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