Add 'per_job_logs' option
[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 (td->o.block_error_hist) {
884                 int len;
885
886                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
887                 f = td->files[0];
888                 len = f->io_size / td->o.bs[DDIR_TRIM];
889                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
890                         log_err("fio: cannot calculate block histogram with "
891                                 "%d trim blocks, maximum %d\n",
892                                 len, MAX_NR_BLOCK_INFOS);
893                         td_verror(td, EINVAL, "block_error_hist");
894                         goto err_out;
895                 }
896
897                 td->ts.nr_block_infos = len;
898                 for (int i = 0; i < len; i++)
899                         td->ts.block_infos[i] =
900                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
901         } else
902                 td->ts.nr_block_infos = 0;
903
904         if (!o->size || (total_size && o->size > total_size))
905                 o->size = total_size;
906
907         if (o->size < td_min_bs(td)) {
908                 log_err("fio: blocksize too large for data set\n");
909                 goto err_out;
910         }
911
912         /*
913          * See if we need to extend some files
914          */
915         if (need_extend) {
916                 temp_stall_ts = 1;
917                 if (output_format == FIO_OUTPUT_NORMAL)
918                         log_info("%s: Laying out IO file(s) (%u file(s) /"
919                                  " %lluMB)\n", o->name, need_extend,
920                                         extend_size >> 20);
921
922                 for_each_file(td, f, i) {
923                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
924
925                         if (!fio_file_extend(f))
926                                 continue;
927
928                         assert(f->filetype == FIO_TYPE_FILE);
929                         fio_file_clear_extend(f);
930                         if (!o->fill_device) {
931                                 old_len = f->real_file_size;
932                                 extend_len = f->io_size + f->file_offset -
933                                                 old_len;
934                         }
935                         f->real_file_size = (f->io_size + f->file_offset);
936                         err = extend_file(td, f);
937                         if (err)
938                                 break;
939
940                         err = __file_invalidate_cache(td, f, old_len,
941                                                                 extend_len);
942
943                         /*
944                          * Shut up static checker
945                          */
946                         if (f->fd != -1)
947                                 close(f->fd);
948
949                         f->fd = -1;
950                         if (err)
951                                 break;
952                 }
953                 temp_stall_ts = 0;
954         }
955
956         if (err)
957                 goto err_out;
958
959         if (!o->zone_size)
960                 o->zone_size = o->size;
961
962         /*
963          * iolog already set the total io size, if we read back
964          * stored entries.
965          */
966         if (!o->read_iolog_file) {
967                 if (o->io_limit)
968                         td->total_io_size = o->io_limit * o->loops;
969                 else
970                         td->total_io_size = o->size * o->loops;
971         }
972
973 done:
974         if (o->create_only)
975                 td->done = 1;
976
977         td_restore_runstate(td, old_state);
978         return 0;
979 err_offset:
980         log_err("%s: you need to specify valid offset=\n", o->name);
981 err_out:
982         td_restore_runstate(td, old_state);
983         return 1;
984 }
985
986 int pre_read_files(struct thread_data *td)
987 {
988         struct fio_file *f;
989         unsigned int i;
990
991         dprint(FD_FILE, "pre_read files\n");
992
993         for_each_file(td, f, i) {
994                 pre_read_file(td, f);
995         }
996
997         return 1;
998 }
999
1000 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1001 {
1002         unsigned int range_size, seed;
1003         unsigned long nranges;
1004         uint64_t fsize;
1005
1006         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1007         fsize = min(f->real_file_size, f->io_size);
1008
1009         nranges = (fsize + range_size - 1) / range_size;
1010
1011         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1012         if (!td->o.rand_repeatable)
1013                 seed = td->rand_seeds[4];
1014
1015         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1016                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1017         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1018                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1019         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1020                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1021
1022         return 1;
1023 }
1024
1025 static int init_rand_distribution(struct thread_data *td)
1026 {
1027         struct fio_file *f;
1028         unsigned int i;
1029         int state;
1030
1031         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1032                 return 0;
1033
1034         state = td_bump_runstate(td, TD_SETTING_UP);
1035
1036         for_each_file(td, f, i)
1037                 __init_rand_distribution(td, f);
1038
1039         td_restore_runstate(td, state);
1040
1041         return 1;
1042 }
1043
1044 int init_random_map(struct thread_data *td)
1045 {
1046         unsigned long long blocks;
1047         struct fio_file *f;
1048         unsigned int i;
1049
1050         if (init_rand_distribution(td))
1051                 return 0;
1052         if (!td_random(td))
1053                 return 0;
1054
1055         for_each_file(td, f, i) {
1056                 uint64_t fsize = min(f->real_file_size, f->io_size);
1057
1058                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1059
1060                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1061                         unsigned long seed;
1062
1063                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1064
1065                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1066                                 fio_file_set_lfsr(f);
1067                                 continue;
1068                         }
1069                 } else if (!td->o.norandommap) {
1070                         f->io_axmap = axmap_new(blocks);
1071                         if (f->io_axmap) {
1072                                 fio_file_set_axmap(f);
1073                                 continue;
1074                         }
1075                 } else if (td->o.norandommap)
1076                         continue;
1077
1078                 if (!td->o.softrandommap) {
1079                         log_err("fio: failed allocating random map. If running"
1080                                 " a large number of jobs, try the 'norandommap'"
1081                                 " option or set 'softrandommap'. Or give"
1082                                 " a larger --alloc-size to fio.\n");
1083                         return 1;
1084                 }
1085
1086                 log_info("fio: file %s failed allocating random map. Running "
1087                          "job without.\n", f->file_name);
1088         }
1089
1090         return 0;
1091 }
1092
1093 void close_files(struct thread_data *td)
1094 {
1095         struct fio_file *f;
1096         unsigned int i;
1097
1098         for_each_file(td, f, i) {
1099                 if (fio_file_open(f))
1100                         td_io_close_file(td, f);
1101         }
1102 }
1103
1104 void close_and_free_files(struct thread_data *td)
1105 {
1106         struct fio_file *f;
1107         unsigned int i;
1108
1109         dprint(FD_FILE, "close files\n");
1110
1111         for_each_file(td, f, i) {
1112                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1113                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1114                         td_io_unlink_file(td, f);
1115                 }
1116
1117                 if (fio_file_open(f))
1118                         td_io_close_file(td, f);
1119
1120                 remove_file_hash(f);
1121
1122                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1123                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1124                         td_io_unlink_file(td, f);
1125                 }
1126
1127                 sfree(f->file_name);
1128                 f->file_name = NULL;
1129                 if (fio_file_axmap(f)) {
1130                         axmap_free(f->io_axmap);
1131                         f->io_axmap = NULL;
1132                 }
1133                 sfree(f);
1134         }
1135
1136         td->o.filename = NULL;
1137         free(td->files);
1138         free(td->file_locks);
1139         td->files_index = 0;
1140         td->files = NULL;
1141         td->file_locks = NULL;
1142         td->o.file_lock_mode = FILE_LOCK_NONE;
1143         td->o.nr_files = 0;
1144 }
1145
1146 static void get_file_type(struct fio_file *f)
1147 {
1148         struct stat sb;
1149
1150         if (!strcmp(f->file_name, "-"))
1151                 f->filetype = FIO_TYPE_PIPE;
1152         else
1153                 f->filetype = FIO_TYPE_FILE;
1154
1155         /* \\.\ is the device namespace in Windows, where every file is
1156          * a block device */
1157         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1158                 f->filetype = FIO_TYPE_BD;
1159
1160         if (!stat(f->file_name, &sb)) {
1161                 if (S_ISBLK(sb.st_mode))
1162                         f->filetype = FIO_TYPE_BD;
1163                 else if (S_ISCHR(sb.st_mode))
1164                         f->filetype = FIO_TYPE_CHAR;
1165                 else if (S_ISFIFO(sb.st_mode))
1166                         f->filetype = FIO_TYPE_PIPE;
1167         }
1168 }
1169
1170 static int __is_already_allocated(const char *fname)
1171 {
1172         struct flist_head *entry;
1173         char *filename;
1174
1175         if (flist_empty(&filename_list))
1176                 return 0;
1177
1178         flist_for_each(entry, &filename_list) {
1179                 filename = flist_entry(entry, struct file_name, list)->filename;
1180
1181                 if (strcmp(filename, fname) == 0)
1182                         return 1;
1183         }
1184
1185         return 0;
1186 }
1187
1188 static int is_already_allocated(const char *fname)
1189 {
1190         int ret;
1191
1192         fio_file_hash_lock();
1193         ret = __is_already_allocated(fname);
1194         fio_file_hash_unlock();
1195         return ret;
1196 }
1197
1198 static void set_already_allocated(const char *fname)
1199 {
1200         struct file_name *fn;
1201
1202         fn = malloc(sizeof(struct file_name));
1203         fn->filename = strdup(fname);
1204
1205         fio_file_hash_lock();
1206         if (!__is_already_allocated(fname)) {
1207                 flist_add_tail(&fn->list, &filename_list);
1208                 fn = NULL;
1209         }
1210         fio_file_hash_unlock();
1211
1212         if (fn) {
1213                 free(fn->filename);
1214                 free(fn);
1215         }
1216 }
1217
1218
1219 static void free_already_allocated(void)
1220 {
1221         struct flist_head *entry, *tmp;
1222         struct file_name *fn;
1223
1224         if (flist_empty(&filename_list))
1225                 return;
1226
1227         fio_file_hash_lock();
1228         flist_for_each_safe(entry, tmp, &filename_list) {
1229                 fn = flist_entry(entry, struct file_name, list);
1230                 free(fn->filename);
1231                 flist_del(&fn->list);
1232                 free(fn);
1233         }
1234
1235         fio_file_hash_unlock();
1236 }
1237
1238 static struct fio_file *alloc_new_file(struct thread_data *td)
1239 {
1240         struct fio_file *f;
1241
1242         f = smalloc(sizeof(*f));
1243         if (!f) {
1244                 log_err("fio: smalloc OOM\n");
1245                 assert(0);
1246                 return NULL;
1247         }
1248
1249         f->fd = -1;
1250         f->shadow_fd = -1;
1251         fio_file_reset(td, f);
1252         return f;
1253 }
1254
1255 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1256 {
1257         int cur_files = td->files_index;
1258         char file_name[PATH_MAX];
1259         struct fio_file *f;
1260         int len = 0;
1261
1262         dprint(FD_FILE, "add file %s\n", fname);
1263
1264         if (td->o.directory)
1265                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob);
1266
1267         sprintf(file_name + len, "%s", fname);
1268
1269         /* clean cloned siblings using existing files */
1270         if (numjob && is_already_allocated(file_name))
1271                 return 0;
1272
1273         f = alloc_new_file(td);
1274
1275         if (td->files_size <= td->files_index) {
1276                 unsigned int new_size = td->o.nr_files + 1;
1277
1278                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1279
1280                 td->files = realloc(td->files, new_size * sizeof(f));
1281                 if (td->files == NULL) {
1282                         log_err("fio: realloc OOM\n");
1283                         assert(0);
1284                 }
1285                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1286                         td->file_locks = realloc(td->file_locks, new_size);
1287                         if (!td->file_locks) {
1288                                 log_err("fio: realloc OOM\n");
1289                                 assert(0);
1290                         }
1291                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1292                 }
1293                 td->files_size = new_size;
1294         }
1295         td->files[cur_files] = f;
1296         f->fileno = cur_files;
1297
1298         /*
1299          * init function, io engine may not be loaded yet
1300          */
1301         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1302                 f->real_file_size = -1ULL;
1303
1304         f->file_name = smalloc_strdup(file_name);
1305         if (!f->file_name) {
1306                 log_err("fio: smalloc OOM\n");
1307                 assert(0);
1308         }
1309
1310         get_file_type(f);
1311
1312         switch (td->o.file_lock_mode) {
1313         case FILE_LOCK_NONE:
1314                 break;
1315         case FILE_LOCK_READWRITE:
1316                 f->rwlock = fio_rwlock_init();
1317                 break;
1318         case FILE_LOCK_EXCLUSIVE:
1319                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1320                 break;
1321         default:
1322                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1323                 assert(0);
1324         }
1325
1326         td->files_index++;
1327         if (f->filetype == FIO_TYPE_FILE)
1328                 td->nr_normal_files++;
1329
1330         set_already_allocated(file_name);
1331
1332         if (inc)
1333                 td->o.nr_files++;
1334
1335         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1336                                                         cur_files);
1337
1338         return cur_files;
1339 }
1340
1341 int add_file_exclusive(struct thread_data *td, const char *fname)
1342 {
1343         struct fio_file *f;
1344         unsigned int i;
1345
1346         for_each_file(td, f, i) {
1347                 if (!strcmp(f->file_name, fname))
1348                         return i;
1349         }
1350
1351         return add_file(td, fname, 0, 1);
1352 }
1353
1354 void get_file(struct fio_file *f)
1355 {
1356         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1357         assert(fio_file_open(f));
1358         f->references++;
1359 }
1360
1361 int put_file(struct thread_data *td, struct fio_file *f)
1362 {
1363         int f_ret = 0, ret = 0;
1364
1365         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1366
1367         if (!fio_file_open(f)) {
1368                 assert(f->fd == -1);
1369                 return 0;
1370         }
1371
1372         assert(f->references);
1373         if (--f->references)
1374                 return 0;
1375
1376         if (should_fsync(td) && td->o.fsync_on_close) {
1377                 f_ret = fsync(f->fd);
1378                 if (f_ret < 0)
1379                         f_ret = errno;
1380         }
1381
1382         if (td->io_ops->close_file)
1383                 ret = td->io_ops->close_file(td, f);
1384
1385         if (!ret)
1386                 ret = f_ret;
1387
1388         td->nr_open_files--;
1389         fio_file_clear_open(f);
1390         assert(f->fd == -1);
1391         return ret;
1392 }
1393
1394 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1395 {
1396         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1397                 return;
1398
1399         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1400                 if (ddir == DDIR_READ)
1401                         fio_rwlock_read(f->rwlock);
1402                 else
1403                         fio_rwlock_write(f->rwlock);
1404         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1405                 fio_mutex_down(f->lock);
1406
1407         td->file_locks[f->fileno] = td->o.file_lock_mode;
1408 }
1409
1410 void unlock_file(struct thread_data *td, struct fio_file *f)
1411 {
1412         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1413                 return;
1414
1415         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1416                 fio_rwlock_unlock(f->rwlock);
1417         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1418                 fio_mutex_up(f->lock);
1419
1420         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1421 }
1422
1423 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1424 {
1425         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1426                 return;
1427         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1428                 unlock_file(td, f);
1429 }
1430
1431 static int recurse_dir(struct thread_data *td, const char *dirname)
1432 {
1433         struct dirent *dir;
1434         int ret = 0;
1435         DIR *D;
1436
1437         D = opendir(dirname);
1438         if (!D) {
1439                 char buf[FIO_VERROR_SIZE];
1440
1441                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1442                 td_verror(td, errno, buf);
1443                 return 1;
1444         }
1445
1446         while ((dir = readdir(D)) != NULL) {
1447                 char full_path[PATH_MAX];
1448                 struct stat sb;
1449
1450                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1451                         continue;
1452
1453                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1454
1455                 if (lstat(full_path, &sb) == -1) {
1456                         if (errno != ENOENT) {
1457                                 td_verror(td, errno, "stat");
1458                                 ret = 1;
1459                                 break;
1460                         }
1461                 }
1462
1463                 if (S_ISREG(sb.st_mode)) {
1464                         add_file(td, full_path, 0, 1);
1465                         continue;
1466                 }
1467                 if (!S_ISDIR(sb.st_mode))
1468                         continue;
1469
1470                 ret = recurse_dir(td, full_path);
1471                 if (ret)
1472                         break;
1473         }
1474
1475         closedir(D);
1476         return ret;
1477 }
1478
1479 int add_dir_files(struct thread_data *td, const char *path)
1480 {
1481         int ret = recurse_dir(td, path);
1482
1483         if (!ret)
1484                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1485
1486         return ret;
1487 }
1488
1489 void dup_files(struct thread_data *td, struct thread_data *org)
1490 {
1491         struct fio_file *f;
1492         unsigned int i;
1493
1494         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1495
1496         if (!org->files)
1497                 return;
1498
1499         td->files = malloc(org->files_index * sizeof(f));
1500
1501         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1502                 td->file_locks = malloc(org->files_index);
1503
1504         for_each_file(org, f, i) {
1505                 struct fio_file *__f;
1506
1507                 __f = alloc_new_file(td);
1508
1509                 if (f->file_name) {
1510                         __f->file_name = smalloc_strdup(f->file_name);
1511                         if (!__f->file_name) {
1512                                 log_err("fio: smalloc OOM\n");
1513                                 assert(0);
1514                         }
1515
1516                         __f->filetype = f->filetype;
1517                 }
1518
1519                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1520                         __f->lock = f->lock;
1521                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1522                         __f->rwlock = f->rwlock;
1523
1524                 td->files[i] = __f;
1525         }
1526 }
1527
1528 /*
1529  * Returns the index that matches the filename, or -1 if not there
1530  */
1531 int get_fileno(struct thread_data *td, const char *fname)
1532 {
1533         struct fio_file *f;
1534         unsigned int i;
1535
1536         for_each_file(td, f, i)
1537                 if (!strcmp(f->file_name, fname))
1538                         return i;
1539
1540         return -1;
1541 }
1542
1543 /*
1544  * For log usage, where we add/open/close files automatically
1545  */
1546 void free_release_files(struct thread_data *td)
1547 {
1548         close_files(td);
1549         td->o.nr_files = 0;
1550         td->o.open_files = 0;
1551         td->files_index = 0;
1552         td->nr_normal_files = 0;
1553 }
1554
1555 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1556 {
1557         int i;
1558
1559         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1560                 f->last_pos[i] = f->file_offset;
1561                 f->last_start[i] = -1ULL;
1562         }
1563
1564         if (fio_file_axmap(f))
1565                 axmap_reset(f->io_axmap);
1566         else if (fio_file_lfsr(f))
1567                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1568 }
1569
1570 int fio_files_done(struct thread_data *td)
1571 {
1572         struct fio_file *f;
1573         unsigned int i;
1574
1575         for_each_file(td, f, i)
1576                 if (!fio_file_done(f))
1577                         return 0;
1578
1579         return 1;
1580 }
1581
1582 /* free memory used in initialization phase only */
1583 void filesetup_mem_free(void)
1584 {
1585         free_already_allocated();
1586 }