Revert "Minor style fixups"
[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
9 #include "fio.h"
10 #include "smalloc.h"
11 #include "filehash.h"
12 #include "options.h"
13 #include "os/os.h"
14 #include "hash.h"
15 #include "lib/axmap.h"
16 #include "rwlock.h"
17 #include "zbd.h"
18
19 #ifdef CONFIG_LINUX_FALLOCATE
20 #include <linux/falloc.h>
21 #endif
22
23 static FLIST_HEAD(filename_list);
24
25 /*
26  * List entry for filename_list
27  */
28 struct file_name {
29         struct flist_head list;
30         char *filename;
31 };
32
33 static inline void clear_error(struct thread_data *td)
34 {
35         td->error = 0;
36         td->verror[0] = '\0';
37 }
38
39 static int native_fallocate(struct thread_data *td, struct fio_file *f)
40 {
41         bool success;
42
43         success = fio_fallocate(f, 0, f->real_file_size);
44         dprint(FD_FILE, "native fallocate of file %s size %llu was "
45                         "%ssuccessful\n", f->file_name,
46                         (unsigned long long) f->real_file_size,
47                         !success ? "un": "");
48
49         if (success)
50                 return false;
51
52         if (errno == ENOSYS)
53                 dprint(FD_FILE, "native fallocate is not implemented\n");
54
55         return true;
56 }
57
58 static void fallocate_file(struct thread_data *td, struct fio_file *f)
59 {
60         if (td->o.fill_device)
61                 return;
62
63         switch (td->o.fallocate_mode) {
64         case FIO_FALLOCATE_NATIVE:
65                 native_fallocate(td, f);
66                 break;
67         case FIO_FALLOCATE_NONE:
68                 break;
69 #ifdef CONFIG_POSIX_FALLOCATE
70         case FIO_FALLOCATE_POSIX: {
71                 int r;
72
73                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
74                                  f->file_name,
75                                  (unsigned long long) f->real_file_size);
76
77                 r = posix_fallocate(f->fd, 0, f->real_file_size);
78                 if (r > 0)
79                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
80                 break;
81                 }
82 #endif /* CONFIG_POSIX_FALLOCATE */
83 #ifdef CONFIG_LINUX_FALLOCATE
84         case FIO_FALLOCATE_KEEP_SIZE: {
85                 int r;
86
87                 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
88                                 "file %s size %llu\n", f->file_name,
89                                 (unsigned long long) f->real_file_size);
90
91                 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
92                 if (r != 0)
93                         td_verror(td, errno, "fallocate");
94
95                 break;
96                 }
97 #endif /* CONFIG_LINUX_FALLOCATE */
98         case FIO_FALLOCATE_TRUNCATE: {
99                 int r;
100
101                 dprint(FD_FILE, "ftruncate file %s size %llu\n",
102                                 f->file_name,
103                                 (unsigned long long) f->real_file_size);
104                 r = ftruncate(f->fd, f->real_file_size);
105                 if (r != 0)
106                         td_verror(td, errno, "ftruncate");
107
108                 break;
109         }
110         default:
111                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
112                 assert(0);
113         }
114 }
115
116 /*
117  * Leaves f->fd open on success, caller must close
118  */
119 static int extend_file(struct thread_data *td, struct fio_file *f)
120 {
121         int new_layout = 0, unlink_file = 0, flags;
122         unsigned long long left;
123         unsigned long long bs;
124         char *b = NULL;
125
126         if (read_only) {
127                 log_err("fio: refusing extend of file due to read-only\n");
128                 return 0;
129         }
130
131         /*
132          * check if we need to lay the file out complete again. fio
133          * does that for operations involving reads, or for writes
134          * where overwrite is set
135          */
136         if (td_read(td) ||
137            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
138             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
139                 new_layout = 1;
140         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
141                 unlink_file = 1;
142
143         if (unlink_file || new_layout) {
144                 int ret;
145
146                 dprint(FD_FILE, "layout %d unlink %d %s\n", new_layout, unlink_file, f->file_name);
147
148                 ret = td_io_unlink_file(td, f);
149                 if (ret != 0 && ret != ENOENT) {
150                         td_verror(td, errno, "unlink");
151                         return 1;
152                 }
153         }
154
155         flags = O_WRONLY;
156         if (td->o.allow_create)
157                 flags |= O_CREAT;
158         if (new_layout)
159                 flags |= O_TRUNC;
160
161 #ifdef WIN32
162         flags |= _O_BINARY;
163 #endif
164
165         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
166         f->fd = open(f->file_name, flags, 0644);
167         if (f->fd < 0) {
168                 int err = errno;
169
170                 if (err == ENOENT && !td->o.allow_create)
171                         log_err("fio: file creation disallowed by "
172                                         "allow_file_create=0\n");
173                 else
174                         td_verror(td, err, "open");
175                 return 1;
176         }
177
178         fallocate_file(td, f);
179
180         /*
181          * If our jobs don't require regular files initially, we're done.
182          */
183         if (!new_layout)
184                 goto done;
185
186         /*
187          * The size will be -1ULL when fill_device is used, so don't truncate
188          * or fallocate this file, just write it
189          */
190         if (!td->o.fill_device) {
191                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
192                                         (unsigned long long) f->real_file_size);
193                 if (ftruncate(f->fd, f->real_file_size) == -1) {
194                         if (errno != EFBIG) {
195                                 td_verror(td, errno, "ftruncate");
196                                 goto err;
197                         }
198                 }
199         }
200
201
202         dprint(FD_FILE, "fill file %s, size %llu\n", f->file_name, (unsigned long long) f->real_file_size);
203
204         left = f->real_file_size;
205         bs = td->o.max_bs[DDIR_WRITE];
206         if (bs > left)
207                 bs = left;
208
209         b = malloc(bs);
210         if (!b) {
211                 td_verror(td, errno, "malloc");
212                 goto err;
213         }
214
215         while (left && !td->terminate) {
216                 ssize_t r;
217
218                 if (bs > left)
219                         bs = left;
220
221                 fill_io_buffer(td, b, bs, bs);
222
223                 r = write(f->fd, b, bs);
224
225                 if (r > 0) {
226                         left -= r;
227                         continue;
228                 } else {
229                         if (r < 0) {
230                                 int __e = errno;
231
232                                 if (__e == ENOSPC || __e == EDQUOT) {
233                                         const char *__e_name;
234                                         if (td->o.fill_device)
235                                                 break;
236                                         if (__e == ENOSPC)
237                                                 __e_name = "ENOSPC";
238                                         else
239                                                 __e_name = "EDQUOT";
240                                         log_info("fio: %s on laying out "
241                                                  "file, stopping\n", __e_name);
242                                 }
243                                 td_verror(td, errno, "write");
244                         } else
245                                 td_verror(td, EIO, "write");
246
247                         goto err;
248                 }
249         }
250
251         if (td->terminate) {
252                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
253                 td_io_unlink_file(td, f);
254         } else if (td->o.create_fsync) {
255                 if (fsync(f->fd) < 0) {
256                         td_verror(td, errno, "fsync");
257                         goto err;
258                 }
259         }
260         if (td->o.fill_device && !td_write(td)) {
261                 fio_file_clear_size_known(f);
262                 if (td_io_get_file_size(td, f))
263                         goto err;
264                 if (f->io_size > f->real_file_size)
265                         f->io_size = f->real_file_size;
266         }
267
268         free(b);
269 done:
270         return 0;
271 err:
272         close(f->fd);
273         f->fd = -1;
274         if (b)
275                 free(b);
276         return 1;
277 }
278
279 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
280 {
281         int r, did_open = 0, old_runstate;
282         unsigned long long left;
283         unsigned long long bs;
284         bool ret = true;
285         char *b;
286
287         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
288             td_ioengine_flagged(td, FIO_NOIO))
289                 return true;
290
291         if (f->filetype == FIO_TYPE_CHAR)
292                 return true;
293
294         if (!fio_file_open(f)) {
295                 if (td->io_ops->open_file(td, f)) {
296                         log_err("fio: cannot pre-read, failed to open file\n");
297                         return false;
298                 }
299                 did_open = 1;
300         }
301
302         old_runstate = td_bump_runstate(td, TD_PRE_READING);
303
304         left = f->io_size;
305         bs = td->o.max_bs[DDIR_READ];
306         if (bs > left)
307                 bs = left;
308
309         b = malloc(bs);
310         if (!b) {
311                 td_verror(td, errno, "malloc");
312                 ret = false;
313                 goto error;
314         }
315         memset(b, 0, bs);
316
317         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
318                 td_verror(td, errno, "lseek");
319                 log_err("fio: failed to lseek pre-read file\n");
320                 ret = false;
321                 goto error;
322         }
323
324         while (left && !td->terminate) {
325                 if (bs > left)
326                         bs = left;
327
328                 r = read(f->fd, b, bs);
329
330                 if (r == (int) bs) {
331                         left -= bs;
332                         continue;
333                 } else {
334                         td_verror(td, EIO, "pre_read");
335                         break;
336                 }
337         }
338
339 error:
340         td_restore_runstate(td, old_runstate);
341
342         if (did_open)
343                 td->io_ops->close_file(td, f);
344
345         free(b);
346         return ret;
347 }
348
349 /*
350  * Generic function to prepopulate regular file with data.
351  * Useful if you want to make sure I/O engine has data to read.
352  * Leaves f->fd open on success, caller must close.
353  */
354 int generic_prepopulate_file(struct thread_data *td, struct fio_file *f)
355 {
356         int flags;
357         unsigned long long left, bs;
358         char *b = NULL;
359
360         /* generic function for regular files only */
361         assert(f->filetype == FIO_TYPE_FILE);
362
363         if (read_only) {
364                 log_err("fio: refusing to write a file due to read-only\n");
365                 return 0;
366         }
367
368         flags = O_WRONLY;
369         if (td->o.allow_create)
370                 flags |= O_CREAT;
371
372 #ifdef WIN32
373         flags |= _O_BINARY;
374 #endif
375
376         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
377         f->fd = open(f->file_name, flags, 0644);
378         if (f->fd < 0) {
379                 int err = errno;
380
381                 if (err == ENOENT && !td->o.allow_create)
382                         log_err("fio: file creation disallowed by "
383                                         "allow_file_create=0\n");
384                 else
385                         td_verror(td, err, "open");
386                 return 1;
387         }
388
389         left = f->real_file_size;
390         bs = td->o.max_bs[DDIR_WRITE];
391         if (bs > left)
392                 bs = left;
393
394         b = malloc(bs);
395         if (!b) {
396                 td_verror(td, errno, "malloc");
397                 goto err;
398         }
399
400         while (left && !td->terminate) {
401                 ssize_t r;
402
403                 if (bs > left)
404                         bs = left;
405
406                 fill_io_buffer(td, b, bs, bs);
407
408                 r = write(f->fd, b, bs);
409
410                 if (r > 0) {
411                         left -= r;
412                 } else {
413                         td_verror(td, errno, "write");
414                         goto err;
415                 }
416         }
417
418         if (td->terminate) {
419                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
420                 td_io_unlink_file(td, f);
421         } else if (td->o.create_fsync) {
422                 if (fsync(f->fd) < 0) {
423                         td_verror(td, errno, "fsync");
424                         goto err;
425                 }
426         }
427
428         free(b);
429         return 0;
430 err:
431         close(f->fd);
432         f->fd = -1;
433         if (b)
434                 free(b);
435         return 1;
436 }
437
438 unsigned long long get_rand_file_size(struct thread_data *td)
439 {
440         unsigned long long ret, sized;
441         uint64_t frand_max;
442         uint64_t r;
443
444         frand_max = rand_max(&td->file_size_state);
445         r = __rand(&td->file_size_state);
446         sized = td->o.file_size_high - td->o.file_size_low;
447         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
448         ret += td->o.file_size_low;
449         ret -= (ret % td->o.rw_min_bs);
450         return ret;
451 }
452
453 static int file_size(struct thread_data *td, struct fio_file *f)
454 {
455         struct stat st;
456
457         if (stat(f->file_name, &st) == -1) {
458                 td_verror(td, errno, "fstat");
459                 return 1;
460         }
461
462         f->real_file_size = st.st_size;
463         return 0;
464 }
465
466 static int bdev_size(struct thread_data *td, struct fio_file *f)
467 {
468         unsigned long long bytes = 0;
469         int r;
470
471         if (td->io_ops->open_file(td, f)) {
472                 log_err("fio: failed opening blockdev %s for size check\n",
473                         f->file_name);
474                 return 1;
475         }
476
477         r = blockdev_size(f, &bytes);
478         if (r) {
479                 td_verror(td, r, "blockdev_size");
480                 goto err;
481         }
482
483         if (!bytes) {
484                 log_err("%s: zero sized block device?\n", f->file_name);
485                 goto err;
486         }
487
488         f->real_file_size = bytes;
489         td->io_ops->close_file(td, f);
490         return 0;
491 err:
492         td->io_ops->close_file(td, f);
493         return 1;
494 }
495
496 static int char_size(struct thread_data *td, struct fio_file *f)
497 {
498 #ifdef FIO_HAVE_CHARDEV_SIZE
499         unsigned long long bytes = 0;
500         int r;
501
502         if (td->io_ops->open_file(td, f)) {
503                 log_err("fio: failed opening chardev %s for size check\n",
504                         f->file_name);
505                 return 1;
506         }
507
508         r = chardev_size(f, &bytes);
509         if (r) {
510                 td_verror(td, r, "chardev_size");
511                 goto err;
512         }
513
514         if (!bytes) {
515                 log_err("%s: zero sized char device?\n", f->file_name);
516                 goto err;
517         }
518
519         f->real_file_size = bytes;
520         td->io_ops->close_file(td, f);
521         return 0;
522 err:
523         td->io_ops->close_file(td, f);
524         return 1;
525 #else
526         f->real_file_size = -1ULL;
527         return 0;
528 #endif
529 }
530
531 static int get_file_size(struct thread_data *td, struct fio_file *f)
532 {
533         int ret = 0;
534
535         if (fio_file_size_known(f))
536                 return 0;
537
538         if (f->filetype == FIO_TYPE_FILE)
539                 ret = file_size(td, f);
540         else if (f->filetype == FIO_TYPE_BLOCK)
541                 ret = bdev_size(td, f);
542         else if (f->filetype == FIO_TYPE_CHAR)
543                 ret = char_size(td, f);
544         else {
545                 f->real_file_size = -1;
546                 log_info("%s: failed to get file size of %s\n", td->o.name,
547                                         f->file_name);
548                 return 1; /* avoid offset extends end error message */
549         }
550
551         /*
552          * Leave ->real_file_size with 0 since it could be expectation
553          * of initial setup for regular files.
554          */
555         if (ret)
556                 return ret;
557
558         /*
559          * ->file_offset normally hasn't been initialized yet, so this
560          * is basically always false unless ->real_file_size is -1, but
561          * if ->real_file_size is -1 this message doesn't make sense.
562          * As a result, this message is basically useless.
563          */
564         if (f->file_offset > f->real_file_size) {
565                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
566                                         (unsigned long long) f->file_offset,
567                                         (unsigned long long) f->real_file_size);
568                 return 1;
569         }
570
571         fio_file_set_size_known(f);
572         return 0;
573 }
574
575 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
576                                    unsigned long long off,
577                                    unsigned long long len)
578 {
579         int errval = 0, ret = 0;
580
581 #ifdef CONFIG_ESX
582         return 0;
583 #endif
584
585         if (len == -1ULL)
586                 len = f->io_size;
587         if (off == -1ULL)
588                 off = f->file_offset;
589
590         if (len == -1ULL || off == -1ULL)
591                 return 0;
592
593         if (td->io_ops->invalidate) {
594                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
595                         f->file_name);
596                 ret = td->io_ops->invalidate(td, f);
597                 if (ret < 0)
598                         errval = -ret;
599         } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
600                 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
601                        td->io_ops->name);
602         } else if (f->filetype == FIO_TYPE_FILE) {
603                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
604                         f->file_name, off, len);
605                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
606                 if (ret)
607                         errval = ret;
608         } else if (f->filetype == FIO_TYPE_BLOCK) {
609                 int retry_count = 0;
610
611                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
612                 ret = blockdev_invalidate_cache(f);
613                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
614                         /*
615                          * Linux multipath devices reject ioctl while
616                          * the maps are being updated. That window can
617                          * last tens of milliseconds; we'll try up to
618                          * a quarter of a second.
619                          */
620                         usleep(10000);
621                         ret = blockdev_invalidate_cache(f);
622                 }
623                 if (ret < 0 && errno == EACCES && geteuid()) {
624                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
625                                 log_err("fio: only root may flush block "
626                                         "devices. Cache flush bypassed!\n");
627                         }
628                 }
629                 if (ret < 0)
630                         errval = errno;
631         } else if (f->filetype == FIO_TYPE_CHAR ||
632                    f->filetype == FIO_TYPE_PIPE) {
633                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
634         }
635
636         /*
637          * Cache flushing isn't a fatal condition, and we know it will
638          * happen on some platforms where we don't have the proper
639          * function to flush eg block device caches. So just warn and
640          * continue on our way.
641          */
642         if (errval)
643                 log_info("fio: cache invalidation of %s failed: %s\n",
644                          f->file_name, strerror(errval));
645
646         return 0;
647
648 }
649
650 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
651 {
652         if (!fio_file_open(f))
653                 return 0;
654
655         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
656 }
657
658 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
659 {
660         int ret = 0;
661
662         dprint(FD_FILE, "fd close %s\n", f->file_name);
663
664         remove_file_hash(f);
665
666         if (close(f->fd) < 0)
667                 ret = errno;
668
669         f->fd = -1;
670
671         if (f->shadow_fd != -1) {
672                 close(f->shadow_fd);
673                 f->shadow_fd = -1;
674         }
675
676         f->engine_pos = 0;
677         return ret;
678 }
679
680 int file_lookup_open(struct fio_file *f, int flags)
681 {
682         struct fio_file *__f;
683         int from_hash;
684
685         __f = lookup_file_hash(f->file_name);
686         if (__f) {
687                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
688                 f->lock = __f->lock;
689                 from_hash = 1;
690         } else {
691                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
692                 from_hash = 0;
693         }
694
695 #ifdef WIN32
696         flags |= _O_BINARY;
697 #endif
698
699         f->fd = open(f->file_name, flags, 0600);
700         return from_hash;
701 }
702
703 static int file_close_shadow_fds(struct thread_data *td)
704 {
705         struct fio_file *f;
706         int num_closed = 0;
707         unsigned int i;
708
709         for_each_file(td, f, i) {
710                 if (f->shadow_fd == -1)
711                         continue;
712
713                 close(f->shadow_fd);
714                 f->shadow_fd = -1;
715                 num_closed++;
716         }
717
718         return num_closed;
719 }
720
721 int generic_open_file(struct thread_data *td, struct fio_file *f)
722 {
723         int is_std = 0;
724         int flags = 0;
725         int from_hash = 0;
726
727         dprint(FD_FILE, "fd open %s\n", f->file_name);
728
729         if (!strcmp(f->file_name, "-")) {
730                 if (td_rw(td)) {
731                         log_err("fio: can't read/write to stdin/out\n");
732                         return 1;
733                 }
734                 is_std = 1;
735
736                 /*
737                  * move output logging to stderr, if we are writing to stdout
738                  */
739                 if (td_write(td))
740                         f_out = stderr;
741         }
742
743         if (td_trim(td))
744                 goto skip_flags;
745         if (td->o.odirect)
746                 flags |= OS_O_DIRECT;
747         if (td->o.oatomic) {
748                 if (!FIO_O_ATOMIC) {
749                         td_verror(td, EINVAL, "OS does not support atomic IO");
750                         return 1;
751                 }
752                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
753         }
754         flags |= td->o.sync_io;
755         if (td->o.create_on_open && td->o.allow_create)
756                 flags |= O_CREAT;
757 skip_flags:
758         if (f->filetype != FIO_TYPE_FILE)
759                 flags |= FIO_O_NOATIME;
760
761 open_again:
762         if (td_write(td)) {
763                 if (!read_only)
764                         flags |= O_RDWR;
765
766                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
767                         flags |= O_CREAT;
768
769                 if (is_std)
770                         f->fd = dup(STDOUT_FILENO);
771                 else
772                         from_hash = file_lookup_open(f, flags);
773         } else if (td_read(td)) {
774                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
775                         flags |= O_RDWR;
776                 else
777                         flags |= O_RDONLY;
778
779                 if (is_std)
780                         f->fd = dup(STDIN_FILENO);
781                 else
782                         from_hash = file_lookup_open(f, flags);
783         } else if (td_trim(td)) {
784                 assert(!td_rw(td)); /* should have matched above */
785                 if (!read_only)
786                         flags |= O_RDWR;
787                 from_hash = file_lookup_open(f, flags);
788         }
789
790         if (f->fd == -1) {
791                 char buf[FIO_VERROR_SIZE];
792                 int __e = errno;
793
794                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
795                         flags &= ~FIO_O_NOATIME;
796                         goto open_again;
797                 }
798                 if (__e == EMFILE && file_close_shadow_fds(td))
799                         goto open_again;
800
801                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
802
803                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
804                         log_err("fio: looks like your file system does not " \
805                                 "support direct=1/buffered=0\n");
806                 }
807
808                 td_verror(td, __e, buf);
809                 return 1;
810         }
811
812         if (!from_hash && f->fd != -1) {
813                 if (add_file_hash(f)) {
814                         int fio_unused ret;
815
816                         /*
817                          * Stash away descriptor for later close. This is to
818                          * work-around a "feature" on Linux, where a close of
819                          * an fd that has been opened for write will trigger
820                          * udev to call blkid to check partitions, fs id, etc.
821                          * That pollutes the device cache, which can slow down
822                          * unbuffered accesses.
823                          */
824                         if (f->shadow_fd == -1)
825                                 f->shadow_fd = f->fd;
826                         else {
827                                 /*
828                                  * OK to ignore, we haven't done anything
829                                  * with it
830                                  */
831                                 ret = generic_close_file(td, f);
832                         }
833                         goto open_again;
834                 }
835         }
836
837         return 0;
838 }
839
840 /*
841  * This function i.e. get_file_size() is the default .get_file_size
842  * implementation of majority of I/O engines.
843  */
844 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
845 {
846         return get_file_size(td, f);
847 }
848
849 /*
850  * open/close all files, so that ->real_file_size gets set
851  */
852 static int get_file_sizes(struct thread_data *td)
853 {
854         struct fio_file *f;
855         unsigned int i;
856         int err = 0;
857
858         for_each_file(td, f, i) {
859                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
860                                                                 f->file_name);
861
862                 if (td_io_get_file_size(td, f)) {
863                         if (td->error != ENOENT) {
864                                 log_err("%s\n", td->verror);
865                                 err = 1;
866                                 break;
867                         }
868                         clear_error(td);
869                 }
870
871                 /*
872                  * There are corner cases where we end up with -1 for
873                  * ->real_file_size due to unsupported file type, etc.
874                  * We then just set to size option value divided by number
875                  * of files, similar to the way file ->io_size is set.
876                  * stat(2) failure doesn't set ->real_file_size to -1.
877                  */
878                 if (f->real_file_size == -1ULL && td->o.size)
879                         f->real_file_size = td->o.size / td->o.nr_files;
880         }
881
882         return err;
883 }
884
885 struct fio_mount {
886         struct flist_head list;
887         const char *base;
888         char __base[256];
889         unsigned int key;
890 };
891
892 /*
893  * Get free number of bytes for each file on each unique mount.
894  */
895 static unsigned long long get_fs_free_counts(struct thread_data *td)
896 {
897         struct flist_head *n, *tmp;
898         unsigned long long ret = 0;
899         struct fio_mount *fm;
900         FLIST_HEAD(list);
901         struct fio_file *f;
902         unsigned int i;
903
904         for_each_file(td, f, i) {
905                 struct stat sb;
906                 char buf[256];
907
908                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
909                         if (f->real_file_size != -1ULL)
910                                 ret += f->real_file_size;
911                         continue;
912                 } else if (f->filetype != FIO_TYPE_FILE)
913                         continue;
914
915                 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
916
917                 if (stat(buf, &sb) < 0) {
918                         if (errno != ENOENT)
919                                 break;
920                         strcpy(buf, ".");
921                         if (stat(buf, &sb) < 0)
922                                 break;
923                 }
924
925                 fm = NULL;
926                 flist_for_each(n, &list) {
927                         fm = flist_entry(n, struct fio_mount, list);
928                         if (fm->key == sb.st_dev)
929                                 break;
930
931                         fm = NULL;
932                 }
933
934                 if (fm)
935                         continue;
936
937                 fm = calloc(1, sizeof(*fm));
938                 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
939                 fm->base = basename(fm->__base);
940                 fm->key = sb.st_dev;
941                 flist_add(&fm->list, &list);
942         }
943
944         flist_for_each_safe(n, tmp, &list) {
945                 unsigned long long sz;
946
947                 fm = flist_entry(n, struct fio_mount, list);
948                 flist_del(&fm->list);
949
950                 sz = get_fs_free_size(fm->base);
951                 if (sz && sz != -1ULL)
952                         ret += sz;
953
954                 free(fm);
955         }
956
957         return ret;
958 }
959
960 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
961 {
962         bool align = false;
963         struct thread_options *o = &td->o;
964         unsigned long long align_bs;
965         unsigned long long offset;
966         unsigned long long increment;
967
968         if (o->file_append && f->filetype == FIO_TYPE_FILE)
969                 return f->real_file_size;
970
971         if (o->offset_increment_percent) {
972                 assert(!o->offset_increment);
973                 increment = o->offset_increment_percent * f->real_file_size / 100;
974                 align = true;
975         } else
976                 increment = o->offset_increment;
977
978         if (o->start_offset_percent > 0) {
979                 /* calculate the raw offset */
980                 offset = (f->real_file_size * o->start_offset_percent / 100) +
981                         (td->subjob_number * increment);
982
983                 align = true;
984         } else {
985                 /* start_offset_percent not set */
986                 offset = o->start_offset +
987                                 td->subjob_number * increment;
988         }
989
990         if (align) {
991                 /*
992                  * if offset_align is provided, use it
993                  */
994                 if (fio_option_is_set(o, start_offset_align)) {
995                         align_bs = o->start_offset_align;
996                 } else {
997                         /* else take the minimum block size */
998                         align_bs = td_min_bs(td);
999                 }
1000
1001                 /*
1002                  * block align the offset at the next available boundary at
1003                  * ceiling(offset / align_bs) * align_bs
1004                  */
1005                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
1006         }
1007
1008         return offset;
1009 }
1010
1011 /*
1012  * Find longest path component that exists and return its length
1013  */
1014 int longest_existing_path(char *path) {
1015         char buf[PATH_MAX];
1016         bool done;
1017         char *buf_pos;
1018         int offset;
1019 #ifdef WIN32
1020         DWORD dwAttr;
1021 #else
1022         struct stat sb;
1023 #endif
1024
1025         sprintf(buf, "%s", path);
1026         done = false;
1027         while (!done) {
1028                 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1029                 if (!buf_pos) {
1030                         offset = 0;
1031                         break;
1032                 }
1033
1034                 *(buf_pos + 1) = '\0';
1035
1036 #ifdef WIN32
1037                 dwAttr = GetFileAttributesA(buf);
1038                 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1039                         done = true;
1040                 }
1041 #else
1042                 if (stat(buf, &sb) == 0)
1043                         done = true;
1044 #endif
1045                 if (done)
1046                         offset = buf_pos - buf;
1047                 else
1048                         *buf_pos = '\0';
1049         }
1050
1051         return offset;
1052 }
1053
1054 static bool create_work_dirs(struct thread_data *td, const char *fname)
1055 {
1056         char path[PATH_MAX];
1057         char *start, *end;
1058         int offset;
1059
1060         snprintf(path, PATH_MAX, "%s", fname);
1061         start = path;
1062
1063         offset = longest_existing_path(path);
1064         end = start + offset;
1065         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1066                 if (end == start) {
1067                         end++;
1068                         continue;
1069                 }
1070                 *end = '\0';
1071                 errno = 0;
1072                 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1073                         log_err("fio: failed to create dir (%s): %s\n",
1074                                 start, strerror(errno));
1075                         return false;
1076                 }
1077                 *end = FIO_OS_PATH_SEPARATOR;
1078                 end++;
1079         }
1080         td->flags |= TD_F_DIRS_CREATED;
1081         return true;
1082 }
1083
1084 int setup_shared_file(struct thread_data *td)
1085 {
1086         struct fio_file *f;
1087         uint64_t file_size;
1088         int err = 0;
1089
1090         if (td->o.nr_files > 1) {
1091                 log_err("fio: shared file setup called for multiple files\n");
1092                 return -1;
1093         }
1094
1095         get_file_sizes(td);
1096
1097         f = td->files[0];
1098
1099         if (f == NULL) {
1100                 log_err("fio: NULL shared file\n");
1101                 return -1;
1102         }
1103
1104         file_size = thread_number * td->o.size;
1105         dprint(FD_FILE, "shared setup %s real_file_size=%llu, desired=%llu\n", 
1106                         f->file_name, (unsigned long long)f->real_file_size, (unsigned long long)file_size);
1107
1108         if (f->real_file_size < file_size) {
1109                 dprint(FD_FILE, "fio: extending shared file\n");
1110                 f->real_file_size = file_size;
1111                 err = extend_file(td, f);
1112                 if (!err) {
1113                         err = __file_invalidate_cache(td, f, 0, f->real_file_size);
1114                 }
1115                 get_file_sizes(td);
1116                 dprint(FD_FILE, "shared setup new real_file_size=%llu\n", 
1117                                 (unsigned long long)f->real_file_size);
1118         }
1119
1120         return err;
1121 }
1122
1123 /*
1124  * Open the files and setup files sizes, creating files if necessary.
1125  */
1126 int setup_files(struct thread_data *td)
1127 {
1128         unsigned long long total_size, extend_size;
1129         struct thread_options *o = &td->o;
1130         struct fio_file *f;
1131         unsigned int i, nr_fs_extra = 0;
1132         int err = 0, need_extend;
1133         int old_state;
1134         const unsigned long long bs = td_min_bs(td);
1135         uint64_t fs = 0;
1136
1137         dprint(FD_FILE, "setup files (thread_number=%d, subjob_number=%d)\n", td->thread_number, td->subjob_number);
1138
1139         old_state = td_bump_runstate(td, TD_SETTING_UP);
1140
1141         for_each_file(td, f, i) {
1142                 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1143                     strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1144                     !(td->flags & TD_F_DIRS_CREATED) &&
1145                     !create_work_dirs(td, f->file_name))
1146                         goto err_out;
1147         }
1148
1149         /*
1150          * Find out physical size of files or devices for this thread,
1151          * before we determine I/O size and range of our targets.
1152          * If ioengine defines a setup() method, it's responsible for
1153          * opening the files and setting f->real_file_size to indicate
1154          * the valid range for that file.
1155          */
1156         if (td->io_ops->setup)
1157                 err = td->io_ops->setup(td);
1158         else
1159                 err = get_file_sizes(td);
1160
1161         if (err)
1162                 goto err_out;
1163
1164         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1165                 err = zbd_init_files(td);
1166                 if (err)
1167                         goto err_out;
1168         }
1169         zbd_recalc_options_with_zone_granularity(td);
1170
1171         if (o->read_iolog_file)
1172                 goto done;
1173
1174         /*
1175          * check sizes. if the files/devices do not exist and the size
1176          * isn't passed to fio, abort.
1177          */
1178         total_size = 0;
1179         for_each_file(td, f, i) {
1180                 f->fileno = i;
1181                 if (f->real_file_size == -1ULL)
1182                         total_size = -1ULL;
1183                 else
1184                         total_size += f->real_file_size;
1185         }
1186
1187         if (o->fill_device)
1188                 td->fill_device_size = get_fs_free_counts(td);
1189
1190         /*
1191          * device/file sizes are zero and no size given, punt
1192          */
1193         if ((!total_size || total_size == -1ULL) && !o->size &&
1194             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1195             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1196                 log_err("%s: you need to specify size=\n", o->name);
1197                 td_verror(td, EINVAL, "total_file_size");
1198                 goto err_out;
1199         }
1200
1201         /*
1202          * Calculate per-file size and potential extra size for the
1203          * first files, if needed (i.e. if we don't have a fixed size).
1204          */
1205         if (!o->file_size_low && o->nr_files) {
1206                 uint64_t all_fs;
1207
1208                 fs = o->size / o->nr_files;
1209                 all_fs = fs * o->nr_files;
1210
1211                 if (all_fs < o->size)
1212                         nr_fs_extra = (o->size - all_fs) / bs;
1213         }
1214
1215         /*
1216          * now file sizes are known, so we can set ->io_size. if size= is
1217          * not given, ->io_size is just equal to ->real_file_size. if size
1218          * is given, ->io_size is size / nr_files.
1219          */
1220         extend_size = total_size = 0;
1221         need_extend = 0;
1222         for_each_file(td, f, i) {
1223                 f->file_offset = get_start_offset(td, f);
1224
1225                 /*
1226                  * Update ->io_size depending on options specified.
1227                  * ->file_size_low being 0 means filesize option isn't set.
1228                  * Non zero ->file_size_low equals ->file_size_high means
1229                  * filesize option is set in a fixed size format.
1230                  * Non zero ->file_size_low not equals ->file_size_high means
1231                  * filesize option is set in a range format.
1232                  */
1233                 if (!o->file_size_low) {
1234                         /*
1235                          * no file size or range given, file size is equal to
1236                          * total size divided by number of files. If the size
1237                          * doesn't divide nicely with the min blocksize,
1238                          * make the first files bigger.
1239                          */
1240                         f->io_size = fs;
1241                         if (nr_fs_extra) {
1242                                 nr_fs_extra--;
1243                                 f->io_size += bs;
1244                         }
1245
1246                         /*
1247                          * We normally don't come here for regular files, but
1248                          * if the result is 0 for a regular file, set it to the
1249                          * real file size. This could be size of the existing
1250                          * one if it already exists, but otherwise will be set
1251                          * to 0. A new file won't be created because
1252                          * ->io_size + ->file_offset equals ->real_file_size.
1253                          */
1254                         if (!f->io_size) {
1255                                 if (f->file_offset > f->real_file_size)
1256                                         goto err_offset;
1257                                 f->io_size = f->real_file_size - f->file_offset;
1258                                 if (!f->io_size)
1259                                         log_info("fio: file %s may be ignored\n",
1260                                                 f->file_name);
1261                         }
1262                 } else if (f->real_file_size < o->file_size_low ||
1263                            f->real_file_size > o->file_size_high) {
1264                         if (f->file_offset > o->file_size_low)
1265                                 goto err_offset;
1266                         /*
1267                          * file size given. if it's fixed, use that. if it's a
1268                          * range, generate a random size in-between.
1269                          */
1270                         if (o->file_size_low == o->file_size_high)
1271                                 f->io_size = o->file_size_low - f->file_offset;
1272                         else {
1273                                 f->io_size = get_rand_file_size(td)
1274                                                 - f->file_offset;
1275                         }
1276                 } else
1277                         f->io_size = f->real_file_size - f->file_offset;
1278
1279                 if (f->io_size == -1ULL)
1280                         total_size = -1ULL;
1281                 else {
1282                         uint64_t io_size;
1283
1284                         if (o->size_percent && o->size_percent != 100) {
1285                                 uint64_t file_size;
1286
1287                                 file_size = f->io_size + f->file_offset;
1288                                 f->io_size = (file_size *
1289                                               o->size_percent) / 100;
1290                                 if (f->io_size > (file_size - f->file_offset))
1291                                         f->io_size = file_size - f->file_offset;
1292
1293                                 f->io_size -= (f->io_size % td_min_bs(td));
1294                         }
1295
1296                         io_size = f->io_size;
1297                         if (o->io_size_percent && o->io_size_percent != 100) {
1298                                 io_size *= o->io_size_percent;
1299                                 io_size /= 100;
1300                         }
1301
1302                         total_size += io_size;
1303                 }
1304
1305                 if (f->filetype == FIO_TYPE_FILE &&
1306                     (f->io_size + f->file_offset) > f->real_file_size) {
1307                         if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1308                             !o->create_on_open) {
1309                                 need_extend++;
1310                                 extend_size += (f->io_size + f->file_offset);
1311                                 fio_file_set_extend(f);
1312                         } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1313                                    (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1314                                     td_ioengine_flagged(td, FIO_FAKEIO)))
1315                                 f->real_file_size = f->io_size + f->file_offset;
1316                 }
1317         }
1318
1319         if (td->o.block_error_hist) {
1320                 int len;
1321
1322                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1323                 f = td->files[0];
1324                 len = f->io_size / td->o.bs[DDIR_TRIM];
1325                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1326                         log_err("fio: cannot calculate block histogram with "
1327                                 "%d trim blocks, maximum %d\n",
1328                                 len, MAX_NR_BLOCK_INFOS);
1329                         td_verror(td, EINVAL, "block_error_hist");
1330                         goto err_out;
1331                 }
1332
1333                 td->ts.nr_block_infos = len;
1334                 for (i = 0; i < len; i++)
1335                         td->ts.block_infos[i] =
1336                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1337         } else
1338                 td->ts.nr_block_infos = 0;
1339
1340         if (!o->size || (total_size && o->size > total_size))
1341                 o->size = total_size;
1342
1343         if (o->size < td_min_bs(td)) {
1344                 log_err("fio: blocksize is larger than data set range\n");
1345                 goto err_out;
1346         }
1347
1348         /*
1349          * See if we need to extend some files, typically needed when our
1350          * target regular files don't exist yet, but our jobs require them
1351          * initially due to read I/Os.
1352          */
1353         if (need_extend) {
1354                 temp_stall_ts = 1;
1355                 if (output_format & FIO_OUTPUT_NORMAL) {
1356                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1357                                  o->name,
1358                                  need_extend > 1 ? "s" : "",
1359                                  need_extend,
1360                                  need_extend > 1 ? "s" : "",
1361                                  need_extend > 1 ? "total " : "",
1362                                  extend_size >> 20);
1363                 }
1364
1365                 for_each_file(td, f, i) {
1366                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1367
1368                         if (!fio_file_extend(f))
1369                                 continue;
1370
1371                         assert(f->filetype == FIO_TYPE_FILE);
1372                         fio_file_clear_extend(f);
1373                         if (!o->fill_device) {
1374                                 old_len = f->real_file_size;
1375                                 extend_len = f->io_size + f->file_offset -
1376                                                 old_len;
1377                         }
1378                         f->real_file_size = (f->io_size + f->file_offset);
1379                         err = extend_file(td, f);
1380                         if (err)
1381                                 break;
1382
1383                         err = __file_invalidate_cache(td, f, old_len,
1384                                                                 extend_len);
1385
1386                         /*
1387                          * Shut up static checker
1388                          */
1389                         if (f->fd != -1)
1390                                 close(f->fd);
1391
1392                         f->fd = -1;
1393                         if (err)
1394                                 break;
1395                 }
1396                 temp_stall_ts = 0;
1397         }
1398
1399         if (err)
1400                 goto err_out;
1401
1402         /*
1403          * Prepopulate files with data. It might be expected to read some
1404          * "real" data instead of zero'ed files (if no writes to file occurred
1405          * prior to a read job). Engine has to provide a way to do that.
1406          */
1407         if (td->io_ops->prepopulate_file) {
1408                 temp_stall_ts = 1;
1409
1410                 for_each_file(td, f, i) {
1411                         if (output_format & FIO_OUTPUT_NORMAL) {
1412                                 log_info("%s: Prepopulating IO file (%s)\n",
1413                                                         o->name, f->file_name);
1414                         }
1415
1416                         err = td->io_ops->prepopulate_file(td, f);
1417                         if (err)
1418                                 break;
1419
1420                         err = __file_invalidate_cache(td, f, f->file_offset,
1421                                                                 f->io_size);
1422
1423                         /*
1424                          * Shut up static checker
1425                          */
1426                         if (f->fd != -1)
1427                                 close(f->fd);
1428
1429                         f->fd = -1;
1430                         if (err)
1431                                 break;
1432                 }
1433                 temp_stall_ts = 0;
1434         }
1435
1436         if (err)
1437                 goto err_out;
1438
1439         /*
1440          * iolog already set the total io size, if we read back
1441          * stored entries.
1442          */
1443         if (!o->read_iolog_file) {
1444                 if (o->io_size)
1445                         td->total_io_size = o->io_size * o->loops;
1446                 else
1447                         td->total_io_size = o->size * o->loops;
1448         }
1449
1450 done:
1451         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1452                 err = zbd_setup_files(td);
1453                 if (err)
1454                         goto err_out;
1455         }
1456
1457         if (o->create_only)
1458                 td->done = 1;
1459
1460         td_restore_runstate(td, old_state);
1461
1462         return 0;
1463
1464 err_offset:
1465         log_err("%s: you need to specify valid offset=\n", o->name);
1466 err_out:
1467         td_restore_runstate(td, old_state);
1468         return 1;
1469 }
1470
1471 bool pre_read_files(struct thread_data *td)
1472 {
1473         struct fio_file *f;
1474         unsigned int i;
1475
1476         dprint(FD_FILE, "pre_read files\n");
1477
1478         for_each_file(td, f, i) {
1479                 if (!pre_read_file(td, f))
1480                         return false;
1481         }
1482
1483         return true;
1484 }
1485
1486 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1487 {
1488         unsigned int range_size, seed;
1489         uint64_t nranges;
1490         uint64_t fsize;
1491
1492         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1493         fsize = min(f->real_file_size, f->io_size);
1494
1495         nranges = (fsize + range_size - 1ULL) / range_size;
1496
1497         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1498         if (!td->o.rand_repeatable)
1499                 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1500
1501         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1502                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1503         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1504                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1505         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1506                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1507 }
1508
1509 static bool init_rand_distribution(struct thread_data *td)
1510 {
1511         struct fio_file *f;
1512         unsigned int i;
1513         int state;
1514
1515         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1516             td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1517             td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1518                 return false;
1519
1520         state = td_bump_runstate(td, TD_SETTING_UP);
1521
1522         for_each_file(td, f, i)
1523                 __init_rand_distribution(td, f);
1524
1525         td_restore_runstate(td, state);
1526         return true;
1527 }
1528
1529 /*
1530  * Check if the number of blocks exceeds the randomness capability of
1531  * the selected generator. Tausworthe is 32-bit, the others are fully
1532  * 64-bit capable.
1533  */
1534 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1535                                  uint64_t blocks)
1536 {
1537         if (blocks <= FRAND32_MAX)
1538                 return 0;
1539         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1540                 return 0;
1541
1542         /*
1543          * If the user hasn't specified a random generator, switch
1544          * to tausworthe64 with informational warning. If the user did
1545          * specify one, just warn.
1546          */
1547         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1548                         f->file_name);
1549
1550         if (!fio_option_is_set(&td->o, random_generator)) {
1551                 log_info("fio: Switching to tausworthe64. Use the "
1552                          "random_generator= option to get rid of this "
1553                          "warning.\n");
1554                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1555                 return 0;
1556         }
1557
1558         /*
1559          * Just make this information to avoid breaking scripts.
1560          */
1561         log_info("fio: Use the random_generator= option to switch to lfsr or "
1562                          "tausworthe64.\n");
1563         return 0;
1564 }
1565
1566 bool init_random_map(struct thread_data *td)
1567 {
1568         unsigned long long blocks;
1569         struct fio_file *f;
1570         unsigned int i;
1571
1572         if (init_rand_distribution(td))
1573                 return true;
1574         if (!td_random(td))
1575                 return true;
1576
1577         for_each_file(td, f, i) {
1578                 uint64_t fsize = min(f->real_file_size, f->io_size);
1579
1580                 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1581                         fsize = td->o.zone_range;
1582
1583                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1584
1585                 if (check_rand_gen_limits(td, f, blocks))
1586                         return false;
1587
1588                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1589                         uint64_t seed;
1590
1591                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1592
1593                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1594                                 fio_file_set_lfsr(f);
1595                                 continue;
1596                         } else {
1597                                 log_err("fio: failed initializing LFSR\n");
1598                                 return false;
1599                         }
1600                 } else if (!td->o.norandommap) {
1601                         f->io_axmap = axmap_new(blocks);
1602                         if (f->io_axmap) {
1603                                 fio_file_set_axmap(f);
1604                                 continue;
1605                         }
1606                 } else if (td->o.norandommap)
1607                         continue;
1608
1609                 if (!td->o.softrandommap) {
1610                         log_err("fio: failed allocating random map. If running"
1611                                 " a large number of jobs, try the 'norandommap'"
1612                                 " option or set 'softrandommap'. Or give"
1613                                 " a larger --alloc-size to fio.\n");
1614                         return false;
1615                 }
1616
1617                 log_info("fio: file %s failed allocating random map. Running "
1618                          "job without.\n", f->file_name);
1619         }
1620
1621         return true;
1622 }
1623
1624 void close_files(struct thread_data *td)
1625 {
1626         struct fio_file *f;
1627         unsigned int i;
1628
1629         for_each_file(td, f, i) {
1630                 if (fio_file_open(f))
1631                         td_io_close_file(td, f);
1632         }
1633 }
1634
1635 void fio_file_free(struct fio_file *f)
1636 {
1637         if (fio_file_axmap(f))
1638                 axmap_free(f->io_axmap);
1639         if (!fio_file_smalloc(f)) {
1640                 free(f->file_name);
1641                 free(f);
1642         } else {
1643                 sfree(f->file_name);
1644                 sfree(f);
1645         }
1646 }
1647
1648 void close_and_free_files(struct thread_data *td)
1649 {
1650         struct fio_file *f;
1651         unsigned int i;
1652
1653         dprint(FD_FILE, "close files\n");
1654
1655         for_each_file(td, f, i) {
1656                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1657                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1658                         td_io_unlink_file(td, f);
1659                 }
1660
1661                 if (fio_file_open(f))
1662                         td_io_close_file(td, f);
1663
1664                 remove_file_hash(f);
1665
1666                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1667                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1668                         td_io_unlink_file(td, f);
1669                 }
1670
1671                 zbd_close_file(f);
1672                 fio_file_free(f);
1673         }
1674
1675         td->o.filename = NULL;
1676         free(td->files);
1677         free(td->file_locks);
1678         td->files_index = 0;
1679         td->files = NULL;
1680         td->file_locks = NULL;
1681         td->o.file_lock_mode = FILE_LOCK_NONE;
1682         td->o.nr_files = 0;
1683 }
1684
1685 static void get_file_type(struct fio_file *f)
1686 {
1687         struct stat sb;
1688
1689         if (!strcmp(f->file_name, "-"))
1690                 f->filetype = FIO_TYPE_PIPE;
1691         else
1692                 f->filetype = FIO_TYPE_FILE;
1693
1694 #ifdef WIN32
1695         /* \\.\ is the device namespace in Windows, where every file is
1696          * a block device */
1697         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1698                 f->filetype = FIO_TYPE_BLOCK;
1699 #endif
1700
1701         if (!stat(f->file_name, &sb)) {
1702                 if (S_ISBLK(sb.st_mode))
1703                         f->filetype = FIO_TYPE_BLOCK;
1704                 else if (S_ISCHR(sb.st_mode))
1705                         f->filetype = FIO_TYPE_CHAR;
1706                 else if (S_ISFIFO(sb.st_mode))
1707                         f->filetype = FIO_TYPE_PIPE;
1708         }
1709 }
1710
1711 static bool __is_already_allocated(const char *fname, bool set)
1712 {
1713         struct flist_head *entry;
1714         bool ret;
1715
1716         ret = file_bloom_exists(fname, set);
1717         if (!ret)
1718                 return ret;
1719
1720         flist_for_each(entry, &filename_list) {
1721                 struct file_name *fn;
1722
1723                 fn = flist_entry(entry, struct file_name, list);
1724
1725                 if (!strcmp(fn->filename, fname))
1726                         return true;
1727         }
1728
1729         return false;
1730 }
1731
1732 static bool is_already_allocated(const char *fname)
1733 {
1734         bool ret;
1735
1736         fio_file_hash_lock();
1737         ret = __is_already_allocated(fname, false);
1738         fio_file_hash_unlock();
1739
1740         return ret;
1741 }
1742
1743 static void set_already_allocated(const char *fname)
1744 {
1745         struct file_name *fn;
1746
1747         fn = malloc(sizeof(struct file_name));
1748         fn->filename = strdup(fname);
1749
1750         fio_file_hash_lock();
1751         if (!__is_already_allocated(fname, true)) {
1752                 flist_add_tail(&fn->list, &filename_list);
1753                 fn = NULL;
1754         }
1755         fio_file_hash_unlock();
1756
1757         if (fn) {
1758                 free(fn->filename);
1759                 free(fn);
1760         }
1761 }
1762
1763 static void free_already_allocated(void)
1764 {
1765         struct flist_head *entry, *tmp;
1766         struct file_name *fn;
1767
1768         if (flist_empty(&filename_list))
1769                 return;
1770
1771         fio_file_hash_lock();
1772         flist_for_each_safe(entry, tmp, &filename_list) {
1773                 fn = flist_entry(entry, struct file_name, list);
1774                 free(fn->filename);
1775                 flist_del(&fn->list);
1776                 free(fn);
1777         }
1778
1779         fio_file_hash_unlock();
1780 }
1781
1782 static struct fio_file *alloc_new_file(struct thread_data *td)
1783 {
1784         struct fio_file *f;
1785
1786         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1787                 f = calloc(1, sizeof(*f));
1788         else
1789                 f = scalloc(1, sizeof(*f));
1790         if (!f) {
1791                 assert(0);
1792                 return NULL;
1793         }
1794
1795         f->fd = -1;
1796         f->shadow_fd = -1;
1797         fio_file_reset(td, f);
1798         if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1799                 fio_file_set_smalloc(f);
1800         return f;
1801 }
1802
1803 bool exists_and_not_regfile(const char *filename)
1804 {
1805         struct stat sb;
1806
1807         if (lstat(filename, &sb) == -1)
1808                 return false;
1809
1810 #ifndef WIN32 /* NOT Windows */
1811         if (S_ISREG(sb.st_mode))
1812                 return false;
1813 #else
1814         /* \\.\ is the device namespace in Windows, where every file
1815          * is a device node */
1816         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1817                 return false;
1818 #endif
1819
1820         return true;
1821 }
1822
1823 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1824 {
1825         int cur_files = td->files_index;
1826         char file_name[PATH_MAX];
1827         struct fio_file *f;
1828         int len = 0;
1829
1830         dprint(FD_FILE, "add file %s\n", fname);
1831
1832         if (td->o.directory)
1833                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1834                                         td->o.unique_filename);
1835
1836         sprintf(file_name + len, "%s", fname);
1837
1838         /* clean cloned siblings using existing files */
1839         if (numjob && is_already_allocated(file_name) &&
1840             !exists_and_not_regfile(fname))
1841                 return 0;
1842
1843         f = alloc_new_file(td);
1844
1845         if (td->files_size <= td->files_index) {
1846                 unsigned int new_size = td->o.nr_files + 1;
1847
1848                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1849
1850                 td->files = realloc(td->files, new_size * sizeof(f));
1851                 if (td->files == NULL) {
1852                         log_err("fio: realloc OOM\n");
1853                         assert(0);
1854                 }
1855                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1856                         td->file_locks = realloc(td->file_locks, new_size);
1857                         if (!td->file_locks) {
1858                                 log_err("fio: realloc OOM\n");
1859                                 assert(0);
1860                         }
1861                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1862                 }
1863                 td->files_size = new_size;
1864         }
1865         td->files[cur_files] = f;
1866         f->fileno = cur_files;
1867
1868         /*
1869          * init function, io engine may not be loaded yet
1870          */
1871         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1872                 f->real_file_size = -1ULL;
1873
1874         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1875                 f->file_name = strdup(file_name);
1876         else
1877                 f->file_name = smalloc_strdup(file_name);
1878
1879         /* can't handle smalloc failure from here */
1880         assert(f->file_name);
1881
1882         get_file_type(f);
1883
1884         switch (td->o.file_lock_mode) {
1885         case FILE_LOCK_NONE:
1886                 break;
1887         case FILE_LOCK_READWRITE:
1888                 f->rwlock = fio_rwlock_init();
1889                 break;
1890         case FILE_LOCK_EXCLUSIVE:
1891                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1892                 break;
1893         default:
1894                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1895                 assert(0);
1896         }
1897
1898         td->files_index++;
1899
1900         if (td->o.numjobs > 1)
1901                 set_already_allocated(file_name);
1902
1903         if (inc)
1904                 td->o.nr_files++;
1905
1906         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1907                                                         cur_files);
1908
1909         return cur_files;
1910 }
1911
1912 int add_file_exclusive(struct thread_data *td, const char *fname)
1913 {
1914         struct fio_file *f;
1915         unsigned int i;
1916
1917         for_each_file(td, f, i) {
1918                 if (!strcmp(f->file_name, fname))
1919                         return i;
1920         }
1921
1922         return add_file(td, fname, 0, 1);
1923 }
1924
1925 void get_file(struct fio_file *f)
1926 {
1927         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1928         assert(fio_file_open(f));
1929         f->references++;
1930 }
1931
1932 int put_file(struct thread_data *td, struct fio_file *f)
1933 {
1934         int f_ret = 0, ret = 0;
1935
1936         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1937
1938         if (!fio_file_open(f)) {
1939                 assert(f->fd == -1);
1940                 return 0;
1941         }
1942
1943         assert(f->references);
1944         if (--f->references)
1945                 return 0;
1946
1947         disk_util_dec(f->du);
1948
1949         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1950                 unlock_file_all(td, f);
1951
1952         if (should_fsync(td) && td->o.fsync_on_close) {
1953                 f_ret = fsync(f->fd);
1954                 if (f_ret < 0)
1955                         f_ret = errno;
1956         }
1957
1958         if (td->io_ops->close_file)
1959                 ret = td->io_ops->close_file(td, f);
1960
1961         if (!ret)
1962                 ret = f_ret;
1963
1964         td->nr_open_files--;
1965         fio_file_clear_closing(f);
1966         fio_file_clear_open(f);
1967         assert(f->fd == -1);
1968         return ret;
1969 }
1970
1971 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1972 {
1973         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1974                 return;
1975
1976         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1977                 if (ddir == DDIR_READ)
1978                         fio_rwlock_read(f->rwlock);
1979                 else
1980                         fio_rwlock_write(f->rwlock);
1981         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1982                 fio_sem_down(f->lock);
1983
1984         td->file_locks[f->fileno] = td->o.file_lock_mode;
1985 }
1986
1987 void unlock_file(struct thread_data *td, struct fio_file *f)
1988 {
1989         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1990                 return;
1991
1992         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1993                 fio_rwlock_unlock(f->rwlock);
1994         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1995                 fio_sem_up(f->lock);
1996
1997         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1998 }
1999
2000 void unlock_file_all(struct thread_data *td, struct fio_file *f)
2001 {
2002         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
2003                 return;
2004         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
2005                 unlock_file(td, f);
2006 }
2007
2008 static bool recurse_dir(struct thread_data *td, const char *dirname)
2009 {
2010         struct dirent *dir;
2011         bool ret = false;
2012         DIR *D;
2013
2014         D = opendir(dirname);
2015         if (!D) {
2016                 char buf[FIO_VERROR_SIZE];
2017
2018                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
2019                 td_verror(td, errno, buf);
2020                 return true;
2021         }
2022
2023         while ((dir = readdir(D)) != NULL) {
2024                 char full_path[PATH_MAX];
2025                 struct stat sb;
2026
2027                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
2028                         continue;
2029
2030                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
2031
2032                 if (lstat(full_path, &sb) == -1) {
2033                         if (errno != ENOENT) {
2034                                 td_verror(td, errno, "stat");
2035                                 ret = true;
2036                                 break;
2037                         }
2038                 }
2039
2040                 if (S_ISREG(sb.st_mode)) {
2041                         add_file(td, full_path, 0, 1);
2042                         continue;
2043                 }
2044                 if (!S_ISDIR(sb.st_mode))
2045                         continue;
2046
2047                 ret = recurse_dir(td, full_path);
2048                 if (ret)
2049                         break;
2050         }
2051
2052         closedir(D);
2053         return ret;
2054 }
2055
2056 int add_dir_files(struct thread_data *td, const char *path)
2057 {
2058         int ret = recurse_dir(td, path);
2059
2060         if (!ret)
2061                 log_info("fio: opendir added %d files\n", td->o.nr_files);
2062
2063         return ret;
2064 }
2065
2066 void dup_files(struct thread_data *td, struct thread_data *org)
2067 {
2068         struct fio_file *f;
2069         unsigned int i;
2070
2071         dprint(FD_FILE, "dup files: %d\n", org->files_index);
2072
2073         if (!org->files)
2074                 return;
2075
2076         td->files = calloc(org->files_index, sizeof(f));
2077
2078         if (td->o.file_lock_mode != FILE_LOCK_NONE)
2079                 td->file_locks = malloc(org->files_index);
2080
2081         assert(org->files_index >= org->o.nr_files);
2082         for_each_file(org, f, i) {
2083                 struct fio_file *__f;
2084
2085                 __f = alloc_new_file(td);
2086
2087                 if (f->file_name) {
2088                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
2089                                 __f->file_name = strdup(f->file_name);
2090                         else
2091                                 __f->file_name = smalloc_strdup(f->file_name);
2092
2093                         /* can't handle smalloc failure from here */
2094                         assert(__f->file_name);
2095                         __f->filetype = f->filetype;
2096                 }
2097
2098                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
2099                         __f->lock = f->lock;
2100                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
2101                         __f->rwlock = f->rwlock;
2102
2103                 td->files[i] = __f;
2104         }
2105 }
2106
2107 /*
2108  * Returns the index that matches the filename, or -1 if not there
2109  */
2110 int get_fileno(struct thread_data *td, const char *fname)
2111 {
2112         struct fio_file *f;
2113         unsigned int i;
2114
2115         for_each_file(td, f, i)
2116                 if (!strcmp(f->file_name, fname))
2117                         return i;
2118
2119         return -1;
2120 }
2121
2122 /*
2123  * For log usage, where we add/open/close files automatically
2124  */
2125 void free_release_files(struct thread_data *td)
2126 {
2127         close_files(td);
2128         td->o.nr_files = 0;
2129         td->o.open_files = 0;
2130         td->files_index = 0;
2131 }
2132
2133 void fio_file_reset(struct thread_data *td, struct fio_file *f)
2134 {
2135         int i;
2136
2137         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2138                 f->last_pos[i] = f->file_offset;
2139                 f->last_start[i] = -1ULL;
2140         }
2141
2142         if (fio_file_axmap(f))
2143                 axmap_reset(f->io_axmap);
2144         else if (fio_file_lfsr(f))
2145                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
2146
2147         zbd_file_reset(td, f);
2148 }
2149
2150 bool fio_files_done(struct thread_data *td)
2151 {
2152         struct fio_file *f;
2153         unsigned int i;
2154
2155         for_each_file(td, f, i)
2156                 if (!fio_file_done(f))
2157                         return false;
2158
2159         return true;
2160 }
2161
2162 /* free memory used in initialization phase only */
2163 void filesetup_mem_free(void)
2164 {
2165         free_already_allocated();
2166 }
2167
2168 /*
2169  * This function is for platforms which support direct I/O but not O_DIRECT.
2170  */
2171 int fio_set_directio(struct thread_data *td, struct fio_file *f)
2172 {
2173 #ifdef FIO_OS_DIRECTIO
2174         int ret = fio_set_odirect(f);
2175
2176         if (ret) {
2177                 td_verror(td, ret, "fio_set_directio");
2178 #if defined(__sun__)
2179                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2180                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2181                 } else {
2182                         log_err("fio: the file system does not seem to support direct IO\n");
2183                 }
2184 #else
2185                 log_err("fio: the file system does not seem to support direct IO\n");
2186 #endif
2187                 return -1;
2188         }
2189
2190         return 0;
2191 #else
2192         log_err("fio: direct IO is not supported on this host operating system\n");
2193         return -1;
2194 #endif
2195 }