Merge branch 'master' of ssh://git.kernel.dk/data/git/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 static inline void clear_error(struct thread_data *td)
28 {
29         td->error = 0;
30         td->verror[0] = '\0';
31 }
32
33 /*
34  * Leaves f->fd open on success, caller must close
35  */
36 static int extend_file(struct thread_data *td, struct fio_file *f)
37 {
38         int r, new_layout = 0, unlink_file = 0, flags;
39         unsigned long long left;
40         unsigned int bs;
41         char *b = NULL;
42
43         if (read_only) {
44                 log_err("fio: refusing extend of file due to read-only\n");
45                 return 0;
46         }
47
48         /*
49          * check if we need to lay the file out complete again. fio
50          * does that for operations involving reads, or for writes
51          * where overwrite is set
52          */
53         if (td_read(td) ||
54            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
55             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
56                 new_layout = 1;
57         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
58                 unlink_file = 1;
59
60         if (unlink_file || new_layout) {
61                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
62                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
63                         td_verror(td, errno, "unlink");
64                         return 1;
65                 }
66         }
67
68         flags = O_WRONLY | O_CREAT;
69         if (new_layout)
70                 flags |= O_TRUNC;
71
72         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
73         f->fd = open(f->file_name, flags, 0644);
74         if (f->fd < 0) {
75                 td_verror(td, errno, "open");
76                 return 1;
77         }
78
79 #ifdef CONFIG_POSIX_FALLOCATE
80         if (!td->o.fill_device) {
81                 switch (td->o.fallocate_mode) {
82                 case FIO_FALLOCATE_NONE:
83                         break;
84                 case FIO_FALLOCATE_POSIX:
85                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
86                                  f->file_name,
87                                  (unsigned long long) f->real_file_size);
88
89                         r = posix_fallocate(f->fd, 0, f->real_file_size);
90                         if (r > 0) {
91                                 log_err("fio: posix_fallocate fails: %s\n",
92                                                 strerror(r));
93                         }
94                         break;
95 #ifdef CONFIG_LINUX_FALLOCATE
96                 case FIO_FALLOCATE_KEEP_SIZE:
97                         dprint(FD_FILE,
98                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
99                                 "file %s size %llu\n", f->file_name,
100                                 (unsigned long long) f->real_file_size);
101
102                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
103                                         f->real_file_size);
104                         if (r != 0)
105                                 td_verror(td, errno, "fallocate");
106
107                         break;
108 #endif /* CONFIG_LINUX_FALLOCATE */
109                 default:
110                         log_err("fio: unknown fallocate mode: %d\n",
111                                 td->o.fallocate_mode);
112                         assert(0);
113                 }
114         }
115 #endif /* CONFIG_POSIX_FALLOCATE */
116
117         if (!new_layout)
118                 goto done;
119
120         /*
121          * The size will be -1ULL when fill_device is used, so don't truncate
122          * or fallocate this file, just write it
123          */
124         if (!td->o.fill_device) {
125                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
126                                         (unsigned long long) f->real_file_size);
127                 if (ftruncate(f->fd, f->real_file_size) == -1) {
128                         if (errno != EFBIG) {
129                                 td_verror(td, errno, "ftruncate");
130                                 goto err;
131                         }
132                 }
133         }
134
135         b = malloc(td->o.max_bs[DDIR_WRITE]);
136
137         left = f->real_file_size;
138         while (left && !td->terminate) {
139                 bs = td->o.max_bs[DDIR_WRITE];
140                 if (bs > left)
141                         bs = left;
142
143                 fill_io_buffer(td, b, bs, bs);
144
145                 r = write(f->fd, b, bs);
146
147                 if (r > 0) {
148                         left -= r;
149                         continue;
150                 } else {
151                         if (r < 0) {
152                                 int __e = errno;
153
154                                 if (__e == ENOSPC) {
155                                         if (td->o.fill_device)
156                                                 break;
157                                         log_info("fio: ENOSPC on laying out "
158                                                  "file, stopping\n");
159                                         break;
160                                 }
161                                 td_verror(td, errno, "write");
162                         } else
163                                 td_verror(td, EIO, "write");
164
165                         break;
166                 }
167         }
168
169         if (td->terminate) {
170                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
171                 unlink(f->file_name);
172         } else if (td->o.create_fsync) {
173                 if (fsync(f->fd) < 0) {
174                         td_verror(td, errno, "fsync");
175                         goto err;
176                 }
177         }
178         if (td->o.fill_device && !td_write(td)) {
179                 fio_file_clear_size_known(f);
180                 if (td_io_get_file_size(td, f))
181                         goto err;
182                 if (f->io_size > f->real_file_size)
183                         f->io_size = f->real_file_size;
184         }
185
186         free(b);
187 done:
188         return 0;
189 err:
190         close(f->fd);
191         f->fd = -1;
192         if (b)
193                 free(b);
194         return 1;
195 }
196
197 static int pre_read_file(struct thread_data *td, struct fio_file *f)
198 {
199         int ret = 0, r, did_open = 0, old_runstate;
200         unsigned long long left;
201         unsigned int bs;
202         char *b;
203
204         if (td->io_ops->flags & FIO_PIPEIO)
205                 return 0;
206
207         if (!fio_file_open(f)) {
208                 if (td->io_ops->open_file(td, f)) {
209                         log_err("fio: cannot pre-read, failed to open file\n");
210                         return 1;
211                 }
212                 did_open = 1;
213         }
214
215         old_runstate = td_bump_runstate(td, TD_PRE_READING);
216
217         bs = td->o.max_bs[DDIR_READ];
218         b = malloc(bs);
219         memset(b, 0, bs);
220
221         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
222                 td_verror(td, errno, "lseek");
223                 log_err("fio: failed to lseek pre-read file\n");
224                 ret = 1;
225                 goto error;
226         }
227
228         left = f->io_size;
229
230         while (left && !td->terminate) {
231                 if (bs > left)
232                         bs = left;
233
234                 r = read(f->fd, b, bs);
235
236                 if (r == (int) bs) {
237                         left -= bs;
238                         continue;
239                 } else {
240                         td_verror(td, EIO, "pre_read");
241                         break;
242                 }
243         }
244
245 error:
246         td_restore_runstate(td, old_runstate);
247
248         if (did_open)
249                 td->io_ops->close_file(td, f);
250
251         free(b);
252         return ret;
253 }
254
255 static unsigned long long get_rand_file_size(struct thread_data *td)
256 {
257         unsigned long long ret, sized;
258         unsigned long r;
259
260         if (td->o.use_os_rand) {
261                 r = os_random_long(&td->file_size_state);
262                 sized = td->o.file_size_high - td->o.file_size_low;
263                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
264         } else {
265                 r = __rand(&td->__file_size_state);
266                 sized = td->o.file_size_high - td->o.file_size_low;
267                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
268         }
269
270         ret += td->o.file_size_low;
271         ret -= (ret % td->o.rw_min_bs);
272         return ret;
273 }
274
275 static int file_size(struct thread_data *td, struct fio_file *f)
276 {
277         struct stat st;
278
279         if (stat(f->file_name, &st) == -1) {
280                 td_verror(td, errno, "fstat");
281                 return 1;
282         }
283
284         f->real_file_size = st.st_size;
285         return 0;
286 }
287
288 static int bdev_size(struct thread_data *td, struct fio_file *f)
289 {
290         unsigned long long bytes = 0;
291         int r;
292
293         if (td->io_ops->open_file(td, f)) {
294                 log_err("fio: failed opening blockdev %s for size check\n",
295                         f->file_name);
296                 return 1;
297         }
298
299         r = blockdev_size(f, &bytes);
300         if (r) {
301                 td_verror(td, r, "blockdev_size");
302                 goto err;
303         }
304
305         if (!bytes) {
306                 log_err("%s: zero sized block device?\n", f->file_name);
307                 goto err;
308         }
309
310         f->real_file_size = bytes;
311         td->io_ops->close_file(td, f);
312         return 0;
313 err:
314         td->io_ops->close_file(td, f);
315         return 1;
316 }
317
318 static int char_size(struct thread_data *td, struct fio_file *f)
319 {
320 #ifdef FIO_HAVE_CHARDEV_SIZE
321         unsigned long long bytes = 0;
322         int r;
323
324         if (td->io_ops->open_file(td, f)) {
325                 log_err("fio: failed opening blockdev %s for size check\n",
326                         f->file_name);
327                 return 1;
328         }
329
330         r = chardev_size(f, &bytes);
331         if (r) {
332                 td_verror(td, r, "chardev_size");
333                 goto err;
334         }
335
336         if (!bytes) {
337                 log_err("%s: zero sized char device?\n", f->file_name);
338                 goto err;
339         }
340
341         f->real_file_size = bytes;
342         td->io_ops->close_file(td, f);
343         return 0;
344 err:
345         td->io_ops->close_file(td, f);
346         return 1;
347 #else
348         f->real_file_size = -1ULL;
349         return 0;
350 #endif
351 }
352
353 static int get_file_size(struct thread_data *td, struct fio_file *f)
354 {
355         int ret = 0;
356
357         if (fio_file_size_known(f))
358                 return 0;
359
360         if (f->filetype == FIO_TYPE_FILE)
361                 ret = file_size(td, f);
362         else if (f->filetype == FIO_TYPE_BD)
363                 ret = bdev_size(td, f);
364         else if (f->filetype == FIO_TYPE_CHAR)
365                 ret = char_size(td, f);
366         else
367                 f->real_file_size = -1;
368
369         if (ret)
370                 return ret;
371
372         if (f->file_offset > f->real_file_size) {
373                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
374                                         (unsigned long long) f->file_offset,
375                                         (unsigned long long) f->real_file_size);
376                 return 1;
377         }
378
379         fio_file_set_size_known(f);
380         return 0;
381 }
382
383 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
384                                    unsigned long long off,
385                                    unsigned long long len)
386 {
387         int ret = 0;
388
389         if (len == -1ULL)
390                 len = f->io_size;
391         if (off == -1ULL)
392                 off = f->file_offset;
393
394         if (len == -1ULL || off == -1ULL)
395                 return 0;
396
397         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
398                                                                 len);
399
400         if (f->mmap_ptr) {
401                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
402 #ifdef FIO_MADV_FREE
403                 if (f->filetype == FIO_TYPE_BD)
404                         (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
405 #endif
406         } else if (f->filetype == FIO_TYPE_FILE) {
407                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
408         } else if (f->filetype == FIO_TYPE_BD) {
409                 ret = blockdev_invalidate_cache(f);
410                 if (ret < 0 && errno == EACCES && geteuid()) {
411                         if (!root_warn) {
412                                 log_err("fio: only root may flush block "
413                                         "devices. Cache flush bypassed!\n");
414                                 root_warn = 1;
415                         }
416                         ret = 0;
417                 }
418         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
419                 ret = 0;
420
421         /*
422          * Cache flushing isn't a fatal condition, and we know it will
423          * happen on some platforms where we don't have the proper
424          * function to flush eg block device caches. So just warn and
425          * continue on our way.
426          */
427         if (ret) {
428                 log_info("fio: cache invalidation of %s failed: %s\n", f->file_name, strerror(errno));
429                 ret = 0;
430         }
431
432         return 0;
433
434 }
435
436 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
437 {
438         if (!fio_file_open(f))
439                 return 0;
440
441         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
442 }
443
444 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
445 {
446         int ret = 0;
447
448         dprint(FD_FILE, "fd close %s\n", f->file_name);
449
450         remove_file_hash(f);
451
452         if (close(f->fd) < 0)
453                 ret = errno;
454
455         f->fd = -1;
456
457         if (f->shadow_fd != -1) {
458                 close(f->shadow_fd);
459                 f->shadow_fd = -1;
460         }
461
462         f->engine_data = 0;
463         return ret;
464 }
465
466 int file_lookup_open(struct fio_file *f, int flags)
467 {
468         struct fio_file *__f;
469         int from_hash;
470
471         __f = lookup_file_hash(f->file_name);
472         if (__f) {
473                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
474                 /*
475                  * racy, need the __f->lock locked
476                  */
477                 f->lock = __f->lock;
478                 from_hash = 1;
479         } else {
480                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
481                 from_hash = 0;
482         }
483
484         f->fd = open(f->file_name, flags, 0600);
485         return from_hash;
486 }
487
488 static int file_close_shadow_fds(struct thread_data *td)
489 {
490         struct fio_file *f;
491         int num_closed = 0;
492         unsigned int i;
493
494         for_each_file(td, f, i) {
495                 if (f->shadow_fd == -1)
496                         continue;
497
498                 close(f->shadow_fd);
499                 f->shadow_fd = -1;
500                 num_closed++;
501         }
502
503         return num_closed;
504 }
505
506 int generic_open_file(struct thread_data *td, struct fio_file *f)
507 {
508         int is_std = 0;
509         int flags = 0;
510         int from_hash = 0;
511
512         dprint(FD_FILE, "fd open %s\n", f->file_name);
513
514         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
515                 log_err("fio: trim only applies to block device\n");
516                 return 1;
517         }
518
519         if (!strcmp(f->file_name, "-")) {
520                 if (td_rw(td)) {
521                         log_err("fio: can't read/write to stdin/out\n");
522                         return 1;
523                 }
524                 is_std = 1;
525
526                 /*
527                  * move output logging to stderr, if we are writing to stdout
528                  */
529                 if (td_write(td))
530                         f_out = stderr;
531         }
532
533         if (td_trim(td))
534                 goto skip_flags;
535         if (td->o.odirect)
536                 flags |= OS_O_DIRECT;
537         if (td->o.oatomic) {
538                 if (!FIO_O_ATOMIC) {
539                         td_verror(td, EINVAL, "OS does not support atomic IO");
540                         return 1;
541                 }
542                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
543         }
544         if (td->o.sync_io)
545                 flags |= O_SYNC;
546         if (td->o.create_on_open)
547                 flags |= O_CREAT;
548 skip_flags:
549         if (f->filetype != FIO_TYPE_FILE)
550                 flags |= FIO_O_NOATIME;
551
552 open_again:
553         if (td_write(td)) {
554                 if (!read_only)
555                         flags |= O_RDWR;
556
557                 if (f->filetype == FIO_TYPE_FILE)
558                         flags |= O_CREAT;
559
560                 if (is_std)
561                         f->fd = dup(STDOUT_FILENO);
562                 else
563                         from_hash = file_lookup_open(f, flags);
564         } else if (td_read(td)) {
565                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
566                         flags |= O_RDWR;
567                 else
568                         flags |= O_RDONLY;
569
570                 if (is_std)
571                         f->fd = dup(STDIN_FILENO);
572                 else
573                         from_hash = file_lookup_open(f, flags);
574         } else { //td trim
575                 flags |= O_RDWR;
576                 from_hash = file_lookup_open(f, flags);
577         }
578
579         if (f->fd == -1) {
580                 char buf[FIO_VERROR_SIZE];
581                 int __e = errno;
582
583                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
584                         flags &= ~FIO_O_NOATIME;
585                         goto open_again;
586                 }
587                 if (__e == EMFILE && file_close_shadow_fds(td))
588                         goto open_again;
589
590                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
591
592                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
593                         log_err("fio: looks like your file system does not " \
594                                 "support direct=1/buffered=0\n");
595                 }
596
597                 td_verror(td, __e, buf);
598         }
599
600         if (!from_hash && f->fd != -1) {
601                 if (add_file_hash(f)) {
602                         int fio_unused ret;
603
604                         /*
605                          * Stash away descriptor for later close. This is to
606                          * work-around a "feature" on Linux, where a close of
607                          * an fd that has been opened for write will trigger
608                          * udev to call blkid to check partitions, fs id, etc.
609                          * That pollutes the device cache, which can slow down
610                          * unbuffered accesses.
611                          */
612                         if (f->shadow_fd == -1)
613                                 f->shadow_fd = f->fd;
614                         else {
615                                 /*
616                                  * OK to ignore, we haven't done anything
617                                  * with it
618                                  */
619                                 ret = generic_close_file(td, f);
620                         }
621                         goto open_again;
622                 }
623         }
624
625         return 0;
626 }
627
628 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
629 {
630         return get_file_size(td, f);
631 }
632
633 /*
634  * open/close all files, so that ->real_file_size gets set
635  */
636 static int get_file_sizes(struct thread_data *td)
637 {
638         struct fio_file *f;
639         unsigned int i;
640         int err = 0;
641
642         for_each_file(td, f, i) {
643                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
644                                                                 f->file_name);
645
646                 if (td_io_get_file_size(td, f)) {
647                         if (td->error != ENOENT) {
648                                 log_err("%s\n", td->verror);
649                                 err = 1;
650                         }
651                         clear_error(td);
652                 }
653
654                 if (f->real_file_size == -1ULL && td->o.size)
655                         f->real_file_size = td->o.size / td->o.nr_files;
656         }
657
658         return err;
659 }
660
661 struct fio_mount {
662         struct flist_head list;
663         const char *base;
664         char __base[256];
665         unsigned int key;
666 };
667
668 /*
669  * Get free number of bytes for each file on each unique mount.
670  */
671 static unsigned long long get_fs_free_counts(struct thread_data *td)
672 {
673         struct flist_head *n, *tmp;
674         unsigned long long ret = 0;
675         struct fio_mount *fm;
676         FLIST_HEAD(list);
677         struct fio_file *f;
678         unsigned int i;
679
680         for_each_file(td, f, i) {
681                 struct stat sb;
682                 char buf[256];
683
684                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
685                         if (f->real_file_size != -1ULL)
686                                 ret += f->real_file_size;
687                         continue;
688                 } else if (f->filetype != FIO_TYPE_FILE)
689                         continue;
690
691                 buf[255] = '\0';
692                 strncpy(buf, f->file_name, 255);
693
694                 if (stat(buf, &sb) < 0) {
695                         if (errno != ENOENT)
696                                 break;
697                         strcpy(buf, ".");
698                         if (stat(buf, &sb) < 0)
699                                 break;
700                 }
701
702                 fm = NULL;
703                 flist_for_each(n, &list) {
704                         fm = flist_entry(n, struct fio_mount, list);
705                         if (fm->key == sb.st_dev)
706                                 break;
707
708                         fm = NULL;
709                 }
710
711                 if (fm)
712                         continue;
713
714                 fm = calloc(1, sizeof(*fm));
715                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
716                 fm->base = basename(fm->__base);
717                 fm->key = sb.st_dev;
718                 flist_add(&fm->list, &list);
719         }
720
721         flist_for_each_safe(n, tmp, &list) {
722                 unsigned long long sz;
723
724                 fm = flist_entry(n, struct fio_mount, list);
725                 flist_del(&fm->list);
726
727                 sz = get_fs_size(fm->base);
728                 if (sz && sz != -1ULL)
729                         ret += sz;
730
731                 free(fm);
732         }
733
734         return ret;
735 }
736
737 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
738 {
739         struct thread_options *o = &td->o;
740
741         if (o->file_append && f->filetype == FIO_TYPE_FILE)
742                 return f->real_file_size;
743
744         return td->o.start_offset +
745                 (td->thread_number - 1) * td->o.offset_increment;
746 }
747
748 /*
749  * Open the files and setup files sizes, creating files if necessary.
750  */
751 int setup_files(struct thread_data *td)
752 {
753         unsigned long long total_size, extend_size;
754         struct thread_options *o = &td->o;
755         struct fio_file *f;
756         unsigned int i, nr_fs_extra = 0;
757         int err = 0, need_extend;
758         int old_state;
759         const unsigned int bs = td_min_bs(td);
760         uint64_t fs = 0;
761
762         dprint(FD_FILE, "setup files\n");
763
764         old_state = td_bump_runstate(td, TD_SETTING_UP);
765
766         if (o->read_iolog_file)
767                 goto done;
768
769         /*
770          * if ioengine defines a setup() method, it's responsible for
771          * opening the files and setting f->real_file_size to indicate
772          * the valid range for that file.
773          */
774         if (td->io_ops->setup)
775                 err = td->io_ops->setup(td);
776         else
777                 err = get_file_sizes(td);
778
779         if (err)
780                 goto err_out;
781
782         /*
783          * check sizes. if the files/devices do not exist and the size
784          * isn't passed to fio, abort.
785          */
786         total_size = 0;
787         for_each_file(td, f, i) {
788                 if (f->real_file_size == -1ULL)
789                         total_size = -1ULL;
790                 else
791                         total_size += f->real_file_size;
792         }
793
794         if (o->fill_device)
795                 td->fill_device_size = get_fs_free_counts(td);
796
797         /*
798          * device/file sizes are zero and no size given, punt
799          */
800         if ((!total_size || total_size == -1ULL) && !o->size &&
801             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
802             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
803                 log_err("%s: you need to specify size=\n", o->name);
804                 td_verror(td, EINVAL, "total_file_size");
805                 goto err_out;
806         }
807
808         /*
809          * Calculate per-file size and potential extra size for the
810          * first files, if needed.
811          */
812         if (!o->file_size_low && o->nr_files) {
813                 uint64_t all_fs;
814
815                 fs = o->size / o->nr_files;
816                 all_fs = fs * o->nr_files;
817
818                 if (all_fs < o->size)
819                         nr_fs_extra = (o->size - all_fs) / bs;
820         }
821
822         /*
823          * now file sizes are known, so we can set ->io_size. if size= is
824          * not given, ->io_size is just equal to ->real_file_size. if size
825          * is given, ->io_size is size / nr_files.
826          */
827         extend_size = total_size = 0;
828         need_extend = 0;
829         for_each_file(td, f, i) {
830                 f->file_offset = get_start_offset(td, f);
831
832                 if (!o->file_size_low) {
833                         /*
834                          * no file size range given, file size is equal to
835                          * total size divided by number of files. If that is
836                          * zero, set it to the real file size. If the size
837                          * doesn't divide nicely with the min blocksize,
838                          * make the first files bigger.
839                          */
840                         f->io_size = fs;
841                         if (nr_fs_extra) {
842                                 nr_fs_extra--;
843                                 f->io_size += bs;
844                         }
845
846                         if (!f->io_size)
847                                 f->io_size = f->real_file_size - f->file_offset;
848                 } else if (f->real_file_size < o->file_size_low ||
849                            f->real_file_size > o->file_size_high) {
850                         if (f->file_offset > o->file_size_low)
851                                 goto err_offset;
852                         /*
853                          * file size given. if it's fixed, use that. if it's a
854                          * range, generate a random size in-between.
855                          */
856                         if (o->file_size_low == o->file_size_high)
857                                 f->io_size = o->file_size_low - f->file_offset;
858                         else {
859                                 f->io_size = get_rand_file_size(td)
860                                                 - f->file_offset;
861                         }
862                 } else
863                         f->io_size = f->real_file_size - f->file_offset;
864
865                 if (f->io_size == -1ULL)
866                         total_size = -1ULL;
867                 else {
868                         if (o->size_percent)
869                                 f->io_size = (f->io_size * o->size_percent) / 100;
870                         total_size += f->io_size;
871                 }
872
873                 if (f->filetype == FIO_TYPE_FILE &&
874                     (f->io_size + f->file_offset) > f->real_file_size &&
875                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
876                         if (!o->create_on_open) {
877                                 need_extend++;
878                                 extend_size += (f->io_size + f->file_offset);
879                         } else
880                                 f->real_file_size = f->io_size + f->file_offset;
881                         fio_file_set_extend(f);
882                 }
883         }
884
885         if (!o->size || o->size > total_size)
886                 o->size = total_size;
887
888         if (o->size < td_min_bs(td)) {
889                 log_err("fio: blocksize too large for data set\n");
890                 goto err_out;
891         }
892
893         /*
894          * See if we need to extend some files
895          */
896         if (need_extend) {
897                 temp_stall_ts = 1;
898                 if (output_format == FIO_OUTPUT_NORMAL)
899                         log_info("%s: Laying out IO file(s) (%u file(s) /"
900                                  " %lluMB)\n", o->name, need_extend,
901                                         extend_size >> 20);
902
903                 for_each_file(td, f, i) {
904                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
905
906                         if (!fio_file_extend(f))
907                                 continue;
908
909                         assert(f->filetype == FIO_TYPE_FILE);
910                         fio_file_clear_extend(f);
911                         if (!o->fill_device) {
912                                 old_len = f->real_file_size;
913                                 extend_len = f->io_size + f->file_offset -
914                                                 old_len;
915                         }
916                         f->real_file_size = (f->io_size + f->file_offset);
917                         err = extend_file(td, f);
918                         if (err)
919                                 break;
920
921                         err = __file_invalidate_cache(td, f, old_len,
922                                                                 extend_len);
923
924                         /*
925                          * Shut up static checker
926                          */
927                         if (f->fd != -1)
928                                 close(f->fd);
929
930                         f->fd = -1;
931                         if (err)
932                                 break;
933                 }
934                 temp_stall_ts = 0;
935         }
936
937         if (err)
938                 goto err_out;
939
940         if (!o->zone_size)
941                 o->zone_size = o->size;
942
943         /*
944          * iolog already set the total io size, if we read back
945          * stored entries.
946          */
947         if (!o->read_iolog_file)
948                 td->total_io_size = o->size * o->loops;
949
950 done:
951         if (o->create_only)
952                 td->done = 1;
953
954         td_restore_runstate(td, old_state);
955         return 0;
956 err_offset:
957         log_err("%s: you need to specify valid offset=\n", o->name);
958 err_out:
959         td_restore_runstate(td, old_state);
960         return 1;
961 }
962
963 int pre_read_files(struct thread_data *td)
964 {
965         struct fio_file *f;
966         unsigned int i;
967
968         dprint(FD_FILE, "pre_read files\n");
969
970         for_each_file(td, f, i) {
971                 pre_read_file(td, f);
972         }
973
974         return 1;
975 }
976
977 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
978 {
979         unsigned int range_size, seed;
980         unsigned long nranges;
981         uint64_t file_size;
982
983         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
984         file_size = min(f->real_file_size, f->io_size);
985
986         nranges = (file_size + range_size - 1) / range_size;
987
988         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
989         if (!td->o.rand_repeatable)
990                 seed = td->rand_seeds[4];
991
992         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
993                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
994         else
995                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
996
997         return 1;
998 }
999
1000 static int init_rand_distribution(struct thread_data *td)
1001 {
1002         struct fio_file *f;
1003         unsigned int i;
1004         int state;
1005
1006         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1007                 return 0;
1008
1009         state = td_bump_runstate(td, TD_SETTING_UP);
1010
1011         for_each_file(td, f, i)
1012                 __init_rand_distribution(td, f);
1013
1014         td_restore_runstate(td, state);
1015
1016         return 1;
1017 }
1018
1019 int init_random_map(struct thread_data *td)
1020 {
1021         unsigned long long blocks;
1022         struct fio_file *f;
1023         unsigned int i;
1024
1025         if (init_rand_distribution(td))
1026                 return 0;
1027         if (!td_random(td))
1028                 return 0;
1029
1030         for_each_file(td, f, i) {
1031                 uint64_t file_size = min(f->real_file_size, f->io_size);
1032
1033                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
1034
1035                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1036                         unsigned long seed;
1037
1038                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1039                         
1040                         if (!lfsr_init(&f->lfsr, blocks, seed, 0))
1041                                 continue;
1042                 } else if (!td->o.norandommap) {
1043                         f->io_axmap = axmap_new(blocks);
1044                         if (f->io_axmap)
1045                                 continue;
1046                 } else if (td->o.norandommap)
1047                         continue;
1048
1049                 if (!td->o.softrandommap) {
1050                         log_err("fio: failed allocating random map. If running"
1051                                 " a large number of jobs, try the 'norandommap'"
1052                                 " option or set 'softrandommap'. Or give"
1053                                 " a larger --alloc-size to fio.\n");
1054                         return 1;
1055                 }
1056
1057                 log_info("fio: file %s failed allocating random map. Running "
1058                          "job without.\n", f->file_name);
1059         }
1060
1061         return 0;
1062 }
1063
1064 void close_files(struct thread_data *td)
1065 {
1066         struct fio_file *f;
1067         unsigned int i;
1068
1069         for_each_file(td, f, i) {
1070                 if (fio_file_open(f))
1071                         td_io_close_file(td, f);
1072         }
1073 }
1074
1075 void close_and_free_files(struct thread_data *td)
1076 {
1077         struct fio_file *f;
1078         unsigned int i;
1079
1080         dprint(FD_FILE, "close files\n");
1081
1082         for_each_file(td, f, i) {
1083                 if (fio_file_open(f))
1084                         td_io_close_file(td, f);
1085
1086                 remove_file_hash(f);
1087
1088                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1089                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1090                         unlink(f->file_name);
1091                 }
1092
1093                 sfree(f->file_name);
1094                 f->file_name = NULL;
1095                 axmap_free(f->io_axmap);
1096                 f->io_axmap = NULL;
1097                 sfree(f);
1098         }
1099
1100         td->o.filename = NULL;
1101         free(td->files);
1102         free(td->file_locks);
1103         td->files_index = 0;
1104         td->files = NULL;
1105         td->file_locks = NULL;
1106         td->o.file_lock_mode = FILE_LOCK_NONE;
1107         td->o.nr_files = 0;
1108 }
1109
1110 static void get_file_type(struct fio_file *f)
1111 {
1112         struct stat sb;
1113
1114         if (!strcmp(f->file_name, "-"))
1115                 f->filetype = FIO_TYPE_PIPE;
1116         else
1117                 f->filetype = FIO_TYPE_FILE;
1118
1119         /* \\.\ is the device namespace in Windows, where every file is
1120          * a block device */
1121         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1122                 f->filetype = FIO_TYPE_BD;
1123
1124         if (!stat(f->file_name, &sb)) {
1125                 if (S_ISBLK(sb.st_mode))
1126                         f->filetype = FIO_TYPE_BD;
1127                 else if (S_ISCHR(sb.st_mode))
1128                         f->filetype = FIO_TYPE_CHAR;
1129                 else if (S_ISFIFO(sb.st_mode))
1130                         f->filetype = FIO_TYPE_PIPE;
1131         }
1132 }
1133
1134 static int __is_already_allocated(const char *fname)
1135 {
1136         struct flist_head *entry;
1137         char *filename;
1138
1139         if (flist_empty(&filename_list))
1140                 return 0;
1141
1142         flist_for_each(entry, &filename_list) {
1143                 filename = flist_entry(entry, struct file_name, list)->filename;
1144
1145                 if (strcmp(filename, fname) == 0)
1146                         return 1;
1147         }
1148
1149         return 0;
1150 }
1151
1152 static int is_already_allocated(const char *fname)
1153 {
1154         int ret;
1155
1156         fio_file_hash_lock();
1157         ret = __is_already_allocated(fname);
1158         fio_file_hash_unlock();
1159         return ret;
1160 }
1161
1162 static void set_already_allocated(const char *fname)
1163 {
1164         struct file_name *fn;
1165
1166         fn = malloc(sizeof(struct file_name));
1167         fn->filename = strdup(fname);
1168
1169         fio_file_hash_lock();
1170         if (!__is_already_allocated(fname)) {
1171                 flist_add_tail(&fn->list, &filename_list);
1172                 fn = NULL;
1173         }
1174         fio_file_hash_unlock();
1175
1176         if (fn) {
1177                 free(fn->filename);
1178                 free(fn);
1179         }
1180 }
1181
1182
1183 static void free_already_allocated(void)
1184 {
1185         struct flist_head *entry, *tmp;
1186         struct file_name *fn;
1187
1188         if (flist_empty(&filename_list))
1189                 return;
1190
1191         fio_file_hash_lock();
1192         flist_for_each_safe(entry, tmp, &filename_list) {
1193                 fn = flist_entry(entry, struct file_name, list);
1194                 free(fn->filename);
1195                 flist_del(&fn->list);
1196                 free(fn);
1197         }
1198
1199         fio_file_hash_unlock();
1200 }
1201
1202 static struct fio_file *alloc_new_file(struct thread_data *td)
1203 {
1204         struct fio_file *f;
1205
1206         f = smalloc(sizeof(*f));
1207         if (!f) {
1208                 log_err("fio: smalloc OOM\n");
1209                 assert(0);
1210                 return NULL;
1211         }
1212
1213         f->fd = -1;
1214         f->shadow_fd = -1;
1215         fio_file_reset(td, f);
1216         return f;
1217 }
1218
1219 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1220 {
1221         int cur_files = td->files_index;
1222         char file_name[PATH_MAX];
1223         struct fio_file *f;
1224         int len = 0;
1225
1226         dprint(FD_FILE, "add file %s\n", fname);
1227
1228         if (td->o.directory)
1229                 len = set_name_idx(file_name, td->o.directory, numjob);
1230
1231         sprintf(file_name + len, "%s", fname);
1232
1233         /* clean cloned siblings using existing files */
1234         if (numjob && is_already_allocated(file_name))
1235                 return 0;
1236
1237         f = alloc_new_file(td);
1238
1239         if (td->files_size <= td->files_index) {
1240                 unsigned int new_size = td->o.nr_files + 1;
1241
1242                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1243
1244                 td->files = realloc(td->files, new_size * sizeof(f));
1245                 if (td->files == NULL) {
1246                         log_err("fio: realloc OOM\n");
1247                         assert(0);
1248                 }
1249                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1250                         td->file_locks = realloc(td->file_locks, new_size);
1251                         if (!td->file_locks) {
1252                                 log_err("fio: realloc OOM\n");
1253                                 assert(0);
1254                         }
1255                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1256                 }
1257                 td->files_size = new_size;
1258         }
1259         td->files[cur_files] = f;
1260         f->fileno = cur_files;
1261
1262         /*
1263          * init function, io engine may not be loaded yet
1264          */
1265         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1266                 f->real_file_size = -1ULL;
1267
1268         f->file_name = smalloc_strdup(file_name);
1269         if (!f->file_name) {
1270                 log_err("fio: smalloc OOM\n");
1271                 assert(0);
1272         }
1273
1274         get_file_type(f);
1275
1276         switch (td->o.file_lock_mode) {
1277         case FILE_LOCK_NONE:
1278                 break;
1279         case FILE_LOCK_READWRITE:
1280                 f->rwlock = fio_rwlock_init();
1281                 break;
1282         case FILE_LOCK_EXCLUSIVE:
1283                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1284                 break;
1285         default:
1286                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1287                 assert(0);
1288         }
1289
1290         td->files_index++;
1291         if (f->filetype == FIO_TYPE_FILE)
1292                 td->nr_normal_files++;
1293
1294         set_already_allocated(file_name);
1295
1296         if (inc)
1297                 td->o.nr_files++;
1298
1299         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1300                                                         cur_files);
1301
1302         return cur_files;
1303 }
1304
1305 int add_file_exclusive(struct thread_data *td, const char *fname)
1306 {
1307         struct fio_file *f;
1308         unsigned int i;
1309
1310         for_each_file(td, f, i) {
1311                 if (!strcmp(f->file_name, fname))
1312                         return i;
1313         }
1314
1315         return add_file(td, fname, 0, 1);
1316 }
1317
1318 void get_file(struct fio_file *f)
1319 {
1320         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1321         assert(fio_file_open(f));
1322         f->references++;
1323 }
1324
1325 int put_file(struct thread_data *td, struct fio_file *f)
1326 {
1327         int f_ret = 0, ret = 0;
1328
1329         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1330
1331         if (!fio_file_open(f)) {
1332                 assert(f->fd == -1);
1333                 return 0;
1334         }
1335
1336         assert(f->references);
1337         if (--f->references)
1338                 return 0;
1339
1340         if (should_fsync(td) && td->o.fsync_on_close) {
1341                 f_ret = fsync(f->fd);
1342                 if (f_ret < 0)
1343                         f_ret = errno;
1344         }
1345
1346         if (td->io_ops->close_file)
1347                 ret = td->io_ops->close_file(td, f);
1348
1349         if (!ret)
1350                 ret = f_ret;
1351
1352         td->nr_open_files--;
1353         fio_file_clear_open(f);
1354         assert(f->fd == -1);
1355         return ret;
1356 }
1357
1358 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1359 {
1360         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1361                 return;
1362
1363         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1364                 if (ddir == DDIR_READ)
1365                         fio_rwlock_read(f->rwlock);
1366                 else
1367                         fio_rwlock_write(f->rwlock);
1368         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1369                 fio_mutex_down(f->lock);
1370
1371         td->file_locks[f->fileno] = td->o.file_lock_mode;
1372 }
1373
1374 void unlock_file(struct thread_data *td, struct fio_file *f)
1375 {
1376         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1377                 return;
1378
1379         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1380                 fio_rwlock_unlock(f->rwlock);
1381         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1382                 fio_mutex_up(f->lock);
1383
1384         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1385 }
1386
1387 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1388 {
1389         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1390                 return;
1391         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1392                 unlock_file(td, f);
1393 }
1394
1395 static int recurse_dir(struct thread_data *td, const char *dirname)
1396 {
1397         struct dirent *dir;
1398         int ret = 0;
1399         DIR *D;
1400
1401         D = opendir(dirname);
1402         if (!D) {
1403                 char buf[FIO_VERROR_SIZE];
1404
1405                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1406                 td_verror(td, errno, buf);
1407                 return 1;
1408         }
1409
1410         while ((dir = readdir(D)) != NULL) {
1411                 char full_path[PATH_MAX];
1412                 struct stat sb;
1413
1414                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1415                         continue;
1416
1417                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1418
1419                 if (lstat(full_path, &sb) == -1) {
1420                         if (errno != ENOENT) {
1421                                 td_verror(td, errno, "stat");
1422                                 ret = 1;
1423                                 break;
1424                         }
1425                 }
1426
1427                 if (S_ISREG(sb.st_mode)) {
1428                         add_file(td, full_path, 0, 1);
1429                         continue;
1430                 }
1431                 if (!S_ISDIR(sb.st_mode))
1432                         continue;
1433
1434                 ret = recurse_dir(td, full_path);
1435                 if (ret)
1436                         break;
1437         }
1438
1439         closedir(D);
1440         return ret;
1441 }
1442
1443 int add_dir_files(struct thread_data *td, const char *path)
1444 {
1445         int ret = recurse_dir(td, path);
1446
1447         if (!ret)
1448                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1449
1450         return ret;
1451 }
1452
1453 void dup_files(struct thread_data *td, struct thread_data *org)
1454 {
1455         struct fio_file *f;
1456         unsigned int i;
1457
1458         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1459
1460         if (!org->files)
1461                 return;
1462
1463         td->files = malloc(org->files_index * sizeof(f));
1464
1465         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1466                 td->file_locks = malloc(org->files_index);
1467
1468         for_each_file(org, f, i) {
1469                 struct fio_file *__f;
1470
1471                 __f = alloc_new_file(td);
1472
1473                 if (f->file_name) {
1474                         __f->file_name = smalloc_strdup(f->file_name);
1475                         if (!__f->file_name) {
1476                                 log_err("fio: smalloc OOM\n");
1477                                 assert(0);
1478                         }
1479
1480                         __f->filetype = f->filetype;
1481                 }
1482
1483                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1484                         __f->lock = f->lock;
1485                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1486                         __f->rwlock = f->rwlock;
1487
1488                 td->files[i] = __f;
1489         }
1490 }
1491
1492 /*
1493  * Returns the index that matches the filename, or -1 if not there
1494  */
1495 int get_fileno(struct thread_data *td, const char *fname)
1496 {
1497         struct fio_file *f;
1498         unsigned int i;
1499
1500         for_each_file(td, f, i)
1501                 if (!strcmp(f->file_name, fname))
1502                         return i;
1503
1504         return -1;
1505 }
1506
1507 /*
1508  * For log usage, where we add/open/close files automatically
1509  */
1510 void free_release_files(struct thread_data *td)
1511 {
1512         close_files(td);
1513         td->files_index = 0;
1514         td->nr_normal_files = 0;
1515 }
1516
1517 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1518 {
1519         f->last_pos = f->file_offset;
1520         f->last_start = -1ULL;
1521         if (f->io_axmap)
1522                 axmap_reset(f->io_axmap);
1523         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1524                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1525 }
1526
1527 int fio_files_done(struct thread_data *td)
1528 {
1529         struct fio_file *f;
1530         unsigned int i;
1531
1532         for_each_file(td, f, i)
1533                 if (!fio_file_done(f))
1534                         return 0;
1535
1536         return 1;
1537 }
1538
1539 /* free memory used in initialization phase only */
1540 void filesetup_mem_free(void)
1541 {
1542         free_already_allocated();
1543 }