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