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