unix_io.c revision 534a4c3db58064caee4fc3e9e294251240d9d28a
1/*
2 * unix_io.c --- This is the Unix (well, really POSIX) implementation
3 * 	of the I/O manager.
4 *
5 * Implements a one-block write-through cache.
6 *
7 * Includes support for Windows NT support under Cygwin.
8 *
9 * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
10 * 	2002 by Theodore Ts'o.
11 *
12 * %Begin-Header%
13 * This file may be redistributed under the terms of the GNU Library
14 * General Public License, version 2.
15 * %End-Header%
16 */
17
18#define _LARGEFILE_SOURCE
19#define _LARGEFILE64_SOURCE
20#ifndef _GNU_SOURCE
21#define _GNU_SOURCE
22#endif
23
24#include <stdio.h>
25#include <string.h>
26#if HAVE_UNISTD_H
27#include <unistd.h>
28#endif
29#if HAVE_ERRNO_H
30#include <errno.h>
31#endif
32#include <fcntl.h>
33#include <time.h>
34#ifdef __linux__
35#include <sys/utsname.h>
36#endif
37#ifdef HAVE_SYS_IOCTL_H
38#include <sys/ioctl.h>
39#endif
40#ifdef HAVE_SYS_MOUNT_H
41#include <sys/mount.h>
42#endif
43#if HAVE_SYS_STAT_H
44#include <sys/stat.h>
45#endif
46#if HAVE_SYS_TYPES_H
47#include <sys/types.h>
48#endif
49#if HAVE_SYS_RESOURCE_H
50#include <sys/resource.h>
51#endif
52
53#if defined(__linux__) && defined(_IO) && !defined(BLKROGET)
54#define BLKROGET   _IO(0x12, 94) /* Get read-only status (0 = read_write).  */
55#endif
56
57#if defined(__linux__) && defined(_IO) && !defined(BLKSSZGET)
58#define BLKSSZGET  _IO(0x12,104)/* get block device sector size */
59#endif
60
61#undef ALIGN_DEBUG
62
63#include "ext2_fs.h"
64#include "ext2fs.h"
65
66/*
67 * For checking structure magic numbers...
68 */
69
70#define EXT2_CHECK_MAGIC(struct, code) \
71	  if ((struct)->magic != (code)) return (code)
72
73struct unix_cache {
74	char		*buf;
75	unsigned long	block;
76	int		access_time;
77	unsigned	dirty:1;
78	unsigned	in_use:1;
79};
80
81#define CACHE_SIZE 8
82#define WRITE_DIRECT_SIZE 4	/* Must be smaller than CACHE_SIZE */
83#define READ_DIRECT_SIZE 4	/* Should be smaller than CACHE_SIZE */
84
85struct unix_private_data {
86	int	magic;
87	int	dev;
88	int	flags;
89	int	align;
90	int	access_time;
91	ext2_loff_t offset;
92	struct unix_cache cache[CACHE_SIZE];
93	void	*bounce;
94	struct struct_io_stats io_stats;
95};
96
97#define IS_ALIGNED(n, align) ((((unsigned long) n) & \
98			       ((unsigned long) ((align)-1))) == 0)
99
100static errcode_t unix_open(const char *name, int flags, io_channel *channel);
101static errcode_t unix_close(io_channel channel);
102static errcode_t unix_set_blksize(io_channel channel, int blksize);
103static errcode_t unix_read_blk(io_channel channel, unsigned long block,
104			       int count, void *data);
105static errcode_t unix_write_blk(io_channel channel, unsigned long block,
106				int count, const void *data);
107static errcode_t unix_flush(io_channel channel);
108static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
109				int size, const void *data);
110static errcode_t unix_set_option(io_channel channel, const char *option,
111				 const char *arg);
112static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
113;
114static void reuse_cache(io_channel channel, struct unix_private_data *data,
115		 struct unix_cache *cache, unsigned long long block);
116static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
117			       int count, void *data);
118static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
119				int count, const void *data);
120static errcode_t unix_discard(io_channel channel, unsigned long long block,
121			      unsigned long long count);
122
123static struct struct_io_manager struct_unix_manager = {
124	EXT2_ET_MAGIC_IO_MANAGER,
125	"Unix I/O Manager",
126	unix_open,
127	unix_close,
128	unix_set_blksize,
129	unix_read_blk,
130	unix_write_blk,
131	unix_flush,
132	unix_write_byte,
133	unix_set_option,
134	unix_get_stats,
135	unix_read_blk64,
136	unix_write_blk64,
137	unix_discard,
138};
139
140io_manager unix_io_manager = &struct_unix_manager;
141
142static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
143{
144	errcode_t 	retval = 0;
145
146	struct unix_private_data *data;
147
148	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
149	data = (struct unix_private_data *) channel->private_data;
150	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
151
152	if (stats)
153		*stats = &data->io_stats;
154
155	return retval;
156}
157
158/*
159 * Here are the raw I/O functions
160 */
161static errcode_t raw_read_blk(io_channel channel,
162			      struct unix_private_data *data,
163			      unsigned long long block,
164			      int count, void *buf)
165{
166	errcode_t	retval;
167	ssize_t		size;
168	ext2_loff_t	location;
169	int		actual = 0;
170
171	size = (count < 0) ? -count : count * channel->block_size;
172	data->io_stats.bytes_read += size;
173	location = ((ext2_loff_t) block * channel->block_size) + data->offset;
174	if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
175		retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
176		goto error_out;
177	}
178	if ((data->align == 0) ||
179	    ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
180		actual = read(data->dev, buf, size);
181		if (actual != size) {
182		short_read:
183			if (actual < 0)
184				actual = 0;
185			retval = EXT2_ET_SHORT_READ;
186			goto error_out;
187		}
188		return 0;
189	}
190
191#ifdef ALIGN_DEBUG
192	printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
193	       (unsigned long) size);
194#endif
195
196	/*
197	 * The buffer or size which we're trying to read isn't aligned
198	 * to the O_DIRECT rules, so we need to do this the hard way...
199	 */
200	while (size > 0) {
201		actual = read(data->dev, data->bounce, channel->block_size);
202		if (actual != channel->block_size)
203			goto short_read;
204		actual = size;
205		if (size > channel->block_size)
206			actual = channel->block_size;
207		memcpy(buf, data->bounce, actual);
208		size -= actual;
209		buf += actual;
210	}
211	return 0;
212
213error_out:
214	memset((char *) buf+actual, 0, size-actual);
215	if (channel->read_error)
216		retval = (channel->read_error)(channel, block, count, buf,
217					       size, actual, retval);
218	return retval;
219}
220
221static errcode_t raw_write_blk(io_channel channel,
222			       struct unix_private_data *data,
223			       unsigned long long block,
224			       int count, const void *buf)
225{
226	ssize_t		size;
227	ext2_loff_t	location;
228	int		actual = 0;
229	errcode_t	retval;
230
231	if (count == 1)
232		size = channel->block_size;
233	else {
234		if (count < 0)
235			size = -count;
236		else
237			size = count * channel->block_size;
238	}
239	data->io_stats.bytes_written += size;
240
241	location = ((ext2_loff_t) block * channel->block_size) + data->offset;
242	if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
243		retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
244		goto error_out;
245	}
246
247	if ((data->align == 0) ||
248	    ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
249		actual = write(data->dev, buf, size);
250		if (actual != size) {
251		short_write:
252			retval = EXT2_ET_SHORT_WRITE;
253			goto error_out;
254		}
255		return 0;
256	}
257
258#ifdef ALIGN_DEBUG
259	printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
260	       (unsigned long) size);
261#endif
262	/*
263	 * The buffer or size which we're trying to write isn't aligned
264	 * to the O_DIRECT rules, so we need to do this the hard way...
265	 */
266	while (size > 0) {
267		if (size < channel->block_size) {
268			actual = read(data->dev, data->bounce,
269				      channel->block_size);
270			if (actual != channel->block_size) {
271				retval = EXT2_ET_SHORT_READ;
272				goto error_out;
273			}
274		}
275		actual = size;
276		if (size > channel->block_size)
277			actual = channel->block_size;
278		memcpy(data->bounce, buf, actual);
279		actual = write(data->dev, data->bounce, channel->block_size);
280		if (actual != channel->block_size)
281			goto short_write;
282		size -= actual;
283		buf += actual;
284	}
285	return 0;
286
287error_out:
288	if (channel->write_error)
289		retval = (channel->write_error)(channel, block, count, buf,
290						size, actual, retval);
291	return retval;
292}
293
294
295/*
296 * Here we implement the cache functions
297 */
298
299/* Allocate the cache buffers */
300static errcode_t alloc_cache(io_channel channel,
301			     struct unix_private_data *data)
302{
303	errcode_t		retval;
304	struct unix_cache	*cache;
305	int			i;
306
307	data->access_time = 0;
308	for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
309		cache->block = 0;
310		cache->access_time = 0;
311		cache->dirty = 0;
312		cache->in_use = 0;
313		if (cache->buf)
314			ext2fs_free_mem(&cache->buf);
315		retval = ext2fs_get_memalign(channel->block_size,
316					     data->align, &cache->buf);
317		if (retval)
318			return retval;
319	}
320	if (data->align) {
321		if (data->bounce)
322			ext2fs_free_mem(&data->bounce);
323		retval = ext2fs_get_memalign(channel->block_size, data->align,
324					     &data->bounce);
325	}
326	return retval;
327}
328
329/* Free the cache buffers */
330static void free_cache(struct unix_private_data *data)
331{
332	struct unix_cache	*cache;
333	int			i;
334
335	data->access_time = 0;
336	for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
337		cache->block = 0;
338		cache->access_time = 0;
339		cache->dirty = 0;
340		cache->in_use = 0;
341		if (cache->buf)
342			ext2fs_free_mem(&cache->buf);
343	}
344	if (data->bounce)
345		ext2fs_free_mem(&data->bounce);
346}
347
348#ifndef NO_IO_CACHE
349/*
350 * Try to find a block in the cache.  If the block is not found, and
351 * eldest is a non-zero pointer, then fill in eldest with the cache
352 * entry to that should be reused.
353 */
354static struct unix_cache *find_cached_block(struct unix_private_data *data,
355					    unsigned long long block,
356					    struct unix_cache **eldest)
357{
358	struct unix_cache	*cache, *unused_cache, *oldest_cache;
359	int			i;
360
361	unused_cache = oldest_cache = 0;
362	for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
363		if (!cache->in_use) {
364			if (!unused_cache)
365				unused_cache = cache;
366			continue;
367		}
368		if (cache->block == block) {
369			cache->access_time = ++data->access_time;
370			return cache;
371		}
372		if (!oldest_cache ||
373		    (cache->access_time < oldest_cache->access_time))
374			oldest_cache = cache;
375	}
376	if (eldest)
377		*eldest = (unused_cache) ? unused_cache : oldest_cache;
378	return 0;
379}
380
381/*
382 * Reuse a particular cache entry for another block.
383 */
384static void reuse_cache(io_channel channel, struct unix_private_data *data,
385		 struct unix_cache *cache, unsigned long long block)
386{
387	if (cache->dirty && cache->in_use)
388		raw_write_blk(channel, data, cache->block, 1, cache->buf);
389
390	cache->in_use = 1;
391	cache->dirty = 0;
392	cache->block = block;
393	cache->access_time = ++data->access_time;
394}
395
396/*
397 * Flush all of the blocks in the cache
398 */
399static errcode_t flush_cached_blocks(io_channel channel,
400				     struct unix_private_data *data,
401				     int invalidate)
402
403{
404	struct unix_cache	*cache;
405	errcode_t		retval, retval2;
406	int			i;
407
408	retval2 = 0;
409	for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
410		if (!cache->in_use)
411			continue;
412
413		if (invalidate)
414			cache->in_use = 0;
415
416		if (!cache->dirty)
417			continue;
418
419		retval = raw_write_blk(channel, data,
420				       cache->block, 1, cache->buf);
421		if (retval)
422			retval2 = retval;
423		else
424			cache->dirty = 0;
425	}
426	return retval2;
427}
428#endif /* NO_IO_CACHE */
429
430#ifdef __linux__
431#ifndef BLKDISCARDZEROES
432#define BLKDISCARDZEROES _IO(0x12,124)
433#endif
434#endif
435
436static errcode_t unix_open(const char *name, int flags, io_channel *channel)
437{
438	io_channel	io = NULL;
439	struct unix_private_data *data = NULL;
440	errcode_t	retval;
441	int		open_flags, zeroes = 0;
442	struct stat	st;
443#ifdef __linux__
444	struct 		utsname ut;
445#endif
446
447	if (name == 0)
448		return EXT2_ET_BAD_DEVICE_NAME;
449	retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
450	if (retval)
451		return retval;
452	memset(io, 0, sizeof(struct struct_io_channel));
453	io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
454	retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
455	if (retval)
456		goto cleanup;
457
458	io->manager = unix_io_manager;
459	retval = ext2fs_get_mem(strlen(name)+1, &io->name);
460	if (retval)
461		goto cleanup;
462
463	strcpy(io->name, name);
464	io->private_data = data;
465	io->block_size = 1024;
466	io->read_error = 0;
467	io->write_error = 0;
468	io->refcount = 1;
469
470	memset(data, 0, sizeof(struct unix_private_data));
471	data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
472	data->io_stats.num_fields = 2;
473
474	open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
475	if (flags & IO_FLAG_EXCLUSIVE)
476		open_flags |= O_EXCL;
477#ifdef O_DIRECT
478	if (flags & IO_FLAG_DIRECT_IO)
479		open_flags |= O_DIRECT;
480#endif
481	data->flags = flags;
482
483#ifdef HAVE_OPEN64
484	data->dev = open64(io->name, open_flags);
485#else
486	data->dev = open(io->name, open_flags);
487#endif
488	if (data->dev < 0) {
489		retval = errno;
490		goto cleanup;
491	}
492
493#ifdef BLKSSZGET
494	if (flags & IO_FLAG_DIRECT_IO) {
495		if (ioctl(data->dev, BLKSSZGET, &data->align) != 0)
496			data->align = io->block_size;
497	}
498#endif
499
500#ifdef BLKDISCARDZEROES
501	ioctl(data->dev, BLKDISCARDZEROES, &zeroes);
502	if (zeroes)
503		io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
504#endif
505
506#if defined(__CYGWIN__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
507	/*
508	 * Some operating systems require that the buffers be aligned,
509	 * regardless of O_DIRECT
510	 */
511	data->align = 512;
512#endif
513
514
515	if ((retval = alloc_cache(io, data)))
516		goto cleanup;
517
518#ifdef BLKROGET
519	if (flags & IO_FLAG_RW) {
520		int error;
521		int readonly = 0;
522
523		/* Is the block device actually writable? */
524		error = ioctl(data->dev, BLKROGET, &readonly);
525		if (!error && readonly) {
526			close(data->dev);
527			retval = EPERM;
528			goto cleanup;
529		}
530	}
531#endif
532
533#ifdef __linux__
534#undef RLIM_INFINITY
535#if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
536#define RLIM_INFINITY	((unsigned long)(~0UL>>1))
537#else
538#define RLIM_INFINITY  (~0UL)
539#endif
540	/*
541	 * Work around a bug in 2.4.10-2.4.18 kernels where writes to
542	 * block devices are wrongly getting hit by the filesize
543	 * limit.  This workaround isn't perfect, since it won't work
544	 * if glibc wasn't built against 2.2 header files.  (Sigh.)
545	 *
546	 */
547	if ((flags & IO_FLAG_RW) &&
548	    (uname(&ut) == 0) &&
549	    ((ut.release[0] == '2') && (ut.release[1] == '.') &&
550	     (ut.release[2] == '4') && (ut.release[3] == '.') &&
551	     (ut.release[4] == '1') && (ut.release[5] >= '0') &&
552	     (ut.release[5] < '8')) &&
553	    (fstat(data->dev, &st) == 0) &&
554	    (S_ISBLK(st.st_mode))) {
555		struct rlimit	rlim;
556
557		rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
558		setrlimit(RLIMIT_FSIZE, &rlim);
559		getrlimit(RLIMIT_FSIZE, &rlim);
560		if (((unsigned long) rlim.rlim_cur) <
561		    ((unsigned long) rlim.rlim_max)) {
562			rlim.rlim_cur = rlim.rlim_max;
563			setrlimit(RLIMIT_FSIZE, &rlim);
564		}
565	}
566#endif
567	*channel = io;
568	return 0;
569
570cleanup:
571	if (data) {
572		free_cache(data);
573		ext2fs_free_mem(&data);
574	}
575	if (io)
576		ext2fs_free_mem(&io);
577	return retval;
578}
579
580static errcode_t unix_close(io_channel channel)
581{
582	struct unix_private_data *data;
583	errcode_t	retval = 0;
584
585	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
586	data = (struct unix_private_data *) channel->private_data;
587	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
588
589	if (--channel->refcount > 0)
590		return 0;
591
592#ifndef NO_IO_CACHE
593	retval = flush_cached_blocks(channel, data, 0);
594#endif
595
596	if (close(data->dev) < 0)
597		retval = errno;
598	free_cache(data);
599
600	ext2fs_free_mem(&channel->private_data);
601	if (channel->name)
602		ext2fs_free_mem(&channel->name);
603	ext2fs_free_mem(&channel);
604	return retval;
605}
606
607static errcode_t unix_set_blksize(io_channel channel, int blksize)
608{
609	struct unix_private_data *data;
610	errcode_t		retval;
611
612	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
613	data = (struct unix_private_data *) channel->private_data;
614	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
615
616	if (channel->block_size != blksize) {
617#ifndef NO_IO_CACHE
618		if ((retval = flush_cached_blocks(channel, data, 0)))
619			return retval;
620#endif
621
622		channel->block_size = blksize;
623		free_cache(data);
624		if ((retval = alloc_cache(channel, data)))
625			return retval;
626	}
627	return 0;
628}
629
630
631static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
632			       int count, void *buf)
633{
634	struct unix_private_data *data;
635	struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
636	errcode_t	retval;
637	char		*cp;
638	int		i, j;
639
640	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
641	data = (struct unix_private_data *) channel->private_data;
642	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
643
644#ifdef NO_IO_CACHE
645	return raw_read_blk(channel, data, block, count, buf);
646#else
647	/*
648	 * If we're doing an odd-sized read or a very large read,
649	 * flush out the cache and then do a direct read.
650	 */
651	if (count < 0 || count > WRITE_DIRECT_SIZE) {
652		if ((retval = flush_cached_blocks(channel, data, 0)))
653			return retval;
654		return raw_read_blk(channel, data, block, count, buf);
655	}
656
657	cp = buf;
658	while (count > 0) {
659		/* If it's in the cache, use it! */
660		if ((cache = find_cached_block(data, block, &reuse[0]))) {
661#ifdef DEBUG
662			printf("Using cached block %lu\n", block);
663#endif
664			memcpy(cp, cache->buf, channel->block_size);
665			count--;
666			block++;
667			cp += channel->block_size;
668			continue;
669		}
670		if (count == 1) {
671			/*
672			 * Special case where we read directly into the
673			 * cache buffer; important in the O_DIRECT case
674			 */
675			cache = reuse[0];
676			reuse_cache(channel, data, cache, block);
677			if ((retval = raw_read_blk(channel, data, block, 1,
678						   cache->buf))) {
679				cache->in_use = 0;
680				return retval;
681			}
682			memcpy(cp, cache->buf, channel->block_size);
683			return 0;
684		}
685
686		/*
687		 * Find the number of uncached blocks so we can do a
688		 * single read request
689		 */
690		for (i=1; i < count; i++)
691			if (find_cached_block(data, block+i, &reuse[i]))
692				break;
693#ifdef DEBUG
694		printf("Reading %d blocks starting at %lu\n", i, block);
695#endif
696		if ((retval = raw_read_blk(channel, data, block, i, cp)))
697			return retval;
698
699		/* Save the results in the cache */
700		for (j=0; j < i; j++) {
701			count--;
702			cache = reuse[j];
703			reuse_cache(channel, data, cache, block++);
704			memcpy(cache->buf, cp, channel->block_size);
705			cp += channel->block_size;
706		}
707	}
708	return 0;
709#endif /* NO_IO_CACHE */
710}
711
712static errcode_t unix_read_blk(io_channel channel, unsigned long block,
713			       int count, void *buf)
714{
715	return unix_read_blk64(channel, block, count, buf);
716}
717
718static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
719				int count, const void *buf)
720{
721	struct unix_private_data *data;
722	struct unix_cache *cache, *reuse;
723	errcode_t	retval = 0;
724	const char	*cp;
725	int		writethrough;
726
727	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
728	data = (struct unix_private_data *) channel->private_data;
729	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
730
731#ifdef NO_IO_CACHE
732	return raw_write_blk(channel, data, block, count, buf);
733#else
734	/*
735	 * If we're doing an odd-sized write or a very large write,
736	 * flush out the cache completely and then do a direct write.
737	 */
738	if (count < 0 || count > WRITE_DIRECT_SIZE) {
739		if ((retval = flush_cached_blocks(channel, data, 1)))
740			return retval;
741		return raw_write_blk(channel, data, block, count, buf);
742	}
743
744	/*
745	 * For a moderate-sized multi-block write, first force a write
746	 * if we're in write-through cache mode, and then fill the
747	 * cache with the blocks.
748	 */
749	writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
750	if (writethrough)
751		retval = raw_write_blk(channel, data, block, count, buf);
752
753	cp = buf;
754	while (count > 0) {
755		cache = find_cached_block(data, block, &reuse);
756		if (!cache) {
757			cache = reuse;
758			reuse_cache(channel, data, cache, block);
759		}
760		memcpy(cache->buf, cp, channel->block_size);
761		cache->dirty = !writethrough;
762		count--;
763		block++;
764		cp += channel->block_size;
765	}
766	return retval;
767#endif /* NO_IO_CACHE */
768}
769
770static errcode_t unix_write_blk(io_channel channel, unsigned long block,
771				int count, const void *buf)
772{
773	return unix_write_blk64(channel, block, count, buf);
774}
775
776static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
777				 int size, const void *buf)
778{
779	struct unix_private_data *data;
780	errcode_t	retval = 0;
781	ssize_t		actual;
782
783	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
784	data = (struct unix_private_data *) channel->private_data;
785	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
786
787	if (data->align != 0) {
788#ifdef ALIGN_DEBUG
789		printf("unix_write_byte: O_DIRECT fallback\n");
790#endif
791		return EXT2_ET_UNIMPLEMENTED;
792	}
793
794#ifndef NO_IO_CACHE
795	/*
796	 * Flush out the cache completely
797	 */
798	if ((retval = flush_cached_blocks(channel, data, 1)))
799		return retval;
800#endif
801
802	if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
803		return errno;
804
805	actual = write(data->dev, buf, size);
806	if (actual != size)
807		return EXT2_ET_SHORT_WRITE;
808
809	return 0;
810}
811
812/*
813 * Flush data buffers to disk.
814 */
815static errcode_t unix_flush(io_channel channel)
816{
817	struct unix_private_data *data;
818	errcode_t retval = 0;
819
820	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
821	data = (struct unix_private_data *) channel->private_data;
822	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
823
824#ifndef NO_IO_CACHE
825	retval = flush_cached_blocks(channel, data, 0);
826#endif
827	fsync(data->dev);
828	return retval;
829}
830
831static errcode_t unix_set_option(io_channel channel, const char *option,
832				 const char *arg)
833{
834	struct unix_private_data *data;
835	unsigned long long tmp;
836	char *end;
837
838	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
839	data = (struct unix_private_data *) channel->private_data;
840	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
841
842	if (!strcmp(option, "offset")) {
843		if (!arg)
844			return EXT2_ET_INVALID_ARGUMENT;
845
846		tmp = strtoull(arg, &end, 0);
847		if (*end)
848			return EXT2_ET_INVALID_ARGUMENT;
849		data->offset = tmp;
850		if (data->offset < 0)
851			return EXT2_ET_INVALID_ARGUMENT;
852		return 0;
853	}
854	return EXT2_ET_INVALID_ARGUMENT;
855}
856
857#if defined(__linux__) && !defined(BLKDISCARD)
858#define BLKDISCARD	_IO(0x12,119)
859#endif
860
861static errcode_t unix_discard(io_channel channel, unsigned long long block,
862			      unsigned long long count)
863{
864#ifdef BLKDISCARD
865	struct unix_private_data *data;
866	__uint64_t	range[2];
867	int		ret;
868
869	EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
870	data = (struct unix_private_data *) channel->private_data;
871	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
872
873	range[0] = (__uint64_t)(block) * channel->block_size;
874	range[1] = (__uint64_t)(count) * channel->block_size;
875
876	ret = ioctl(data->dev, BLKDISCARD, &range);
877	if (ret < 0)
878		return errno;
879	return 0;
880#else
881	return EXT2_ET_UNIMPLEMENTED;
882#endif
883}
884