filesetup.c revision dfe11fd1e8249a52c42f34c9e9e6b910ab6b05c5
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
23static int root_warn;
24
25static FLIST_HEAD(filename_list);
26
27static 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 */
36static 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;
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 ((unlink(f->file_name) < 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	dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
73	f->fd = open(f->file_name, flags, 0644);
74	if (f->fd < 0) {
75		td_verror(td, errno, "open");
76		return 1;
77	}
78
79#ifdef CONFIG_POSIX_FALLOCATE
80	if (!td->o.fill_device) {
81		switch (td->o.fallocate_mode) {
82		case FIO_FALLOCATE_NONE:
83			break;
84		case FIO_FALLOCATE_POSIX:
85			dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
86				 f->file_name,
87				 (unsigned long long) f->real_file_size);
88
89			r = posix_fallocate(f->fd, 0, f->real_file_size);
90			if (r > 0) {
91				log_err("fio: posix_fallocate fails: %s\n",
92						strerror(r));
93			}
94			break;
95#ifdef CONFIG_LINUX_FALLOCATE
96		case FIO_FALLOCATE_KEEP_SIZE:
97			dprint(FD_FILE,
98				"fallocate(FALLOC_FL_KEEP_SIZE) "
99				"file %s size %llu\n", f->file_name,
100				(unsigned long long) f->real_file_size);
101
102			r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
103					f->real_file_size);
104			if (r != 0)
105				td_verror(td, errno, "fallocate");
106
107			break;
108#endif /* CONFIG_LINUX_FALLOCATE */
109		default:
110			log_err("fio: unknown fallocate mode: %d\n",
111				td->o.fallocate_mode);
112			assert(0);
113		}
114	}
115#endif /* CONFIG_POSIX_FALLOCATE */
116
117	if (!new_layout)
118		goto done;
119
120	/*
121	 * The size will be -1ULL when fill_device is used, so don't truncate
122	 * or fallocate this file, just write it
123	 */
124	if (!td->o.fill_device) {
125		dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
126					(unsigned long long) f->real_file_size);
127		if (ftruncate(f->fd, f->real_file_size) == -1) {
128			if (errno != EFBIG) {
129				td_verror(td, errno, "ftruncate");
130				goto err;
131			}
132		}
133	}
134
135	b = malloc(td->o.max_bs[DDIR_WRITE]);
136
137	left = f->real_file_size;
138	while (left && !td->terminate) {
139		bs = td->o.max_bs[DDIR_WRITE];
140		if (bs > left)
141			bs = left;
142
143		fill_io_buffer(td, b, bs, bs);
144
145		r = write(f->fd, b, bs);
146
147		if (r > 0) {
148			left -= r;
149			continue;
150		} else {
151			if (r < 0) {
152				int __e = errno;
153
154				if (__e == ENOSPC) {
155					if (td->o.fill_device)
156						break;
157					log_info("fio: ENOSPC on laying out "
158						 "file, stopping\n");
159					break;
160				}
161				td_verror(td, errno, "write");
162			} else
163				td_verror(td, EIO, "write");
164
165			break;
166		}
167	}
168
169	if (td->terminate) {
170		dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
171		unlink(f->file_name);
172	} else if (td->o.create_fsync) {
173		if (fsync(f->fd) < 0) {
174			td_verror(td, errno, "fsync");
175			goto err;
176		}
177	}
178	if (td->o.fill_device && !td_write(td)) {
179		fio_file_clear_size_known(f);
180		if (td_io_get_file_size(td, f))
181			goto err;
182		if (f->io_size > f->real_file_size)
183			f->io_size = f->real_file_size;
184	}
185
186	free(b);
187done:
188	return 0;
189err:
190	close(f->fd);
191	f->fd = -1;
192	return 1;
193}
194
195static int pre_read_file(struct thread_data *td, struct fio_file *f)
196{
197	int ret = 0, r, did_open = 0, old_runstate;
198	unsigned long long left;
199	unsigned int bs;
200	char *b;
201
202	if (td->io_ops->flags & FIO_PIPEIO)
203		return 0;
204
205	if (!fio_file_open(f)) {
206		if (td->io_ops->open_file(td, f)) {
207			log_err("fio: cannot pre-read, failed to open file\n");
208			return 1;
209		}
210		did_open = 1;
211	}
212
213	old_runstate = td_bump_runstate(td, TD_PRE_READING);
214
215	bs = td->o.max_bs[DDIR_READ];
216	b = malloc(bs);
217	memset(b, 0, bs);
218
219	if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
220		td_verror(td, errno, "lseek");
221		log_err("fio: failed to lseek pre-read file\n");
222		ret = 1;
223		goto error;
224	}
225
226	left = f->io_size;
227
228	while (left && !td->terminate) {
229		if (bs > left)
230			bs = left;
231
232		r = read(f->fd, b, bs);
233
234		if (r == (int) bs) {
235			left -= bs;
236			continue;
237		} else {
238			td_verror(td, EIO, "pre_read");
239			break;
240		}
241	}
242
243error:
244	td_restore_runstate(td, old_runstate);
245
246	if (did_open)
247		td->io_ops->close_file(td, f);
248
249	free(b);
250	return ret;
251}
252
253static unsigned long long get_rand_file_size(struct thread_data *td)
254{
255	unsigned long long ret, sized;
256	unsigned long r;
257
258	if (td->o.use_os_rand) {
259		r = os_random_long(&td->file_size_state);
260		sized = td->o.file_size_high - td->o.file_size_low;
261		ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
262	} else {
263		r = __rand(&td->__file_size_state);
264		sized = td->o.file_size_high - td->o.file_size_low;
265		ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
266	}
267
268	ret += td->o.file_size_low;
269	ret -= (ret % td->o.rw_min_bs);
270	return ret;
271}
272
273static int file_size(struct thread_data *td, struct fio_file *f)
274{
275	struct stat st;
276
277	if (stat(f->file_name, &st) == -1) {
278		td_verror(td, errno, "fstat");
279		return 1;
280	}
281
282	f->real_file_size = st.st_size;
283	return 0;
284}
285
286static int bdev_size(struct thread_data *td, struct fio_file *f)
287{
288	unsigned long long bytes = 0;
289	int r;
290
291	if (td->io_ops->open_file(td, f)) {
292		log_err("fio: failed opening blockdev %s for size check\n",
293			f->file_name);
294		return 1;
295	}
296
297	r = blockdev_size(f, &bytes);
298	if (r) {
299		td_verror(td, r, "blockdev_size");
300		goto err;
301	}
302
303	if (!bytes) {
304		log_err("%s: zero sized block device?\n", f->file_name);
305		goto err;
306	}
307
308	f->real_file_size = bytes;
309	td->io_ops->close_file(td, f);
310	return 0;
311err:
312	td->io_ops->close_file(td, f);
313	return 1;
314}
315
316static int char_size(struct thread_data *td, struct fio_file *f)
317{
318#ifdef FIO_HAVE_CHARDEV_SIZE
319	unsigned long long bytes = 0;
320	int r;
321
322	if (td->io_ops->open_file(td, f)) {
323		log_err("fio: failed opening blockdev %s for size check\n",
324			f->file_name);
325		return 1;
326	}
327
328	r = chardev_size(f, &bytes);
329	if (r) {
330		td_verror(td, r, "chardev_size");
331		goto err;
332	}
333
334	if (!bytes) {
335		log_err("%s: zero sized char device?\n", f->file_name);
336		goto err;
337	}
338
339	f->real_file_size = bytes;
340	td->io_ops->close_file(td, f);
341	return 0;
342err:
343	td->io_ops->close_file(td, f);
344	return 1;
345#else
346	f->real_file_size = -1ULL;
347	return 0;
348#endif
349}
350
351static int get_file_size(struct thread_data *td, struct fio_file *f)
352{
353	int ret = 0;
354
355	if (fio_file_size_known(f))
356		return 0;
357
358	if (f->filetype == FIO_TYPE_FILE)
359		ret = file_size(td, f);
360	else if (f->filetype == FIO_TYPE_BD)
361		ret = bdev_size(td, f);
362	else if (f->filetype == FIO_TYPE_CHAR)
363		ret = char_size(td, f);
364	else
365		f->real_file_size = -1;
366
367	if (ret)
368		return ret;
369
370	if (f->file_offset > f->real_file_size) {
371		log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
372					(unsigned long long) f->file_offset,
373					(unsigned long long) f->real_file_size);
374		return 1;
375	}
376
377	fio_file_set_size_known(f);
378	return 0;
379}
380
381static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
382				   unsigned long long off,
383				   unsigned long long len)
384{
385	int ret = 0;
386
387	if (len == -1ULL)
388		len = f->io_size;
389	if (off == -1ULL)
390		off = f->file_offset;
391
392	if (len == -1ULL || off == -1ULL)
393		return 0;
394
395	dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
396								len);
397
398	if (f->mmap_ptr) {
399		ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
400#ifdef FIO_MADV_FREE
401		if (f->filetype == FIO_TYPE_BD)
402			(void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
403#endif
404	} else if (f->filetype == FIO_TYPE_FILE) {
405		ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
406	} else if (f->filetype == FIO_TYPE_BD) {
407		ret = blockdev_invalidate_cache(f);
408		if (ret < 0 && errno == EACCES && geteuid()) {
409			if (!root_warn) {
410				log_err("fio: only root may flush block "
411					"devices. Cache flush bypassed!\n");
412				root_warn = 1;
413			}
414			ret = 0;
415		}
416	} else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
417		ret = 0;
418
419	/*
420	 * Cache flushing isn't a fatal condition, and we know it will
421	 * happen on some platforms where we don't have the proper
422	 * function to flush eg block device caches. So just warn and
423	 * continue on our way.
424	 */
425	if (ret) {
426		log_info("fio: cache invalidation of %s failed: %s\n", f->file_name, strerror(errno));
427		ret = 0;
428	}
429
430	return 0;
431
432}
433
434int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
435{
436	if (!fio_file_open(f))
437		return 0;
438
439	return __file_invalidate_cache(td, f, -1ULL, -1ULL);
440}
441
442int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
443{
444	int ret = 0;
445
446	dprint(FD_FILE, "fd close %s\n", f->file_name);
447
448	remove_file_hash(f);
449
450	if (close(f->fd) < 0)
451		ret = errno;
452
453	f->fd = -1;
454
455	if (f->shadow_fd != -1) {
456		close(f->shadow_fd);
457		f->shadow_fd = -1;
458	}
459
460	f->engine_data = 0;
461	return ret;
462}
463
464int file_lookup_open(struct fio_file *f, int flags)
465{
466	struct fio_file *__f;
467	int from_hash;
468
469	__f = lookup_file_hash(f->file_name);
470	if (__f) {
471		dprint(FD_FILE, "found file in hash %s\n", f->file_name);
472		/*
473		 * racy, need the __f->lock locked
474		 */
475		f->lock = __f->lock;
476		from_hash = 1;
477	} else {
478		dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
479		from_hash = 0;
480	}
481
482	f->fd = open(f->file_name, flags, 0600);
483	return from_hash;
484}
485
486static int file_close_shadow_fds(struct thread_data *td)
487{
488	struct fio_file *f;
489	int num_closed = 0;
490	unsigned int i;
491
492	for_each_file(td, f, i) {
493		if (f->shadow_fd == -1)
494			continue;
495
496		close(f->shadow_fd);
497		f->shadow_fd = -1;
498		num_closed++;
499	}
500
501	return num_closed;
502}
503
504int generic_open_file(struct thread_data *td, struct fio_file *f)
505{
506	int is_std = 0;
507	int flags = 0;
508	int from_hash = 0;
509
510	dprint(FD_FILE, "fd open %s\n", f->file_name);
511
512	if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
513		log_err("fio: trim only applies to block device\n");
514		return 1;
515	}
516
517	if (!strcmp(f->file_name, "-")) {
518		if (td_rw(td)) {
519			log_err("fio: can't read/write to stdin/out\n");
520			return 1;
521		}
522		is_std = 1;
523
524		/*
525		 * move output logging to stderr, if we are writing to stdout
526		 */
527		if (td_write(td))
528			f_out = stderr;
529	}
530
531	if (td_trim(td))
532		goto skip_flags;
533	if (td->o.odirect)
534		flags |= OS_O_DIRECT;
535	if (td->o.oatomic) {
536		if (!FIO_O_ATOMIC) {
537			td_verror(td, EINVAL, "OS does not support atomic IO");
538			return 1;
539		}
540		flags |= OS_O_DIRECT | FIO_O_ATOMIC;
541	}
542	if (td->o.sync_io)
543		flags |= O_SYNC;
544	if (td->o.create_on_open)
545		flags |= O_CREAT;
546skip_flags:
547	if (f->filetype != FIO_TYPE_FILE)
548		flags |= FIO_O_NOATIME;
549
550open_again:
551	if (td_write(td)) {
552		if (!read_only)
553			flags |= O_RDWR;
554
555		if (f->filetype == FIO_TYPE_FILE)
556			flags |= O_CREAT;
557
558		if (is_std)
559			f->fd = dup(STDOUT_FILENO);
560		else
561			from_hash = file_lookup_open(f, flags);
562	} else if (td_read(td)) {
563		if (f->filetype == FIO_TYPE_CHAR && !read_only)
564			flags |= O_RDWR;
565		else
566			flags |= O_RDONLY;
567
568		if (is_std)
569			f->fd = dup(STDIN_FILENO);
570		else
571			from_hash = file_lookup_open(f, flags);
572	} else { //td trim
573		flags |= O_RDWR;
574		from_hash = file_lookup_open(f, flags);
575	}
576
577	if (f->fd == -1) {
578		char buf[FIO_VERROR_SIZE];
579		int __e = errno;
580
581		if (__e == EPERM && (flags & FIO_O_NOATIME)) {
582			flags &= ~FIO_O_NOATIME;
583			goto open_again;
584		}
585		if (__e == EMFILE && file_close_shadow_fds(td))
586			goto open_again;
587
588		snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
589
590		if (__e == EINVAL && (flags & OS_O_DIRECT)) {
591			log_err("fio: looks like your file system does not " \
592				"support direct=1/buffered=0\n");
593		}
594
595		td_verror(td, __e, buf);
596	}
597
598	if (!from_hash && f->fd != -1) {
599		if (add_file_hash(f)) {
600			int fio_unused ret;
601
602			/*
603			 * Stash away descriptor for later close. This is to
604			 * work-around a "feature" on Linux, where a close of
605			 * an fd that has been opened for write will trigger
606			 * udev to call blkid to check partitions, fs id, etc.
607			 * That pollutes the device cache, which can slow down
608			 * unbuffered accesses.
609			 */
610			if (f->shadow_fd == -1)
611				f->shadow_fd = f->fd;
612			else {
613				/*
614			 	 * OK to ignore, we haven't done anything
615				 * with it
616				 */
617				ret = generic_close_file(td, f);
618			}
619			goto open_again;
620		}
621	}
622
623	return 0;
624}
625
626int generic_get_file_size(struct thread_data *td, struct fio_file *f)
627{
628	return get_file_size(td, f);
629}
630
631/*
632 * open/close all files, so that ->real_file_size gets set
633 */
634static int get_file_sizes(struct thread_data *td)
635{
636	struct fio_file *f;
637	unsigned int i;
638	int err = 0;
639
640	for_each_file(td, f, i) {
641		dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
642								f->file_name);
643
644		if (td_io_get_file_size(td, f)) {
645			if (td->error != ENOENT) {
646				log_err("%s\n", td->verror);
647				err = 1;
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
659struct 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 */
669static 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		strcpy(buf, f->file_name);
690
691		if (stat(buf, &sb) < 0) {
692			if (errno != ENOENT)
693				break;
694			strcpy(buf, ".");
695			if (stat(buf, &sb) < 0)
696				break;
697		}
698
699		fm = NULL;
700		flist_for_each(n, &list) {
701			fm = flist_entry(n, struct fio_mount, list);
702			if (fm->key == sb.st_dev)
703				break;
704
705			fm = NULL;
706		}
707
708		if (fm)
709			continue;
710
711		fm = malloc(sizeof(*fm));
712		strcpy(fm->__base, buf);
713		fm->base = basename(fm->__base);
714		fm->key = sb.st_dev;
715		flist_add(&fm->list, &list);
716	}
717
718	flist_for_each_safe(n, tmp, &list) {
719		unsigned long long sz;
720
721		fm = flist_entry(n, struct fio_mount, list);
722		flist_del(&fm->list);
723
724		sz = get_fs_size(fm->base);
725		if (sz && sz != -1ULL)
726			ret += sz;
727
728		free(fm);
729	}
730
731	return ret;
732}
733
734uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
735{
736	struct thread_options *o = &td->o;
737
738	if (o->file_append && f->filetype == FIO_TYPE_FILE)
739		return f->real_file_size;
740
741	return td->o.start_offset +
742		(td->thread_number - 1) * td->o.offset_increment;
743}
744
745/*
746 * Open the files and setup files sizes, creating files if necessary.
747 */
748int setup_files(struct thread_data *td)
749{
750	unsigned long long total_size, extend_size;
751	struct thread_options *o = &td->o;
752	struct fio_file *f;
753	unsigned int i, nr_fs_extra = 0;
754	int err = 0, need_extend;
755	int old_state;
756	const unsigned int bs = td_min_bs(td);
757	uint64_t fs = 0;
758
759	dprint(FD_FILE, "setup files\n");
760
761	old_state = td_bump_runstate(td, TD_SETTING_UP);
762
763	if (o->read_iolog_file)
764		goto done;
765
766	/*
767	 * if ioengine defines a setup() method, it's responsible for
768	 * opening the files and setting f->real_file_size to indicate
769	 * the valid range for that file.
770	 */
771	if (td->io_ops->setup)
772		err = td->io_ops->setup(td);
773	else
774		err = get_file_sizes(td);
775
776	if (err)
777		goto err_out;
778
779	/*
780	 * check sizes. if the files/devices do not exist and the size
781	 * isn't passed to fio, abort.
782	 */
783	total_size = 0;
784	for_each_file(td, f, i) {
785		if (f->real_file_size == -1ULL)
786			total_size = -1ULL;
787		else
788			total_size += f->real_file_size;
789	}
790
791	if (o->fill_device)
792		td->fill_device_size = get_fs_free_counts(td);
793
794	/*
795	 * device/file sizes are zero and no size given, punt
796	 */
797	if ((!total_size || total_size == -1ULL) && !o->size &&
798	    !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
799	    !(o->nr_files && (o->file_size_low || o->file_size_high))) {
800		log_err("%s: you need to specify size=\n", o->name);
801		td_verror(td, EINVAL, "total_file_size");
802		goto err_out;
803	}
804
805	/*
806	 * Calculate per-file size and potential extra size for the
807	 * first files, if needed.
808	 */
809	if (!o->file_size_low && o->nr_files) {
810		uint64_t all_fs;
811
812		fs = o->size / o->nr_files;
813		all_fs = fs * o->nr_files;
814
815		if (all_fs < o->size)
816			nr_fs_extra = (o->size - all_fs) / bs;
817	}
818
819	/*
820	 * now file sizes are known, so we can set ->io_size. if size= is
821	 * not given, ->io_size is just equal to ->real_file_size. if size
822	 * is given, ->io_size is size / nr_files.
823	 */
824	extend_size = total_size = 0;
825	need_extend = 0;
826	for_each_file(td, f, i) {
827		f->file_offset = get_start_offset(td, f);
828
829		if (!o->file_size_low) {
830			/*
831			 * no file size range given, file size is equal to
832			 * total size divided by number of files. If that is
833			 * zero, set it to the real file size. If the size
834			 * doesn't divide nicely with the min blocksize,
835			 * make the first files bigger.
836			 */
837			f->io_size = fs;
838			if (nr_fs_extra) {
839				nr_fs_extra--;
840				f->io_size += bs;
841			}
842
843			if (!f->io_size)
844				f->io_size = f->real_file_size - f->file_offset;
845		} else if (f->real_file_size < o->file_size_low ||
846			   f->real_file_size > o->file_size_high) {
847			if (f->file_offset > o->file_size_low)
848				goto err_offset;
849			/*
850			 * file size given. if it's fixed, use that. if it's a
851			 * range, generate a random size in-between.
852			 */
853			if (o->file_size_low == o->file_size_high)
854				f->io_size = o->file_size_low - f->file_offset;
855			else {
856				f->io_size = get_rand_file_size(td)
857						- f->file_offset;
858			}
859		} else
860			f->io_size = f->real_file_size - f->file_offset;
861
862		if (f->io_size == -1ULL)
863			total_size = -1ULL;
864		else {
865                        if (o->size_percent)
866                                f->io_size = (f->io_size * o->size_percent) / 100;
867			total_size += f->io_size;
868		}
869
870		if (f->filetype == FIO_TYPE_FILE &&
871		    (f->io_size + f->file_offset) > f->real_file_size &&
872		    !(td->io_ops->flags & FIO_DISKLESSIO)) {
873			if (!o->create_on_open) {
874				need_extend++;
875				extend_size += (f->io_size + f->file_offset);
876			} else
877				f->real_file_size = f->io_size + f->file_offset;
878			fio_file_set_extend(f);
879		}
880	}
881
882	if (!o->size || o->size > total_size)
883		o->size = total_size;
884
885	/*
886	 * See if we need to extend some files
887	 */
888	if (need_extend) {
889		temp_stall_ts = 1;
890		if (output_format == FIO_OUTPUT_NORMAL)
891			log_info("%s: Laying out IO file(s) (%u file(s) /"
892				 " %lluMB)\n", o->name, need_extend,
893					extend_size >> 20);
894
895		for_each_file(td, f, i) {
896			unsigned long long old_len = -1ULL, extend_len = -1ULL;
897
898			if (!fio_file_extend(f))
899				continue;
900
901			assert(f->filetype == FIO_TYPE_FILE);
902			fio_file_clear_extend(f);
903			if (!o->fill_device) {
904				old_len = f->real_file_size;
905				extend_len = f->io_size + f->file_offset -
906						old_len;
907			}
908			f->real_file_size = (f->io_size + f->file_offset);
909			err = extend_file(td, f);
910			if (err)
911				break;
912
913			err = __file_invalidate_cache(td, f, old_len,
914								extend_len);
915			close(f->fd);
916			f->fd = -1;
917			if (err)
918				break;
919		}
920		temp_stall_ts = 0;
921	}
922
923	if (err)
924		goto err_out;
925
926	if (!o->zone_size)
927		o->zone_size = o->size;
928
929	/*
930	 * iolog already set the total io size, if we read back
931	 * stored entries.
932	 */
933	if (!o->read_iolog_file)
934		td->total_io_size = o->size * o->loops;
935
936done:
937	if (o->create_only)
938		td->done = 1;
939
940	td_restore_runstate(td, old_state);
941	return 0;
942err_offset:
943	log_err("%s: you need to specify valid offset=\n", o->name);
944err_out:
945	td_restore_runstate(td, old_state);
946	return 1;
947}
948
949int pre_read_files(struct thread_data *td)
950{
951	struct fio_file *f;
952	unsigned int i;
953
954	dprint(FD_FILE, "pre_read files\n");
955
956	for_each_file(td, f, i) {
957		pre_read_file(td, f);
958	}
959
960	return 1;
961}
962
963static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
964{
965	unsigned int range_size, seed;
966	unsigned long nranges;
967	uint64_t file_size;
968
969	range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
970	file_size = min(f->real_file_size, f->io_size);
971
972	nranges = (file_size + range_size - 1) / range_size;
973
974	seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
975	if (!td->o.rand_repeatable)
976		seed = td->rand_seeds[4];
977
978	if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
979		zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
980	else
981		pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
982
983	return 1;
984}
985
986static int init_rand_distribution(struct thread_data *td)
987{
988	struct fio_file *f;
989	unsigned int i;
990	int state;
991
992	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
993		return 0;
994
995	state = td_bump_runstate(td, TD_SETTING_UP);
996
997	for_each_file(td, f, i)
998		__init_rand_distribution(td, f);
999
1000	td_restore_runstate(td, state);
1001
1002	return 1;
1003}
1004
1005int init_random_map(struct thread_data *td)
1006{
1007	unsigned long long blocks;
1008	struct fio_file *f;
1009	unsigned int i;
1010
1011	if (init_rand_distribution(td))
1012		return 0;
1013	if (!td_random(td))
1014		return 0;
1015
1016	for_each_file(td, f, i) {
1017		uint64_t file_size = min(f->real_file_size, f->io_size);
1018
1019		blocks = file_size / (unsigned long long) td->o.rw_min_bs;
1020
1021		if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1022			unsigned long seed;
1023
1024			seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1025
1026			if (!lfsr_init(&f->lfsr, blocks, seed, 0))
1027				continue;
1028		} else if (!td->o.norandommap) {
1029			f->io_axmap = axmap_new(blocks);
1030			if (f->io_axmap)
1031				continue;
1032		} else if (td->o.norandommap)
1033			continue;
1034
1035		if (!td->o.softrandommap) {
1036			log_err("fio: failed allocating random map. If running"
1037				" a large number of jobs, try the 'norandommap'"
1038				" option or set 'softrandommap'. Or give"
1039				" a larger --alloc-size to fio.\n");
1040			return 1;
1041		}
1042
1043		log_info("fio: file %s failed allocating random map. Running "
1044			 "job without.\n", f->file_name);
1045	}
1046
1047	return 0;
1048}
1049
1050void close_files(struct thread_data *td)
1051{
1052	struct fio_file *f;
1053	unsigned int i;
1054
1055	for_each_file(td, f, i) {
1056		if (fio_file_open(f))
1057			td_io_close_file(td, f);
1058	}
1059}
1060
1061void close_and_free_files(struct thread_data *td)
1062{
1063	struct fio_file *f;
1064	unsigned int i;
1065
1066	dprint(FD_FILE, "close files\n");
1067
1068	for_each_file(td, f, i) {
1069		if (fio_file_open(f))
1070			td_io_close_file(td, f);
1071
1072		remove_file_hash(f);
1073
1074		if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1075			dprint(FD_FILE, "free unlink %s\n", f->file_name);
1076			unlink(f->file_name);
1077		}
1078
1079		sfree(f->file_name);
1080		f->file_name = NULL;
1081		axmap_free(f->io_axmap);
1082		f->io_axmap = NULL;
1083		sfree(f);
1084	}
1085
1086	td->o.filename = NULL;
1087	free(td->files);
1088	free(td->file_locks);
1089	td->files_index = 0;
1090	td->files = NULL;
1091	td->file_locks = NULL;
1092	td->o.file_lock_mode = FILE_LOCK_NONE;
1093	td->o.nr_files = 0;
1094}
1095
1096static void get_file_type(struct fio_file *f)
1097{
1098	struct stat sb;
1099
1100	if (!strcmp(f->file_name, "-"))
1101		f->filetype = FIO_TYPE_PIPE;
1102	else
1103		f->filetype = FIO_TYPE_FILE;
1104
1105	/* \\.\ is the device namespace in Windows, where every file is
1106	 * a block device */
1107	if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1108		f->filetype = FIO_TYPE_BD;
1109
1110	if (!stat(f->file_name, &sb)) {
1111		if (S_ISBLK(sb.st_mode))
1112			f->filetype = FIO_TYPE_BD;
1113		else if (S_ISCHR(sb.st_mode))
1114			f->filetype = FIO_TYPE_CHAR;
1115		else if (S_ISFIFO(sb.st_mode))
1116			f->filetype = FIO_TYPE_PIPE;
1117	}
1118}
1119
1120static int __is_already_allocated(const char *fname)
1121{
1122	struct flist_head *entry;
1123	char *filename;
1124
1125	if (flist_empty(&filename_list))
1126		return 0;
1127
1128	flist_for_each(entry, &filename_list) {
1129		filename = flist_entry(entry, struct file_name, list)->filename;
1130
1131		if (strcmp(filename, fname) == 0)
1132			return 1;
1133	}
1134
1135	return 0;
1136}
1137
1138static int is_already_allocated(const char *fname)
1139{
1140	int ret;
1141
1142	fio_file_hash_lock();
1143	ret = __is_already_allocated(fname);
1144	fio_file_hash_unlock();
1145	return ret;
1146}
1147
1148static void set_already_allocated(const char *fname)
1149{
1150	struct file_name *fn;
1151
1152	fn = malloc(sizeof(struct file_name));
1153	fn->filename = strdup(fname);
1154
1155	fio_file_hash_lock();
1156	if (!__is_already_allocated(fname)) {
1157		flist_add_tail(&fn->list, &filename_list);
1158		fn = NULL;
1159	}
1160	fio_file_hash_unlock();
1161
1162	if (fn) {
1163		free(fn->filename);
1164		free(fn);
1165	}
1166}
1167
1168
1169static void free_already_allocated(void)
1170{
1171	struct flist_head *entry, *tmp;
1172	struct file_name *fn;
1173
1174	if (flist_empty(&filename_list))
1175		return;
1176
1177	fio_file_hash_lock();
1178	flist_for_each_safe(entry, tmp, &filename_list) {
1179		fn = flist_entry(entry, struct file_name, list);
1180		free(fn->filename);
1181		flist_del(&fn->list);
1182		free(fn);
1183	}
1184
1185	fio_file_hash_unlock();
1186}
1187
1188static struct fio_file *alloc_new_file(struct thread_data *td)
1189{
1190	struct fio_file *f;
1191
1192	f = smalloc(sizeof(*f));
1193	if (!f) {
1194		log_err("fio: smalloc OOM\n");
1195		assert(0);
1196		return NULL;
1197	}
1198
1199	f->fd = -1;
1200	f->shadow_fd = -1;
1201	fio_file_reset(td, f);
1202	return f;
1203}
1204
1205int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1206{
1207	int cur_files = td->files_index;
1208	char file_name[PATH_MAX];
1209	struct fio_file *f;
1210	int len = 0;
1211
1212	dprint(FD_FILE, "add file %s\n", fname);
1213
1214	if (td->o.directory)
1215		len = set_name_idx(file_name, td->o.directory, numjob);
1216
1217	sprintf(file_name + len, "%s", fname);
1218
1219	/* clean cloned siblings using existing files */
1220	if (numjob && is_already_allocated(file_name))
1221		return 0;
1222
1223	f = alloc_new_file(td);
1224
1225	if (td->files_size <= td->files_index) {
1226		unsigned int new_size = td->o.nr_files + 1;
1227
1228		dprint(FD_FILE, "resize file array to %d files\n", new_size);
1229
1230		td->files = realloc(td->files, new_size * sizeof(f));
1231		if (td->files == NULL) {
1232			log_err("fio: realloc OOM\n");
1233			assert(0);
1234		}
1235		if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1236			td->file_locks = realloc(td->file_locks, new_size);
1237			if (!td->file_locks) {
1238				log_err("fio: realloc OOM\n");
1239				assert(0);
1240			}
1241			td->file_locks[cur_files] = FILE_LOCK_NONE;
1242		}
1243		td->files_size = new_size;
1244	}
1245	td->files[cur_files] = f;
1246	f->fileno = cur_files;
1247
1248	/*
1249	 * init function, io engine may not be loaded yet
1250	 */
1251	if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1252		f->real_file_size = -1ULL;
1253
1254	f->file_name = smalloc_strdup(file_name);
1255	if (!f->file_name) {
1256		log_err("fio: smalloc OOM\n");
1257		assert(0);
1258	}
1259
1260	get_file_type(f);
1261
1262	switch (td->o.file_lock_mode) {
1263	case FILE_LOCK_NONE:
1264		break;
1265	case FILE_LOCK_READWRITE:
1266		f->rwlock = fio_rwlock_init();
1267		break;
1268	case FILE_LOCK_EXCLUSIVE:
1269		f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1270		break;
1271	default:
1272		log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1273		assert(0);
1274	}
1275
1276	td->files_index++;
1277	if (f->filetype == FIO_TYPE_FILE)
1278		td->nr_normal_files++;
1279
1280	set_already_allocated(file_name);
1281
1282	if (inc)
1283		td->o.nr_files++;
1284
1285	dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1286							cur_files);
1287
1288	return cur_files;
1289}
1290
1291int add_file_exclusive(struct thread_data *td, const char *fname)
1292{
1293	struct fio_file *f;
1294	unsigned int i;
1295
1296	for_each_file(td, f, i) {
1297		if (!strcmp(f->file_name, fname))
1298			return i;
1299	}
1300
1301	return add_file(td, fname, 0, 1);
1302}
1303
1304void get_file(struct fio_file *f)
1305{
1306	dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1307	assert(fio_file_open(f));
1308	f->references++;
1309}
1310
1311int put_file(struct thread_data *td, struct fio_file *f)
1312{
1313	int f_ret = 0, ret = 0;
1314
1315	dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1316
1317	if (!fio_file_open(f)) {
1318		assert(f->fd == -1);
1319		return 0;
1320	}
1321
1322	assert(f->references);
1323	if (--f->references)
1324		return 0;
1325
1326	if (should_fsync(td) && td->o.fsync_on_close)
1327		f_ret = fsync(f->fd);
1328
1329	if (td->io_ops->close_file)
1330		ret = td->io_ops->close_file(td, f);
1331
1332	if (!ret)
1333		ret = f_ret;
1334
1335	td->nr_open_files--;
1336	fio_file_clear_open(f);
1337	assert(f->fd == -1);
1338	return ret;
1339}
1340
1341void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1342{
1343	if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1344		return;
1345
1346	if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1347		if (ddir == DDIR_READ)
1348			fio_rwlock_read(f->rwlock);
1349		else
1350			fio_rwlock_write(f->rwlock);
1351	} else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1352		fio_mutex_down(f->lock);
1353
1354	td->file_locks[f->fileno] = td->o.file_lock_mode;
1355}
1356
1357void unlock_file(struct thread_data *td, struct fio_file *f)
1358{
1359	if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1360		return;
1361
1362	if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1363		fio_rwlock_unlock(f->rwlock);
1364	else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1365		fio_mutex_up(f->lock);
1366
1367	td->file_locks[f->fileno] = FILE_LOCK_NONE;
1368}
1369
1370void unlock_file_all(struct thread_data *td, struct fio_file *f)
1371{
1372	if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1373		return;
1374	if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1375		unlock_file(td, f);
1376}
1377
1378static int recurse_dir(struct thread_data *td, const char *dirname)
1379{
1380	struct dirent *dir;
1381	int ret = 0;
1382	DIR *D;
1383
1384	D = opendir(dirname);
1385	if (!D) {
1386		char buf[FIO_VERROR_SIZE];
1387
1388		snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1389		td_verror(td, errno, buf);
1390		return 1;
1391	}
1392
1393	while ((dir = readdir(D)) != NULL) {
1394		char full_path[PATH_MAX];
1395		struct stat sb;
1396
1397		if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1398			continue;
1399
1400		sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1401
1402		if (lstat(full_path, &sb) == -1) {
1403			if (errno != ENOENT) {
1404				td_verror(td, errno, "stat");
1405				return 1;
1406			}
1407		}
1408
1409		if (S_ISREG(sb.st_mode)) {
1410			add_file(td, full_path, 0, 1);
1411			continue;
1412		}
1413		if (!S_ISDIR(sb.st_mode))
1414			continue;
1415
1416		ret = recurse_dir(td, full_path);
1417		if (ret)
1418			break;
1419	}
1420
1421	closedir(D);
1422	return ret;
1423}
1424
1425int add_dir_files(struct thread_data *td, const char *path)
1426{
1427	int ret = recurse_dir(td, path);
1428
1429	if (!ret)
1430		log_info("fio: opendir added %d files\n", td->o.nr_files);
1431
1432	return ret;
1433}
1434
1435void dup_files(struct thread_data *td, struct thread_data *org)
1436{
1437	struct fio_file *f;
1438	unsigned int i;
1439
1440	dprint(FD_FILE, "dup files: %d\n", org->files_index);
1441
1442	if (!org->files)
1443		return;
1444
1445	td->files = malloc(org->files_index * sizeof(f));
1446
1447	if (td->o.file_lock_mode != FILE_LOCK_NONE)
1448		td->file_locks = malloc(org->files_index);
1449
1450	for_each_file(org, f, i) {
1451		struct fio_file *__f;
1452
1453		__f = alloc_new_file(td);
1454
1455		if (f->file_name) {
1456			__f->file_name = smalloc_strdup(f->file_name);
1457			if (!__f->file_name) {
1458				log_err("fio: smalloc OOM\n");
1459				assert(0);
1460			}
1461
1462			__f->filetype = f->filetype;
1463		}
1464
1465		if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1466			__f->lock = f->lock;
1467		else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1468			__f->rwlock = f->rwlock;
1469
1470		td->files[i] = __f;
1471	}
1472}
1473
1474/*
1475 * Returns the index that matches the filename, or -1 if not there
1476 */
1477int get_fileno(struct thread_data *td, const char *fname)
1478{
1479	struct fio_file *f;
1480	unsigned int i;
1481
1482	for_each_file(td, f, i)
1483		if (!strcmp(f->file_name, fname))
1484			return i;
1485
1486	return -1;
1487}
1488
1489/*
1490 * For log usage, where we add/open/close files automatically
1491 */
1492void free_release_files(struct thread_data *td)
1493{
1494	close_files(td);
1495	td->files_index = 0;
1496	td->nr_normal_files = 0;
1497}
1498
1499void fio_file_reset(struct thread_data *td, struct fio_file *f)
1500{
1501	f->last_pos = f->file_offset;
1502	f->last_start = -1ULL;
1503	if (f->io_axmap)
1504		axmap_reset(f->io_axmap);
1505	if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1506		lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1507}
1508
1509int fio_files_done(struct thread_data *td)
1510{
1511	struct fio_file *f;
1512	unsigned int i;
1513
1514	for_each_file(td, f, i)
1515		if (!fio_file_done(f))
1516			return 0;
1517
1518	return 1;
1519}
1520
1521/* free memory used in initialization phase only */
1522void filesetup_mem_free(void)
1523{
1524	free_already_allocated();
1525}
1526