filesetup.c revision b5f4d8baefc6eb3e13235b7d9042b7ffecdb23dd
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) || (td_write(td) && td->o.overwrite) ||
54	    (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
55		new_layout = 1;
56	if (td_write(td) && !td->o.overwrite)
57		unlink_file = 1;
58
59	if (unlink_file || new_layout) {
60		dprint(FD_FILE, "layout unlink %s\n", f->file_name);
61		if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
62			td_verror(td, errno, "unlink");
63			return 1;
64		}
65	}
66
67	flags = O_WRONLY | O_CREAT;
68	if (new_layout)
69		flags |= O_TRUNC;
70
71	dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
72	f->fd = open(f->file_name, flags, 0644);
73	if (f->fd < 0) {
74		td_verror(td, errno, "open");
75		return 1;
76	}
77
78#ifdef CONFIG_POSIX_FALLOCATE
79	if (!td->o.fill_device) {
80		switch (td->o.fallocate_mode) {
81		case FIO_FALLOCATE_NONE:
82			break;
83		case FIO_FALLOCATE_POSIX:
84			dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
85				 f->file_name,
86				 (unsigned long long) f->real_file_size);
87
88			r = posix_fallocate(f->fd, 0, f->real_file_size);
89			if (r > 0) {
90				log_err("fio: posix_fallocate fails: %s\n",
91						strerror(r));
92			}
93			break;
94#ifdef CONFIG_LINUX_FALLOCATE
95		case FIO_FALLOCATE_KEEP_SIZE:
96			dprint(FD_FILE,
97				"fallocate(FALLOC_FL_KEEP_SIZE) "
98				"file %s size %llu\n", f->file_name,
99				(unsigned long long) f->real_file_size);
100
101			r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
102					f->real_file_size);
103			if (r != 0)
104				td_verror(td, errno, "fallocate");
105
106			break;
107#endif /* CONFIG_LINUX_FALLOCATE */
108		default:
109			log_err("fio: unknown fallocate mode: %d\n",
110				td->o.fallocate_mode);
111			assert(0);
112		}
113	}
114#endif /* CONFIG_POSIX_FALLOCATE */
115
116	if (!new_layout)
117		goto done;
118
119	/*
120	 * The size will be -1ULL when fill_device is used, so don't truncate
121	 * or fallocate this file, just write it
122	 */
123	if (!td->o.fill_device) {
124		dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
125					(unsigned long long) f->real_file_size);
126		if (ftruncate(f->fd, f->real_file_size) == -1) {
127			if (errno != EFBIG) {
128				td_verror(td, errno, "ftruncate");
129				goto err;
130			}
131		}
132	}
133
134	b = malloc(td->o.max_bs[DDIR_WRITE]);
135
136	left = f->real_file_size;
137	while (left && !td->terminate) {
138		bs = td->o.max_bs[DDIR_WRITE];
139		if (bs > left)
140			bs = left;
141
142		fill_io_buffer(td, b, bs, bs);
143
144		r = write(f->fd, b, bs);
145
146		if (r > 0) {
147			left -= r;
148			continue;
149		} else {
150			if (r < 0) {
151				int __e = errno;
152
153				if (__e == ENOSPC) {
154					if (td->o.fill_device)
155						break;
156					log_info("fio: ENOSPC on laying out "
157						 "file, stopping\n");
158					break;
159				}
160				td_verror(td, errno, "write");
161			} else
162				td_verror(td, EIO, "write");
163
164			break;
165		}
166	}
167
168	if (td->terminate) {
169		dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
170		unlink(f->file_name);
171	} else if (td->o.create_fsync) {
172		if (fsync(f->fd) < 0) {
173			td_verror(td, errno, "fsync");
174			goto err;
175		}
176	}
177	if (td->o.fill_device && !td_write(td)) {
178		fio_file_clear_size_known(f);
179		if (td_io_get_file_size(td, f))
180			goto err;
181		if (f->io_size > f->real_file_size)
182			f->io_size = f->real_file_size;
183	}
184
185	free(b);
186done:
187	return 0;
188err:
189	close(f->fd);
190	f->fd = -1;
191	return 1;
192}
193
194static int pre_read_file(struct thread_data *td, struct fio_file *f)
195{
196	int r, did_open = 0, old_runstate;
197	unsigned long long left;
198	unsigned int bs;
199	char *b;
200
201	if (td->io_ops->flags & FIO_PIPEIO)
202		return 0;
203
204	if (!fio_file_open(f)) {
205		if (td->io_ops->open_file(td, f)) {
206			log_err("fio: cannot pre-read, failed to open file\n");
207			return 1;
208		}
209		did_open = 1;
210	}
211
212	old_runstate = td_bump_runstate(td, TD_PRE_READING);
213
214	bs = td->o.max_bs[DDIR_READ];
215	b = malloc(bs);
216	memset(b, 0, bs);
217
218	lseek(f->fd, f->file_offset, SEEK_SET);
219	left = f->io_size;
220
221	while (left && !td->terminate) {
222		if (bs > left)
223			bs = left;
224
225		r = read(f->fd, b, bs);
226
227		if (r == (int) bs) {
228			left -= bs;
229			continue;
230		} else {
231			td_verror(td, EIO, "pre_read");
232			break;
233		}
234	}
235
236	td_restore_runstate(td, old_runstate);
237
238	if (did_open)
239		td->io_ops->close_file(td, f);
240	free(b);
241	return 0;
242}
243
244static unsigned long long get_rand_file_size(struct thread_data *td)
245{
246	unsigned long long ret, sized;
247	unsigned long r;
248
249	if (td->o.use_os_rand) {
250		r = os_random_long(&td->file_size_state);
251		sized = td->o.file_size_high - td->o.file_size_low;
252		ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
253	} else {
254		r = __rand(&td->__file_size_state);
255		sized = td->o.file_size_high - td->o.file_size_low;
256		ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
257	}
258
259	ret += td->o.file_size_low;
260	ret -= (ret % td->o.rw_min_bs);
261	return ret;
262}
263
264static int file_size(struct thread_data *td, struct fio_file *f)
265{
266	struct stat st;
267
268	if (stat(f->file_name, &st) == -1) {
269		td_verror(td, errno, "fstat");
270		return 1;
271	}
272
273	f->real_file_size = st.st_size;
274	return 0;
275}
276
277static int bdev_size(struct thread_data *td, struct fio_file *f)
278{
279	unsigned long long bytes = 0;
280	int r;
281
282	if (td->io_ops->open_file(td, f)) {
283		log_err("fio: failed opening blockdev %s for size check\n",
284			f->file_name);
285		return 1;
286	}
287
288	r = blockdev_size(f, &bytes);
289	if (r) {
290		td_verror(td, r, "blockdev_size");
291		goto err;
292	}
293
294	if (!bytes) {
295		log_err("%s: zero sized block device?\n", f->file_name);
296		goto err;
297	}
298
299	f->real_file_size = bytes;
300	td->io_ops->close_file(td, f);
301	return 0;
302err:
303	td->io_ops->close_file(td, f);
304	return 1;
305}
306
307static int char_size(struct thread_data *td, struct fio_file *f)
308{
309#ifdef FIO_HAVE_CHARDEV_SIZE
310	unsigned long long bytes = 0;
311	int r;
312
313	if (td->io_ops->open_file(td, f)) {
314		log_err("fio: failed opening blockdev %s for size check\n",
315			f->file_name);
316		return 1;
317	}
318
319	r = chardev_size(f, &bytes);
320	if (r) {
321		td_verror(td, r, "chardev_size");
322		goto err;
323	}
324
325	if (!bytes) {
326		log_err("%s: zero sized char device?\n", f->file_name);
327		goto err;
328	}
329
330	f->real_file_size = bytes;
331	td->io_ops->close_file(td, f);
332	return 0;
333err:
334	td->io_ops->close_file(td, f);
335	return 1;
336#else
337	f->real_file_size = -1ULL;
338	return 0;
339#endif
340}
341
342static int get_file_size(struct thread_data *td, struct fio_file *f)
343{
344	int ret = 0;
345
346	if (fio_file_size_known(f))
347		return 0;
348
349	if (f->filetype == FIO_TYPE_FILE)
350		ret = file_size(td, f);
351	else if (f->filetype == FIO_TYPE_BD)
352		ret = bdev_size(td, f);
353	else if (f->filetype == FIO_TYPE_CHAR)
354		ret = char_size(td, f);
355	else
356		f->real_file_size = -1;
357
358	if (ret)
359		return ret;
360
361	if (f->file_offset > f->real_file_size) {
362		log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
363					(unsigned long long) f->file_offset,
364					(unsigned long long) f->real_file_size);
365		return 1;
366	}
367
368	fio_file_set_size_known(f);
369	return 0;
370}
371
372static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
373				   unsigned long long off,
374				   unsigned long long len)
375{
376	int ret = 0;
377
378	if (len == -1ULL)
379		len = f->io_size;
380	if (off == -1ULL)
381		off = f->file_offset;
382
383	if (len == -1ULL || off == -1ULL)
384		return 0;
385
386	dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
387								len);
388
389	/*
390	 * FIXME: add blockdev flushing too
391	 */
392	if (f->mmap_ptr) {
393		ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
394#ifdef FIO_MADV_FREE
395		if (f->filetype == FIO_TYPE_BD)
396			(void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
397#endif
398	} else if (f->filetype == FIO_TYPE_FILE) {
399		ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
400	} else if (f->filetype == FIO_TYPE_BD) {
401		ret = blockdev_invalidate_cache(f);
402		if (ret < 0 && errno == EACCES && geteuid()) {
403			if (!root_warn) {
404				log_err("fio: only root may flush block "
405					"devices. Cache flush bypassed!\n");
406				root_warn = 1;
407			}
408			ret = 0;
409		}
410	} else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
411		ret = 0;
412
413	if (ret < 0) {
414		td_verror(td, errno, "invalidate_cache");
415		return 1;
416	} else if (ret > 0) {
417		td_verror(td, ret, "invalidate_cache");
418		return 1;
419	}
420
421	return ret;
422
423}
424
425int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
426{
427	if (!fio_file_open(f))
428		return 0;
429
430	return __file_invalidate_cache(td, f, -1ULL, -1ULL);
431}
432
433int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
434{
435	int ret = 0;
436
437	dprint(FD_FILE, "fd close %s\n", f->file_name);
438
439	remove_file_hash(f);
440
441	if (close(f->fd) < 0)
442		ret = errno;
443
444	f->fd = -1;
445
446	if (f->shadow_fd != -1) {
447		close(f->shadow_fd);
448		f->shadow_fd = -1;
449	}
450
451	f->engine_data = 0;
452	return ret;
453}
454
455int file_lookup_open(struct fio_file *f, int flags)
456{
457	struct fio_file *__f;
458	int from_hash;
459
460	__f = lookup_file_hash(f->file_name);
461	if (__f) {
462		dprint(FD_FILE, "found file in hash %s\n", f->file_name);
463		/*
464		 * racy, need the __f->lock locked
465		 */
466		f->lock = __f->lock;
467		from_hash = 1;
468	} else {
469		dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
470		from_hash = 0;
471	}
472
473	f->fd = open(f->file_name, flags, 0600);
474	return from_hash;
475}
476
477static int file_close_shadow_fds(struct thread_data *td)
478{
479	struct fio_file *f;
480	int num_closed = 0;
481	unsigned int i;
482
483	for_each_file(td, f, i) {
484		if (f->shadow_fd == -1)
485			continue;
486
487		close(f->shadow_fd);
488		f->shadow_fd = -1;
489		num_closed++;
490	}
491
492	return num_closed;
493}
494
495int generic_open_file(struct thread_data *td, struct fio_file *f)
496{
497	int is_std = 0;
498	int flags = 0;
499	int from_hash = 0;
500
501	dprint(FD_FILE, "fd open %s\n", f->file_name);
502
503	if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
504		log_err("fio: trim only applies to block device\n");
505		return 1;
506	}
507
508	if (!strcmp(f->file_name, "-")) {
509		if (td_rw(td)) {
510			log_err("fio: can't read/write to stdin/out\n");
511			return 1;
512		}
513		is_std = 1;
514
515		/*
516		 * move output logging to stderr, if we are writing to stdout
517		 */
518		if (td_write(td))
519			f_out = stderr;
520	}
521
522	if (td_trim(td))
523		goto skip_flags;
524	if (td->o.odirect)
525		flags |= OS_O_DIRECT;
526	if (td->o.oatomic) {
527		if (!FIO_O_ATOMIC) {
528			td_verror(td, EINVAL, "OS does not support atomic IO");
529			return 1;
530		}
531		flags |= OS_O_DIRECT | FIO_O_ATOMIC;
532	}
533	if (td->o.sync_io)
534		flags |= O_SYNC;
535	if (td->o.create_on_open)
536		flags |= O_CREAT;
537skip_flags:
538	if (f->filetype != FIO_TYPE_FILE)
539		flags |= FIO_O_NOATIME;
540
541open_again:
542	if (td_write(td)) {
543		if (!read_only)
544			flags |= O_RDWR;
545
546		if (f->filetype == FIO_TYPE_FILE)
547			flags |= O_CREAT;
548
549		if (is_std)
550			f->fd = dup(STDOUT_FILENO);
551		else
552			from_hash = file_lookup_open(f, flags);
553	} else if (td_read(td)) {
554		if (f->filetype == FIO_TYPE_CHAR && !read_only)
555			flags |= O_RDWR;
556		else
557			flags |= O_RDONLY;
558
559		if (is_std)
560			f->fd = dup(STDIN_FILENO);
561		else
562			from_hash = file_lookup_open(f, flags);
563	} else { //td trim
564		flags |= O_RDWR;
565		from_hash = file_lookup_open(f, flags);
566	}
567
568	if (f->fd == -1) {
569		char buf[FIO_VERROR_SIZE];
570		int __e = errno;
571
572		if (__e == EPERM && (flags & FIO_O_NOATIME)) {
573			flags &= ~FIO_O_NOATIME;
574			goto open_again;
575		}
576		if (__e == EMFILE && file_close_shadow_fds(td))
577			goto open_again;
578
579		snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
580
581		if (__e == EINVAL && (flags & OS_O_DIRECT)) {
582			log_err("fio: looks like your file system does not " \
583				"support direct=1/buffered=0\n");
584		}
585
586		td_verror(td, __e, buf);
587	}
588
589	if (!from_hash && f->fd != -1) {
590		if (add_file_hash(f)) {
591			int fio_unused ret;
592
593			/*
594			 * Stash away descriptor for later close. This is to
595			 * work-around a "feature" on Linux, where a close of
596			 * an fd that has been opened for write will trigger
597			 * udev to call blkid to check partitions, fs id, etc.
598			 * That pollutes the device cache, which can slow down
599			 * unbuffered accesses.
600			 */
601			if (f->shadow_fd == -1)
602				f->shadow_fd = f->fd;
603			else {
604				/*
605			 	 * OK to ignore, we haven't done anything
606				 * with it
607				 */
608				ret = generic_close_file(td, f);
609			}
610			goto open_again;
611		}
612	}
613
614	return 0;
615}
616
617int generic_get_file_size(struct thread_data *td, struct fio_file *f)
618{
619	return get_file_size(td, f);
620}
621
622/*
623 * open/close all files, so that ->real_file_size gets set
624 */
625static int get_file_sizes(struct thread_data *td)
626{
627	struct fio_file *f;
628	unsigned int i;
629	int err = 0;
630
631	for_each_file(td, f, i) {
632		dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
633								f->file_name);
634
635		if (td_io_get_file_size(td, f)) {
636			if (td->error != ENOENT) {
637				log_err("%s\n", td->verror);
638				err = 1;
639			}
640			clear_error(td);
641		}
642
643		if (f->real_file_size == -1ULL && td->o.size)
644			f->real_file_size = td->o.size / td->o.nr_files;
645	}
646
647	return err;
648}
649
650struct fio_mount {
651	struct flist_head list;
652	const char *base;
653	char __base[256];
654	unsigned int key;
655};
656
657/*
658 * Get free number of bytes for each file on each unique mount.
659 */
660static unsigned long long get_fs_free_counts(struct thread_data *td)
661{
662	struct flist_head *n, *tmp;
663	unsigned long long ret = 0;
664	struct fio_mount *fm;
665	FLIST_HEAD(list);
666	struct fio_file *f;
667	unsigned int i;
668
669	for_each_file(td, f, i) {
670		struct stat sb;
671		char buf[256];
672
673		if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
674			if (f->real_file_size != -1ULL)
675				ret += f->real_file_size;
676			continue;
677		} else if (f->filetype != FIO_TYPE_FILE)
678			continue;
679
680		strcpy(buf, f->file_name);
681
682		if (stat(buf, &sb) < 0) {
683			if (errno != ENOENT)
684				break;
685			strcpy(buf, ".");
686			if (stat(buf, &sb) < 0)
687				break;
688		}
689
690		fm = NULL;
691		flist_for_each(n, &list) {
692			fm = flist_entry(n, struct fio_mount, list);
693			if (fm->key == sb.st_dev)
694				break;
695
696			fm = NULL;
697		}
698
699		if (fm)
700			continue;
701
702		fm = malloc(sizeof(*fm));
703		strcpy(fm->__base, buf);
704		fm->base = basename(fm->__base);
705		fm->key = sb.st_dev;
706		flist_add(&fm->list, &list);
707	}
708
709	flist_for_each_safe(n, tmp, &list) {
710		unsigned long long sz;
711
712		fm = flist_entry(n, struct fio_mount, list);
713		flist_del(&fm->list);
714
715		sz = get_fs_size(fm->base);
716		if (sz && sz != -1ULL)
717			ret += sz;
718
719		free(fm);
720	}
721
722	return ret;
723}
724
725uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
726{
727	struct thread_options *o = &td->o;
728
729	if (o->file_append && f->filetype == FIO_TYPE_FILE)
730		return f->real_file_size;
731
732	return td->o.start_offset +
733		(td->thread_number - 1) * td->o.offset_increment;
734}
735
736/*
737 * Open the files and setup files sizes, creating files if necessary.
738 */
739int setup_files(struct thread_data *td)
740{
741	unsigned long long total_size, extend_size;
742	struct thread_options *o = &td->o;
743	struct fio_file *f;
744	unsigned int i, nr_fs_extra = 0;
745	int err = 0, need_extend;
746	int old_state;
747	const unsigned int bs = td_min_bs(td);
748	uint64_t fs = 0;
749
750	dprint(FD_FILE, "setup files\n");
751
752	old_state = td_bump_runstate(td, TD_SETTING_UP);
753
754	if (o->read_iolog_file)
755		goto done;
756
757	/*
758	 * if ioengine defines a setup() method, it's responsible for
759	 * opening the files and setting f->real_file_size to indicate
760	 * the valid range for that file.
761	 */
762	if (td->io_ops->setup)
763		err = td->io_ops->setup(td);
764	else
765		err = get_file_sizes(td);
766
767	if (err)
768		goto err_out;
769
770	/*
771	 * check sizes. if the files/devices do not exist and the size
772	 * isn't passed to fio, abort.
773	 */
774	total_size = 0;
775	for_each_file(td, f, i) {
776		if (f->real_file_size == -1ULL)
777			total_size = -1ULL;
778		else
779			total_size += f->real_file_size;
780	}
781
782	if (o->fill_device)
783		td->fill_device_size = get_fs_free_counts(td);
784
785	/*
786	 * device/file sizes are zero and no size given, punt
787	 */
788	if ((!total_size || total_size == -1ULL) && !o->size &&
789	    !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
790	    !(o->nr_files && (o->file_size_low || o->file_size_high))) {
791		log_err("%s: you need to specify size=\n", o->name);
792		td_verror(td, EINVAL, "total_file_size");
793		goto err_out;
794	}
795
796	/*
797	 * Calculate per-file size and potential extra size for the
798	 * first files, if needed.
799	 */
800	if (!o->file_size_low) {
801		uint64_t all_fs;
802
803		fs = o->size / o->nr_files;
804		all_fs = fs * o->nr_files;
805
806		if (all_fs < o->size)
807			nr_fs_extra = (o->size - all_fs) / bs;
808	}
809
810	/*
811	 * now file sizes are known, so we can set ->io_size. if size= is
812	 * not given, ->io_size is just equal to ->real_file_size. if size
813	 * is given, ->io_size is size / nr_files.
814	 */
815	extend_size = total_size = 0;
816	need_extend = 0;
817	for_each_file(td, f, i) {
818		f->file_offset = get_start_offset(td, f);
819
820		if (!o->file_size_low) {
821			/*
822			 * no file size range given, file size is equal to
823			 * total size divided by number of files. If that is
824			 * zero, set it to the real file size. If the size
825			 * doesn't divide nicely with the min blocksize,
826			 * make the first files bigger.
827			 */
828			f->io_size = fs;
829			if (nr_fs_extra) {
830				nr_fs_extra--;
831				f->io_size += bs;
832			}
833
834			if (!f->io_size)
835				f->io_size = f->real_file_size - f->file_offset;
836		} else if (f->real_file_size < o->file_size_low ||
837			   f->real_file_size > o->file_size_high) {
838			if (f->file_offset > o->file_size_low)
839				goto err_offset;
840			/*
841			 * file size given. if it's fixed, use that. if it's a
842			 * range, generate a random size in-between.
843			 */
844			if (o->file_size_low == o->file_size_high)
845				f->io_size = o->file_size_low - f->file_offset;
846			else {
847				f->io_size = get_rand_file_size(td)
848						- f->file_offset;
849			}
850		} else
851			f->io_size = f->real_file_size - f->file_offset;
852
853		if (f->io_size == -1ULL)
854			total_size = -1ULL;
855		else {
856                        if (o->size_percent)
857                                f->io_size = (f->io_size * o->size_percent) / 100;
858			total_size += f->io_size;
859		}
860
861		if (f->filetype == FIO_TYPE_FILE &&
862		    (f->io_size + f->file_offset) > f->real_file_size &&
863		    !(td->io_ops->flags & FIO_DISKLESSIO)) {
864			if (!o->create_on_open) {
865				need_extend++;
866				extend_size += (f->io_size + f->file_offset);
867			} else
868				f->real_file_size = f->io_size + f->file_offset;
869			fio_file_set_extend(f);
870		}
871	}
872
873	if (!o->size || o->size > total_size)
874		o->size = total_size;
875
876	/*
877	 * See if we need to extend some files
878	 */
879	if (need_extend) {
880		temp_stall_ts = 1;
881		if (output_format == FIO_OUTPUT_NORMAL)
882			log_info("%s: Laying out IO file(s) (%u file(s) /"
883				 " %lluMB)\n", o->name, need_extend,
884					extend_size >> 20);
885
886		for_each_file(td, f, i) {
887			unsigned long long old_len = -1ULL, extend_len = -1ULL;
888
889			if (!fio_file_extend(f))
890				continue;
891
892			assert(f->filetype == FIO_TYPE_FILE);
893			fio_file_clear_extend(f);
894			if (!o->fill_device) {
895				old_len = f->real_file_size;
896				extend_len = f->io_size + f->file_offset -
897						old_len;
898			}
899			f->real_file_size = (f->io_size + f->file_offset);
900			err = extend_file(td, f);
901			if (err)
902				break;
903
904			err = __file_invalidate_cache(td, f, old_len,
905								extend_len);
906			close(f->fd);
907			f->fd = -1;
908			if (err)
909				break;
910		}
911		temp_stall_ts = 0;
912	}
913
914	if (err)
915		goto err_out;
916
917	if (!o->zone_size)
918		o->zone_size = o->size;
919
920	/*
921	 * iolog already set the total io size, if we read back
922	 * stored entries.
923	 */
924	if (!o->read_iolog_file)
925		td->total_io_size = o->size * o->loops;
926
927done:
928	if (o->create_only)
929		td->done = 1;
930
931	td_restore_runstate(td, old_state);
932	return 0;
933err_offset:
934	log_err("%s: you need to specify valid offset=\n", o->name);
935err_out:
936	td_restore_runstate(td, old_state);
937	return 1;
938}
939
940int pre_read_files(struct thread_data *td)
941{
942	struct fio_file *f;
943	unsigned int i;
944
945	dprint(FD_FILE, "pre_read files\n");
946
947	for_each_file(td, f, i) {
948		pre_read_file(td, f);
949	}
950
951	return 1;
952}
953
954static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
955{
956	unsigned int range_size, seed;
957	unsigned long nranges;
958	uint64_t file_size;
959
960	range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
961	file_size = min(f->real_file_size, f->io_size);
962
963	nranges = (file_size + range_size - 1) / range_size;
964
965	seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
966	if (!td->o.rand_repeatable)
967		seed = td->rand_seeds[4];
968
969	if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
970		zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
971	else
972		pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
973
974	return 1;
975}
976
977static int init_rand_distribution(struct thread_data *td)
978{
979	struct fio_file *f;
980	unsigned int i;
981	int state;
982
983	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
984		return 0;
985
986	state = td_bump_runstate(td, TD_SETTING_UP);
987
988	for_each_file(td, f, i)
989		__init_rand_distribution(td, f);
990
991	td_restore_runstate(td, state);
992
993	return 1;
994}
995
996int init_random_map(struct thread_data *td)
997{
998	unsigned long long blocks;
999	struct fio_file *f;
1000	unsigned int i;
1001
1002	if (init_rand_distribution(td))
1003		return 0;
1004	if (!td_random(td))
1005		return 0;
1006
1007	for_each_file(td, f, i) {
1008		uint64_t file_size = min(f->real_file_size, f->io_size);
1009
1010		blocks = file_size / (unsigned long long) td->o.rw_min_bs;
1011
1012		if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1013			unsigned long seed;
1014
1015			seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1016
1017			if (!lfsr_init(&f->lfsr, blocks, seed, 0))
1018				continue;
1019		} else if (!td->o.norandommap) {
1020			f->io_axmap = axmap_new(blocks);
1021			if (f->io_axmap)
1022				continue;
1023		} else if (td->o.norandommap)
1024			continue;
1025
1026		if (!td->o.softrandommap) {
1027			log_err("fio: failed allocating random map. If running"
1028				" a large number of jobs, try the 'norandommap'"
1029				" option or set 'softrandommap'. Or give"
1030				" a larger --alloc-size to fio.\n");
1031			return 1;
1032		}
1033
1034		log_info("fio: file %s failed allocating random map. Running "
1035			 "job without.\n", f->file_name);
1036	}
1037
1038	return 0;
1039}
1040
1041void close_files(struct thread_data *td)
1042{
1043	struct fio_file *f;
1044	unsigned int i;
1045
1046	for_each_file(td, f, i) {
1047		if (fio_file_open(f))
1048			td_io_close_file(td, f);
1049	}
1050}
1051
1052void close_and_free_files(struct thread_data *td)
1053{
1054	struct fio_file *f;
1055	unsigned int i;
1056
1057	dprint(FD_FILE, "close files\n");
1058
1059	for_each_file(td, f, i) {
1060		if (fio_file_open(f))
1061			td_io_close_file(td, f);
1062
1063		remove_file_hash(f);
1064
1065		if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1066			dprint(FD_FILE, "free unlink %s\n", f->file_name);
1067			unlink(f->file_name);
1068		}
1069
1070		sfree(f->file_name);
1071		f->file_name = NULL;
1072		axmap_free(f->io_axmap);
1073		f->io_axmap = NULL;
1074		sfree(f);
1075	}
1076
1077	td->o.filename = NULL;
1078	free(td->files);
1079	free(td->file_locks);
1080	td->files_index = 0;
1081	td->files = NULL;
1082	td->file_locks = NULL;
1083	td->o.file_lock_mode = FILE_LOCK_NONE;
1084	td->o.nr_files = 0;
1085}
1086
1087static void get_file_type(struct fio_file *f)
1088{
1089	struct stat sb;
1090
1091	if (!strcmp(f->file_name, "-"))
1092		f->filetype = FIO_TYPE_PIPE;
1093	else
1094		f->filetype = FIO_TYPE_FILE;
1095
1096	/* \\.\ is the device namespace in Windows, where every file is
1097	 * a block device */
1098	if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1099		f->filetype = FIO_TYPE_BD;
1100
1101	if (!stat(f->file_name, &sb)) {
1102		if (S_ISBLK(sb.st_mode))
1103			f->filetype = FIO_TYPE_BD;
1104		else if (S_ISCHR(sb.st_mode))
1105			f->filetype = FIO_TYPE_CHAR;
1106		else if (S_ISFIFO(sb.st_mode))
1107			f->filetype = FIO_TYPE_PIPE;
1108	}
1109}
1110
1111static void set_already_allocated(const char *fname) {
1112	struct file_name *fn;
1113
1114	fn = malloc(sizeof(struct file_name));
1115	fn->filename = strdup(fname);
1116	flist_add_tail(&fn->list, &filename_list);
1117}
1118
1119static int is_already_allocated(const char *fname)
1120{
1121	struct flist_head *entry;
1122	char *filename;
1123
1124	if (!flist_empty(&filename_list))
1125	{
1126		flist_for_each(entry, &filename_list) {
1127			filename = flist_entry(entry, struct file_name, list)->filename;
1128
1129			if (strcmp(filename, fname) == 0)
1130				return 1;
1131		}
1132	}
1133
1134	return 0;
1135}
1136
1137static void free_already_allocated() {
1138	struct flist_head *entry, *tmp;
1139	struct file_name *fn;
1140
1141	if (!flist_empty(&filename_list))
1142	{
1143		flist_for_each_safe(entry, tmp, &filename_list) {
1144			fn = flist_entry(entry, struct file_name, list);
1145			free(fn->filename);
1146			flist_del(&fn->list);
1147			free(fn);
1148		}
1149	}
1150}
1151
1152int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1153{
1154	int cur_files = td->files_index;
1155	char file_name[PATH_MAX];
1156	struct fio_file *f;
1157	int len = 0;
1158
1159	dprint(FD_FILE, "add file %s\n", fname);
1160
1161	if (td->o.directory)
1162		len = set_name_idx(file_name, td->o.directory, numjob);
1163
1164	sprintf(file_name + len, "%s", fname);
1165
1166	/* clean cloned siblings using existing files */
1167	if (numjob && is_already_allocated(file_name))
1168		return 0;
1169
1170	f = smalloc(sizeof(*f));
1171	if (!f) {
1172		log_err("fio: smalloc OOM\n");
1173		assert(0);
1174	}
1175
1176	f->fd = -1;
1177	f->shadow_fd = -1;
1178	fio_file_reset(td, f);
1179
1180	if (td->files_size <= td->files_index) {
1181		unsigned int new_size = td->o.nr_files + 1;
1182
1183		dprint(FD_FILE, "resize file array to %d files\n", new_size);
1184
1185		td->files = realloc(td->files, new_size * sizeof(f));
1186		if (td->files == NULL) {
1187			log_err("fio: realloc OOM\n");
1188			assert(0);
1189		}
1190		if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1191			td->file_locks = realloc(td->file_locks, new_size);
1192			if (!td->file_locks) {
1193				log_err("fio: realloc OOM\n");
1194				assert(0);
1195			}
1196			td->file_locks[cur_files] = FILE_LOCK_NONE;
1197		}
1198		td->files_size = new_size;
1199	}
1200	td->files[cur_files] = f;
1201	f->fileno = cur_files;
1202
1203	/*
1204	 * init function, io engine may not be loaded yet
1205	 */
1206	if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1207		f->real_file_size = -1ULL;
1208
1209	f->file_name = smalloc_strdup(file_name);
1210	if (!f->file_name) {
1211		log_err("fio: smalloc OOM\n");
1212		assert(0);
1213	}
1214
1215	get_file_type(f);
1216
1217	switch (td->o.file_lock_mode) {
1218	case FILE_LOCK_NONE:
1219		break;
1220	case FILE_LOCK_READWRITE:
1221		f->rwlock = fio_rwlock_init();
1222		break;
1223	case FILE_LOCK_EXCLUSIVE:
1224		f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1225		break;
1226	default:
1227		log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1228		assert(0);
1229	}
1230
1231	td->files_index++;
1232	if (f->filetype == FIO_TYPE_FILE)
1233		td->nr_normal_files++;
1234
1235	set_already_allocated(file_name);
1236
1237	/*
1238	 * For adding files after the fact - if openfiles= isn't
1239	 * given as an option, ensure we allow at least one file open
1240	 */
1241	if (!td->o.open_files)
1242		td->o.open_files = 1;
1243
1244	if (inc)
1245		td->o.nr_files++;
1246
1247	dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1248							cur_files);
1249
1250	return cur_files;
1251}
1252
1253int add_file_exclusive(struct thread_data *td, const char *fname)
1254{
1255	struct fio_file *f;
1256	unsigned int i;
1257
1258	for_each_file(td, f, i) {
1259		if (!strcmp(f->file_name, fname))
1260			return i;
1261	}
1262
1263	return add_file(td, fname, 0, 1);
1264}
1265
1266void get_file(struct fio_file *f)
1267{
1268	dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1269	assert(fio_file_open(f));
1270	f->references++;
1271}
1272
1273int put_file(struct thread_data *td, struct fio_file *f)
1274{
1275	int f_ret = 0, ret = 0;
1276
1277	dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1278
1279	if (!fio_file_open(f)) {
1280		assert(f->fd == -1);
1281		return 0;
1282	}
1283
1284	assert(f->references);
1285	if (--f->references)
1286		return 0;
1287
1288	if (should_fsync(td) && td->o.fsync_on_close)
1289		f_ret = fsync(f->fd);
1290
1291	if (td->io_ops->close_file)
1292		ret = td->io_ops->close_file(td, f);
1293
1294	if (!ret)
1295		ret = f_ret;
1296
1297	td->nr_open_files--;
1298	fio_file_clear_open(f);
1299	assert(f->fd == -1);
1300	return ret;
1301}
1302
1303void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1304{
1305	if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1306		return;
1307
1308	if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1309		if (ddir == DDIR_READ)
1310			fio_rwlock_read(f->rwlock);
1311		else
1312			fio_rwlock_write(f->rwlock);
1313	} else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1314		fio_mutex_down(f->lock);
1315
1316	td->file_locks[f->fileno] = td->o.file_lock_mode;
1317}
1318
1319void unlock_file(struct thread_data *td, struct fio_file *f)
1320{
1321	if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1322		return;
1323
1324	if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1325		fio_rwlock_unlock(f->rwlock);
1326	else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1327		fio_mutex_up(f->lock);
1328
1329	td->file_locks[f->fileno] = FILE_LOCK_NONE;
1330}
1331
1332void unlock_file_all(struct thread_data *td, struct fio_file *f)
1333{
1334	if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1335		return;
1336	if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1337		unlock_file(td, f);
1338}
1339
1340static int recurse_dir(struct thread_data *td, const char *dirname)
1341{
1342	struct dirent *dir;
1343	int ret = 0;
1344	DIR *D;
1345
1346	D = opendir(dirname);
1347	if (!D) {
1348		char buf[FIO_VERROR_SIZE];
1349
1350		snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1351		td_verror(td, errno, buf);
1352		return 1;
1353	}
1354
1355	while ((dir = readdir(D)) != NULL) {
1356		char full_path[PATH_MAX];
1357		struct stat sb;
1358
1359		if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1360			continue;
1361
1362		sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1363
1364		if (lstat(full_path, &sb) == -1) {
1365			if (errno != ENOENT) {
1366				td_verror(td, errno, "stat");
1367				return 1;
1368			}
1369		}
1370
1371		if (S_ISREG(sb.st_mode)) {
1372			add_file(td, full_path, 0, 1);
1373			continue;
1374		}
1375		if (!S_ISDIR(sb.st_mode))
1376			continue;
1377
1378		ret = recurse_dir(td, full_path);
1379		if (ret)
1380			break;
1381	}
1382
1383	closedir(D);
1384	return ret;
1385}
1386
1387int add_dir_files(struct thread_data *td, const char *path)
1388{
1389	int ret = recurse_dir(td, path);
1390
1391	if (!ret)
1392		log_info("fio: opendir added %d files\n", td->o.nr_files);
1393
1394	return ret;
1395}
1396
1397void dup_files(struct thread_data *td, struct thread_data *org)
1398{
1399	struct fio_file *f;
1400	unsigned int i;
1401
1402	dprint(FD_FILE, "dup files: %d\n", org->files_index);
1403
1404	if (!org->files)
1405		return;
1406
1407	td->files = malloc(org->files_index * sizeof(f));
1408
1409	if (td->o.file_lock_mode != FILE_LOCK_NONE)
1410		td->file_locks = malloc(org->files_index);
1411
1412	for_each_file(org, f, i) {
1413		struct fio_file *__f;
1414
1415		__f = smalloc(sizeof(*__f));
1416		if (!__f) {
1417			log_err("fio: smalloc OOM\n");
1418			assert(0);
1419		}
1420		__f->fd = -1;
1421		fio_file_reset(td, __f);
1422
1423		if (f->file_name) {
1424			__f->file_name = smalloc_strdup(f->file_name);
1425			if (!__f->file_name) {
1426				log_err("fio: smalloc OOM\n");
1427				assert(0);
1428			}
1429
1430			__f->filetype = f->filetype;
1431		}
1432
1433		if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1434			__f->lock = f->lock;
1435		else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1436			__f->rwlock = f->rwlock;
1437
1438		td->files[i] = __f;
1439	}
1440}
1441
1442/*
1443 * Returns the index that matches the filename, or -1 if not there
1444 */
1445int get_fileno(struct thread_data *td, const char *fname)
1446{
1447	struct fio_file *f;
1448	unsigned int i;
1449
1450	for_each_file(td, f, i)
1451		if (!strcmp(f->file_name, fname))
1452			return i;
1453
1454	return -1;
1455}
1456
1457/*
1458 * For log usage, where we add/open/close files automatically
1459 */
1460void free_release_files(struct thread_data *td)
1461{
1462	close_files(td);
1463	td->files_index = 0;
1464	td->nr_normal_files = 0;
1465}
1466
1467void fio_file_reset(struct thread_data *td, struct fio_file *f)
1468{
1469	f->last_pos = f->file_offset;
1470	f->last_start = -1ULL;
1471	if (f->io_axmap)
1472		axmap_reset(f->io_axmap);
1473	if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1474		lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1475}
1476
1477int fio_files_done(struct thread_data *td)
1478{
1479	struct fio_file *f;
1480	unsigned int i;
1481
1482	for_each_file(td, f, i)
1483		if (!fio_file_done(f))
1484			return 0;
1485
1486	return 1;
1487}
1488
1489/* free memory used in initialization phase only */
1490void filesetup_mem_free() {
1491	free_already_allocated();
1492}
1493