12a43b1fa14459439dd73013dd4ef018e5f6bfba
[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 #ifdef WIN32
73         flags |= _O_BINARY;
74 #endif
75
76         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
77         f->fd = open(f->file_name, flags, 0644);
78         if (f->fd < 0) {
79                 td_verror(td, errno, "open");
80                 return 1;
81         }
82
83 #ifdef CONFIG_POSIX_FALLOCATE
84         if (!td->o.fill_device) {
85                 switch (td->o.fallocate_mode) {
86                 case FIO_FALLOCATE_NONE:
87                         break;
88                 case FIO_FALLOCATE_POSIX:
89                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
90                                  f->file_name,
91                                  (unsigned long long) f->real_file_size);
92
93                         r = posix_fallocate(f->fd, 0, f->real_file_size);
94                         if (r > 0) {
95                                 log_err("fio: posix_fallocate fails: %s\n",
96                                                 strerror(r));
97                         }
98                         break;
99 #ifdef CONFIG_LINUX_FALLOCATE
100                 case FIO_FALLOCATE_KEEP_SIZE:
101                         dprint(FD_FILE,
102                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
103                                 "file %s size %llu\n", f->file_name,
104                                 (unsigned long long) f->real_file_size);
105
106                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
107                                         f->real_file_size);
108                         if (r != 0)
109                                 td_verror(td, errno, "fallocate");
110
111                         break;
112 #endif /* CONFIG_LINUX_FALLOCATE */
113                 default:
114                         log_err("fio: unknown fallocate mode: %d\n",
115                                 td->o.fallocate_mode);
116                         assert(0);
117                 }
118         }
119 #endif /* CONFIG_POSIX_FALLOCATE */
120
121         if (!new_layout)
122                 goto done;
123
124         /*
125          * The size will be -1ULL when fill_device is used, so don't truncate
126          * or fallocate this file, just write it
127          */
128         if (!td->o.fill_device) {
129                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
130                                         (unsigned long long) f->real_file_size);
131                 if (ftruncate(f->fd, f->real_file_size) == -1) {
132                         if (errno != EFBIG) {
133                                 td_verror(td, errno, "ftruncate");
134                                 goto err;
135                         }
136                 }
137         }
138
139         b = malloc(td->o.max_bs[DDIR_WRITE]);
140
141         left = f->real_file_size;
142         while (left && !td->terminate) {
143                 bs = td->o.max_bs[DDIR_WRITE];
144                 if (bs > left)
145                         bs = left;
146
147                 fill_io_buffer(td, b, bs, bs);
148
149                 r = write(f->fd, b, bs);
150
151                 if (r > 0) {
152                         left -= r;
153                         continue;
154                 } else {
155                         if (r < 0) {
156                                 int __e = errno;
157
158                                 if (__e == ENOSPC) {
159                                         if (td->o.fill_device)
160                                                 break;
161                                         log_info("fio: ENOSPC on laying out "
162                                                  "file, stopping\n");
163                                         break;
164                                 }
165                                 td_verror(td, errno, "write");
166                         } else
167                                 td_verror(td, EIO, "write");
168
169                         break;
170                 }
171         }
172
173         if (td->terminate) {
174                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
175                 unlink(f->file_name);
176         } else if (td->o.create_fsync) {
177                 if (fsync(f->fd) < 0) {
178                         td_verror(td, errno, "fsync");
179                         goto err;
180                 }
181         }
182         if (td->o.fill_device && !td_write(td)) {
183                 fio_file_clear_size_known(f);
184                 if (td_io_get_file_size(td, f))
185                         goto err;
186                 if (f->io_size > f->real_file_size)
187                         f->io_size = f->real_file_size;
188         }
189
190         free(b);
191 done:
192         return 0;
193 err:
194         close(f->fd);
195         f->fd = -1;
196         if (b)
197                 free(b);
198         return 1;
199 }
200
201 static int pre_read_file(struct thread_data *td, struct fio_file *f)
202 {
203         int ret = 0, r, did_open = 0, old_runstate;
204         unsigned long long left;
205         unsigned int bs;
206         char *b;
207
208         if (td->io_ops->flags & FIO_PIPEIO)
209                 return 0;
210
211         if (!fio_file_open(f)) {
212                 if (td->io_ops->open_file(td, f)) {
213                         log_err("fio: cannot pre-read, failed to open file\n");
214                         return 1;
215                 }
216                 did_open = 1;
217         }
218
219         old_runstate = td_bump_runstate(td, TD_PRE_READING);
220
221         bs = td->o.max_bs[DDIR_READ];
222         b = malloc(bs);
223         memset(b, 0, bs);
224
225         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
226                 td_verror(td, errno, "lseek");
227                 log_err("fio: failed to lseek pre-read file\n");
228                 ret = 1;
229                 goto error;
230         }
231
232         left = f->io_size;
233
234         while (left && !td->terminate) {
235                 if (bs > left)
236                         bs = left;
237
238                 r = read(f->fd, b, bs);
239
240                 if (r == (int) bs) {
241                         left -= bs;
242                         continue;
243                 } else {
244                         td_verror(td, EIO, "pre_read");
245                         break;
246                 }
247         }
248
249 error:
250         td_restore_runstate(td, old_runstate);
251
252         if (did_open)
253                 td->io_ops->close_file(td, f);
254
255         free(b);
256         return ret;
257 }
258
259 static unsigned long long get_rand_file_size(struct thread_data *td)
260 {
261         unsigned long long ret, sized;
262         unsigned long r;
263
264         if (td->o.use_os_rand) {
265                 r = os_random_long(&td->file_size_state);
266                 sized = td->o.file_size_high - td->o.file_size_low;
267                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
268         } else {
269                 r = __rand(&td->__file_size_state);
270                 sized = td->o.file_size_high - td->o.file_size_low;
271                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
272         }
273
274         ret += td->o.file_size_low;
275         ret -= (ret % td->o.rw_min_bs);
276         return ret;
277 }
278
279 static int file_size(struct thread_data *td, struct fio_file *f)
280 {
281         struct stat st;
282
283         if (stat(f->file_name, &st) == -1) {
284                 td_verror(td, errno, "fstat");
285                 return 1;
286         }
287
288         f->real_file_size = st.st_size;
289         return 0;
290 }
291
292 static int bdev_size(struct thread_data *td, struct fio_file *f)
293 {
294         unsigned long long bytes = 0;
295         int r;
296
297         if (td->io_ops->open_file(td, f)) {
298                 log_err("fio: failed opening blockdev %s for size check\n",
299                         f->file_name);
300                 return 1;
301         }
302
303         r = blockdev_size(f, &bytes);
304         if (r) {
305                 td_verror(td, r, "blockdev_size");
306                 goto err;
307         }
308
309         if (!bytes) {
310                 log_err("%s: zero sized block device?\n", f->file_name);
311                 goto err;
312         }
313
314         f->real_file_size = bytes;
315         td->io_ops->close_file(td, f);
316         return 0;
317 err:
318         td->io_ops->close_file(td, f);
319         return 1;
320 }
321
322 static int char_size(struct thread_data *td, struct fio_file *f)
323 {
324 #ifdef FIO_HAVE_CHARDEV_SIZE
325         unsigned long long bytes = 0;
326         int r;
327
328         if (td->io_ops->open_file(td, f)) {
329                 log_err("fio: failed opening blockdev %s for size check\n",
330                         f->file_name);
331                 return 1;
332         }
333
334         r = chardev_size(f, &bytes);
335         if (r) {
336                 td_verror(td, r, "chardev_size");
337                 goto err;
338         }
339
340         if (!bytes) {
341                 log_err("%s: zero sized char device?\n", f->file_name);
342                 goto err;
343         }
344
345         f->real_file_size = bytes;
346         td->io_ops->close_file(td, f);
347         return 0;
348 err:
349         td->io_ops->close_file(td, f);
350         return 1;
351 #else
352         f->real_file_size = -1ULL;
353         return 0;
354 #endif
355 }
356
357 static int get_file_size(struct thread_data *td, struct fio_file *f)
358 {
359         int ret = 0;
360
361         if (fio_file_size_known(f))
362                 return 0;
363
364         if (f->filetype == FIO_TYPE_FILE)
365                 ret = file_size(td, f);
366         else if (f->filetype == FIO_TYPE_BD)
367                 ret = bdev_size(td, f);
368         else if (f->filetype == FIO_TYPE_CHAR)
369                 ret = char_size(td, f);
370         else
371                 f->real_file_size = -1;
372
373         if (ret)
374                 return ret;
375
376         if (f->file_offset > f->real_file_size) {
377                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
378                                         (unsigned long long) f->file_offset,
379                                         (unsigned long long) f->real_file_size);
380                 return 1;
381         }
382
383         fio_file_set_size_known(f);
384         return 0;
385 }
386
387 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
388                                    unsigned long long off,
389                                    unsigned long long len)
390 {
391         int ret = 0;
392
393 #ifdef CONFIG_ESX
394         return 0;
395 #endif
396
397         if (len == -1ULL)
398                 len = f->io_size;
399         if (off == -1ULL)
400                 off = f->file_offset;
401
402         if (len == -1ULL || off == -1ULL)
403                 return 0;
404
405         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
406                                                                 len);
407
408         if (td->io_ops->invalidate)
409                 ret = td->io_ops->invalidate(td, f);
410         else if (f->mmap_ptr) {
411                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
412 #ifdef FIO_MADV_FREE
413                 if (f->filetype == FIO_TYPE_BD)
414                         (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
415 #endif
416         } else if (f->filetype == FIO_TYPE_FILE) {
417                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
418         } else if (f->filetype == FIO_TYPE_BD) {
419                 ret = blockdev_invalidate_cache(f);
420                 if (ret < 0 && errno == EACCES && geteuid()) {
421                         if (!root_warn) {
422                                 log_err("fio: only root may flush block "
423                                         "devices. Cache flush bypassed!\n");
424                                 root_warn = 1;
425                         }
426                         ret = 0;
427                 }
428         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
429                 ret = 0;
430
431         /*
432          * Cache flushing isn't a fatal condition, and we know it will
433          * happen on some platforms where we don't have the proper
434          * function to flush eg block device caches. So just warn and
435          * continue on our way.
436          */
437         if (ret) {
438                 log_info("fio: cache invalidation of %s failed: %s\n", f->file_name, strerror(errno));
439                 ret = 0;
440         }
441
442         return 0;
443
444 }
445
446 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
447 {
448         if (!fio_file_open(f))
449                 return 0;
450
451         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
452 }
453
454 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
455 {
456         int ret = 0;
457
458         dprint(FD_FILE, "fd close %s\n", f->file_name);
459
460         remove_file_hash(f);
461
462         if (close(f->fd) < 0)
463                 ret = errno;
464
465         f->fd = -1;
466
467         if (f->shadow_fd != -1) {
468                 close(f->shadow_fd);
469                 f->shadow_fd = -1;
470         }
471
472         f->engine_data = 0;
473         return ret;
474 }
475
476 int file_lookup_open(struct fio_file *f, int flags)
477 {
478         struct fio_file *__f;
479         int from_hash;
480
481         __f = lookup_file_hash(f->file_name);
482         if (__f) {
483                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
484                 /*
485                  * racy, need the __f->lock locked
486                  */
487                 f->lock = __f->lock;
488                 from_hash = 1;
489         } else {
490                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
491                 from_hash = 0;
492         }
493
494 #ifdef WIN32
495         flags |= _O_BINARY;
496 #endif
497
498         f->fd = open(f->file_name, flags, 0600);
499         return from_hash;
500 }
501
502 static int file_close_shadow_fds(struct thread_data *td)
503 {
504         struct fio_file *f;
505         int num_closed = 0;
506         unsigned int i;
507
508         for_each_file(td, f, i) {
509                 if (f->shadow_fd == -1)
510                         continue;
511
512                 close(f->shadow_fd);
513                 f->shadow_fd = -1;
514                 num_closed++;
515         }
516
517         return num_closed;
518 }
519
520 int generic_open_file(struct thread_data *td, struct fio_file *f)
521 {
522         int is_std = 0;
523         int flags = 0;
524         int from_hash = 0;
525
526         dprint(FD_FILE, "fd open %s\n", f->file_name);
527
528         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
529                 log_err("fio: trim only applies to block device\n");
530                 return 1;
531         }
532
533         if (!strcmp(f->file_name, "-")) {
534                 if (td_rw(td)) {
535                         log_err("fio: can't read/write to stdin/out\n");
536                         return 1;
537                 }
538                 is_std = 1;
539
540                 /*
541                  * move output logging to stderr, if we are writing to stdout
542                  */
543                 if (td_write(td))
544                         f_out = stderr;
545         }
546
547         if (td_trim(td))
548                 goto skip_flags;
549         if (td->o.odirect)
550                 flags |= OS_O_DIRECT;
551         if (td->o.oatomic) {
552                 if (!FIO_O_ATOMIC) {
553                         td_verror(td, EINVAL, "OS does not support atomic IO");
554                         return 1;
555                 }
556                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
557         }
558         if (td->o.sync_io)
559                 flags |= O_SYNC;
560         if (td->o.create_on_open)
561                 flags |= O_CREAT;
562 skip_flags:
563         if (f->filetype != FIO_TYPE_FILE)
564                 flags |= FIO_O_NOATIME;
565
566 open_again:
567         if (td_write(td)) {
568                 if (!read_only)
569                         flags |= O_RDWR;
570
571                 if (f->filetype == FIO_TYPE_FILE)
572                         flags |= O_CREAT;
573
574                 if (is_std)
575                         f->fd = dup(STDOUT_FILENO);
576                 else
577                         from_hash = file_lookup_open(f, flags);
578         } else if (td_read(td)) {
579                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
580                         flags |= O_RDWR;
581                 else
582                         flags |= O_RDONLY;
583
584                 if (is_std)
585                         f->fd = dup(STDIN_FILENO);
586                 else
587                         from_hash = file_lookup_open(f, flags);
588         } else { //td trim
589                 flags |= O_RDWR;
590                 from_hash = file_lookup_open(f, flags);
591         }
592
593         if (f->fd == -1) {
594                 char buf[FIO_VERROR_SIZE];
595                 int __e = errno;
596
597                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
598                         flags &= ~FIO_O_NOATIME;
599                         goto open_again;
600                 }
601                 if (__e == EMFILE && file_close_shadow_fds(td))
602                         goto open_again;
603
604                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
605
606                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
607                         log_err("fio: looks like your file system does not " \
608                                 "support direct=1/buffered=0\n");
609                 }
610
611                 td_verror(td, __e, buf);
612                 return 1;
613         }
614
615         if (!from_hash && f->fd != -1) {
616                 if (add_file_hash(f)) {
617                         int fio_unused ret;
618
619                         /*
620                          * Stash away descriptor for later close. This is to
621                          * work-around a "feature" on Linux, where a close of
622                          * an fd that has been opened for write will trigger
623                          * udev to call blkid to check partitions, fs id, etc.
624                          * That pollutes the device cache, which can slow down
625                          * unbuffered accesses.
626                          */
627                         if (f->shadow_fd == -1)
628                                 f->shadow_fd = f->fd;
629                         else {
630                                 /*
631                                  * OK to ignore, we haven't done anything
632                                  * with it
633                                  */
634                                 ret = generic_close_file(td, f);
635                         }
636                         goto open_again;
637                 }
638         }
639
640         return 0;
641 }
642
643 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
644 {
645         return get_file_size(td, f);
646 }
647
648 /*
649  * open/close all files, so that ->real_file_size gets set
650  */
651 static int get_file_sizes(struct thread_data *td)
652 {
653         struct fio_file *f;
654         unsigned int i;
655         int err = 0;
656
657         for_each_file(td, f, i) {
658                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
659                                                                 f->file_name);
660
661                 if (td_io_get_file_size(td, f)) {
662                         if (td->error != ENOENT) {
663                                 log_err("%s\n", td->verror);
664                                 err = 1;
665                                 break;
666                         }
667                         clear_error(td);
668                 }
669
670                 if (f->real_file_size == -1ULL && td->o.size)
671                         f->real_file_size = td->o.size / td->o.nr_files;
672         }
673
674         return err;
675 }
676
677 struct fio_mount {
678         struct flist_head list;
679         const char *base;
680         char __base[256];
681         unsigned int key;
682 };
683
684 /*
685  * Get free number of bytes for each file on each unique mount.
686  */
687 static unsigned long long get_fs_free_counts(struct thread_data *td)
688 {
689         struct flist_head *n, *tmp;
690         unsigned long long ret = 0;
691         struct fio_mount *fm;
692         FLIST_HEAD(list);
693         struct fio_file *f;
694         unsigned int i;
695
696         for_each_file(td, f, i) {
697                 struct stat sb;
698                 char buf[256];
699
700                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
701                         if (f->real_file_size != -1ULL)
702                                 ret += f->real_file_size;
703                         continue;
704                 } else if (f->filetype != FIO_TYPE_FILE)
705                         continue;
706
707                 buf[255] = '\0';
708                 strncpy(buf, f->file_name, 255);
709
710                 if (stat(buf, &sb) < 0) {
711                         if (errno != ENOENT)
712                                 break;
713                         strcpy(buf, ".");
714                         if (stat(buf, &sb) < 0)
715                                 break;
716                 }
717
718                 fm = NULL;
719                 flist_for_each(n, &list) {
720                         fm = flist_entry(n, struct fio_mount, list);
721                         if (fm->key == sb.st_dev)
722                                 break;
723
724                         fm = NULL;
725                 }
726
727                 if (fm)
728                         continue;
729
730                 fm = calloc(1, sizeof(*fm));
731                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
732                 fm->base = basename(fm->__base);
733                 fm->key = sb.st_dev;
734                 flist_add(&fm->list, &list);
735         }
736
737         flist_for_each_safe(n, tmp, &list) {
738                 unsigned long long sz;
739
740                 fm = flist_entry(n, struct fio_mount, list);
741                 flist_del(&fm->list);
742
743                 sz = get_fs_size(fm->base);
744                 if (sz && sz != -1ULL)
745                         ret += sz;
746
747                 free(fm);
748         }
749
750         return ret;
751 }
752
753 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
754 {
755         struct thread_options *o = &td->o;
756
757         if (o->file_append && f->filetype == FIO_TYPE_FILE)
758                 return f->real_file_size;
759
760         return td->o.start_offset +
761                 td->subjob_number * td->o.offset_increment;
762 }
763
764 /*
765  * Open the files and setup files sizes, creating files if necessary.
766  */
767 int setup_files(struct thread_data *td)
768 {
769         unsigned long long total_size, extend_size;
770         struct thread_options *o = &td->o;
771         struct fio_file *f;
772         unsigned int i, nr_fs_extra = 0;
773         int err = 0, need_extend;
774         int old_state;
775         const unsigned int bs = td_min_bs(td);
776         uint64_t fs = 0;
777
778         dprint(FD_FILE, "setup files\n");
779
780         old_state = td_bump_runstate(td, TD_SETTING_UP);
781
782         if (o->read_iolog_file)
783                 goto done;
784
785         /*
786          * if ioengine defines a setup() method, it's responsible for
787          * opening the files and setting f->real_file_size to indicate
788          * the valid range for that file.
789          */
790         if (td->io_ops->setup)
791                 err = td->io_ops->setup(td);
792         else
793                 err = get_file_sizes(td);
794
795         if (err)
796                 goto err_out;
797
798         /*
799          * check sizes. if the files/devices do not exist and the size
800          * isn't passed to fio, abort.
801          */
802         total_size = 0;
803         for_each_file(td, f, i) {
804                 if (f->real_file_size == -1ULL)
805                         total_size = -1ULL;
806                 else
807                         total_size += f->real_file_size;
808         }
809
810         if (o->fill_device)
811                 td->fill_device_size = get_fs_free_counts(td);
812
813         /*
814          * device/file sizes are zero and no size given, punt
815          */
816         if ((!total_size || total_size == -1ULL) && !o->size &&
817             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
818             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
819                 log_err("%s: you need to specify size=\n", o->name);
820                 td_verror(td, EINVAL, "total_file_size");
821                 goto err_out;
822         }
823
824         /*
825          * Calculate per-file size and potential extra size for the
826          * first files, if needed.
827          */
828         if (!o->file_size_low && o->nr_files) {
829                 uint64_t all_fs;
830
831                 fs = o->size / o->nr_files;
832                 all_fs = fs * o->nr_files;
833
834                 if (all_fs < o->size)
835                         nr_fs_extra = (o->size - all_fs) / bs;
836         }
837
838         /*
839          * now file sizes are known, so we can set ->io_size. if size= is
840          * not given, ->io_size is just equal to ->real_file_size. if size
841          * is given, ->io_size is size / nr_files.
842          */
843         extend_size = total_size = 0;
844         need_extend = 0;
845         for_each_file(td, f, i) {
846                 f->file_offset = get_start_offset(td, f);
847
848                 if (!o->file_size_low) {
849                         /*
850                          * no file size range given, file size is equal to
851                          * total size divided by number of files. If that is
852                          * zero, set it to the real file size. If the size
853                          * doesn't divide nicely with the min blocksize,
854                          * make the first files bigger.
855                          */
856                         f->io_size = fs;
857                         if (nr_fs_extra) {
858                                 nr_fs_extra--;
859                                 f->io_size += bs;
860                         }
861
862                         if (!f->io_size)
863                                 f->io_size = f->real_file_size - f->file_offset;
864                 } else if (f->real_file_size < o->file_size_low ||
865                            f->real_file_size > o->file_size_high) {
866                         if (f->file_offset > o->file_size_low)
867                                 goto err_offset;
868                         /*
869                          * file size given. if it's fixed, use that. if it's a
870                          * range, generate a random size in-between.
871                          */
872                         if (o->file_size_low == o->file_size_high)
873                                 f->io_size = o->file_size_low - f->file_offset;
874                         else {
875                                 f->io_size = get_rand_file_size(td)
876                                                 - f->file_offset;
877                         }
878                 } else
879                         f->io_size = f->real_file_size - f->file_offset;
880
881                 if (f->io_size == -1ULL)
882                         total_size = -1ULL;
883                 else {
884                         if (o->size_percent)
885                                 f->io_size = (f->io_size * o->size_percent) / 100;
886                         total_size += f->io_size;
887                 }
888
889                 if (f->filetype == FIO_TYPE_FILE &&
890                     (f->io_size + f->file_offset) > f->real_file_size &&
891                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
892                         if (!o->create_on_open) {
893                                 need_extend++;
894                                 extend_size += (f->io_size + f->file_offset);
895                         } else
896                                 f->real_file_size = f->io_size + f->file_offset;
897                         fio_file_set_extend(f);
898                 }
899         }
900
901         if (!o->size || o->size > total_size)
902                 o->size = total_size;
903
904         if (o->size < td_min_bs(td)) {
905                 log_err("fio: blocksize too large for data set\n");
906                 goto err_out;
907         }
908
909         /*
910          * See if we need to extend some files
911          */
912         if (need_extend) {
913                 temp_stall_ts = 1;
914                 if (output_format == FIO_OUTPUT_NORMAL)
915                         log_info("%s: Laying out IO file(s) (%u file(s) /"
916                                  " %lluMB)\n", o->name, need_extend,
917                                         extend_size >> 20);
918
919                 for_each_file(td, f, i) {
920                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
921
922                         if (!fio_file_extend(f))
923                                 continue;
924
925                         assert(f->filetype == FIO_TYPE_FILE);
926                         fio_file_clear_extend(f);
927                         if (!o->fill_device) {
928                                 old_len = f->real_file_size;
929                                 extend_len = f->io_size + f->file_offset -
930                                                 old_len;
931                         }
932                         f->real_file_size = (f->io_size + f->file_offset);
933                         err = extend_file(td, f);
934                         if (err)
935                                 break;
936
937                         err = __file_invalidate_cache(td, f, old_len,
938                                                                 extend_len);
939
940                         /*
941                          * Shut up static checker
942                          */
943                         if (f->fd != -1)
944                                 close(f->fd);
945
946                         f->fd = -1;
947                         if (err)
948                                 break;
949                 }
950                 temp_stall_ts = 0;
951         }
952
953         if (err)
954                 goto err_out;
955
956         if (!o->zone_size)
957                 o->zone_size = o->size;
958
959         /*
960          * iolog already set the total io size, if we read back
961          * stored entries.
962          */
963         if (!o->read_iolog_file) {
964                 if (o->io_limit)
965                         td->total_io_size = o->io_limit * o->loops;
966                 else
967                         td->total_io_size = o->size * o->loops;
968         }
969
970 done:
971         if (o->create_only)
972                 td->done = 1;
973
974         td_restore_runstate(td, old_state);
975         return 0;
976 err_offset:
977         log_err("%s: you need to specify valid offset=\n", o->name);
978 err_out:
979         td_restore_runstate(td, old_state);
980         return 1;
981 }
982
983 int pre_read_files(struct thread_data *td)
984 {
985         struct fio_file *f;
986         unsigned int i;
987
988         dprint(FD_FILE, "pre_read files\n");
989
990         for_each_file(td, f, i) {
991                 pre_read_file(td, f);
992         }
993
994         return 1;
995 }
996
997 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
998 {
999         unsigned int range_size, seed;
1000         unsigned long nranges;
1001         uint64_t file_size;
1002
1003         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1004         file_size = min(f->real_file_size, f->io_size);
1005
1006         nranges = (file_size + range_size - 1) / range_size;
1007
1008         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1009         if (!td->o.rand_repeatable)
1010                 seed = td->rand_seeds[4];
1011
1012         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1013                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1014         else
1015                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1016
1017         return 1;
1018 }
1019
1020 static int init_rand_distribution(struct thread_data *td)
1021 {
1022         struct fio_file *f;
1023         unsigned int i;
1024         int state;
1025
1026         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1027                 return 0;
1028
1029         state = td_bump_runstate(td, TD_SETTING_UP);
1030
1031         for_each_file(td, f, i)
1032                 __init_rand_distribution(td, f);
1033
1034         td_restore_runstate(td, state);
1035
1036         return 1;
1037 }
1038
1039 int init_random_map(struct thread_data *td)
1040 {
1041         unsigned long long blocks;
1042         struct fio_file *f;
1043         unsigned int i;
1044
1045         if (init_rand_distribution(td))
1046                 return 0;
1047         if (!td_random(td))
1048                 return 0;
1049
1050         for_each_file(td, f, i) {
1051                 uint64_t file_size = min(f->real_file_size, f->io_size);
1052
1053                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
1054
1055                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1056                         unsigned long seed;
1057
1058                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1059
1060                         if (!lfsr_init(&f->lfsr, blocks, seed, 0))
1061                                 continue;
1062                 } else if (!td->o.norandommap) {
1063                         f->io_axmap = axmap_new(blocks);
1064                         if (f->io_axmap)
1065                                 continue;
1066                 } else if (td->o.norandommap)
1067                         continue;
1068
1069                 if (!td->o.softrandommap) {
1070                         log_err("fio: failed allocating random map. If running"
1071                                 " a large number of jobs, try the 'norandommap'"
1072                                 " option or set 'softrandommap'. Or give"
1073                                 " a larger --alloc-size to fio.\n");
1074                         return 1;
1075                 }
1076
1077                 log_info("fio: file %s failed allocating random map. Running "
1078                          "job without.\n", f->file_name);
1079         }
1080
1081         return 0;
1082 }
1083
1084 void close_files(struct thread_data *td)
1085 {
1086         struct fio_file *f;
1087         unsigned int i;
1088
1089         for_each_file(td, f, i) {
1090                 if (fio_file_open(f))
1091                         td_io_close_file(td, f);
1092         }
1093 }
1094
1095 void close_and_free_files(struct thread_data *td)
1096 {
1097         struct fio_file *f;
1098         unsigned int i;
1099
1100         dprint(FD_FILE, "close files\n");
1101
1102         for_each_file(td, f, i) {
1103                 if (fio_file_open(f))
1104                         td_io_close_file(td, f);
1105
1106                 remove_file_hash(f);
1107
1108                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1109                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1110                         unlink(f->file_name);
1111                 }
1112
1113                 sfree(f->file_name);
1114                 f->file_name = NULL;
1115                 axmap_free(f->io_axmap);
1116                 f->io_axmap = NULL;
1117                 sfree(f);
1118         }
1119
1120         td->o.filename = NULL;
1121         free(td->files);
1122         free(td->file_locks);
1123         td->files_index = 0;
1124         td->files = NULL;
1125         td->file_locks = NULL;
1126         td->o.file_lock_mode = FILE_LOCK_NONE;
1127         td->o.nr_files = 0;
1128 }
1129
1130 static void get_file_type(struct fio_file *f)
1131 {
1132         struct stat sb;
1133
1134         if (!strcmp(f->file_name, "-"))
1135                 f->filetype = FIO_TYPE_PIPE;
1136         else
1137                 f->filetype = FIO_TYPE_FILE;
1138
1139         /* \\.\ is the device namespace in Windows, where every file is
1140          * a block device */
1141         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1142                 f->filetype = FIO_TYPE_BD;
1143
1144         if (!stat(f->file_name, &sb)) {
1145                 if (S_ISBLK(sb.st_mode))
1146                         f->filetype = FIO_TYPE_BD;
1147                 else if (S_ISCHR(sb.st_mode))
1148                         f->filetype = FIO_TYPE_CHAR;
1149                 else if (S_ISFIFO(sb.st_mode))
1150                         f->filetype = FIO_TYPE_PIPE;
1151         }
1152 }
1153
1154 static int __is_already_allocated(const char *fname)
1155 {
1156         struct flist_head *entry;
1157         char *filename;
1158
1159         if (flist_empty(&filename_list))
1160                 return 0;
1161
1162         flist_for_each(entry, &filename_list) {
1163                 filename = flist_entry(entry, struct file_name, list)->filename;
1164
1165                 if (strcmp(filename, fname) == 0)
1166                         return 1;
1167         }
1168
1169         return 0;
1170 }
1171
1172 static int is_already_allocated(const char *fname)
1173 {
1174         int ret;
1175
1176         fio_file_hash_lock();
1177         ret = __is_already_allocated(fname);
1178         fio_file_hash_unlock();
1179         return ret;
1180 }
1181
1182 static void set_already_allocated(const char *fname)
1183 {
1184         struct file_name *fn;
1185
1186         fn = malloc(sizeof(struct file_name));
1187         fn->filename = strdup(fname);
1188
1189         fio_file_hash_lock();
1190         if (!__is_already_allocated(fname)) {
1191                 flist_add_tail(&fn->list, &filename_list);
1192                 fn = NULL;
1193         }
1194         fio_file_hash_unlock();
1195
1196         if (fn) {
1197                 free(fn->filename);
1198                 free(fn);
1199         }
1200 }
1201
1202
1203 static void free_already_allocated(void)
1204 {
1205         struct flist_head *entry, *tmp;
1206         struct file_name *fn;
1207
1208         if (flist_empty(&filename_list))
1209                 return;
1210
1211         fio_file_hash_lock();
1212         flist_for_each_safe(entry, tmp, &filename_list) {
1213                 fn = flist_entry(entry, struct file_name, list);
1214                 free(fn->filename);
1215                 flist_del(&fn->list);
1216                 free(fn);
1217         }
1218
1219         fio_file_hash_unlock();
1220 }
1221
1222 static struct fio_file *alloc_new_file(struct thread_data *td)
1223 {
1224         struct fio_file *f;
1225
1226         f = smalloc(sizeof(*f));
1227         if (!f) {
1228                 log_err("fio: smalloc OOM\n");
1229                 assert(0);
1230                 return NULL;
1231         }
1232
1233         f->fd = -1;
1234         f->shadow_fd = -1;
1235         fio_file_reset(td, f);
1236         return f;
1237 }
1238
1239 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1240 {
1241         int cur_files = td->files_index;
1242         char file_name[PATH_MAX];
1243         struct fio_file *f;
1244         int len = 0;
1245
1246         dprint(FD_FILE, "add file %s\n", fname);
1247
1248         if (td->o.directory)
1249                 len = set_name_idx(file_name, td->o.directory, numjob);
1250
1251         sprintf(file_name + len, "%s", fname);
1252
1253         /* clean cloned siblings using existing files */
1254         if (numjob && is_already_allocated(file_name))
1255                 return 0;
1256
1257         f = alloc_new_file(td);
1258
1259         if (td->files_size <= td->files_index) {
1260                 unsigned int new_size = td->o.nr_files + 1;
1261
1262                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1263
1264                 td->files = realloc(td->files, new_size * sizeof(f));
1265                 if (td->files == NULL) {
1266                         log_err("fio: realloc OOM\n");
1267                         assert(0);
1268                 }
1269                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1270                         td->file_locks = realloc(td->file_locks, new_size);
1271                         if (!td->file_locks) {
1272                                 log_err("fio: realloc OOM\n");
1273                                 assert(0);
1274                         }
1275                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1276                 }
1277                 td->files_size = new_size;
1278         }
1279         td->files[cur_files] = f;
1280         f->fileno = cur_files;
1281
1282         /*
1283          * init function, io engine may not be loaded yet
1284          */
1285         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1286                 f->real_file_size = -1ULL;
1287
1288         f->file_name = smalloc_strdup(file_name);
1289         if (!f->file_name) {
1290                 log_err("fio: smalloc OOM\n");
1291                 assert(0);
1292         }
1293
1294         get_file_type(f);
1295
1296         switch (td->o.file_lock_mode) {
1297         case FILE_LOCK_NONE:
1298                 break;
1299         case FILE_LOCK_READWRITE:
1300                 f->rwlock = fio_rwlock_init();
1301                 break;
1302         case FILE_LOCK_EXCLUSIVE:
1303                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1304                 break;
1305         default:
1306                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1307                 assert(0);
1308         }
1309
1310         td->files_index++;
1311         if (f->filetype == FIO_TYPE_FILE)
1312                 td->nr_normal_files++;
1313
1314         set_already_allocated(file_name);
1315
1316         if (inc)
1317                 td->o.nr_files++;
1318
1319         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1320                                                         cur_files);
1321
1322         return cur_files;
1323 }
1324
1325 int add_file_exclusive(struct thread_data *td, const char *fname)
1326 {
1327         struct fio_file *f;
1328         unsigned int i;
1329
1330         for_each_file(td, f, i) {
1331                 if (!strcmp(f->file_name, fname))
1332                         return i;
1333         }
1334
1335         return add_file(td, fname, 0, 1);
1336 }
1337
1338 void get_file(struct fio_file *f)
1339 {
1340         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1341         assert(fio_file_open(f));
1342         f->references++;
1343 }
1344
1345 int put_file(struct thread_data *td, struct fio_file *f)
1346 {
1347         int f_ret = 0, ret = 0;
1348
1349         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1350
1351         if (!fio_file_open(f)) {
1352                 assert(f->fd == -1);
1353                 return 0;
1354         }
1355
1356         assert(f->references);
1357         if (--f->references)
1358                 return 0;
1359
1360         if (should_fsync(td) && td->o.fsync_on_close) {
1361                 f_ret = fsync(f->fd);
1362                 if (f_ret < 0)
1363                         f_ret = errno;
1364         }
1365
1366         if (td->io_ops->close_file)
1367                 ret = td->io_ops->close_file(td, f);
1368
1369         if (!ret)
1370                 ret = f_ret;
1371
1372         td->nr_open_files--;
1373         fio_file_clear_open(f);
1374         assert(f->fd == -1);
1375         return ret;
1376 }
1377
1378 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1379 {
1380         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1381                 return;
1382
1383         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1384                 if (ddir == DDIR_READ)
1385                         fio_rwlock_read(f->rwlock);
1386                 else
1387                         fio_rwlock_write(f->rwlock);
1388         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1389                 fio_mutex_down(f->lock);
1390
1391         td->file_locks[f->fileno] = td->o.file_lock_mode;
1392 }
1393
1394 void unlock_file(struct thread_data *td, struct fio_file *f)
1395 {
1396         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1397                 return;
1398
1399         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1400                 fio_rwlock_unlock(f->rwlock);
1401         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1402                 fio_mutex_up(f->lock);
1403
1404         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1405 }
1406
1407 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1408 {
1409         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1410                 return;
1411         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1412                 unlock_file(td, f);
1413 }
1414
1415 static int recurse_dir(struct thread_data *td, const char *dirname)
1416 {
1417         struct dirent *dir;
1418         int ret = 0;
1419         DIR *D;
1420
1421         D = opendir(dirname);
1422         if (!D) {
1423                 char buf[FIO_VERROR_SIZE];
1424
1425                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1426                 td_verror(td, errno, buf);
1427                 return 1;
1428         }
1429
1430         while ((dir = readdir(D)) != NULL) {
1431                 char full_path[PATH_MAX];
1432                 struct stat sb;
1433
1434                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1435                         continue;
1436
1437                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1438
1439                 if (lstat(full_path, &sb) == -1) {
1440                         if (errno != ENOENT) {
1441                                 td_verror(td, errno, "stat");
1442                                 ret = 1;
1443                                 break;
1444                         }
1445                 }
1446
1447                 if (S_ISREG(sb.st_mode)) {
1448                         add_file(td, full_path, 0, 1);
1449                         continue;
1450                 }
1451                 if (!S_ISDIR(sb.st_mode))
1452                         continue;
1453
1454                 ret = recurse_dir(td, full_path);
1455                 if (ret)
1456                         break;
1457         }
1458
1459         closedir(D);
1460         return ret;
1461 }
1462
1463 int add_dir_files(struct thread_data *td, const char *path)
1464 {
1465         int ret = recurse_dir(td, path);
1466
1467         if (!ret)
1468                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1469
1470         return ret;
1471 }
1472
1473 void dup_files(struct thread_data *td, struct thread_data *org)
1474 {
1475         struct fio_file *f;
1476         unsigned int i;
1477
1478         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1479
1480         if (!org->files)
1481                 return;
1482
1483         td->files = malloc(org->files_index * sizeof(f));
1484
1485         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1486                 td->file_locks = malloc(org->files_index);
1487
1488         for_each_file(org, f, i) {
1489                 struct fio_file *__f;
1490
1491                 __f = alloc_new_file(td);
1492
1493                 if (f->file_name) {
1494                         __f->file_name = smalloc_strdup(f->file_name);
1495                         if (!__f->file_name) {
1496                                 log_err("fio: smalloc OOM\n");
1497                                 assert(0);
1498                         }
1499
1500                         __f->filetype = f->filetype;
1501                 }
1502
1503                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1504                         __f->lock = f->lock;
1505                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1506                         __f->rwlock = f->rwlock;
1507
1508                 td->files[i] = __f;
1509         }
1510 }
1511
1512 /*
1513  * Returns the index that matches the filename, or -1 if not there
1514  */
1515 int get_fileno(struct thread_data *td, const char *fname)
1516 {
1517         struct fio_file *f;
1518         unsigned int i;
1519
1520         for_each_file(td, f, i)
1521                 if (!strcmp(f->file_name, fname))
1522                         return i;
1523
1524         return -1;
1525 }
1526
1527 /*
1528  * For log usage, where we add/open/close files automatically
1529  */
1530 void free_release_files(struct thread_data *td)
1531 {
1532         close_files(td);
1533         td->o.nr_files = 0;
1534         td->o.open_files = 0;
1535         td->files_index = 0;
1536         td->nr_normal_files = 0;
1537 }
1538
1539 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1540 {
1541         f->last_pos = f->file_offset;
1542         f->last_start = -1ULL;
1543         if (f->io_axmap)
1544                 axmap_reset(f->io_axmap);
1545         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1546                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1547 }
1548
1549 int fio_files_done(struct thread_data *td)
1550 {
1551         struct fio_file *f;
1552         unsigned int i;
1553
1554         for_each_file(td, f, i)
1555                 if (!fio_file_done(f))
1556                         return 0;
1557
1558         return 1;
1559 }
1560
1561 /* free memory used in initialization phase only */
1562 void filesetup_mem_free(void)
1563 {
1564         free_already_allocated();
1565 }