filesetup.c revision 4d4e80f2b4260f2c8b37a8612ce655502a799f7a
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <assert.h>
5#include <dirent.h>
6#include <sys/stat.h>
7#include <sys/mman.h>
8#include <sys/types.h>
9
10#include "fio.h"
11#include "smalloc.h"
12#include "filehash.h"
13
14static int root_warn;
15
16static int extend_file(struct thread_data *td, struct fio_file *f)
17{
18	int r, new_layout = 0, unlink_file = 0, flags;
19	unsigned long long left;
20	unsigned int bs;
21	char *b;
22
23	if (read_only) {
24		log_err("fio: refusing extend of file due to read-only\n");
25		return 0;
26	}
27
28	/*
29	 * check if we need to lay the file out complete again. fio
30	 * does that for operations involving reads, or for writes
31	 * where overwrite is set
32	 */
33	if (td_read(td) || (td_write(td) && td->o.overwrite))
34		new_layout = 1;
35	if (td_write(td) && !td->o.overwrite)
36		unlink_file = 1;
37
38	if (unlink_file || new_layout) {
39		if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
40			td_verror(td, errno, "unlink");
41			return 1;
42		}
43	}
44
45	flags = O_WRONLY | O_CREAT;
46	if (new_layout)
47		flags |= O_TRUNC;
48
49	dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
50	f->fd = open(f->file_name, flags, 0644);
51	if (f->fd < 0) {
52		td_verror(td, errno, "open");
53		return 1;
54	}
55
56	if (!new_layout)
57		goto done;
58
59	dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
60							f->real_file_size);
61	if (ftruncate(f->fd, f->real_file_size) == -1) {
62		td_verror(td, errno, "ftruncate");
63		goto err;
64	}
65
66	dprint(FD_FILE, "fallocate file %s, size %llu\n", f->file_name,
67							f->real_file_size);
68	if (posix_fallocate(f->fd, 0, f->real_file_size) < 0) {
69		td_verror(td, errno, "posix_fallocate");
70		goto err;
71	}
72
73	b = malloc(td->o.max_bs[DDIR_WRITE]);
74	memset(b, 0, td->o.max_bs[DDIR_WRITE]);
75
76	left = f->real_file_size;
77	while (left && !td->terminate) {
78		bs = td->o.max_bs[DDIR_WRITE];
79		if (bs > left)
80			bs = left;
81
82		r = write(f->fd, b, bs);
83
84		if (r == (int) bs) {
85			left -= bs;
86			continue;
87		} else {
88			if (r < 0)
89				td_verror(td, errno, "write");
90			else
91				td_verror(td, EIO, "write");
92
93			break;
94		}
95	}
96
97	if (td->terminate)
98		unlink(f->file_name);
99	else if (td->o.create_fsync)
100		fsync(f->fd);
101
102	free(b);
103done:
104	close(f->fd);
105	f->fd = -1;
106	return 0;
107err:
108	close(f->fd);
109	f->fd = -1;
110	return 1;
111}
112
113static unsigned long long get_rand_file_size(struct thread_data *td)
114{
115	unsigned long long ret;
116	long r;
117
118	r = os_random_long(&td->file_size_state);
119	ret = td->o.file_size_low + (unsigned long long) ((double) (td->o.file_size_high - td->o.file_size_low) * (r / (RAND_MAX + 1.0)));
120	ret -= (ret % td->o.rw_min_bs);
121	return ret;
122}
123
124static int file_size(struct thread_data *td, struct fio_file *f)
125{
126	struct stat st;
127
128	if (fstat(f->fd, &st) == -1) {
129		td_verror(td, errno, "fstat");
130		return 1;
131	}
132
133	f->real_file_size = st.st_size;
134	return 0;
135}
136
137static int bdev_size(struct thread_data *td, struct fio_file *f)
138{
139	unsigned long long bytes;
140	int r;
141
142	r = blockdev_size(f->fd, &bytes);
143	if (r) {
144		td_verror(td, r, "blockdev_size");
145		return 1;
146	}
147
148	f->real_file_size = bytes;
149	return 0;
150}
151
152static int get_file_size(struct thread_data *td, struct fio_file *f)
153{
154	int ret = 0;
155
156	if (f->flags & FIO_SIZE_KNOWN)
157		return 0;
158
159	if (f->filetype == FIO_TYPE_FILE)
160		ret = file_size(td, f);
161	else if (f->filetype == FIO_TYPE_BD)
162		ret = bdev_size(td, f);
163	else
164		f->real_file_size = -1;
165
166	if (ret)
167		return ret;
168
169	if (f->file_offset > f->real_file_size) {
170		log_err("%s: offset extends end (%Lu > %Lu)\n", td->o.name, f->file_offset, f->real_file_size);
171		return 1;
172	}
173
174	f->flags |= FIO_SIZE_KNOWN;
175	return 0;
176}
177
178int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
179{
180	int ret = 0;
181
182	dprint(FD_IO, "invalidate cache (%d)\n", td->o.odirect);
183
184	if (td->o.odirect)
185		return 0;
186
187	/*
188	 * FIXME: add blockdev flushing too
189	 */
190	if (f->mmap)
191		ret = madvise(f->mmap, f->io_size, MADV_DONTNEED);
192	else if (f->filetype == FIO_TYPE_FILE)
193		ret = fadvise(f->fd, f->file_offset, f->io_size, POSIX_FADV_DONTNEED);
194	else if (f->filetype == FIO_TYPE_BD) {
195		ret = blockdev_invalidate_cache(f->fd);
196		if (ret < 0 && errno == EACCES && geteuid()) {
197			if (!root_warn) {
198				log_err("fio: only root may flush block devices. Cache flush bypassed!\n");
199				root_warn = 1;
200			}
201			ret = 0;
202		}
203	} else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
204		ret = 0;
205
206	if (ret < 0) {
207		td_verror(td, errno, "invalidate_cache");
208		return 1;
209	}
210
211	return ret;
212}
213
214int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
215{
216	int ret = 0;
217
218	dprint(FD_FILE, "fd close %s\n", f->file_name);
219
220	remove_file_hash(f);
221
222	if (close(f->fd) < 0)
223		ret = errno;
224
225	f->fd = -1;
226	return ret;
227}
228
229static int file_lookup_open(struct fio_file *f, int flags)
230{
231	struct fio_file *__f;
232	int from_hash;
233
234	__f = lookup_file_hash(f->file_name);
235	if (__f) {
236		/*
237		 * racy, need the __f->lock locked
238		 */
239		f->lock = __f->lock;
240		f->lock_owner = __f->lock_owner;
241		f->lock_batch = __f->lock_batch;
242		f->lock_ddir = __f->lock_ddir;
243		f->fd = dup(__f->fd);
244		f->references++;
245		from_hash = 1;
246	} else {
247		f->fd = open(f->file_name, flags, 0600);
248		from_hash = 0;
249	}
250
251	return from_hash;
252}
253
254int generic_open_file(struct thread_data *td, struct fio_file *f)
255{
256	int is_std = 0;
257	int flags = 0;
258	int from_hash = 0;
259
260	dprint(FD_FILE, "fd open %s\n", f->file_name);
261
262	if (!strcmp(f->file_name, "-")) {
263		if (td_rw(td)) {
264			log_err("fio: can't read/write to stdin/out\n");
265			return 1;
266		}
267		is_std = 1;
268
269		/*
270		 * move output logging to stderr, if we are writing to stdout
271		 */
272		if (td_write(td))
273			f_out = stderr;
274	}
275
276	if (td->o.odirect)
277		flags |= OS_O_DIRECT;
278	if (td->o.sync_io)
279		flags |= O_SYNC;
280	if (f->filetype != FIO_TYPE_FILE)
281		flags |= O_NOATIME;
282
283open_again:
284	if (td_write(td)) {
285		assert(!read_only);
286
287		flags |= O_RDWR;
288
289		if (f->filetype == FIO_TYPE_FILE)
290			flags |= O_CREAT;
291
292		if (is_std)
293			f->fd = dup(STDOUT_FILENO);
294		else
295			from_hash = file_lookup_open(f, flags);
296	} else {
297		if (f->filetype == FIO_TYPE_CHAR && !read_only)
298			flags |= O_RDWR;
299		else
300			flags |= O_RDONLY;
301
302		if (is_std)
303			f->fd = dup(STDIN_FILENO);
304		else
305			from_hash = file_lookup_open(f, flags);
306	}
307
308	if (f->fd == -1) {
309		char buf[FIO_VERROR_SIZE];
310		int __e = errno;
311
312		if (errno == EPERM && (flags & O_NOATIME)) {
313			flags &= ~O_NOATIME;
314			goto open_again;
315		}
316
317		snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
318
319		td_verror(td, __e, buf);
320	}
321
322	if (get_file_size(td, f))
323		goto err;
324
325	if (!from_hash && f->fd != -1) {
326		if (add_file_hash(f)) {
327			int ret;
328
329			/*
330			 * OK to ignore, we haven't done anything with it
331			 */
332			ret = generic_close_file(td, f);
333			goto open_again;
334		}
335	}
336
337	return 0;
338err:
339	close(f->fd);
340	return 1;
341}
342
343int open_files(struct thread_data *td)
344{
345	struct fio_file *f;
346	unsigned int i;
347	int err = 0;
348
349	dprint(FD_FILE, "open files\n");
350
351	for_each_file(td, f, i) {
352		err = td_io_open_file(td, f);
353		if (err) {
354			if (td->error == EMFILE) {
355				log_err("fio: limited open files to: %d\n", td->nr_open_files);
356				td->o.open_files = td->nr_open_files;
357				err = 0;
358				clear_error(td);
359			}
360			break;
361		}
362
363		if (td->o.open_files == td->nr_open_files)
364			break;
365	}
366
367	if (!err)
368		return 0;
369
370	for_each_file(td, f, i)
371		td_io_close_file(td, f);
372
373	return err;
374}
375
376/*
377 * open/close all files, so that ->real_file_size gets set
378 */
379static int get_file_sizes(struct thread_data *td)
380{
381	struct fio_file *f;
382	unsigned int i;
383	int err = 0;
384
385	for_each_file(td, f, i) {
386		if (td->io_ops->open_file(td, f)) {
387			if (td->error != ENOENT) {
388				log_err("%s\n", td->verror);
389				err = 1;
390			}
391			clear_error(td);
392		} else {
393			if (td->io_ops->close_file)
394				td->io_ops->close_file(td, f);
395		}
396
397		if (f->real_file_size == -1ULL && td->o.size)
398			f->real_file_size = td->o.size / td->o.nr_files;
399	}
400
401	return err;
402}
403
404/*
405 * Open the files and setup files sizes, creating files if necessary.
406 */
407int setup_files(struct thread_data *td)
408{
409	unsigned long long total_size, extend_size;
410	struct fio_file *f;
411	unsigned int i;
412	int err = 0, need_extend;
413
414	dprint(FD_FILE, "setup files\n");
415
416	/*
417	 * if ioengine defines a setup() method, it's responsible for
418	 * opening the files and setting f->real_file_size to indicate
419	 * the valid range for that file.
420	 */
421	if (td->io_ops->setup)
422		err = td->io_ops->setup(td);
423	else
424		err = get_file_sizes(td);
425
426	if (err)
427		return err;
428
429	/*
430	 * check sizes. if the files/devices do not exist and the size
431	 * isn't passed to fio, abort.
432	 */
433	total_size = 0;
434	for_each_file(td, f, i) {
435		if (f->real_file_size == -1ULL)
436			total_size = -1ULL;
437		else
438			total_size += f->real_file_size;
439	}
440
441	/*
442	 * device/file sizes are zero and no size given, punt
443	 */
444	if ((!total_size || total_size == -1ULL) && !td->o.size &&
445	    !(td->io_ops->flags & FIO_NOIO) && !td->o.fill_device) {
446		log_err("%s: you need to specify size=\n", td->o.name);
447		td_verror(td, EINVAL, "total_file_size");
448		return 1;
449	}
450
451	/*
452	 * now file sizes are known, so we can set ->io_size. if size= is
453	 * not given, ->io_size is just equal to ->real_file_size. if size
454	 * is given, ->io_size is size / nr_files.
455	 */
456	extend_size = total_size = 0;
457	need_extend = 0;
458	for_each_file(td, f, i) {
459		f->file_offset = td->o.start_offset;
460
461		if (!td->o.file_size_low) {
462			/*
463			 * no file size range given, file size is equal to
464			 * total size divided by number of files. if that is
465			 * zero, set it to the real file size.
466			 */
467			f->io_size = td->o.size / td->o.nr_files;
468			if (!f->io_size)
469				f->io_size = f->real_file_size - f->file_offset;
470		} else if (f->real_file_size < td->o.file_size_low ||
471			   f->real_file_size > td->o.file_size_high) {
472			if (f->file_offset > td->o.file_size_low)
473				goto err_offset;
474			/*
475			 * file size given. if it's fixed, use that. if it's a
476			 * range, generate a random size in-between.
477			 */
478			if (td->o.file_size_low == td->o.file_size_high)
479				f->io_size = td->o.file_size_low - f->file_offset;
480			else
481				f->io_size = get_rand_file_size(td) - f->file_offset;
482		} else
483			f->io_size = f->real_file_size - f->file_offset;
484
485		if (f->io_size == -1ULL)
486			total_size = -1ULL;
487		else
488			total_size += f->io_size;
489
490		if (f->filetype == FIO_TYPE_FILE &&
491		    (f->io_size + f->file_offset) > f->real_file_size &&
492		    !(td->io_ops->flags & FIO_DISKLESSIO)) {
493			need_extend++;
494			extend_size += (f->io_size + f->file_offset);
495			f->flags |= FIO_FILE_EXTEND;
496		}
497	}
498
499	if (!td->o.size || td->o.size > total_size)
500		td->o.size = total_size;
501
502	/*
503	 * See if we need to extend some files
504	 */
505	if (need_extend) {
506		temp_stall_ts = 1;
507		log_info("%s: Laying out IO file(s) (%u file(s) / %LuMiB)\n",
508			td->o.name, need_extend, extend_size >> 20);
509
510		for_each_file(td, f, i) {
511			if (!(f->flags & FIO_FILE_EXTEND))
512				continue;
513
514			assert(f->filetype == FIO_TYPE_FILE);
515			f->flags &= ~FIO_FILE_EXTEND;
516			f->real_file_size = (f->io_size + f->file_offset);
517			err = extend_file(td, f);
518			if (err)
519				break;
520		}
521		temp_stall_ts = 0;
522	}
523
524	if (err)
525		return err;
526
527	if (!td->o.zone_size)
528		td->o.zone_size = td->o.size;
529
530	/*
531	 * iolog already set the total io size, if we read back
532	 * stored entries.
533	 */
534	if (!td->o.read_iolog_file)
535		td->total_io_size = td->o.size * td->o.loops;
536	return 0;
537err_offset:
538	log_err("%s: you need to specify valid offset=\n", td->o.name);
539	return 1;
540}
541
542int init_random_map(struct thread_data *td)
543{
544	unsigned long long blocks, num_maps;
545	struct fio_file *f;
546	unsigned int i;
547
548	if (td->o.norandommap || !td_random(td))
549		return 0;
550
551	for_each_file(td, f, i) {
552		blocks = (f->real_file_size + td->o.rw_min_bs - 1) / (unsigned long long) td->o.rw_min_bs;
553		num_maps = (blocks + BLOCKS_PER_MAP-1)/ (unsigned long long) BLOCKS_PER_MAP;
554		f->file_map = smalloc(num_maps * sizeof(long));
555		if (!f->file_map) {
556			log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
557			return 1;
558		}
559		f->num_maps = num_maps;
560	}
561
562	return 0;
563}
564
565void close_files(struct thread_data *td)
566{
567	struct fio_file *f;
568	unsigned int i;
569
570	for_each_file(td, f, i)
571		td_io_close_file(td, f);
572}
573
574void close_and_free_files(struct thread_data *td)
575{
576	struct fio_file *f;
577	unsigned int i;
578
579	dprint(FD_FILE, "close files\n");
580
581	for_each_file(td, f, i) {
582		if (td->o.unlink && f->filetype == FIO_TYPE_FILE)
583			unlink(f->file_name);
584
585		td_io_close_file(td, f);
586
587		sfree(f->file_name);
588		f->file_name = NULL;
589
590		if (f->file_map) {
591			sfree(f->file_map);
592			f->file_map = NULL;
593		}
594		sfree(f);
595	}
596
597	td->o.filename = NULL;
598	free(td->files);
599	td->files = NULL;
600	td->o.nr_files = 0;
601}
602
603static void get_file_type(struct fio_file *f)
604{
605	struct stat sb;
606
607	if (!strcmp(f->file_name, "-"))
608		f->filetype = FIO_TYPE_PIPE;
609	else
610		f->filetype = FIO_TYPE_FILE;
611
612	if (!lstat(f->file_name, &sb)) {
613		if (S_ISBLK(sb.st_mode))
614			f->filetype = FIO_TYPE_BD;
615		else if (S_ISCHR(sb.st_mode))
616			f->filetype = FIO_TYPE_CHAR;
617		else if (S_ISFIFO(sb.st_mode))
618			f->filetype = FIO_TYPE_PIPE;
619	}
620}
621
622int add_file(struct thread_data *td, const char *fname)
623{
624	int cur_files = td->files_index;
625	char file_name[PATH_MAX];
626	struct fio_file *f;
627	int len = 0;
628
629	dprint(FD_FILE, "add file %s\n", fname);
630
631	f = smalloc(sizeof(*f));
632	f->fd = -1;
633
634	td->files = realloc(td->files, (cur_files + 1) * sizeof(f));
635
636	td->files[cur_files] = f;
637
638	/*
639	 * init function, io engine may not be loaded yet
640	 */
641	if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
642		f->real_file_size = -1ULL;
643
644	if (td->o.directory)
645		len = sprintf(file_name, "%s/", td->o.directory);
646
647	sprintf(file_name + len, "%s", fname);
648	f->file_name = smalloc_strdup(file_name);
649
650	get_file_type(f);
651
652	switch (td->o.file_lock_mode) {
653	case FILE_LOCK_NONE:
654		break;
655	case FILE_LOCK_READWRITE:
656		f->lock = fio_mutex_rw_init();
657		break;
658	case FILE_LOCK_EXCLUSIVE:
659		f->lock = fio_mutex_init(1);
660		break;
661	default:
662		log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
663		assert(0);
664	}
665
666	td->files_index++;
667	if (f->filetype == FIO_TYPE_FILE)
668		td->nr_normal_files++;
669
670	return cur_files;
671}
672
673void get_file(struct fio_file *f)
674{
675	dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
676	assert(f->flags & FIO_FILE_OPEN);
677	f->references++;
678}
679
680int put_file(struct thread_data *td, struct fio_file *f)
681{
682	int ret = 0;
683
684	dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
685
686	if (!(f->flags & FIO_FILE_OPEN))
687		return 0;
688
689	assert(f->references);
690	if (--f->references)
691		return 0;
692
693	if (should_fsync(td) && td->o.fsync_on_close)
694		fsync(f->fd);
695
696	if (td->io_ops->close_file)
697		ret = td->io_ops->close_file(td, f);
698
699	td->nr_open_files--;
700	f->flags &= ~FIO_FILE_OPEN;
701	return ret;
702}
703
704void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
705{
706	if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
707		return;
708
709	if (f->lock_owner == td && f->lock_batch--)
710		return;
711
712	if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
713		if (ddir == DDIR_READ)
714			fio_mutex_down_read(f->lock);
715		else
716			fio_mutex_down_write(f->lock);
717	} else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
718		fio_mutex_down(f->lock);
719
720	f->lock_owner = td;
721	f->lock_batch = td->o.lockfile_batch;
722	f->lock_ddir = ddir;
723}
724
725void unlock_file(struct thread_data *td, struct fio_file *f)
726{
727	if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
728		return;
729	if (f->lock_batch)
730		return;
731
732	if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
733		const int is_read = f->lock_ddir == DDIR_READ;
734		int val = fio_mutex_getval(f->lock);
735
736		if ((is_read && val == 1) || (!is_read && val == -1))
737			f->lock_owner = NULL;
738
739		if (is_read)
740			fio_mutex_up_read(f->lock);
741		else
742			fio_mutex_up_write(f->lock);
743	} else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) {
744		int val = fio_mutex_getval(f->lock);
745
746		if (val == 0)
747			f->lock_owner = NULL;
748
749		fio_mutex_up(f->lock);
750	}
751}
752
753void unlock_file_all(struct thread_data *td, struct fio_file *f)
754{
755	if (f->lock_owner != td)
756		return;
757
758	f->lock_batch = 0;
759	unlock_file(td, f);
760}
761
762static int recurse_dir(struct thread_data *td, const char *dirname)
763{
764	struct dirent *dir;
765	int ret = 0;
766	DIR *D;
767
768	D = opendir(dirname);
769	if (!D) {
770		char buf[FIO_VERROR_SIZE];
771
772		snprintf(buf, FIO_VERROR_SIZE - 1, "opendir(%s)", dirname);
773		td_verror(td, errno, buf);
774		return 1;
775	}
776
777	while ((dir = readdir(D)) != NULL) {
778		char full_path[PATH_MAX];
779		struct stat sb;
780
781		if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
782			continue;
783
784		sprintf(full_path, "%s/%s", dirname, dir->d_name);
785
786		if (lstat(full_path, &sb) == -1) {
787			if (errno != ENOENT) {
788				td_verror(td, errno, "stat");
789				return 1;
790			}
791		}
792
793		if (S_ISREG(sb.st_mode)) {
794			add_file(td, full_path);
795			td->o.nr_files++;
796			continue;
797		}
798		if (!S_ISDIR(sb.st_mode))
799			continue;
800
801		if ((ret = recurse_dir(td, full_path)) != 0)
802			break;
803	}
804
805	closedir(D);
806	return ret;
807}
808
809int add_dir_files(struct thread_data *td, const char *path)
810{
811	int ret = recurse_dir(td, path);
812
813	if (!ret)
814		log_info("fio: opendir added %d files\n", td->o.nr_files);
815
816	return ret;
817}
818
819void dup_files(struct thread_data *td, struct thread_data *org)
820{
821	struct fio_file *f;
822	unsigned int i;
823	size_t bytes;
824
825	if (!org->files)
826		return;
827
828	bytes = org->files_index * sizeof(f);
829	td->files = malloc(bytes);
830	memcpy(td->files, org->files, bytes);
831
832	for_each_file(td, f, i) {
833		struct fio_file *__f;
834
835		__f = smalloc(sizeof(*__f));
836
837		if (f->file_name)
838			__f->file_name = smalloc_strdup(f->file_name);
839
840		td->files[i] = __f;
841	}
842}
843
844/*
845 * Returns the index that matches the filename, or -1 if not there
846 */
847int get_fileno(struct thread_data *td, const char *fname)
848{
849	struct fio_file *f;
850	unsigned int i;
851
852	for_each_file(td, f, i)
853		if (!strcmp(f->file_name, fname))
854			return i;
855
856	return -1;
857}
858
859/*
860 * For log usage, where we add/open/close files automatically
861 */
862void free_release_files(struct thread_data *td)
863{
864	close_files(td);
865	td->files_index = 0;
866	td->nr_normal_files = 0;
867}
868