
604                 }
605                 temp_stall_ts = 0;
606         }
607
608         if (err)
609                 return err;
610
611         if (!td->o.zone_size)
612                 td->o.zone_size = td->o.size;
613
614         /*
615          * iolog already set the total io size, if we read back
616          * stored entries.
617          */
618         if (!td->o.read_iolog_file)
619                 td->total_io_size = td->o.size * td->o.loops;
620         return 0;
621 err_offset:
622         log_err("%s: you need to specify valid offset=\n", td->o.name);
623         return 1;
624 }
625
626 int pre_read_files(struct thread_data *td)
627 {
628         struct fio_file *f;
629         unsigned int i;
630
631         dprint(FD_FILE, "pre_read files\n");
632
633         for_each_file(td, f, i) {
634                 pre_read_file(td, f);
635         }
636
637         return 1;
638 }
639
640 int init_random_map(struct thread_data *td)
641 {
642         unsigned long long blocks, num_maps;
643         struct fio_file *f;
644         unsigned int i;
645
646         if (td->o.norandommap || !td_random(td))
647                 return 0;
648
649         for_each_file(td, f, i) {
650                 blocks = (f->real_file_size + td->o.rw_min_bs - 1) /
651                                 (unsigned long long) td->o.rw_min_bs;
652                 num_maps = (blocks + BLOCKS_PER_MAP - 1) /
653                                 (unsigned long long) BLOCKS_PER_MAP;
654                 f->file_map = smalloc(num_maps * sizeof(int));
655                 if (f->file_map) {
656                         f->num_maps = num_maps;
657                         continue;
658                 }
659                 if (!td->o.softrandommap) {
660                         log_err("fio: failed allocating random map. If running"
661                                 " a large number of jobs, try the 'norandommap'"
662                                 " option or set 'softrandommap'. Or give"
663                                 " a larger --alloc-size to fio.\n");
664                         return 1;
665                 }
666
667                 log_info("fio: file %s failed allocating random map. Running "
668                          "job without.\n", f->file_name);
669                 f->num_maps = 0;
670         }
671
672         return 0;
673 }
674
675 void close_files(struct thread_data *td)
676 {
677         struct fio_file *f;
678         unsigned int i;
679
680         for_each_file(td, f, i)
681                 td_io_close_file(td, f);
682 }
683
684 void close_and_free_files(struct thread_data *td)
685 {
686         struct fio_file *f;
687         unsigned int i;
688
689         dprint(FD_FILE, "close files\n");
690
691         for_each_file(td, f, i) {
692                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
693                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
694                         unlink(f->file_name);
695                 }
696
697                 td_io_close_file(td, f);
698                 remove_file_hash(f);
699
700                 sfree(f->file_name);
701                 f->file_name = NULL;
702
703                 if (f->file_map) {
704                         sfree(f->file_map);
705                         f->file_map = NULL;
706                 }
707                 sfree(f);
708         }
709
710         td->o.filename = NULL;
711         free(td->files);
712         td->files_index = 0;
713         td->files = NULL;
714         td->o.nr_files = 0;
715 }
716
717 static void get_file_type(struct fio_file *f)
718 {
719         struct stat sb;
720
721         if (!strcmp(f->file_name, "-"))
722                 f->filetype = FIO_TYPE_PIPE;
723         else
724                 f->filetype = FIO_TYPE_FILE;
725
726         if (!lstat(f->file_name, &sb)) {
727                 if (S_ISBLK(sb.st_mode))
728                         f->filetype = FIO_TYPE_BD;
729                 else if (S_ISCHR(sb.st_mode))
730                         f->filetype = FIO_TYPE_CHAR;
731                 else if (S_ISFIFO(sb.st_mode))
732                         f->filetype = FIO_TYPE_PIPE;
733         }
734 }
735
736 int add_file(struct thread_data *td, const char *fname)
737 {
738         int cur_files = td->files_index;
739         char file_name[PATH_MAX];
740         struct fio_file *f;
741         int len = 0;
742
743         dprint(FD_FILE, "add file %s\n", fname);
744
745         f = smalloc(sizeof(*f));
746         if (!f) {
747                 log_err("fio: smalloc OOM\n");
748                 assert(0);
749         }
750                 
751         f->fd = -1;
752
753         if (td->files_size <= td->files_index) {
754                 int new_size = td->o.nr_files + 1;
755
756                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
757
758                 td->files = realloc(td->files, new_size * sizeof(f));
759                 td->files_size = new_size;
760         }
761         td->files[cur_files] = f;
762
763         /*
764          * init function, io engine may not be loaded yet
765          */
766         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
767                 f->real_file_size = -1ULL;
768
769         if (td->o.directory)
770                 len = sprintf(file_name, "%s/", td->o.directory);
771
772         sprintf(file_name + len, "%s", fname);
773         f->file_name = smalloc_strdup(file_name);
774         if (!f->file_name) {
775                 log_err("fio: smalloc OOM\n");
776                 assert(0);
777         }
778         
779         get_file_type(f);
780
781         switch (td->o.file_lock_mode) {
782         case FILE_LOCK_NONE:
783                 break;
784         case FILE_LOCK_READWRITE:
785                 f->lock = fio_mutex_rw_init();
786                 break;
787         case FILE_LOCK_EXCLUSIVE:
788                 f->lock = fio_mutex_init(1);
789                 break;
790         default:
791                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
792                 assert(0);
793         }
794
795         td->files_index++;
796         if (f->filetype == FIO_TYPE_FILE)
797                 td->nr_normal_files++;
798
799         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
800                                                         cur_files);
801
802         return cur_files;
803 }
804
805 void get_file(struct fio_file *f)
806 {
807         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
808         assert(f->flags & FIO_FILE_OPEN);
809         f->references++;
810 }
811
812 int put_file(struct thread_data *td, struct fio_file *f)
813 {
814         int f_ret = 0, ret = 0;
815
816         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
817
818         if (!(f->flags & FIO_FILE_OPEN))
819                 return 0;
820
821         assert(f->references);
822         if (--f->references)
823                 return 0;
824
825         if (should_fsync(td) && td->o.fsync_on_close)
826                 f_ret = fsync(f->fd);
827
828         if (td->io_ops->close_file)
829                 ret = td->io_ops->close_file(td, f);
830
831         if (!ret)
832                 ret = f_ret;
833
834         td->nr_open_files--;
835         f->flags &= ~FIO_FILE_OPEN;
836         return ret;
837 }
838
839 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
840 {
841         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
842                 return;
843
844         if (f->lock_owner == td && f->lock_batch--)
845                 return;
846
847         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
848                 if (ddir == DDIR_READ)
849                         fio_mutex_down_read(f->lock);
850                 else
851                         fio_mutex_down_write(f->lock);
852         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
853                 fio_mutex_down(f->lock);
854
855         f->lock_owner = td;
856         f->lock_batch = td->o.lockfile_batch;
857         f->lock_ddir = ddir;
858 }
859
860 void unlock_file(struct thread_data *td, struct fio_file *f)
861 {
862         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
863                 return;
864         if (f->lock_batch)
865                 return;
866
867         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
868                 const int is_read = f->lock_ddir == DDIR_READ;
869                 int val = fio_mutex_getval(f->lock);
870
871                 if ((is_read && val == 1) || (!is_read && val == -1))
872                         f->lock_owner = NULL;
873
874                 if (is_read)
875                         fio_mutex_up_read(f->lock);
876                 else
877                         fio_mutex_up_write(f->lock);
878         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) {
879                 int val = fio_mutex_getval(f->lock);
880
881                 if (val == 0)
882                         f->lock_owner = NULL;
883
884                 fio_mutex_up(f->lock);
885         }
886 }
887
888 void unlock_file_all(struct thread_data *td, struct fio_file *f)
889 {
890         if (f->lock_owner != td)
891                 return;
892
893         f->lock_batch = 0;
894         unlock_file(td, f);
895 }
896
897 static int recurse_dir(struct thread_data *td, const char *dirname)
898 {
899         struct dirent *dir;
900         int ret = 0;
901         DIR *D;
902
903         D = opendir(dirname);
904         if (!D) {
905                 char buf[FIO_VERROR_SIZE];
906
907                 snprintf(buf, FIO_VERROR_SIZE - 1, "opendir(%s)", dirname);
908                 td_verror(td, errno, buf);
909                 return 1;
910         }
911
912         while ((dir = readdir(D)) != NULL) {
913                 char full_path[PATH_MAX];
914                 struct stat sb;
915
916                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
917                         continue;
918
919                 sprintf(full_path, "%s/%s", dirname, dir->d_name);
920
921                 if (lstat(full_path, &sb) == -1) {
922                         if (errno != ENOENT) {
923                                 td_verror(td, errno, "stat");
924                                 return 1;
925                         }
926                 }
927
928                 if (S_ISREG(sb.st_mode)) {
929                         add_file(td, full_path);
930                         td->o.nr_files++;
931                         continue;
932                 }
933                 if (!S_ISDIR(sb.st_mode))
934                         continue;
935
936                 ret = recurse_dir(td, full_path);
937                 if (ret)
938                         break;
939         }
940
941         closedir(D);
942         return ret;
943 }
944
945 int add_dir_files(struct thread_data *td, const char *path)
946 {
947         int ret = recurse_dir(td, path);
948
949         if (!ret)
950                 log_info("fio: opendir added %d files\n", td->o.nr_files);
951
952         return ret;
953 }
954
955 void dup_files(struct thread_data *td, struct thread_data *org)
956 {
957         struct fio_file *f;
958         unsigned int i;
959
960         dprint(FD_FILE, "dup files: %d\n", org->files_index);
961
962         if (!org->files)
963                 return;
964
965         td->files = malloc(org->files_index * sizeof(f));
966
967         for_each_file(org, f, i) {
968                 struct fio_file *__f;
969
970                 __f = smalloc(sizeof(*__f));
971                 if (!__f) {
972                         log_err("fio: smalloc OOM\n");
973                         assert(0);
974                 }
975         
976                 if (f->file_name) {
977                         __f->file_name = smalloc_strdup(f->file_name);
978                         if (!__f->file_name) {
979                                 log_err("fio: smalloc OOM\n");
980                                 assert(0);
981                         }
982         
983                         __f->filetype = f->filetype;
984                 }
985
986                 td->files[i] = __f;
987         }
988 }
989
990 /*
991  * Returns the index that matches the filename, or -1 if not there
992  */
993 int get_fileno(struct thread_data *td, const char *fname)
994 {
995         struct fio_file *f;
996         unsigned int i;
997
998         for_each_file(td, f, i)
999                 if (!strcmp(f->file_name, fname))
1000                         return i;
1001
1002         return -1;
1003 }
1004
1005 /*
1006  * For log usage, where we add/open/close files automatically
1007  */
1008 void free_release_files(struct thread_data *td)
1009 {
1010         close_files(td);
1011         td->files_index = 0;
1012         td->nr_normal_files = 0;
1013 }