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