Add total latency log
[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
15 static int root_warn;
16
17 static inline void clear_error(struct thread_data *td)
18 {
19         td->error = 0;
20         td->verror[0] = '\0';
21 }
22
23 /*
24  * Leaves f->fd open on success, caller must close
25  */
26 static int extend_file(struct thread_data *td, struct fio_file *f)
27 {
28         int r, new_layout = 0, unlink_file = 0, flags;
29         unsigned long long left;
30         unsigned int bs;
31         char *b;
32
33         if (read_only) {
34                 log_err("fio: refusing extend of file due to read-only\n");
35                 return 0;
36         }
37
38         /*
39          * check if we need to lay the file out complete again. fio
40          * does that for operations involving reads, or for writes
41          * where overwrite is set
42          */
43         if (td_read(td) || (td_write(td) && td->o.overwrite) ||
44             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
45                 new_layout = 1;
46         if (td_write(td) && !td->o.overwrite)
47                 unlink_file = 1;
48
49         if (unlink_file || new_layout) {
50                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
51                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
52                         td_verror(td, errno, "unlink");
53                         return 1;
54                 }
55         }
56
57         flags = O_WRONLY | O_CREAT;
58         if (new_layout)
59                 flags |= O_TRUNC;
60
61         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
62         f->fd = open(f->file_name, flags, 0644);
63         if (f->fd < 0) {
64                 td_verror(td, errno, "open");
65                 return 1;
66         }
67
68 #ifdef FIO_HAVE_FALLOCATE
69         if (td->o.fallocate && !td->o.fill_device) {
70                 dprint(FD_FILE, "fallocate file %s size %llu\n", f->file_name,
71                                                         f->real_file_size);
72
73                 r = posix_fallocate(f->fd, 0, f->real_file_size);
74                 if (r < 0) {
75                         log_err("fio: posix_fallocate fails: %s\n",
76                                         strerror(-r));
77                 }
78         }
79 #endif
80         
81         if (!new_layout)
82                 goto done;
83
84         /*
85          * The size will be -1ULL when fill_device is used, so don't truncate
86          * or fallocate this file, just write it
87          */
88         if (!td->o.fill_device) {
89                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
90                                                         f->real_file_size);
91                 if (ftruncate(f->fd, f->real_file_size) == -1) {
92                         td_verror(td, errno, "ftruncate");
93                         goto err;
94                 }
95         }
96
97         b = malloc(td->o.max_bs[DDIR_WRITE]);
98         memset(b, 0, td->o.max_bs[DDIR_WRITE]);
99
100         left = f->real_file_size;
101         while (left && !td->terminate) {
102                 bs = td->o.max_bs[DDIR_WRITE];
103                 if (bs > left)
104                         bs = left;
105
106                 r = write(f->fd, b, bs);
107
108                 if (r > 0) {
109                         left -= r;
110                         continue;
111                 } else {
112                         if (r < 0) {
113                                 int __e = errno;
114
115                                 if (__e == ENOSPC) {
116                                         if (td->o.fill_device)
117                                                 break;
118                                         log_info("fio: ENOSPC on laying out "
119                                                  "file, stopping\n");
120                                         break;
121                                 }
122                                 td_verror(td, errno, "write");
123                         } else
124                                 td_verror(td, EIO, "write");
125
126                         break;
127                 }
128         }
129
130         if (td->terminate) {
131                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
132                 unlink(f->file_name);
133         } else if (td->o.create_fsync) {
134                 if (fsync(f->fd) < 0) {
135                         td_verror(td, errno, "fsync");
136                         goto err;
137                 }
138         }
139         if (td->o.fill_device && !td_write(td)) {
140                 fio_file_clear_size_known(f);
141                 if (td_io_get_file_size(td, f))
142                         goto err;
143                 if (f->io_size > f->real_file_size)
144                         f->io_size = f->real_file_size;
145         }
146
147         free(b);
148 done:
149         return 0;
150 err:
151         close(f->fd);
152         f->fd = -1;
153         return 1;
154 }
155
156 static int pre_read_file(struct thread_data *td, struct fio_file *f)
157 {
158         int r, did_open = 0, old_runstate;
159         unsigned long long left;
160         unsigned int bs;
161         char *b;
162
163         if (td->io_ops->flags & FIO_PIPEIO)
164                 return 0;
165
166         if (!fio_file_open(f)) {
167                 if (td->io_ops->open_file(td, f)) {
168                         log_err("fio: cannot pre-read, failed to open file\n");
169                         return 1;
170                 }
171                 did_open = 1;
172         }
173
174         old_runstate = td->runstate;
175         td_set_runstate(td, TD_PRE_READING);
176
177         bs = td->o.max_bs[DDIR_READ];
178         b = malloc(bs);
179         memset(b, 0, bs);
180
181         lseek(f->fd, f->file_offset, SEEK_SET);
182         left = f->io_size;
183
184         while (left && !td->terminate) {
185                 if (bs > left)
186                         bs = left;
187
188                 r = read(f->fd, b, bs);
189
190                 if (r == (int) bs) {
191                         left -= bs;
192                         continue;
193                 } else {
194                         td_verror(td, EIO, "pre_read");
195                         break;
196                 }
197         }
198
199         td_set_runstate(td, old_runstate);
200
201         if (did_open)
202                 td->io_ops->close_file(td, f);
203         free(b);
204         return 0;
205 }
206
207 static unsigned long long get_rand_file_size(struct thread_data *td)
208 {
209         unsigned long long ret, sized;
210         long r;
211
212         r = os_random_long(&td->file_size_state);
213         sized = td->o.file_size_high - td->o.file_size_low;
214         ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
215         ret += td->o.file_size_low;
216         ret -= (ret % td->o.rw_min_bs);
217         return ret;
218 }
219
220 static int file_size(struct thread_data *td, struct fio_file *f)
221 {
222         struct stat st;
223
224         if (stat(f->file_name, &st) == -1) {
225                 td_verror(td, errno, "fstat");
226                 return 1;
227         }
228
229         f->real_file_size = st.st_size;
230         return 0;
231 }
232
233 static int bdev_size(struct thread_data *td, struct fio_file *f)
234 {
235         unsigned long long bytes;
236         int r;
237
238         if (td->io_ops->open_file(td, f)) {
239                 log_err("fio: failed opening blockdev %s for size check\n",
240                         f->file_name);
241                 return 1;
242         }
243
244         r = blockdev_size(f->fd, &bytes);
245         if (r) {
246                 td_verror(td, r, "blockdev_size");
247                 goto err;
248         }
249
250         if (!bytes) {
251                 log_err("%s: zero sized block device?\n", f->file_name);
252                 goto err;
253         }
254
255         f->real_file_size = bytes;
256         td->io_ops->close_file(td, f);
257         return 0;
258 err:
259         td->io_ops->close_file(td, f);
260         return 1;
261 }
262
263 static int get_file_size(struct thread_data *td, struct fio_file *f)
264 {
265         int ret = 0;
266
267         if (fio_file_size_known(f))
268                 return 0;
269
270         if (f->filetype == FIO_TYPE_FILE)
271                 ret = file_size(td, f);
272         else if (f->filetype == FIO_TYPE_BD)
273                 ret = bdev_size(td, f);
274         else
275                 f->real_file_size = -1;
276
277         if (ret)
278                 return ret;
279
280         if (f->file_offset > f->real_file_size) {
281                 log_err("%s: offset extends end (%Lu > %Lu)\n", td->o.name,
282                                         f->file_offset, f->real_file_size);
283                 return 1;
284         }
285
286         fio_file_set_size_known(f);
287         return 0;
288 }
289
290 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
291                                    unsigned long long off,
292                                    unsigned long long len)
293 {
294         int ret = 0;
295
296         if (len == -1ULL)
297                 len = f->io_size;
298         if (off == -1ULL)
299                 off = f->file_offset;
300
301         if (len == -1ULL || off == -1ULL)
302                 return 0;
303
304         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
305                                                                 len);
306
307         /*
308          * FIXME: add blockdev flushing too
309          */
310         if (f->mmap_ptr) {
311                 ret = madvise(f->mmap_ptr, f->mmap_sz, MADV_DONTNEED);
312 #ifdef FIO_MADV_FREE
313                 (void) madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
314 #endif
315         } else if (f->filetype == FIO_TYPE_FILE) {
316                 ret = fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
317         } else if (f->filetype == FIO_TYPE_BD) {
318                 ret = blockdev_invalidate_cache(f->fd);
319                 if (ret < 0 && errno == EACCES && geteuid()) {
320                         if (!root_warn) {
321                                 log_err("fio: only root may flush block "
322                                         "devices. Cache flush bypassed!\n");
323                                 root_warn = 1;
324                         }
325                         ret = 0;
326                 }
327         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
328                 ret = 0;
329
330         if (ret < 0) {
331                 td_verror(td, errno, "invalidate_cache");
332                 return 1;
333         } else if (ret > 0) {
334                 td_verror(td, ret, "invalidate_cache");
335                 return 1;
336         }
337
338         return ret;
339
340 }
341
342 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
343 {
344         if (!fio_file_open(f))
345                 return 0;
346
347         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
348 }
349
350 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
351 {
352         int ret = 0;
353
354         dprint(FD_FILE, "fd close %s\n", f->file_name);
355
356         remove_file_hash(f);
357
358         if (close(f->fd) < 0)
359                 ret = errno;
360
361         f->fd = -1;
362         return ret;
363 }
364
365 static int file_lookup_open(struct fio_file *f, int flags)
366 {
367         struct fio_file *__f;
368         int from_hash;
369
370         __f = lookup_file_hash(f->file_name);
371         if (__f) {
372                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
373                 /*
374                  * racy, need the __f->lock locked
375                  */
376                 f->lock = __f->lock;
377                 f->lock_owner = __f->lock_owner;
378                 f->lock_batch = __f->lock_batch;
379                 f->lock_ddir = __f->lock_ddir;
380                 from_hash = 1;
381         } else {
382                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
383                 from_hash = 0;
384         }
385
386         f->fd = open(f->file_name, flags, 0600);
387         return from_hash;
388 }
389
390 int generic_open_file(struct thread_data *td, struct fio_file *f)
391 {
392         int is_std = 0;
393         int flags = 0;
394         int from_hash = 0;
395
396         dprint(FD_FILE, "fd open %s\n", f->file_name);
397
398         if (!strcmp(f->file_name, "-")) {
399                 if (td_rw(td)) {
400                         log_err("fio: can't read/write to stdin/out\n");
401                         return 1;
402                 }
403                 is_std = 1;
404
405                 /*
406                  * move output logging to stderr, if we are writing to stdout
407                  */
408                 if (td_write(td))
409                         f_out = stderr;
410         }
411
412         if (td->o.odirect)
413                 flags |= OS_O_DIRECT;
414         if (td->o.sync_io)
415                 flags |= O_SYNC;
416         if (f->filetype != FIO_TYPE_FILE)
417                 flags |= FIO_O_NOATIME;
418         if (td->o.create_on_open)
419                 flags |= O_CREAT;
420
421 open_again:
422         if (td_write(td)) {
423                 if (!read_only)
424                         flags |= O_RDWR;
425
426                 if (f->filetype == FIO_TYPE_FILE)
427                         flags |= O_CREAT;
428
429                 if (is_std)
430                         f->fd = dup(STDOUT_FILENO);
431                 else
432                         from_hash = file_lookup_open(f, flags);
433         } else {
434                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
435                         flags |= O_RDWR;
436                 else
437                         flags |= O_RDONLY;
438
439                 if (is_std)
440                         f->fd = dup(STDIN_FILENO);
441                 else
442                         from_hash = file_lookup_open(f, flags);
443         }
444
445         if (f->fd == -1) {
446                 char buf[FIO_VERROR_SIZE];
447                 int __e = errno;
448
449                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
450                         flags &= ~FIO_O_NOATIME;
451                         goto open_again;
452                 }
453
454                 snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
455
456                 td_verror(td, __e, buf);
457         }
458
459         if (!from_hash && f->fd != -1) {
460                 if (add_file_hash(f)) {
461                         int ret;
462
463                         /*
464                          * OK to ignore, we haven't done anything with it
465                          */
466                         ret = generic_close_file(td, f);
467                         goto open_again;
468                 }
469         }
470
471         return 0;
472 }
473
474 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
475 {
476         return get_file_size(td, f);
477 }
478
479 /*
480  * open/close all files, so that ->real_file_size gets set
481  */
482 static int get_file_sizes(struct thread_data *td)
483 {
484         struct fio_file *f;
485         unsigned int i;
486         int err = 0;
487
488         for_each_file(td, f, i) {
489                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
490                                                                 f->file_name);
491
492                 if (td_io_get_file_size(td, f)) {
493                         if (td->error != ENOENT) {
494                                 log_err("%s\n", td->verror);
495                                 err = 1;
496                         }
497                         clear_error(td);
498                 }
499
500                 if (f->real_file_size == -1ULL && td->o.size)
501                         f->real_file_size = td->o.size / td->o.nr_files;
502         }
503
504         return err;
505 }
506
507 struct fio_mount {
508         struct flist_head list;
509         const char *base;
510         char __base[256];
511         unsigned int key;
512 };
513
514 /*
515  * Get free number of bytes for each file on each unique mount.
516  */
517 static unsigned long long get_fs_free_counts(struct thread_data *td)
518 {
519         struct flist_head *n, *tmp;
520         unsigned long long ret = 0;
521         struct fio_mount *fm;
522         FLIST_HEAD(list);
523         struct fio_file *f;
524         unsigned int i;
525
526         for_each_file(td, f, i) {
527                 struct stat sb;
528                 char buf[256];
529
530                 if (f->filetype == FIO_TYPE_BD) {
531                         ret += f->real_file_size;
532                         continue;
533                 } else if (f->filetype != FIO_TYPE_FILE)
534                         continue;
535
536                 strcpy(buf, f->file_name);
537
538                 if (stat(buf, &sb) < 0) {
539                         if (errno != ENOENT)
540                                 break;
541                         strcpy(buf, ".");
542                         if (stat(buf, &sb) < 0)
543                                 break;
544                 }
545
546                 fm = NULL;
547                 flist_for_each(n, &list) {
548                         fm = flist_entry(n, struct fio_mount, list);
549                         if (fm->key == sb.st_dev)
550                                 break;
551
552                         fm = NULL;
553                 }
554
555                 if (fm)
556                         continue;
557
558                 fm = malloc(sizeof(*fm));
559                 strcpy(fm->__base, buf);
560                 fm->base = basename(fm->__base);
561                 fm->key = sb.st_dev;
562                 flist_add(&fm->list, &list);
563         }
564
565         flist_for_each_safe(n, tmp, &list) {
566                 unsigned long long sz;
567
568                 fm = flist_entry(n, struct fio_mount, list);
569                 flist_del(&fm->list);
570
571                 sz = get_fs_size(fm->base);
572                 if (sz && sz != -1ULL)
573                         ret += sz;
574
575                 free(fm);
576         }
577
578         return ret;
579 }
580
581 /*
582  * Open the files and setup files sizes, creating files if necessary.
583  */
584 int setup_files(struct thread_data *td)
585 {
586         unsigned long long total_size, extend_size;
587         struct fio_file *f;
588         unsigned int i;
589         int err = 0, need_extend;
590
591         dprint(FD_FILE, "setup files\n");
592
593         if (td->o.read_iolog_file)
594                 return 0;
595
596         /*
597          * if ioengine defines a setup() method, it's responsible for
598          * opening the files and setting f->real_file_size to indicate
599          * the valid range for that file.
600          */
601         if (td->io_ops->setup)
602                 err = td->io_ops->setup(td);
603         else
604                 err = get_file_sizes(td);
605
606         if (err)
607                 return err;
608
609         /*
610          * check sizes. if the files/devices do not exist and the size
611          * isn't passed to fio, abort.
612          */
613         total_size = 0;
614         for_each_file(td, f, i) {
615                 if (f->real_file_size == -1ULL)
616                         total_size = -1ULL;
617                 else
618                         total_size += f->real_file_size;
619         }
620
621         if (td->o.fill_device)
622                 td->fill_device_size = get_fs_free_counts(td);
623
624         /*
625          * device/file sizes are zero and no size given, punt
626          */
627         if ((!total_size || total_size == -1ULL) && !td->o.size &&
628             !(td->io_ops->flags & FIO_NOIO) && !td->o.fill_device) {
629                 log_err("%s: you need to specify size=\n", td->o.name);
630                 td_verror(td, EINVAL, "total_file_size");
631                 return 1;
632         }
633
634         /*
635          * now file sizes are known, so we can set ->io_size. if size= is
636          * not given, ->io_size is just equal to ->real_file_size. if size
637          * is given, ->io_size is size / nr_files.
638          */
639         extend_size = total_size = 0;
640         need_extend = 0;
641         for_each_file(td, f, i) {
642                 f->file_offset = td->o.start_offset;
643
644                 if (!td->o.file_size_low) {
645                         /*
646                          * no file size range given, file size is equal to
647                          * total size divided by number of files. if that is
648                          * zero, set it to the real file size.
649                          */
650                         f->io_size = td->o.size / td->o.nr_files;
651                         if (!f->io_size)
652                                 f->io_size = f->real_file_size - f->file_offset;
653                 } else if (f->real_file_size < td->o.file_size_low ||
654                            f->real_file_size > td->o.file_size_high) {
655                         if (f->file_offset > td->o.file_size_low)
656                                 goto err_offset;
657                         /*
658                          * file size given. if it's fixed, use that. if it's a
659                          * range, generate a random size in-between.
660                          */
661                         if (td->o.file_size_low == td->o.file_size_high) {
662                                 f->io_size = td->o.file_size_low
663                                                 - f->file_offset;
664                         } else {
665                                 f->io_size = get_rand_file_size(td)
666                                                 - f->file_offset;
667                         }
668                 } else
669                         f->io_size = f->real_file_size - f->file_offset;
670
671                 if (f->io_size == -1ULL)
672                         total_size = -1ULL;
673                 else
674                         total_size += f->io_size;
675
676                 if (f->filetype == FIO_TYPE_FILE &&
677                     (f->io_size + f->file_offset) > f->real_file_size &&
678                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
679                         if (!td->o.create_on_open) {
680                                 need_extend++;
681                                 extend_size += (f->io_size + f->file_offset);
682                         } else
683                                 f->real_file_size = f->io_size + f->file_offset;
684                         fio_file_set_extend(f);
685                 }
686         }
687
688         if (!td->o.size || td->o.size > total_size)
689                 td->o.size = total_size;
690
691         /*
692          * See if we need to extend some files
693          */
694         if (need_extend) {
695                 temp_stall_ts = 1;
696                 if (!terse_output)
697                         log_info("%s: Laying out IO file(s) (%u file(s) /"
698                                  " %LuMB)\n", td->o.name, need_extend,
699                                         extend_size >> 20);
700
701                 for_each_file(td, f, i) {
702                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
703
704                         if (!fio_file_extend(f))
705                                 continue;
706
707                         assert(f->filetype == FIO_TYPE_FILE);
708                         fio_file_clear_extend(f);
709                         if (!td->o.fill_device) {
710                                 old_len = f->real_file_size;
711                                 extend_len = f->io_size + f->file_offset -
712                                                 old_len;
713                         }
714                         f->real_file_size = (f->io_size + f->file_offset);
715                         err = extend_file(td, f);
716                         if (err)
717                                 break;
718
719                         err = __file_invalidate_cache(td, f, old_len,
720                                                                 extend_len);
721                         close(f->fd);
722                         f->fd = -1;
723                         if (err)
724                                 break;
725                 }
726                 temp_stall_ts = 0;
727         }
728
729         if (err)
730                 return err;
731
732         if (!td->o.zone_size)
733                 td->o.zone_size = td->o.size;
734
735         /*
736          * iolog already set the total io size, if we read back
737          * stored entries.
738          */
739         if (!td->o.read_iolog_file)
740                 td->total_io_size = td->o.size * td->o.loops;
741         return 0;
742 err_offset:
743         log_err("%s: you need to specify valid offset=\n", td->o.name);
744         return 1;
745 }
746
747 int pre_read_files(struct thread_data *td)
748 {
749         struct fio_file *f;
750         unsigned int i;
751
752         dprint(FD_FILE, "pre_read files\n");
753
754         for_each_file(td, f, i) {
755                 pre_read_file(td, f);
756         }
757
758         return 1;
759 }
760
761 int init_random_map(struct thread_data *td)
762 {
763         unsigned long long blocks, num_maps;
764         struct fio_file *f;
765         unsigned int i;
766
767         if (td->o.norandommap || !td_random(td))
768                 return 0;
769
770         for_each_file(td, f, i) {
771                 blocks = (f->real_file_size + td->o.rw_min_bs - 1) /
772                                 (unsigned long long) td->o.rw_min_bs;
773                 num_maps = (blocks + BLOCKS_PER_MAP - 1) /
774                                 (unsigned long long) BLOCKS_PER_MAP;
775                 f->file_map = smalloc(num_maps * sizeof(int));
776                 if (f->file_map) {
777                         f->num_maps = num_maps;
778                         continue;
779                 }
780                 if (!td->o.softrandommap) {
781                         log_err("fio: failed allocating random map. If running"
782                                 " a large number of jobs, try the 'norandommap'"
783                                 " option or set 'softrandommap'. Or give"
784                                 " a larger --alloc-size to fio.\n");
785                         return 1;
786                 }
787
788                 log_info("fio: file %s failed allocating random map. Running "
789                          "job without.\n", f->file_name);
790                 f->num_maps = 0;
791         }
792
793         return 0;
794 }
795
796 void close_files(struct thread_data *td)
797 {
798         struct fio_file *f;
799         unsigned int i;
800
801         for_each_file(td, f, i) {
802                 if (fio_file_open(f))
803                         td_io_close_file(td, f);
804         }
805 }
806
807 void close_and_free_files(struct thread_data *td)
808 {
809         struct fio_file *f;
810         unsigned int i;
811
812         dprint(FD_FILE, "close files\n");
813
814         for_each_file(td, f, i) {
815                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
816                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
817                         unlink(f->file_name);
818                 }
819
820                 if (fio_file_open(f))
821                         td_io_close_file(td, f);
822
823                 remove_file_hash(f);
824
825                 sfree(f->file_name);
826                 f->file_name = NULL;
827                 sfree(f->file_map);
828                 f->file_map = NULL;
829                 sfree(f);
830         }
831
832         td->o.filename = NULL;
833         free(td->files);
834         td->files_index = 0;
835         td->files = NULL;
836         td->o.nr_files = 0;
837 }
838
839 static void get_file_type(struct fio_file *f)
840 {
841         struct stat sb;
842
843         if (!strcmp(f->file_name, "-"))
844                 f->filetype = FIO_TYPE_PIPE;
845         else
846                 f->filetype = FIO_TYPE_FILE;
847
848         if (!stat(f->file_name, &sb)) {
849                 if (S_ISBLK(sb.st_mode))
850                         f->filetype = FIO_TYPE_BD;
851                 else if (S_ISCHR(sb.st_mode))
852                         f->filetype = FIO_TYPE_CHAR;
853                 else if (S_ISFIFO(sb.st_mode))
854                         f->filetype = FIO_TYPE_PIPE;
855         }
856 }
857
858 int add_file(struct thread_data *td, const char *fname)
859 {
860         int cur_files = td->files_index;
861         char file_name[PATH_MAX];
862         struct fio_file *f;
863         int len = 0;
864
865         dprint(FD_FILE, "add file %s\n", fname);
866
867         f = smalloc(sizeof(*f));
868         if (!f) {
869                 log_err("fio: smalloc OOM\n");
870                 assert(0);
871         }
872
873         f->fd = -1;
874
875         if (td->files_size <= td->files_index) {
876                 int new_size = td->o.nr_files + 1;
877
878                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
879
880                 td->files = realloc(td->files, new_size * sizeof(f));
881                 td->files_size = new_size;
882         }
883         td->files[cur_files] = f;
884
885         /*
886          * init function, io engine may not be loaded yet
887          */
888         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
889                 f->real_file_size = -1ULL;
890
891         if (td->o.directory)
892                 len = sprintf(file_name, "%s/", td->o.directory);
893
894         sprintf(file_name + len, "%s", fname);
895         f->file_name = smalloc_strdup(file_name);
896         if (!f->file_name) {
897                 log_err("fio: smalloc OOM\n");
898                 assert(0);
899         }
900
901         get_file_type(f);
902
903         switch (td->o.file_lock_mode) {
904         case FILE_LOCK_NONE:
905                 break;
906         case FILE_LOCK_READWRITE:
907                 f->lock = fio_mutex_rw_init();
908                 break;
909         case FILE_LOCK_EXCLUSIVE:
910                 f->lock = fio_mutex_init(1);
911                 break;
912         default:
913                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
914                 assert(0);
915         }
916
917         td->files_index++;
918         if (f->filetype == FIO_TYPE_FILE)
919                 td->nr_normal_files++;
920
921         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
922                                                         cur_files);
923
924         return cur_files;
925 }
926
927 void get_file(struct fio_file *f)
928 {
929         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
930         assert(fio_file_open(f));
931         f->references++;
932 }
933
934 int put_file(struct thread_data *td, struct fio_file *f)
935 {
936         int f_ret = 0, ret = 0;
937
938         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
939
940         if (!fio_file_open(f)) {
941                 assert(f->fd == -1);
942                 return 0;
943         }
944
945         assert(f->references);
946         if (--f->references)
947                 return 0;
948
949         if (should_fsync(td) && td->o.fsync_on_close)
950                 f_ret = fsync(f->fd);
951
952         if (td->io_ops->close_file)
953                 ret = td->io_ops->close_file(td, f);
954
955         if (!ret)
956                 ret = f_ret;
957
958         td->nr_open_files--;
959         fio_file_clear_open(f);
960         assert(f->fd == -1);
961         return ret;
962 }
963
964 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
965 {
966         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
967                 return;
968
969         if (f->lock_owner == td && f->lock_batch--)
970                 return;
971
972         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
973                 if (ddir == DDIR_READ)
974                         fio_mutex_down_read(f->lock);
975                 else
976                         fio_mutex_down_write(f->lock);
977         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
978                 fio_mutex_down(f->lock);
979
980         f->lock_owner = td;
981         f->lock_batch = td->o.lockfile_batch;
982         f->lock_ddir = ddir;
983 }
984
985 void unlock_file(struct thread_data *td, struct fio_file *f)
986 {
987         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
988                 return;
989         if (f->lock_batch)
990                 return;
991
992         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
993                 const int is_read = f->lock_ddir == DDIR_READ;
994                 int val = fio_mutex_getval(f->lock);
995
996                 if ((is_read && val == 1) || (!is_read && val == -1))
997                         f->lock_owner = NULL;
998
999                 if (is_read)
1000                         fio_mutex_up_read(f->lock);
1001                 else
1002                         fio_mutex_up_write(f->lock);
1003         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) {
1004                 int val = fio_mutex_getval(f->lock);
1005
1006                 if (val == 0)
1007                         f->lock_owner = NULL;
1008
1009                 fio_mutex_up(f->lock);
1010         }
1011 }
1012
1013 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1014 {
1015         if (f->lock_owner != td)
1016                 return;
1017
1018         f->lock_batch = 0;
1019         unlock_file(td, f);
1020 }
1021
1022 static int recurse_dir(struct thread_data *td, const char *dirname)
1023 {
1024         struct dirent *dir;
1025         int ret = 0;
1026         DIR *D;
1027
1028         D = opendir(dirname);
1029         if (!D) {
1030                 char buf[FIO_VERROR_SIZE];
1031
1032                 snprintf(buf, FIO_VERROR_SIZE - 1, "opendir(%s)", dirname);
1033                 td_verror(td, errno, buf);
1034                 return 1;
1035         }
1036
1037         while ((dir = readdir(D)) != NULL) {
1038                 char full_path[PATH_MAX];
1039                 struct stat sb;
1040
1041                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1042                         continue;
1043
1044                 sprintf(full_path, "%s/%s", dirname, dir->d_name);
1045
1046                 if (lstat(full_path, &sb) == -1) {
1047                         if (errno != ENOENT) {
1048                                 td_verror(td, errno, "stat");
1049                                 return 1;
1050                         }
1051                 }
1052
1053                 if (S_ISREG(sb.st_mode)) {
1054                         add_file(td, full_path);
1055                         td->o.nr_files++;
1056                         continue;
1057                 }
1058                 if (!S_ISDIR(sb.st_mode))
1059                         continue;
1060
1061                 ret = recurse_dir(td, full_path);
1062                 if (ret)
1063                         break;
1064         }
1065
1066         closedir(D);
1067         return ret;
1068 }
1069
1070 int add_dir_files(struct thread_data *td, const char *path)
1071 {
1072         int ret = recurse_dir(td, path);
1073
1074         if (!ret)
1075                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1076
1077         return ret;
1078 }
1079
1080 void dup_files(struct thread_data *td, struct thread_data *org)
1081 {
1082         struct fio_file *f;
1083         unsigned int i;
1084
1085         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1086
1087         if (!org->files)
1088                 return;
1089
1090         td->files = malloc(org->files_index * sizeof(f));
1091
1092         for_each_file(org, f, i) {
1093                 struct fio_file *__f;
1094
1095                 __f = smalloc(sizeof(*__f));
1096                 if (!__f) {
1097                         log_err("fio: smalloc OOM\n");
1098                         assert(0);
1099                 }
1100                 __f->fd = -1;
1101
1102                 if (f->file_name) {
1103                         __f->file_name = smalloc_strdup(f->file_name);
1104                         if (!__f->file_name) {
1105                                 log_err("fio: smalloc OOM\n");
1106                                 assert(0);
1107                         }
1108
1109                         __f->filetype = f->filetype;
1110                 }
1111
1112                 td->files[i] = __f;
1113         }
1114 }
1115
1116 /*
1117  * Returns the index that matches the filename, or -1 if not there
1118  */
1119 int get_fileno(struct thread_data *td, const char *fname)
1120 {
1121         struct fio_file *f;
1122         unsigned int i;
1123
1124         for_each_file(td, f, i)
1125                 if (!strcmp(f->file_name, fname))
1126                         return i;
1127
1128         return -1;
1129 }
1130
1131 /*
1132  * For log usage, where we add/open/close files automatically
1133  */
1134 void free_release_files(struct thread_data *td)
1135 {
1136         close_files(td);
1137         td->files_index = 0;
1138         td->nr_normal_files = 0;
1139 }