filesetup.c revision 90426237af4f3c9c7628aebfd4421fbe43d68c2a
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 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	lseek(f->fd, f->file_offset, SEEK_SET);
220	left = f->io_size;
221
222	while (left && !td->terminate) {
223		if (bs > left)
224			bs = left;
225
226		r = read(f->fd, b, bs);
227
228		if (r == (int) bs) {
229			left -= bs;
230			continue;
231		} else {
232			td_verror(td, EIO, "pre_read");
233			break;
234		}
235	}
236
237	td_restore_runstate(td, old_runstate);
238
239	if (did_open)
240		td->io_ops->close_file(td, f);
241	free(b);
242	return 0;
243}
244
245static unsigned long long get_rand_file_size(struct thread_data *td)
246{
247	unsigned long long ret, sized;
248	unsigned long r;
249
250	if (td->o.use_os_rand) {
251		r = os_random_long(&td->file_size_state);
252		sized = td->o.file_size_high - td->o.file_size_low;
253		ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
254	} else {
255		r = __rand(&td->__file_size_state);
256		sized = td->o.file_size_high - td->o.file_size_low;
257		ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
258	}
259
260	ret += td->o.file_size_low;
261	ret -= (ret % td->o.rw_min_bs);
262	return ret;
263}
264
265static int file_size(struct thread_data *td, struct fio_file *f)
266{
267	struct stat st;
268
269	if (stat(f->file_name, &st) == -1) {
270		td_verror(td, errno, "fstat");
271		return 1;
272	}
273
274	f->real_file_size = st.st_size;
275	return 0;
276}
277
278static int bdev_size(struct thread_data *td, struct fio_file *f)
279{
280	unsigned long long bytes = 0;
281	int r;
282
283	if (td->io_ops->open_file(td, f)) {
284		log_err("fio: failed opening blockdev %s for size check\n",
285			f->file_name);
286		return 1;
287	}
288
289	r = blockdev_size(f, &bytes);
290	if (r) {
291		td_verror(td, r, "blockdev_size");
292		goto err;
293	}
294
295	if (!bytes) {
296		log_err("%s: zero sized block device?\n", f->file_name);
297		goto err;
298	}
299
300	f->real_file_size = bytes;
301	td->io_ops->close_file(td, f);
302	return 0;
303err:
304	td->io_ops->close_file(td, f);
305	return 1;
306}
307
308static int char_size(struct thread_data *td, struct fio_file *f)
309{
310#ifdef FIO_HAVE_CHARDEV_SIZE
311	unsigned long long bytes = 0;
312	int r;
313
314	if (td->io_ops->open_file(td, f)) {
315		log_err("fio: failed opening blockdev %s for size check\n",
316			f->file_name);
317		return 1;
318	}
319
320	r = chardev_size(f, &bytes);
321	if (r) {
322		td_verror(td, r, "chardev_size");
323		goto err;
324	}
325
326	if (!bytes) {
327		log_err("%s: zero sized char device?\n", f->file_name);
328		goto err;
329	}
330
331	f->real_file_size = bytes;
332	td->io_ops->close_file(td, f);
333	return 0;
334err:
335	td->io_ops->close_file(td, f);
336	return 1;
337#else
338	f->real_file_size = -1ULL;
339	return 0;
340#endif
341}
342
343static int get_file_size(struct thread_data *td, struct fio_file *f)
344{
345	int ret = 0;
346
347	if (fio_file_size_known(f))
348		return 0;
349
350	if (f->filetype == FIO_TYPE_FILE)
351		ret = file_size(td, f);
352	else if (f->filetype == FIO_TYPE_BD)
353		ret = bdev_size(td, f);
354	else if (f->filetype == FIO_TYPE_CHAR)
355		ret = char_size(td, f);
356	else
357		f->real_file_size = -1;
358
359	if (ret)
360		return ret;
361
362	if (f->file_offset > f->real_file_size) {
363		log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
364					(unsigned long long) f->file_offset,
365					(unsigned long long) f->real_file_size);
366		return 1;
367	}
368
369	fio_file_set_size_known(f);
370	return 0;
371}
372
373static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
374				   unsigned long long off,
375				   unsigned long long len)
376{
377	int ret = 0;
378
379	if (len == -1ULL)
380		len = f->io_size;
381	if (off == -1ULL)
382		off = f->file_offset;
383
384	if (len == -1ULL || off == -1ULL)
385		return 0;
386
387	dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
388								len);
389
390	/*
391	 * FIXME: add blockdev flushing too
392	 */
393	if (f->mmap_ptr) {
394		ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
395#ifdef FIO_MADV_FREE
396		if (f->filetype == FIO_TYPE_BD)
397			(void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
398#endif
399	} else if (f->filetype == FIO_TYPE_FILE) {
400		ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
401	} else if (f->filetype == FIO_TYPE_BD) {
402		ret = blockdev_invalidate_cache(f);
403		if (ret < 0 && errno == EACCES && geteuid()) {
404			if (!root_warn) {
405				log_err("fio: only root may flush block "
406					"devices. Cache flush bypassed!\n");
407				root_warn = 1;
408			}
409			ret = 0;
410		}
411	} else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
412		ret = 0;
413
414	if (ret < 0) {
415		td_verror(td, errno, "invalidate_cache");
416		return 1;
417	} else if (ret > 0) {
418		td_verror(td, ret, "invalidate_cache");
419		return 1;
420	}
421
422	return ret;
423
424}
425
426int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
427{
428	if (!fio_file_open(f))
429		return 0;
430
431	return __file_invalidate_cache(td, f, -1ULL, -1ULL);
432}
433
434int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
435{
436	int ret = 0;
437
438	dprint(FD_FILE, "fd close %s\n", f->file_name);
439
440	remove_file_hash(f);
441
442	if (close(f->fd) < 0)
443		ret = errno;
444
445	f->fd = -1;
446
447	if (f->shadow_fd != -1) {
448		close(f->shadow_fd);
449		f->shadow_fd = -1;
450	}
451
452	f->engine_data = 0;
453	return ret;
454}
455
456int file_lookup_open(struct fio_file *f, int flags)
457{
458	struct fio_file *__f;
459	int from_hash;
460
461	__f = lookup_file_hash(f->file_name);
462	if (__f) {
463		dprint(FD_FILE, "found file in hash %s\n", f->file_name);
464		/*
465		 * racy, need the __f->lock locked
466		 */
467		f->lock = __f->lock;
468		from_hash = 1;
469	} else {
470		dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
471		from_hash = 0;
472	}
473
474	f->fd = open(f->file_name, flags, 0600);
475	return from_hash;
476}
477
478static int file_close_shadow_fds(struct thread_data *td)
479{
480	struct fio_file *f;
481	int num_closed = 0;
482	unsigned int i;
483
484	for_each_file(td, f, i) {
485		if (f->shadow_fd == -1)
486			continue;
487
488		close(f->shadow_fd);
489		f->shadow_fd = -1;
490		num_closed++;
491	}
492
493	return num_closed;
494}
495
496int generic_open_file(struct thread_data *td, struct fio_file *f)
497{
498	int is_std = 0;
499	int flags = 0;
500	int from_hash = 0;
501
502	dprint(FD_FILE, "fd open %s\n", f->file_name);
503
504	if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
505		log_err("fio: trim only applies to block device\n");
506		return 1;
507	}
508
509	if (!strcmp(f->file_name, "-")) {
510		if (td_rw(td)) {
511			log_err("fio: can't read/write to stdin/out\n");
512			return 1;
513		}
514		is_std = 1;
515
516		/*
517		 * move output logging to stderr, if we are writing to stdout
518		 */
519		if (td_write(td))
520			f_out = stderr;
521	}
522
523	if (td_trim(td))
524		goto skip_flags;
525	if (td->o.odirect)
526		flags |= OS_O_DIRECT;
527	if (td->o.oatomic) {
528		if (!FIO_O_ATOMIC) {
529			td_verror(td, EINVAL, "OS does not support atomic IO");
530			return 1;
531		}
532		flags |= OS_O_DIRECT | FIO_O_ATOMIC;
533	}
534	if (td->o.sync_io)
535		flags |= O_SYNC;
536	if (td->o.create_on_open)
537		flags |= O_CREAT;
538skip_flags:
539	if (f->filetype != FIO_TYPE_FILE)
540		flags |= FIO_O_NOATIME;
541
542open_again:
543	if (td_write(td)) {
544		if (!read_only)
545			flags |= O_RDWR;
546
547		if (f->filetype == FIO_TYPE_FILE)
548			flags |= O_CREAT;
549
550		if (is_std)
551			f->fd = dup(STDOUT_FILENO);
552		else
553			from_hash = file_lookup_open(f, flags);
554	} else if (td_read(td)) {
555		if (f->filetype == FIO_TYPE_CHAR && !read_only)
556			flags |= O_RDWR;
557		else
558			flags |= O_RDONLY;
559
560		if (is_std)
561			f->fd = dup(STDIN_FILENO);
562		else
563			from_hash = file_lookup_open(f, flags);
564	} else { //td trim
565		flags |= O_RDWR;
566		from_hash = file_lookup_open(f, flags);
567	}
568
569	if (f->fd == -1) {
570		char buf[FIO_VERROR_SIZE];
571		int __e = errno;
572
573		if (__e == EPERM && (flags & FIO_O_NOATIME)) {
574			flags &= ~FIO_O_NOATIME;
575			goto open_again;
576		}
577		if (__e == EMFILE && file_close_shadow_fds(td))
578			goto open_again;
579
580		snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
581
582		if (__e == EINVAL && (flags & OS_O_DIRECT)) {
583			log_err("fio: looks like your file system does not " \
584				"support direct=1/buffered=0\n");
585		}
586
587		td_verror(td, __e, buf);
588	}
589
590	if (!from_hash && f->fd != -1) {
591		if (add_file_hash(f)) {
592			int fio_unused ret;
593
594			/*
595			 * Stash away descriptor for later close. This is to
596			 * work-around a "feature" on Linux, where a close of
597			 * an fd that has been opened for write will trigger
598			 * udev to call blkid to check partitions, fs id, etc.
599			 * That pollutes the device cache, which can slow down
600			 * unbuffered accesses.
601			 */
602			if (f->shadow_fd == -1)
603				f->shadow_fd = f->fd;
604			else {
605				/*
606			 	 * OK to ignore, we haven't done anything
607				 * with it
608				 */
609				ret = generic_close_file(td, f);
610			}
611			goto open_again;
612		}
613	}
614
615	return 0;
616}
617
618int generic_get_file_size(struct thread_data *td, struct fio_file *f)
619{
620	return get_file_size(td, f);
621}
622
623/*
624 * open/close all files, so that ->real_file_size gets set
625 */
626static int get_file_sizes(struct thread_data *td)
627{
628	struct fio_file *f;
629	unsigned int i;
630	int err = 0;
631
632	for_each_file(td, f, i) {
633		dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
634								f->file_name);
635
636		if (td_io_get_file_size(td, f)) {
637			if (td->error != ENOENT) {
638				log_err("%s\n", td->verror);
639				err = 1;
640			}
641			clear_error(td);
642		}
643
644		if (f->real_file_size == -1ULL && td->o.size)
645			f->real_file_size = td->o.size / td->o.nr_files;
646	}
647
648	return err;
649}
650
651struct fio_mount {
652	struct flist_head list;
653	const char *base;
654	char __base[256];
655	unsigned int key;
656};
657
658/*
659 * Get free number of bytes for each file on each unique mount.
660 */
661static unsigned long long get_fs_free_counts(struct thread_data *td)
662{
663	struct flist_head *n, *tmp;
664	unsigned long long ret = 0;
665	struct fio_mount *fm;
666	FLIST_HEAD(list);
667	struct fio_file *f;
668	unsigned int i;
669
670	for_each_file(td, f, i) {
671		struct stat sb;
672		char buf[256];
673
674		if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
675			if (f->real_file_size != -1ULL)
676				ret += f->real_file_size;
677			continue;
678		} else if (f->filetype != FIO_TYPE_FILE)
679			continue;
680
681		strcpy(buf, f->file_name);
682
683		if (stat(buf, &sb) < 0) {
684			if (errno != ENOENT)
685				break;
686			strcpy(buf, ".");
687			if (stat(buf, &sb) < 0)
688				break;
689		}
690
691		fm = NULL;
692		flist_for_each(n, &list) {
693			fm = flist_entry(n, struct fio_mount, list);
694			if (fm->key == sb.st_dev)
695				break;
696
697			fm = NULL;
698		}
699
700		if (fm)
701			continue;
702
703		fm = malloc(sizeof(*fm));
704		strcpy(fm->__base, buf);
705		fm->base = basename(fm->__base);
706		fm->key = sb.st_dev;
707		flist_add(&fm->list, &list);
708	}
709
710	flist_for_each_safe(n, tmp, &list) {
711		unsigned long long sz;
712
713		fm = flist_entry(n, struct fio_mount, list);
714		flist_del(&fm->list);
715
716		sz = get_fs_size(fm->base);
717		if (sz && sz != -1ULL)
718			ret += sz;
719
720		free(fm);
721	}
722
723	return ret;
724}
725
726uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
727{
728	struct thread_options *o = &td->o;
729
730	if (o->file_append && f->filetype == FIO_TYPE_FILE)
731		return f->real_file_size;
732
733	return td->o.start_offset +
734		(td->thread_number - 1) * td->o.offset_increment;
735}
736
737/*
738 * Open the files and setup files sizes, creating files if necessary.
739 */
740int setup_files(struct thread_data *td)
741{
742	unsigned long long total_size, extend_size;
743	struct thread_options *o = &td->o;
744	struct fio_file *f;
745	unsigned int i, nr_fs_extra = 0;
746	int err = 0, need_extend;
747	int old_state;
748	const unsigned int bs = td_min_bs(td);
749	uint64_t fs = 0;
750
751	dprint(FD_FILE, "setup files\n");
752
753	old_state = td_bump_runstate(td, TD_SETTING_UP);
754
755	if (o->read_iolog_file)
756		goto done;
757
758	/*
759	 * if ioengine defines a setup() method, it's responsible for
760	 * opening the files and setting f->real_file_size to indicate
761	 * the valid range for that file.
762	 */
763	if (td->io_ops->setup)
764		err = td->io_ops->setup(td);
765	else
766		err = get_file_sizes(td);
767
768	if (err)
769		goto err_out;
770
771	/*
772	 * check sizes. if the files/devices do not exist and the size
773	 * isn't passed to fio, abort.
774	 */
775	total_size = 0;
776	for_each_file(td, f, i) {
777		if (f->real_file_size == -1ULL)
778			total_size = -1ULL;
779		else
780			total_size += f->real_file_size;
781	}
782
783	if (o->fill_device)
784		td->fill_device_size = get_fs_free_counts(td);
785
786	/*
787	 * device/file sizes are zero and no size given, punt
788	 */
789	if ((!total_size || total_size == -1ULL) && !o->size &&
790	    !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
791	    !(o->nr_files && (o->file_size_low || o->file_size_high))) {
792		log_err("%s: you need to specify size=\n", o->name);
793		td_verror(td, EINVAL, "total_file_size");
794		goto err_out;
795	}
796
797	/*
798	 * Calculate per-file size and potential extra size for the
799	 * first files, if needed.
800	 */
801	if (!o->file_size_low) {
802		uint64_t all_fs;
803
804		fs = o->size / o->nr_files;
805		all_fs = fs * o->nr_files;
806
807		if (all_fs < o->size)
808			nr_fs_extra = (o->size - all_fs) / bs;
809	}
810
811	/*
812	 * now file sizes are known, so we can set ->io_size. if size= is
813	 * not given, ->io_size is just equal to ->real_file_size. if size
814	 * is given, ->io_size is size / nr_files.
815	 */
816	extend_size = total_size = 0;
817	need_extend = 0;
818	for_each_file(td, f, i) {
819		f->file_offset = get_start_offset(td, f);
820
821		if (!o->file_size_low) {
822			/*
823			 * no file size range given, file size is equal to
824			 * total size divided by number of files. If that is
825			 * zero, set it to the real file size. If the size
826			 * doesn't divide nicely with the min blocksize,
827			 * make the first files bigger.
828			 */
829			f->io_size = fs;
830			if (nr_fs_extra) {
831				nr_fs_extra--;
832				f->io_size += bs;
833			}
834
835			if (!f->io_size)
836				f->io_size = f->real_file_size - f->file_offset;
837		} else if (f->real_file_size < o->file_size_low ||
838			   f->real_file_size > o->file_size_high) {
839			if (f->file_offset > o->file_size_low)
840				goto err_offset;
841			/*
842			 * file size given. if it's fixed, use that. if it's a
843			 * range, generate a random size in-between.
844			 */
845			if (o->file_size_low == o->file_size_high)
846				f->io_size = o->file_size_low - f->file_offset;
847			else {
848				f->io_size = get_rand_file_size(td)
849						- f->file_offset;
850			}
851		} else
852			f->io_size = f->real_file_size - f->file_offset;
853
854		if (f->io_size == -1ULL)
855			total_size = -1ULL;
856		else {
857                        if (o->size_percent)
858                                f->io_size = (f->io_size * o->size_percent) / 100;
859			total_size += f->io_size;
860		}
861
862		if (f->filetype == FIO_TYPE_FILE &&
863		    (f->io_size + f->file_offset) > f->real_file_size &&
864		    !(td->io_ops->flags & FIO_DISKLESSIO)) {
865			if (!o->create_on_open) {
866				need_extend++;
867				extend_size += (f->io_size + f->file_offset);
868			} else
869				f->real_file_size = f->io_size + f->file_offset;
870			fio_file_set_extend(f);
871		}
872	}
873
874	if (!o->size || o->size > total_size)
875		o->size = total_size;
876
877	/*
878	 * See if we need to extend some files
879	 */
880	if (need_extend) {
881		temp_stall_ts = 1;
882		if (output_format == FIO_OUTPUT_NORMAL)
883			log_info("%s: Laying out IO file(s) (%u file(s) /"
884				 " %lluMB)\n", o->name, need_extend,
885					extend_size >> 20);
886
887		for_each_file(td, f, i) {
888			unsigned long long old_len = -1ULL, extend_len = -1ULL;
889
890			if (!fio_file_extend(f))
891				continue;
892
893			assert(f->filetype == FIO_TYPE_FILE);
894			fio_file_clear_extend(f);
895			if (!o->fill_device) {
896				old_len = f->real_file_size;
897				extend_len = f->io_size + f->file_offset -
898						old_len;
899			}
900			f->real_file_size = (f->io_size + f->file_offset);
901			err = extend_file(td, f);
902			if (err)
903				break;
904
905			err = __file_invalidate_cache(td, f, old_len,
906								extend_len);
907			close(f->fd);
908			f->fd = -1;
909			if (err)
910				break;
911		}
912		temp_stall_ts = 0;
913	}
914
915	if (err)
916		goto err_out;
917
918	if (!o->zone_size)
919		o->zone_size = o->size;
920
921	/*
922	 * iolog already set the total io size, if we read back
923	 * stored entries.
924	 */
925	if (!o->read_iolog_file)
926		td->total_io_size = o->size * o->loops;
927
928done:
929	if (o->create_only)
930		td->done = 1;
931
932	td_restore_runstate(td, old_state);
933	return 0;
934err_offset:
935	log_err("%s: you need to specify valid offset=\n", o->name);
936err_out:
937	td_restore_runstate(td, old_state);
938	return 1;
939}
940
941int pre_read_files(struct thread_data *td)
942{
943	struct fio_file *f;
944	unsigned int i;
945
946	dprint(FD_FILE, "pre_read files\n");
947
948	for_each_file(td, f, i) {
949		pre_read_file(td, f);
950	}
951
952	return 1;
953}
954
955static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
956{
957	unsigned int range_size, seed;
958	unsigned long nranges;
959	uint64_t file_size;
960
961	range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
962	file_size = min(f->real_file_size, f->io_size);
963
964	nranges = (file_size + range_size - 1) / range_size;
965
966	seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
967	if (!td->o.rand_repeatable)
968		seed = td->rand_seeds[4];
969
970	if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
971		zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
972	else
973		pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
974
975	return 1;
976}
977
978static int init_rand_distribution(struct thread_data *td)
979{
980	struct fio_file *f;
981	unsigned int i;
982	int state;
983
984	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
985		return 0;
986
987	state = td_bump_runstate(td, TD_SETTING_UP);
988
989	for_each_file(td, f, i)
990		__init_rand_distribution(td, f);
991
992	td_restore_runstate(td, state);
993
994	return 1;
995}
996
997int init_random_map(struct thread_data *td)
998{
999	unsigned long long blocks;
1000	struct fio_file *f;
1001	unsigned int i;
1002
1003	if (init_rand_distribution(td))
1004		return 0;
1005	if (!td_random(td))
1006		return 0;
1007
1008	for_each_file(td, f, i) {
1009		uint64_t file_size = min(f->real_file_size, f->io_size);
1010
1011		blocks = file_size / (unsigned long long) td->o.rw_min_bs;
1012
1013		if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1014			unsigned long seed;
1015
1016			seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1017
1018			if (!lfsr_init(&f->lfsr, blocks, seed, 0))
1019				continue;
1020		} else if (!td->o.norandommap) {
1021			f->io_axmap = axmap_new(blocks);
1022			if (f->io_axmap)
1023				continue;
1024		} else if (td->o.norandommap)
1025			continue;
1026
1027		if (!td->o.softrandommap) {
1028			log_err("fio: failed allocating random map. If running"
1029				" a large number of jobs, try the 'norandommap'"
1030				" option or set 'softrandommap'. Or give"
1031				" a larger --alloc-size to fio.\n");
1032			return 1;
1033		}
1034
1035		log_info("fio: file %s failed allocating random map. Running "
1036			 "job without.\n", f->file_name);
1037	}
1038
1039	return 0;
1040}
1041
1042void close_files(struct thread_data *td)
1043{
1044	struct fio_file *f;
1045	unsigned int i;
1046
1047	for_each_file(td, f, i) {
1048		if (fio_file_open(f))
1049			td_io_close_file(td, f);
1050	}
1051}
1052
1053void close_and_free_files(struct thread_data *td)
1054{
1055	struct fio_file *f;
1056	unsigned int i;
1057
1058	dprint(FD_FILE, "close files\n");
1059
1060	for_each_file(td, f, i) {
1061		if (fio_file_open(f))
1062			td_io_close_file(td, f);
1063
1064		remove_file_hash(f);
1065
1066		if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1067			dprint(FD_FILE, "free unlink %s\n", f->file_name);
1068			unlink(f->file_name);
1069		}
1070
1071		sfree(f->file_name);
1072		f->file_name = NULL;
1073		axmap_free(f->io_axmap);
1074		f->io_axmap = NULL;
1075		sfree(f);
1076	}
1077
1078	td->o.filename = NULL;
1079	free(td->files);
1080	free(td->file_locks);
1081	td->files_index = 0;
1082	td->files = NULL;
1083	td->file_locks = NULL;
1084	td->o.file_lock_mode = FILE_LOCK_NONE;
1085	td->o.nr_files = 0;
1086}
1087
1088static void get_file_type(struct fio_file *f)
1089{
1090	struct stat sb;
1091
1092	if (!strcmp(f->file_name, "-"))
1093		f->filetype = FIO_TYPE_PIPE;
1094	else
1095		f->filetype = FIO_TYPE_FILE;
1096
1097	/* \\.\ is the device namespace in Windows, where every file is
1098	 * a block device */
1099	if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1100		f->filetype = FIO_TYPE_BD;
1101
1102	if (!stat(f->file_name, &sb)) {
1103		if (S_ISBLK(sb.st_mode))
1104			f->filetype = FIO_TYPE_BD;
1105		else if (S_ISCHR(sb.st_mode))
1106			f->filetype = FIO_TYPE_CHAR;
1107		else if (S_ISFIFO(sb.st_mode))
1108			f->filetype = FIO_TYPE_PIPE;
1109	}
1110}
1111
1112static int __is_already_allocated(const char *fname)
1113{
1114	struct flist_head *entry;
1115	char *filename;
1116
1117	if (flist_empty(&filename_list))
1118		return 0;
1119
1120	flist_for_each(entry, &filename_list) {
1121		filename = flist_entry(entry, struct file_name, list)->filename;
1122
1123		if (strcmp(filename, fname) == 0)
1124			return 1;
1125	}
1126
1127	return 0;
1128}
1129
1130static int is_already_allocated(const char *fname)
1131{
1132	int ret;
1133
1134	fio_file_hash_lock();
1135	ret = __is_already_allocated(fname);
1136	fio_file_hash_unlock();
1137	return ret;
1138}
1139
1140static void set_already_allocated(const char *fname)
1141{
1142	struct file_name *fn;
1143
1144	fn = malloc(sizeof(struct file_name));
1145	fn->filename = strdup(fname);
1146
1147	fio_file_hash_lock();
1148	if (!__is_already_allocated(fname)) {
1149		flist_add_tail(&fn->list, &filename_list);
1150		fn = NULL;
1151	}
1152	fio_file_hash_unlock();
1153
1154	if (fn) {
1155		free(fn->filename);
1156		free(fn);
1157	}
1158}
1159
1160
1161static void free_already_allocated(void)
1162{
1163	struct flist_head *entry, *tmp;
1164	struct file_name *fn;
1165
1166	if (flist_empty(&filename_list))
1167		return;
1168
1169	fio_file_hash_lock();
1170	flist_for_each_safe(entry, tmp, &filename_list) {
1171		fn = flist_entry(entry, struct file_name, list);
1172		free(fn->filename);
1173		flist_del(&fn->list);
1174		free(fn);
1175	}
1176
1177	fio_file_hash_unlock();
1178}
1179
1180int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1181{
1182	int cur_files = td->files_index;
1183	char file_name[PATH_MAX];
1184	struct fio_file *f;
1185	int len = 0;
1186
1187	dprint(FD_FILE, "add file %s\n", fname);
1188
1189	if (td->o.directory)
1190		len = set_name_idx(file_name, td->o.directory, numjob);
1191
1192	sprintf(file_name + len, "%s", fname);
1193
1194	/* clean cloned siblings using existing files */
1195	if (numjob && is_already_allocated(file_name))
1196		return 0;
1197
1198	f = smalloc(sizeof(*f));
1199	if (!f) {
1200		log_err("fio: smalloc OOM\n");
1201		assert(0);
1202	}
1203
1204	f->fd = -1;
1205	f->shadow_fd = -1;
1206	fio_file_reset(td, f);
1207
1208	if (td->files_size <= td->files_index) {
1209		unsigned int new_size = td->o.nr_files + 1;
1210
1211		dprint(FD_FILE, "resize file array to %d files\n", new_size);
1212
1213		td->files = realloc(td->files, new_size * sizeof(f));
1214		if (td->files == NULL) {
1215			log_err("fio: realloc OOM\n");
1216			assert(0);
1217		}
1218		if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1219			td->file_locks = realloc(td->file_locks, new_size);
1220			if (!td->file_locks) {
1221				log_err("fio: realloc OOM\n");
1222				assert(0);
1223			}
1224			td->file_locks[cur_files] = FILE_LOCK_NONE;
1225		}
1226		td->files_size = new_size;
1227	}
1228	td->files[cur_files] = f;
1229	f->fileno = cur_files;
1230
1231	/*
1232	 * init function, io engine may not be loaded yet
1233	 */
1234	if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1235		f->real_file_size = -1ULL;
1236
1237	f->file_name = smalloc_strdup(file_name);
1238	if (!f->file_name) {
1239		log_err("fio: smalloc OOM\n");
1240		assert(0);
1241	}
1242
1243	get_file_type(f);
1244
1245	switch (td->o.file_lock_mode) {
1246	case FILE_LOCK_NONE:
1247		break;
1248	case FILE_LOCK_READWRITE:
1249		f->rwlock = fio_rwlock_init();
1250		break;
1251	case FILE_LOCK_EXCLUSIVE:
1252		f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1253		break;
1254	default:
1255		log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1256		assert(0);
1257	}
1258
1259	td->files_index++;
1260	if (f->filetype == FIO_TYPE_FILE)
1261		td->nr_normal_files++;
1262
1263	set_already_allocated(file_name);
1264
1265	/*
1266	 * For adding files after the fact - if openfiles= isn't
1267	 * given as an option, ensure we allow at least one file open
1268	 */
1269	if (!td->o.open_files)
1270		td->o.open_files = 1;
1271
1272	if (inc)
1273		td->o.nr_files++;
1274
1275	dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1276							cur_files);
1277
1278	return cur_files;
1279}
1280
1281int add_file_exclusive(struct thread_data *td, const char *fname)
1282{
1283	struct fio_file *f;
1284	unsigned int i;
1285
1286	for_each_file(td, f, i) {
1287		if (!strcmp(f->file_name, fname))
1288			return i;
1289	}
1290
1291	return add_file(td, fname, 0, 1);
1292}
1293
1294void get_file(struct fio_file *f)
1295{
1296	dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1297	assert(fio_file_open(f));
1298	f->references++;
1299}
1300
1301int put_file(struct thread_data *td, struct fio_file *f)
1302{
1303	int f_ret = 0, ret = 0;
1304
1305	dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1306
1307	if (!fio_file_open(f)) {
1308		assert(f->fd == -1);
1309		return 0;
1310	}
1311
1312	assert(f->references);
1313	if (--f->references)
1314		return 0;
1315
1316	if (should_fsync(td) && td->o.fsync_on_close)
1317		f_ret = fsync(f->fd);
1318
1319	if (td->io_ops->close_file)
1320		ret = td->io_ops->close_file(td, f);
1321
1322	if (!ret)
1323		ret = f_ret;
1324
1325	td->nr_open_files--;
1326	fio_file_clear_open(f);
1327	assert(f->fd == -1);
1328	return ret;
1329}
1330
1331void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1332{
1333	if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1334		return;
1335
1336	if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1337		if (ddir == DDIR_READ)
1338			fio_rwlock_read(f->rwlock);
1339		else
1340			fio_rwlock_write(f->rwlock);
1341	} else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1342		fio_mutex_down(f->lock);
1343
1344	td->file_locks[f->fileno] = td->o.file_lock_mode;
1345}
1346
1347void unlock_file(struct thread_data *td, struct fio_file *f)
1348{
1349	if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1350		return;
1351
1352	if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1353		fio_rwlock_unlock(f->rwlock);
1354	else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1355		fio_mutex_up(f->lock);
1356
1357	td->file_locks[f->fileno] = FILE_LOCK_NONE;
1358}
1359
1360void unlock_file_all(struct thread_data *td, struct fio_file *f)
1361{
1362	if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1363		return;
1364	if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1365		unlock_file(td, f);
1366}
1367
1368static int recurse_dir(struct thread_data *td, const char *dirname)
1369{
1370	struct dirent *dir;
1371	int ret = 0;
1372	DIR *D;
1373
1374	D = opendir(dirname);
1375	if (!D) {
1376		char buf[FIO_VERROR_SIZE];
1377
1378		snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1379		td_verror(td, errno, buf);
1380		return 1;
1381	}
1382
1383	while ((dir = readdir(D)) != NULL) {
1384		char full_path[PATH_MAX];
1385		struct stat sb;
1386
1387		if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1388			continue;
1389
1390		sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1391
1392		if (lstat(full_path, &sb) == -1) {
1393			if (errno != ENOENT) {
1394				td_verror(td, errno, "stat");
1395				return 1;
1396			}
1397		}
1398
1399		if (S_ISREG(sb.st_mode)) {
1400			add_file(td, full_path, 0, 1);
1401			continue;
1402		}
1403		if (!S_ISDIR(sb.st_mode))
1404			continue;
1405
1406		ret = recurse_dir(td, full_path);
1407		if (ret)
1408			break;
1409	}
1410
1411	closedir(D);
1412	return ret;
1413}
1414
1415int add_dir_files(struct thread_data *td, const char *path)
1416{
1417	int ret = recurse_dir(td, path);
1418
1419	if (!ret)
1420		log_info("fio: opendir added %d files\n", td->o.nr_files);
1421
1422	return ret;
1423}
1424
1425void dup_files(struct thread_data *td, struct thread_data *org)
1426{
1427	struct fio_file *f;
1428	unsigned int i;
1429
1430	dprint(FD_FILE, "dup files: %d\n", org->files_index);
1431
1432	if (!org->files)
1433		return;
1434
1435	td->files = malloc(org->files_index * sizeof(f));
1436
1437	if (td->o.file_lock_mode != FILE_LOCK_NONE)
1438		td->file_locks = malloc(org->files_index);
1439
1440	for_each_file(org, f, i) {
1441		struct fio_file *__f;
1442
1443		__f = smalloc(sizeof(*__f));
1444		if (!__f) {
1445			log_err("fio: smalloc OOM\n");
1446			assert(0);
1447		}
1448		__f->fd = -1;
1449		__f->shadow_fd = -1;
1450		fio_file_reset(td, __f);
1451
1452		if (f->file_name) {
1453			__f->file_name = smalloc_strdup(f->file_name);
1454			if (!__f->file_name) {
1455				log_err("fio: smalloc OOM\n");
1456				assert(0);
1457			}
1458
1459			__f->filetype = f->filetype;
1460		}
1461
1462		if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1463			__f->lock = f->lock;
1464		else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1465			__f->rwlock = f->rwlock;
1466
1467		td->files[i] = __f;
1468	}
1469}
1470
1471/*
1472 * Returns the index that matches the filename, or -1 if not there
1473 */
1474int get_fileno(struct thread_data *td, const char *fname)
1475{
1476	struct fio_file *f;
1477	unsigned int i;
1478
1479	for_each_file(td, f, i)
1480		if (!strcmp(f->file_name, fname))
1481			return i;
1482
1483	return -1;
1484}
1485
1486/*
1487 * For log usage, where we add/open/close files automatically
1488 */
1489void free_release_files(struct thread_data *td)
1490{
1491	close_files(td);
1492	td->files_index = 0;
1493	td->nr_normal_files = 0;
1494}
1495
1496void fio_file_reset(struct thread_data *td, struct fio_file *f)
1497{
1498	f->last_pos = f->file_offset;
1499	f->last_start = -1ULL;
1500	if (f->io_axmap)
1501		axmap_reset(f->io_axmap);
1502	if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1503		lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1504}
1505
1506int fio_files_done(struct thread_data *td)
1507{
1508	struct fio_file *f;
1509	unsigned int i;
1510
1511	for_each_file(td, f, i)
1512		if (!fio_file_done(f))
1513			return 0;
1514
1515	return 1;
1516}
1517
1518/* free memory used in initialization phase only */
1519void filesetup_mem_free(void)
1520{
1521	free_already_allocated();
1522}
1523