Merge branch 'fgp_fixes' of https://github.com/sitsofe/fio
[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 <libgen.h>
7 #include <sys/stat.h>
8 #include <sys/mman.h>
9 #include <sys/types.h>
10
11 #include "fio.h"
12 #include "smalloc.h"
13 #include "filehash.h"
14 #include "options.h"
15 #include "os/os.h"
16 #include "hash.h"
17 #include "lib/axmap.h"
18
19 #ifdef CONFIG_LINUX_FALLOCATE
20 #include <linux/falloc.h>
21 #endif
22
23 static int root_warn;
24
25 static FLIST_HEAD(filename_list);
26
27 /*
28  * List entry for filename_list
29  */
30 struct file_name {
31         struct flist_head list;
32         char *filename;
33 };
34
35 static inline void clear_error(struct thread_data *td)
36 {
37         td->error = 0;
38         td->verror[0] = '\0';
39 }
40
41 static inline int native_fallocate(struct thread_data *td, struct fio_file *f)
42 {
43         bool success;
44
45         success = fio_fallocate(f, 0, f->real_file_size);
46         dprint(FD_FILE, "native fallocate of file %s size %llu was "
47                         "%ssuccessful\n", f->file_name,
48                         (unsigned long long) f->real_file_size,
49                         !success ? "un": "");
50
51         if (success)
52                 return 0;
53
54         if (errno == ENOSYS)
55                 dprint(FD_FILE, "native fallocate is not implemented\n");
56
57         return -1;
58 }
59
60 static void fallocate_file(struct thread_data *td, struct fio_file *f)
61 {
62         int r;
63
64         if (td->o.fill_device)
65                 return;
66
67         switch (td->o.fallocate_mode) {
68         case FIO_FALLOCATE_NATIVE:
69                 r = native_fallocate(td, f);
70                 if (r != 0 && errno != ENOSYS)
71                         log_err("fio: native_fallocate call failed: %s\n",
72                                         strerror(errno));
73                 break;
74         case FIO_FALLOCATE_NONE:
75                 break;
76 #ifdef CONFIG_POSIX_FALLOCATE
77         case FIO_FALLOCATE_POSIX:
78                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
79                                  f->file_name,
80                                  (unsigned long long) f->real_file_size);
81
82                 r = posix_fallocate(f->fd, 0, f->real_file_size);
83                 if (r > 0)
84                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
85                 break;
86 #endif /* CONFIG_POSIX_FALLOCATE */
87 #ifdef CONFIG_LINUX_FALLOCATE
88         case FIO_FALLOCATE_KEEP_SIZE:
89                 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
90                                 "file %s size %llu\n", f->file_name,
91                                 (unsigned long long) f->real_file_size);
92
93                 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
94                 if (r != 0)
95                         td_verror(td, errno, "fallocate");
96
97                 break;
98 #endif /* CONFIG_LINUX_FALLOCATE */
99         default:
100                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
101                 assert(0);
102         }
103 }
104
105 /*
106  * Leaves f->fd open on success, caller must close
107  */
108 static int extend_file(struct thread_data *td, struct fio_file *f)
109 {
110         int new_layout = 0, unlink_file = 0, flags;
111         unsigned long long left;
112         unsigned int bs;
113         char *b = NULL;
114
115         if (read_only) {
116                 log_err("fio: refusing extend of file due to read-only\n");
117                 return 0;
118         }
119
120         /*
121          * check if we need to lay the file out complete again. fio
122          * does that for operations involving reads, or for writes
123          * where overwrite is set
124          */
125         if (td_read(td) ||
126            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
127             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
128                 new_layout = 1;
129         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
130                 unlink_file = 1;
131
132         if (unlink_file || new_layout) {
133                 int ret;
134
135                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
136
137                 ret = td_io_unlink_file(td, f);
138                 if (ret != 0 && ret != ENOENT) {
139                         td_verror(td, errno, "unlink");
140                         return 1;
141                 }
142         }
143
144         flags = O_WRONLY;
145         if (td->o.allow_create)
146                 flags |= O_CREAT;
147         if (new_layout)
148                 flags |= O_TRUNC;
149
150 #ifdef WIN32
151         flags |= _O_BINARY;
152 #endif
153
154         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
155         f->fd = open(f->file_name, flags, 0644);
156         if (f->fd < 0) {
157                 int err = errno;
158
159                 if (err == ENOENT && !td->o.allow_create)
160                         log_err("fio: file creation disallowed by "
161                                         "allow_file_create=0\n");
162                 else
163                         td_verror(td, err, "open");
164                 return 1;
165         }
166
167         fallocate_file(td, f);
168
169         /*
170          * If our jobs don't require regular files initially, we're done.
171          */
172         if (!new_layout)
173                 goto done;
174
175         /*
176          * The size will be -1ULL when fill_device is used, so don't truncate
177          * or fallocate this file, just write it
178          */
179         if (!td->o.fill_device) {
180                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
181                                         (unsigned long long) f->real_file_size);
182                 if (ftruncate(f->fd, f->real_file_size) == -1) {
183                         if (errno != EFBIG) {
184                                 td_verror(td, errno, "ftruncate");
185                                 goto err;
186                         }
187                 }
188         }
189
190         left = f->real_file_size;
191         bs = td->o.max_bs[DDIR_WRITE];
192         if (bs > left)
193                 bs = left;
194
195         b = malloc(bs);
196         if (!b) {
197                 td_verror(td, errno, "malloc");
198                 goto err;
199         }
200
201         while (left && !td->terminate) {
202                 ssize_t r;
203
204                 if (bs > left)
205                         bs = left;
206
207                 fill_io_buffer(td, b, bs, bs);
208
209                 r = write(f->fd, b, bs);
210
211                 if (r > 0) {
212                         left -= r;
213                         continue;
214                 } else {
215                         if (r < 0) {
216                                 int __e = errno;
217
218                                 if (__e == ENOSPC) {
219                                         if (td->o.fill_device)
220                                                 break;
221                                         log_info("fio: ENOSPC on laying out "
222                                                  "file, stopping\n");
223                                         break;
224                                 }
225                                 td_verror(td, errno, "write");
226                         } else
227                                 td_verror(td, EIO, "write");
228
229                         break;
230                 }
231         }
232
233         if (td->terminate) {
234                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
235                 td_io_unlink_file(td, f);
236         } else if (td->o.create_fsync) {
237                 if (fsync(f->fd) < 0) {
238                         td_verror(td, errno, "fsync");
239                         goto err;
240                 }
241         }
242         if (td->o.fill_device && !td_write(td)) {
243                 fio_file_clear_size_known(f);
244                 if (td_io_get_file_size(td, f))
245                         goto err;
246                 if (f->io_size > f->real_file_size)
247                         f->io_size = f->real_file_size;
248         }
249
250         free(b);
251 done:
252         return 0;
253 err:
254         close(f->fd);
255         f->fd = -1;
256         if (b)
257                 free(b);
258         return 1;
259 }
260
261 static int pre_read_file(struct thread_data *td, struct fio_file *f)
262 {
263         int ret = 0, r, did_open = 0, old_runstate;
264         unsigned long long left;
265         unsigned int bs;
266         char *b;
267
268         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
269             td_ioengine_flagged(td, FIO_NOIO))
270                 return 0;
271
272         if (f->filetype == FIO_TYPE_CHAR)
273                 return 0;
274
275         if (!fio_file_open(f)) {
276                 if (td->io_ops->open_file(td, f)) {
277                         log_err("fio: cannot pre-read, failed to open file\n");
278                         return 1;
279                 }
280                 did_open = 1;
281         }
282
283         old_runstate = td_bump_runstate(td, TD_PRE_READING);
284
285         left = f->io_size;
286         bs = td->o.max_bs[DDIR_READ];
287         if (bs > left)
288                 bs = left;
289
290         b = malloc(bs);
291         if (!b) {
292                 td_verror(td, errno, "malloc");
293                 ret = 1;
294                 goto error;
295         }
296         memset(b, 0, bs);
297
298         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
299                 td_verror(td, errno, "lseek");
300                 log_err("fio: failed to lseek pre-read file\n");
301                 ret = 1;
302                 goto error;
303         }
304
305         while (left && !td->terminate) {
306                 if (bs > left)
307                         bs = left;
308
309                 r = read(f->fd, b, bs);
310
311                 if (r == (int) bs) {
312                         left -= bs;
313                         continue;
314                 } else {
315                         td_verror(td, EIO, "pre_read");
316                         break;
317                 }
318         }
319
320 error:
321         td_restore_runstate(td, old_runstate);
322
323         if (did_open)
324                 td->io_ops->close_file(td, f);
325
326         free(b);
327         return ret;
328 }
329
330 unsigned long long get_rand_file_size(struct thread_data *td)
331 {
332         unsigned long long ret, sized;
333         uint64_t frand_max;
334         unsigned long r;
335
336         frand_max = rand_max(&td->file_size_state);
337         r = __rand(&td->file_size_state);
338         sized = td->o.file_size_high - td->o.file_size_low;
339         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
340         ret += td->o.file_size_low;
341         ret -= (ret % td->o.rw_min_bs);
342         return ret;
343 }
344
345 static int file_size(struct thread_data *td, struct fio_file *f)
346 {
347         struct stat st;
348
349         if (stat(f->file_name, &st) == -1) {
350                 td_verror(td, errno, "fstat");
351                 return 1;
352         }
353
354         f->real_file_size = st.st_size;
355         return 0;
356 }
357
358 static int bdev_size(struct thread_data *td, struct fio_file *f)
359 {
360         unsigned long long bytes = 0;
361         int r;
362
363         if (td->io_ops->open_file(td, f)) {
364                 log_err("fio: failed opening blockdev %s for size check\n",
365                         f->file_name);
366                 return 1;
367         }
368
369         r = blockdev_size(f, &bytes);
370         if (r) {
371                 td_verror(td, r, "blockdev_size");
372                 goto err;
373         }
374
375         if (!bytes) {
376                 log_err("%s: zero sized block device?\n", f->file_name);
377                 goto err;
378         }
379
380         f->real_file_size = bytes;
381         td->io_ops->close_file(td, f);
382         return 0;
383 err:
384         td->io_ops->close_file(td, f);
385         return 1;
386 }
387
388 static int char_size(struct thread_data *td, struct fio_file *f)
389 {
390 #ifdef FIO_HAVE_CHARDEV_SIZE
391         unsigned long long bytes = 0;
392         int r;
393
394         if (td->io_ops->open_file(td, f)) {
395                 log_err("fio: failed opening chardev %s for size check\n",
396                         f->file_name);
397                 return 1;
398         }
399
400         r = chardev_size(f, &bytes);
401         if (r) {
402                 td_verror(td, r, "chardev_size");
403                 goto err;
404         }
405
406         if (!bytes) {
407                 log_err("%s: zero sized char device?\n", f->file_name);
408                 goto err;
409         }
410
411         f->real_file_size = bytes;
412         td->io_ops->close_file(td, f);
413         return 0;
414 err:
415         td->io_ops->close_file(td, f);
416         return 1;
417 #else
418         f->real_file_size = -1ULL;
419         return 0;
420 #endif
421 }
422
423 static int get_file_size(struct thread_data *td, struct fio_file *f)
424 {
425         int ret = 0;
426
427         if (fio_file_size_known(f))
428                 return 0;
429
430         if (f->filetype == FIO_TYPE_FILE)
431                 ret = file_size(td, f);
432         else if (f->filetype == FIO_TYPE_BLOCK)
433                 ret = bdev_size(td, f);
434         else if (f->filetype == FIO_TYPE_CHAR)
435                 ret = char_size(td, f);
436         else
437                 f->real_file_size = -1ULL;
438
439         /*
440          * Leave ->real_file_size with 0 since it could be expectation
441          * of initial setup for regular files.
442          */
443         if (ret)
444                 return ret;
445
446         /*
447          * If ->real_file_size is -1, a conditional for the message
448          * "offset extends end" is always true, but it makes no sense,
449          * so just return the same value here.
450          */
451         if (f->real_file_size == -1ULL) {
452                 log_info("%s: failed to get file size of %s\n", td->o.name,
453                                         f->file_name);
454                 return 1;
455         }
456
457         if (td->o.start_offset && f->file_offset == 0)
458                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
459                                         f->file_name);
460         /*
461          * ->file_offset normally hasn't been initialized yet, so this
462          * is basically always false.
463          */
464         if (f->file_offset > f->real_file_size) {
465                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
466                                         (unsigned long long) f->file_offset,
467                                         (unsigned long long) f->real_file_size);
468                 return 1;
469         }
470
471         fio_file_set_size_known(f);
472         return 0;
473 }
474
475 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
476                                    unsigned long long off,
477                                    unsigned long long len)
478 {
479         int errval = 0, ret = 0;
480
481 #ifdef CONFIG_ESX
482         return 0;
483 #endif
484
485         if (len == -1ULL)
486                 len = f->io_size;
487         if (off == -1ULL)
488                 off = f->file_offset;
489
490         if (len == -1ULL || off == -1ULL)
491                 return 0;
492
493         if (td->io_ops->invalidate) {
494                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
495                         f->file_name);
496                 ret = td->io_ops->invalidate(td, f);
497                 if (ret < 0)
498                         errval = -ret;
499         } else if (f->filetype == FIO_TYPE_FILE) {
500                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
501                         f->file_name, off, len);
502                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
503                 if (ret)
504                         errval = ret;
505         } else if (f->filetype == FIO_TYPE_BLOCK) {
506                 int retry_count = 0;
507
508                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
509                 ret = blockdev_invalidate_cache(f);
510                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
511                         /*
512                          * Linux multipath devices reject ioctl while
513                          * the maps are being updated. That window can
514                          * last tens of milliseconds; we'll try up to
515                          * a quarter of a second.
516                          */
517                         usleep(10000);
518                         ret = blockdev_invalidate_cache(f);
519                 }
520                 if (ret < 0 && errno == EACCES && geteuid()) {
521                         if (!root_warn) {
522                                 log_err("fio: only root may flush block "
523                                         "devices. Cache flush bypassed!\n");
524                                 root_warn = 1;
525                         }
526                         ret = 0;
527                 }
528                 if (ret < 0)
529                         errval = errno;
530         } else if (f->filetype == FIO_TYPE_CHAR ||
531                    f->filetype == FIO_TYPE_PIPE) {
532                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
533                 ret = 0;
534         }
535
536         /*
537          * Cache flushing isn't a fatal condition, and we know it will
538          * happen on some platforms where we don't have the proper
539          * function to flush eg block device caches. So just warn and
540          * continue on our way.
541          */
542         if (errval)
543                 log_info("fio: cache invalidation of %s failed: %s\n",
544                          f->file_name, strerror(errval));
545
546         return 0;
547
548 }
549
550 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
551 {
552         if (!fio_file_open(f))
553                 return 0;
554
555         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
556 }
557
558 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
559 {
560         int ret = 0;
561
562         dprint(FD_FILE, "fd close %s\n", f->file_name);
563
564         remove_file_hash(f);
565
566         if (close(f->fd) < 0)
567                 ret = errno;
568
569         f->fd = -1;
570
571         if (f->shadow_fd != -1) {
572                 close(f->shadow_fd);
573                 f->shadow_fd = -1;
574         }
575
576         f->engine_pos = 0;
577         return ret;
578 }
579
580 int file_lookup_open(struct fio_file *f, int flags)
581 {
582         struct fio_file *__f;
583         int from_hash;
584
585         __f = lookup_file_hash(f->file_name);
586         if (__f) {
587                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
588                 f->lock = __f->lock;
589                 from_hash = 1;
590         } else {
591                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
592                 from_hash = 0;
593         }
594
595 #ifdef WIN32
596         flags |= _O_BINARY;
597 #endif
598
599         f->fd = open(f->file_name, flags, 0600);
600         return from_hash;
601 }
602
603 static int file_close_shadow_fds(struct thread_data *td)
604 {
605         struct fio_file *f;
606         int num_closed = 0;
607         unsigned int i;
608
609         for_each_file(td, f, i) {
610                 if (f->shadow_fd == -1)
611                         continue;
612
613                 close(f->shadow_fd);
614                 f->shadow_fd = -1;
615                 num_closed++;
616         }
617
618         return num_closed;
619 }
620
621 int generic_open_file(struct thread_data *td, struct fio_file *f)
622 {
623         int is_std = 0;
624         int flags = 0;
625         int from_hash = 0;
626
627         dprint(FD_FILE, "fd open %s\n", f->file_name);
628
629         if (!strcmp(f->file_name, "-")) {
630                 if (td_rw(td)) {
631                         log_err("fio: can't read/write to stdin/out\n");
632                         return 1;
633                 }
634                 is_std = 1;
635
636                 /*
637                  * move output logging to stderr, if we are writing to stdout
638                  */
639                 if (td_write(td))
640                         f_out = stderr;
641         }
642
643         if (td_trim(td))
644                 goto skip_flags;
645         if (td->o.odirect)
646                 flags |= OS_O_DIRECT;
647         if (td->o.oatomic) {
648                 if (!FIO_O_ATOMIC) {
649                         td_verror(td, EINVAL, "OS does not support atomic IO");
650                         return 1;
651                 }
652                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
653         }
654         if (td->o.sync_io)
655                 flags |= O_SYNC;
656         if (td->o.create_on_open && td->o.allow_create)
657                 flags |= O_CREAT;
658 skip_flags:
659         if (f->filetype != FIO_TYPE_FILE)
660                 flags |= FIO_O_NOATIME;
661
662 open_again:
663         if (td_write(td)) {
664                 if (!read_only)
665                         flags |= O_RDWR;
666
667                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
668                         flags |= O_CREAT;
669
670                 if (is_std)
671                         f->fd = dup(STDOUT_FILENO);
672                 else
673                         from_hash = file_lookup_open(f, flags);
674         } else if (td_read(td)) {
675                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
676                         flags |= O_RDWR;
677                 else
678                         flags |= O_RDONLY;
679
680                 if (is_std)
681                         f->fd = dup(STDIN_FILENO);
682                 else
683                         from_hash = file_lookup_open(f, flags);
684         } else if (td_trim(td)) {
685                 assert(!td_rw(td)); /* should have matched above */
686                 flags |= O_RDWR;
687                 from_hash = file_lookup_open(f, flags);
688         }
689
690         if (f->fd == -1) {
691                 char buf[FIO_VERROR_SIZE];
692                 int __e = errno;
693
694                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
695                         flags &= ~FIO_O_NOATIME;
696                         goto open_again;
697                 }
698                 if (__e == EMFILE && file_close_shadow_fds(td))
699                         goto open_again;
700
701                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
702
703                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
704                         log_err("fio: looks like your file system does not " \
705                                 "support direct=1/buffered=0\n");
706                 }
707
708                 td_verror(td, __e, buf);
709                 return 1;
710         }
711
712         if (!from_hash && f->fd != -1) {
713                 if (add_file_hash(f)) {
714                         int fio_unused ret;
715
716                         /*
717                          * Stash away descriptor for later close. This is to
718                          * work-around a "feature" on Linux, where a close of
719                          * an fd that has been opened for write will trigger
720                          * udev to call blkid to check partitions, fs id, etc.
721                          * That pollutes the device cache, which can slow down
722                          * unbuffered accesses.
723                          */
724                         if (f->shadow_fd == -1)
725                                 f->shadow_fd = f->fd;
726                         else {
727                                 /*
728                                  * OK to ignore, we haven't done anything
729                                  * with it
730                                  */
731                                 ret = generic_close_file(td, f);
732                         }
733                         goto open_again;
734                 }
735         }
736
737         return 0;
738 }
739
740 /*
741  * This function i.e. get_file_size() is the default .get_file_size
742  * implementation of majority of I/O engines.
743  */
744 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
745 {
746         return get_file_size(td, f);
747 }
748
749 /*
750  * open/close all files, so that ->real_file_size gets set
751  */
752 static int get_file_sizes(struct thread_data *td)
753 {
754         struct fio_file *f;
755         unsigned int i;
756         int err = 0;
757
758         for_each_file(td, f, i) {
759                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
760                                                                 f->file_name);
761
762                 if (td_io_get_file_size(td, f)) {
763                         if (td->error != ENOENT) {
764                                 log_err("%s\n", td->verror);
765                                 err = 1;
766                                 break;
767                         }
768                         clear_error(td);
769                 }
770
771                 /*
772                  * There are corner cases where we end up with -1 for
773                  * ->real_file_size due to unsupported file type, etc.
774                  * We then just set to size option value divided by number
775                  * of files, similar to the way file ->io_size is set.
776                  * stat(2) failure doesn't set ->real_file_size to -1.
777                  */
778                 if (f->real_file_size == -1ULL && td->o.size)
779                         f->real_file_size = td->o.size / td->o.nr_files;
780         }
781
782         return err;
783 }
784
785 struct fio_mount {
786         struct flist_head list;
787         const char *base;
788         char __base[256];
789         unsigned int key;
790 };
791
792 /*
793  * Get free number of bytes for each file on each unique mount.
794  */
795 static unsigned long long get_fs_free_counts(struct thread_data *td)
796 {
797         struct flist_head *n, *tmp;
798         unsigned long long ret = 0;
799         struct fio_mount *fm;
800         FLIST_HEAD(list);
801         struct fio_file *f;
802         unsigned int i;
803
804         for_each_file(td, f, i) {
805                 struct stat sb;
806                 char buf[256];
807
808                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
809                         if (f->real_file_size != -1ULL)
810                                 ret += f->real_file_size;
811                         continue;
812                 } else if (f->filetype != FIO_TYPE_FILE)
813                         continue;
814
815                 buf[255] = '\0';
816                 strncpy(buf, f->file_name, 255);
817
818                 if (stat(buf, &sb) < 0) {
819                         if (errno != ENOENT)
820                                 break;
821                         strcpy(buf, ".");
822                         if (stat(buf, &sb) < 0)
823                                 break;
824                 }
825
826                 fm = NULL;
827                 flist_for_each(n, &list) {
828                         fm = flist_entry(n, struct fio_mount, list);
829                         if (fm->key == sb.st_dev)
830                                 break;
831
832                         fm = NULL;
833                 }
834
835                 if (fm)
836                         continue;
837
838                 fm = calloc(1, sizeof(*fm));
839                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
840                 fm->base = basename(fm->__base);
841                 fm->key = sb.st_dev;
842                 flist_add(&fm->list, &list);
843         }
844
845         flist_for_each_safe(n, tmp, &list) {
846                 unsigned long long sz;
847
848                 fm = flist_entry(n, struct fio_mount, list);
849                 flist_del(&fm->list);
850
851                 sz = get_fs_free_size(fm->base);
852                 if (sz && sz != -1ULL)
853                         ret += sz;
854
855                 free(fm);
856         }
857
858         return ret;
859 }
860
861 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
862 {
863         struct thread_options *o = &td->o;
864         unsigned long long align_bs;
865         unsigned long long offset;
866
867         if (o->file_append && f->filetype == FIO_TYPE_FILE)
868                 return f->real_file_size;
869
870         if (o->start_offset_percent > 0) {
871                 /*
872                  * if blockalign is provided, find the min across read, write,
873                  * and trim
874                  */
875                 if (fio_option_is_set(o, ba)) {
876                         align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
877                         align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
878                 } else {
879                         /* else take the minimum block size */
880                         align_bs = td_min_bs(td);
881                 }
882
883                 /* calculate the raw offset */
884                 offset = (f->real_file_size * o->start_offset_percent / 100) +
885                         (td->subjob_number * o->offset_increment);
886
887                 /*
888                  * block align the offset at the next available boundary at
889                  * ceiling(offset / align_bs) * align_bs
890                  */
891                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
892
893         } else {
894                 /* start_offset_percent not set */
895                 offset = o->start_offset +
896                                 td->subjob_number * o->offset_increment;
897         }
898
899         return offset;
900 }
901
902 /*
903  * Open the files and setup files sizes, creating files if necessary.
904  */
905 int setup_files(struct thread_data *td)
906 {
907         unsigned long long total_size, extend_size;
908         struct thread_options *o = &td->o;
909         struct fio_file *f;
910         unsigned int i, nr_fs_extra = 0;
911         int err = 0, need_extend;
912         int old_state;
913         const unsigned int bs = td_min_bs(td);
914         uint64_t fs = 0;
915
916         dprint(FD_FILE, "setup files\n");
917
918         old_state = td_bump_runstate(td, TD_SETTING_UP);
919
920         if (o->read_iolog_file)
921                 goto done;
922
923         /*
924          * Find out physical size of files or devices for this thread,
925          * before we determine I/O size and range of our targets.
926          * If ioengine defines a setup() method, it's responsible for
927          * opening the files and setting f->real_file_size to indicate
928          * the valid range for that file.
929          */
930         if (td->io_ops->setup)
931                 err = td->io_ops->setup(td);
932         else
933                 err = get_file_sizes(td);
934
935         if (err)
936                 goto err_out;
937
938         /*
939          * check sizes. if the files/devices do not exist and the size
940          * isn't passed to fio, abort.
941          */
942         total_size = 0;
943         for_each_file(td, f, i) {
944                 f->fileno = i;
945                 if (f->real_file_size == -1ULL)
946                         total_size = -1ULL;
947                 else
948                         total_size += f->real_file_size;
949         }
950
951         if (o->fill_device)
952                 td->fill_device_size = get_fs_free_counts(td);
953
954         /*
955          * device/file sizes are zero and no size given, punt
956          */
957         if ((!total_size || total_size == -1ULL) && !o->size &&
958             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
959             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
960                 log_err("%s: you need to specify size=\n", o->name);
961                 td_verror(td, EINVAL, "total_file_size");
962                 goto err_out;
963         }
964
965         /*
966          * Calculate per-file size and potential extra size for the
967          * first files, if needed (i.e. if we don't have a fixed size).
968          */
969         if (!o->file_size_low && o->nr_files) {
970                 uint64_t all_fs;
971
972                 fs = o->size / o->nr_files;
973                 all_fs = fs * o->nr_files;
974
975                 if (all_fs < o->size)
976                         nr_fs_extra = (o->size - all_fs) / bs;
977         }
978
979         /*
980          * now file sizes are known, so we can set ->io_size. if size= is
981          * not given, ->io_size is just equal to ->real_file_size. if size
982          * is given, ->io_size is size / nr_files.
983          */
984         extend_size = total_size = 0;
985         need_extend = 0;
986         for_each_file(td, f, i) {
987                 f->file_offset = get_start_offset(td, f);
988
989                 /*
990                  * Update ->io_size depending on options specified.
991                  * ->file_size_low being 0 means filesize option isn't set.
992                  * Non zero ->file_size_low equals ->file_size_high means
993                  * filesize option is set in a fixed size format.
994                  * Non zero ->file_size_low not equals ->file_size_high means
995                  * filesize option is set in a range format.
996                  */
997                 if (!o->file_size_low) {
998                         /*
999                          * no file size or range given, file size is equal to
1000                          * total size divided by number of files. If the size
1001                          * doesn't divide nicely with the min blocksize,
1002                          * make the first files bigger.
1003                          */
1004                         f->io_size = fs;
1005                         if (nr_fs_extra) {
1006                                 nr_fs_extra--;
1007                                 f->io_size += bs;
1008                         }
1009
1010                         /*
1011                          * We normally don't come here for regular files, but
1012                          * if the result is 0 for a regular file, set it to the
1013                          * real file size. This could be size of the existing
1014                          * one if it already exists, but otherwise will be set
1015                          * to 0. A new file won't be created because
1016                          * ->io_size + ->file_offset equals ->real_file_size.
1017                          */
1018                         if (!f->io_size) {
1019                                 if (f->file_offset > f->real_file_size)
1020                                         goto err_offset;
1021                                 f->io_size = f->real_file_size - f->file_offset;
1022                                 if (!f->io_size)
1023                                         log_info("fio: file %s may be ignored\n",
1024                                                 f->file_name);
1025                         }
1026                 } else if (f->real_file_size < o->file_size_low ||
1027                            f->real_file_size > o->file_size_high) {
1028                         if (f->file_offset > o->file_size_low)
1029                                 goto err_offset;
1030                         /*
1031                          * file size given. if it's fixed, use that. if it's a
1032                          * range, generate a random size in-between.
1033                          */
1034                         if (o->file_size_low == o->file_size_high)
1035                                 f->io_size = o->file_size_low - f->file_offset;
1036                         else {
1037                                 f->io_size = get_rand_file_size(td)
1038                                                 - f->file_offset;
1039                         }
1040                 } else
1041                         f->io_size = f->real_file_size - f->file_offset;
1042
1043                 if (f->io_size == -1ULL)
1044                         total_size = -1ULL;
1045                 else {
1046                         if (o->size_percent) {
1047                                 uint64_t file_size;
1048
1049                                 file_size = f->io_size + f->file_offset;
1050                                 f->io_size = (file_size *
1051                                               o->size_percent) / 100;
1052                                 if (f->io_size > (file_size - f->file_offset))
1053                                         f->io_size = file_size - f->file_offset;
1054
1055                                 f->io_size -= (f->io_size % td_min_bs(td));
1056                         }
1057                         total_size += f->io_size;
1058                 }
1059
1060                 if (f->filetype == FIO_TYPE_FILE &&
1061                     (f->io_size + f->file_offset) > f->real_file_size &&
1062                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1063                         if (!o->create_on_open) {
1064                                 need_extend++;
1065                                 extend_size += (f->io_size + f->file_offset);
1066                                 fio_file_set_extend(f);
1067                         } else
1068                                 f->real_file_size = f->io_size + f->file_offset;
1069                 }
1070         }
1071
1072         if (td->o.block_error_hist) {
1073                 int len;
1074
1075                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1076                 f = td->files[0];
1077                 len = f->io_size / td->o.bs[DDIR_TRIM];
1078                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1079                         log_err("fio: cannot calculate block histogram with "
1080                                 "%d trim blocks, maximum %d\n",
1081                                 len, MAX_NR_BLOCK_INFOS);
1082                         td_verror(td, EINVAL, "block_error_hist");
1083                         goto err_out;
1084                 }
1085
1086                 td->ts.nr_block_infos = len;
1087                 for (i = 0; i < len; i++)
1088                         td->ts.block_infos[i] =
1089                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1090         } else
1091                 td->ts.nr_block_infos = 0;
1092
1093         if (!o->size || (total_size && o->size > total_size))
1094                 o->size = total_size;
1095
1096         if (o->size < td_min_bs(td)) {
1097                 log_err("fio: blocksize too large for data set\n");
1098                 goto err_out;
1099         }
1100
1101         /*
1102          * See if we need to extend some files, typically needed when our
1103          * target regular files don't exist yet, but our jobs require them
1104          * initially due to read I/Os.
1105          */
1106         if (need_extend) {
1107                 temp_stall_ts = 1;
1108                 if (output_format & FIO_OUTPUT_NORMAL) {
1109                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1110                                  o->name,
1111                                  need_extend > 1 ? "s" : "",
1112                                  need_extend,
1113                                  need_extend > 1 ? "s" : "",
1114                                  need_extend > 1 ? "total " : "",
1115                                  extend_size >> 20);
1116                 }
1117
1118                 for_each_file(td, f, i) {
1119                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1120
1121                         if (!fio_file_extend(f))
1122                                 continue;
1123
1124                         assert(f->filetype == FIO_TYPE_FILE);
1125                         fio_file_clear_extend(f);
1126                         if (!o->fill_device) {
1127                                 old_len = f->real_file_size;
1128                                 extend_len = f->io_size + f->file_offset -
1129                                                 old_len;
1130                         }
1131                         f->real_file_size = (f->io_size + f->file_offset);
1132                         err = extend_file(td, f);
1133                         if (err)
1134                                 break;
1135
1136                         err = __file_invalidate_cache(td, f, old_len,
1137                                                                 extend_len);
1138
1139                         /*
1140                          * Shut up static checker
1141                          */
1142                         if (f->fd != -1)
1143                                 close(f->fd);
1144
1145                         f->fd = -1;
1146                         if (err)
1147                                 break;
1148                 }
1149                 temp_stall_ts = 0;
1150         }
1151
1152         if (err)
1153                 goto err_out;
1154
1155         if (!o->zone_size)
1156                 o->zone_size = o->size;
1157
1158         /*
1159          * iolog already set the total io size, if we read back
1160          * stored entries.
1161          */
1162         if (!o->read_iolog_file) {
1163                 if (o->io_size)
1164                         td->total_io_size = o->io_size * o->loops;
1165                 else
1166                         td->total_io_size = o->size * o->loops;
1167         }
1168
1169 done:
1170         if (o->create_only)
1171                 td->done = 1;
1172
1173         td_restore_runstate(td, old_state);
1174         return 0;
1175 err_offset:
1176         log_err("%s: you need to specify valid offset=\n", o->name);
1177 err_out:
1178         td_restore_runstate(td, old_state);
1179         return 1;
1180 }
1181
1182 int pre_read_files(struct thread_data *td)
1183 {
1184         struct fio_file *f;
1185         unsigned int i;
1186
1187         dprint(FD_FILE, "pre_read files\n");
1188
1189         for_each_file(td, f, i) {
1190                 if (pre_read_file(td, f))
1191                         return -1;
1192         }
1193
1194         return 0;
1195 }
1196
1197 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1198 {
1199         unsigned int range_size, seed;
1200         unsigned long nranges;
1201         uint64_t fsize;
1202
1203         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1204         fsize = min(f->real_file_size, f->io_size);
1205
1206         nranges = (fsize + range_size - 1) / range_size;
1207
1208         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1209         if (!td->o.rand_repeatable)
1210                 seed = td->rand_seeds[4];
1211
1212         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1213                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1214         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1215                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1216         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1217                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1218
1219         return 1;
1220 }
1221
1222 static int init_rand_distribution(struct thread_data *td)
1223 {
1224         struct fio_file *f;
1225         unsigned int i;
1226         int state;
1227
1228         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1229                 return 0;
1230
1231         state = td_bump_runstate(td, TD_SETTING_UP);
1232
1233         for_each_file(td, f, i)
1234                 __init_rand_distribution(td, f);
1235
1236         td_restore_runstate(td, state);
1237
1238         return 1;
1239 }
1240
1241 /*
1242  * Check if the number of blocks exceeds the randomness capability of
1243  * the selected generator. Tausworthe is 32-bit, the others are fullly
1244  * 64-bit capable.
1245  */
1246 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1247                                  uint64_t blocks)
1248 {
1249         if (blocks <= FRAND32_MAX)
1250                 return 0;
1251         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1252                 return 0;
1253
1254         /*
1255          * If the user hasn't specified a random generator, switch
1256          * to tausworthe64 with informational warning. If the user did
1257          * specify one, just warn.
1258          */
1259         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1260                         f->file_name);
1261
1262         if (!fio_option_is_set(&td->o, random_generator)) {
1263                 log_info("fio: Switching to tausworthe64. Use the "
1264                          "random_generator= option to get rid of this "
1265                          "warning.\n");
1266                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1267                 return 0;
1268         }
1269
1270         /*
1271          * Just make this information to avoid breaking scripts.
1272          */
1273         log_info("fio: Use the random_generator= option to switch to lfsr or "
1274                          "tausworthe64.\n");
1275         return 0;
1276 }
1277
1278 int init_random_map(struct thread_data *td)
1279 {
1280         unsigned long long blocks;
1281         struct fio_file *f;
1282         unsigned int i;
1283
1284         if (init_rand_distribution(td))
1285                 return 0;
1286         if (!td_random(td))
1287                 return 0;
1288
1289         for_each_file(td, f, i) {
1290                 uint64_t fsize = min(f->real_file_size, f->io_size);
1291
1292                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1293
1294                 if (check_rand_gen_limits(td, f, blocks))
1295                         return 1;
1296
1297                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1298                         unsigned long seed;
1299
1300                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1301
1302                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1303                                 fio_file_set_lfsr(f);
1304                                 continue;
1305                         }
1306                 } else if (!td->o.norandommap) {
1307                         f->io_axmap = axmap_new(blocks);
1308                         if (f->io_axmap) {
1309                                 fio_file_set_axmap(f);
1310                                 continue;
1311                         }
1312                 } else if (td->o.norandommap)
1313                         continue;
1314
1315                 if (!td->o.softrandommap) {
1316                         log_err("fio: failed allocating random map. If running"
1317                                 " a large number of jobs, try the 'norandommap'"
1318                                 " option or set 'softrandommap'. Or give"
1319                                 " a larger --alloc-size to fio.\n");
1320                         return 1;
1321                 }
1322
1323                 log_info("fio: file %s failed allocating random map. Running "
1324                          "job without.\n", f->file_name);
1325         }
1326
1327         return 0;
1328 }
1329
1330 void close_files(struct thread_data *td)
1331 {
1332         struct fio_file *f;
1333         unsigned int i;
1334
1335         for_each_file(td, f, i) {
1336                 if (fio_file_open(f))
1337                         td_io_close_file(td, f);
1338         }
1339 }
1340
1341 void close_and_free_files(struct thread_data *td)
1342 {
1343         struct fio_file *f;
1344         unsigned int i;
1345         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1346
1347         dprint(FD_FILE, "close files\n");
1348
1349         for_each_file(td, f, i) {
1350                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1351                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1352                         td_io_unlink_file(td, f);
1353                 }
1354
1355                 if (fio_file_open(f))
1356                         td_io_close_file(td, f);
1357
1358                 remove_file_hash(f);
1359
1360                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1361                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1362                         td_io_unlink_file(td, f);
1363                 }
1364
1365                 if (use_free)
1366                         free(f->file_name);
1367                 else
1368                         sfree(f->file_name);
1369                 f->file_name = NULL;
1370                 if (fio_file_axmap(f)) {
1371                         axmap_free(f->io_axmap);
1372                         f->io_axmap = NULL;
1373                 }
1374                 if (use_free)
1375                         free(f);
1376                 else
1377                         sfree(f);
1378         }
1379
1380         td->o.filename = NULL;
1381         free(td->files);
1382         free(td->file_locks);
1383         td->files_index = 0;
1384         td->files = NULL;
1385         td->file_locks = NULL;
1386         td->o.file_lock_mode = FILE_LOCK_NONE;
1387         td->o.nr_files = 0;
1388 }
1389
1390 static void get_file_type(struct fio_file *f)
1391 {
1392         struct stat sb;
1393
1394         if (!strcmp(f->file_name, "-"))
1395                 f->filetype = FIO_TYPE_PIPE;
1396         else
1397                 f->filetype = FIO_TYPE_FILE;
1398
1399 #ifdef WIN32
1400         /* \\.\ is the device namespace in Windows, where every file is
1401          * a block device */
1402         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1403                 f->filetype = FIO_TYPE_BLOCK;
1404 #endif
1405
1406         if (!stat(f->file_name, &sb)) {
1407                 if (S_ISBLK(sb.st_mode))
1408                         f->filetype = FIO_TYPE_BLOCK;
1409                 else if (S_ISCHR(sb.st_mode))
1410                         f->filetype = FIO_TYPE_CHAR;
1411                 else if (S_ISFIFO(sb.st_mode))
1412                         f->filetype = FIO_TYPE_PIPE;
1413         }
1414 }
1415
1416 static bool __is_already_allocated(const char *fname, bool set)
1417 {
1418         struct flist_head *entry;
1419         bool ret;
1420
1421         ret = file_bloom_exists(fname, set);
1422         if (!ret)
1423                 return ret;
1424
1425         flist_for_each(entry, &filename_list) {
1426                 struct file_name *fn;
1427
1428                 fn = flist_entry(entry, struct file_name, list);
1429
1430                 if (!strcmp(fn->filename, fname))
1431                         return true;
1432         }
1433
1434         return false;
1435 }
1436
1437 static bool is_already_allocated(const char *fname)
1438 {
1439         bool ret;
1440
1441         fio_file_hash_lock();
1442         ret = __is_already_allocated(fname, false);
1443         fio_file_hash_unlock();
1444
1445         return ret;
1446 }
1447
1448 static void set_already_allocated(const char *fname)
1449 {
1450         struct file_name *fn;
1451
1452         fn = malloc(sizeof(struct file_name));
1453         fn->filename = strdup(fname);
1454
1455         fio_file_hash_lock();
1456         if (!__is_already_allocated(fname, true)) {
1457                 flist_add_tail(&fn->list, &filename_list);
1458                 fn = NULL;
1459         }
1460         fio_file_hash_unlock();
1461
1462         if (fn) {
1463                 free(fn->filename);
1464                 free(fn);
1465         }
1466 }
1467
1468 static void free_already_allocated(void)
1469 {
1470         struct flist_head *entry, *tmp;
1471         struct file_name *fn;
1472
1473         if (flist_empty(&filename_list))
1474                 return;
1475
1476         fio_file_hash_lock();
1477         flist_for_each_safe(entry, tmp, &filename_list) {
1478                 fn = flist_entry(entry, struct file_name, list);
1479                 free(fn->filename);
1480                 flist_del(&fn->list);
1481                 free(fn);
1482         }
1483
1484         fio_file_hash_unlock();
1485 }
1486
1487 static struct fio_file *alloc_new_file(struct thread_data *td)
1488 {
1489         struct fio_file *f;
1490
1491         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1492                 f = calloc(1, sizeof(*f));
1493         else
1494                 f = smalloc(sizeof(*f));
1495         if (!f) {
1496                 assert(0);
1497                 return NULL;
1498         }
1499
1500         f->fd = -1;
1501         f->shadow_fd = -1;
1502         fio_file_reset(td, f);
1503         return f;
1504 }
1505
1506 bool exists_and_not_regfile(const char *filename)
1507 {
1508         struct stat sb;
1509
1510         if (lstat(filename, &sb) == -1)
1511                 return false;
1512
1513 #ifndef WIN32 /* NOT Windows */
1514         if (S_ISREG(sb.st_mode))
1515                 return false;
1516 #else
1517         /* \\.\ is the device namespace in Windows, where every file
1518          * is a device node */
1519         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1520                 return false;
1521 #endif
1522
1523         return true;
1524 }
1525
1526 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1527 {
1528         int cur_files = td->files_index;
1529         char file_name[PATH_MAX];
1530         struct fio_file *f;
1531         int len = 0;
1532
1533         dprint(FD_FILE, "add file %s\n", fname);
1534
1535         if (td->o.directory)
1536                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1537                                         td->o.unique_filename);
1538
1539         sprintf(file_name + len, "%s", fname);
1540
1541         /* clean cloned siblings using existing files */
1542         if (numjob && is_already_allocated(file_name) &&
1543             !exists_and_not_regfile(fname))
1544                 return 0;
1545
1546         f = alloc_new_file(td);
1547
1548         if (td->files_size <= td->files_index) {
1549                 unsigned int new_size = td->o.nr_files + 1;
1550
1551                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1552
1553                 td->files = realloc(td->files, new_size * sizeof(f));
1554                 if (td->files == NULL) {
1555                         log_err("fio: realloc OOM\n");
1556                         assert(0);
1557                 }
1558                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1559                         td->file_locks = realloc(td->file_locks, new_size);
1560                         if (!td->file_locks) {
1561                                 log_err("fio: realloc OOM\n");
1562                                 assert(0);
1563                         }
1564                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1565                 }
1566                 td->files_size = new_size;
1567         }
1568         td->files[cur_files] = f;
1569         f->fileno = cur_files;
1570
1571         /*
1572          * init function, io engine may not be loaded yet
1573          */
1574         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1575                 f->real_file_size = -1ULL;
1576
1577         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1578                 f->file_name = strdup(file_name);
1579         else
1580                 f->file_name = smalloc_strdup(file_name);
1581         if (!f->file_name)
1582                 assert(0);
1583
1584         get_file_type(f);
1585
1586         switch (td->o.file_lock_mode) {
1587         case FILE_LOCK_NONE:
1588                 break;
1589         case FILE_LOCK_READWRITE:
1590                 f->rwlock = fio_rwlock_init();
1591                 break;
1592         case FILE_LOCK_EXCLUSIVE:
1593                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1594                 break;
1595         default:
1596                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1597                 assert(0);
1598         }
1599
1600         td->files_index++;
1601         if (f->filetype == FIO_TYPE_FILE)
1602                 td->nr_normal_files++;
1603
1604         if (td->o.numjobs > 1)
1605                 set_already_allocated(file_name);
1606
1607         if (inc)
1608                 td->o.nr_files++;
1609
1610         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1611                                                         cur_files);
1612
1613         return cur_files;
1614 }
1615
1616 int add_file_exclusive(struct thread_data *td, const char *fname)
1617 {
1618         struct fio_file *f;
1619         unsigned int i;
1620
1621         for_each_file(td, f, i) {
1622                 if (!strcmp(f->file_name, fname))
1623                         return i;
1624         }
1625
1626         return add_file(td, fname, 0, 1);
1627 }
1628
1629 void get_file(struct fio_file *f)
1630 {
1631         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1632         assert(fio_file_open(f));
1633         f->references++;
1634 }
1635
1636 int put_file(struct thread_data *td, struct fio_file *f)
1637 {
1638         int f_ret = 0, ret = 0;
1639
1640         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1641
1642         if (!fio_file_open(f)) {
1643                 assert(f->fd == -1);
1644                 return 0;
1645         }
1646
1647         assert(f->references);
1648         if (--f->references)
1649                 return 0;
1650
1651         if (should_fsync(td) && td->o.fsync_on_close) {
1652                 f_ret = fsync(f->fd);
1653                 if (f_ret < 0)
1654                         f_ret = errno;
1655         }
1656
1657         if (td->io_ops->close_file)
1658                 ret = td->io_ops->close_file(td, f);
1659
1660         if (!ret)
1661                 ret = f_ret;
1662
1663         td->nr_open_files--;
1664         fio_file_clear_open(f);
1665         assert(f->fd == -1);
1666         return ret;
1667 }
1668
1669 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1670 {
1671         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1672                 return;
1673
1674         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1675                 if (ddir == DDIR_READ)
1676                         fio_rwlock_read(f->rwlock);
1677                 else
1678                         fio_rwlock_write(f->rwlock);
1679         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1680                 fio_mutex_down(f->lock);
1681
1682         td->file_locks[f->fileno] = td->o.file_lock_mode;
1683 }
1684
1685 void unlock_file(struct thread_data *td, struct fio_file *f)
1686 {
1687         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1688                 return;
1689
1690         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1691                 fio_rwlock_unlock(f->rwlock);
1692         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1693                 fio_mutex_up(f->lock);
1694
1695         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1696 }
1697
1698 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1699 {
1700         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1701                 return;
1702         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1703                 unlock_file(td, f);
1704 }
1705
1706 static int recurse_dir(struct thread_data *td, const char *dirname)
1707 {
1708         struct dirent *dir;
1709         int ret = 0;
1710         DIR *D;
1711
1712         D = opendir(dirname);
1713         if (!D) {
1714                 char buf[FIO_VERROR_SIZE];
1715
1716                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1717                 td_verror(td, errno, buf);
1718                 return 1;
1719         }
1720
1721         while ((dir = readdir(D)) != NULL) {
1722                 char full_path[PATH_MAX];
1723                 struct stat sb;
1724
1725                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1726                         continue;
1727
1728                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1729
1730                 if (lstat(full_path, &sb) == -1) {
1731                         if (errno != ENOENT) {
1732                                 td_verror(td, errno, "stat");
1733                                 ret = 1;
1734                                 break;
1735                         }
1736                 }
1737
1738                 if (S_ISREG(sb.st_mode)) {
1739                         add_file(td, full_path, 0, 1);
1740                         continue;
1741                 }
1742                 if (!S_ISDIR(sb.st_mode))
1743                         continue;
1744
1745                 ret = recurse_dir(td, full_path);
1746                 if (ret)
1747                         break;
1748         }
1749
1750         closedir(D);
1751         return ret;
1752 }
1753
1754 int add_dir_files(struct thread_data *td, const char *path)
1755 {
1756         int ret = recurse_dir(td, path);
1757
1758         if (!ret)
1759                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1760
1761         return ret;
1762 }
1763
1764 void dup_files(struct thread_data *td, struct thread_data *org)
1765 {
1766         struct fio_file *f;
1767         unsigned int i;
1768
1769         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1770
1771         if (!org->files)
1772                 return;
1773
1774         td->files = malloc(org->files_index * sizeof(f));
1775
1776         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1777                 td->file_locks = malloc(org->files_index);
1778
1779         for_each_file(org, f, i) {
1780                 struct fio_file *__f;
1781
1782                 __f = alloc_new_file(td);
1783
1784                 if (f->file_name) {
1785                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1786                                 __f->file_name = strdup(f->file_name);
1787                         else
1788                                 __f->file_name = smalloc_strdup(f->file_name);
1789                         if (!__f->file_name)
1790                                 assert(0);
1791
1792                         __f->filetype = f->filetype;
1793                 }
1794
1795                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1796                         __f->lock = f->lock;
1797                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1798                         __f->rwlock = f->rwlock;
1799
1800                 td->files[i] = __f;
1801         }
1802 }
1803
1804 /*
1805  * Returns the index that matches the filename, or -1 if not there
1806  */
1807 int get_fileno(struct thread_data *td, const char *fname)
1808 {
1809         struct fio_file *f;
1810         unsigned int i;
1811
1812         for_each_file(td, f, i)
1813                 if (!strcmp(f->file_name, fname))
1814                         return i;
1815
1816         return -1;
1817 }
1818
1819 /*
1820  * For log usage, where we add/open/close files automatically
1821  */
1822 void free_release_files(struct thread_data *td)
1823 {
1824         close_files(td);
1825         td->o.nr_files = 0;
1826         td->o.open_files = 0;
1827         td->files_index = 0;
1828         td->nr_normal_files = 0;
1829 }
1830
1831 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1832 {
1833         int i;
1834
1835         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1836                 f->last_pos[i] = f->file_offset;
1837                 f->last_start[i] = -1ULL;
1838         }
1839
1840         if (fio_file_axmap(f))
1841                 axmap_reset(f->io_axmap);
1842         else if (fio_file_lfsr(f))
1843                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1844 }
1845
1846 bool fio_files_done(struct thread_data *td)
1847 {
1848         struct fio_file *f;
1849         unsigned int i;
1850
1851         for_each_file(td, f, i)
1852                 if (!fio_file_done(f))
1853                         return false;
1854
1855         return true;
1856 }
1857
1858 /* free memory used in initialization phase only */
1859 void filesetup_mem_free(void)
1860 {
1861         free_already_allocated();
1862 }
1863
1864 /*
1865  * This function is for platforms which support direct I/O but not O_DIRECT.
1866  */
1867 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1868 {
1869 #ifdef FIO_OS_DIRECTIO
1870         int ret = fio_set_odirect(f);
1871
1872         if (ret) {
1873                 td_verror(td, ret, "fio_set_directio");
1874 #if defined(__sun__)
1875                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1876                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1877                 } else {
1878                         log_err("fio: the file system does not seem to support direct IO\n");
1879                 }
1880 #else
1881                 log_err("fio: the file system does not seem to support direct IO\n");
1882 #endif
1883                 return -1;
1884         }
1885
1886         return 0;
1887 #else
1888         log_err("fio: direct IO is not supported on this host operating system\n");
1889         return -1;
1890 #endif
1891 }