xref: /openbmc/linux/fs/xfs/xfs_file.c (revision 05bcf503)
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_log.h"
21 #include "xfs_sb.h"
22 #include "xfs_ag.h"
23 #include "xfs_trans.h"
24 #include "xfs_mount.h"
25 #include "xfs_bmap_btree.h"
26 #include "xfs_alloc.h"
27 #include "xfs_dinode.h"
28 #include "xfs_inode.h"
29 #include "xfs_inode_item.h"
30 #include "xfs_bmap.h"
31 #include "xfs_error.h"
32 #include "xfs_vnodeops.h"
33 #include "xfs_da_btree.h"
34 #include "xfs_ioctl.h"
35 #include "xfs_trace.h"
36 
37 #include <linux/dcache.h>
38 #include <linux/falloc.h>
39 #include <linux/pagevec.h>
40 
41 static const struct vm_operations_struct xfs_file_vm_ops;
42 
43 /*
44  * Locking primitives for read and write IO paths to ensure we consistently use
45  * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
46  */
47 static inline void
48 xfs_rw_ilock(
49 	struct xfs_inode	*ip,
50 	int			type)
51 {
52 	if (type & XFS_IOLOCK_EXCL)
53 		mutex_lock(&VFS_I(ip)->i_mutex);
54 	xfs_ilock(ip, type);
55 }
56 
57 static inline void
58 xfs_rw_iunlock(
59 	struct xfs_inode	*ip,
60 	int			type)
61 {
62 	xfs_iunlock(ip, type);
63 	if (type & XFS_IOLOCK_EXCL)
64 		mutex_unlock(&VFS_I(ip)->i_mutex);
65 }
66 
67 static inline void
68 xfs_rw_ilock_demote(
69 	struct xfs_inode	*ip,
70 	int			type)
71 {
72 	xfs_ilock_demote(ip, type);
73 	if (type & XFS_IOLOCK_EXCL)
74 		mutex_unlock(&VFS_I(ip)->i_mutex);
75 }
76 
77 /*
78  *	xfs_iozero
79  *
80  *	xfs_iozero clears the specified range of buffer supplied,
81  *	and marks all the affected blocks as valid and modified.  If
82  *	an affected block is not allocated, it will be allocated.  If
83  *	an affected block is not completely overwritten, and is not
84  *	valid before the operation, it will be read from disk before
85  *	being partially zeroed.
86  */
87 STATIC int
88 xfs_iozero(
89 	struct xfs_inode	*ip,	/* inode			*/
90 	loff_t			pos,	/* offset in file		*/
91 	size_t			count)	/* size of data to zero		*/
92 {
93 	struct page		*page;
94 	struct address_space	*mapping;
95 	int			status;
96 
97 	mapping = VFS_I(ip)->i_mapping;
98 	do {
99 		unsigned offset, bytes;
100 		void *fsdata;
101 
102 		offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
103 		bytes = PAGE_CACHE_SIZE - offset;
104 		if (bytes > count)
105 			bytes = count;
106 
107 		status = pagecache_write_begin(NULL, mapping, pos, bytes,
108 					AOP_FLAG_UNINTERRUPTIBLE,
109 					&page, &fsdata);
110 		if (status)
111 			break;
112 
113 		zero_user(page, offset, bytes);
114 
115 		status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
116 					page, fsdata);
117 		WARN_ON(status <= 0); /* can't return less than zero! */
118 		pos += bytes;
119 		count -= bytes;
120 		status = 0;
121 	} while (count);
122 
123 	return (-status);
124 }
125 
126 /*
127  * Fsync operations on directories are much simpler than on regular files,
128  * as there is no file data to flush, and thus also no need for explicit
129  * cache flush operations, and there are no non-transaction metadata updates
130  * on directories either.
131  */
132 STATIC int
133 xfs_dir_fsync(
134 	struct file		*file,
135 	loff_t			start,
136 	loff_t			end,
137 	int			datasync)
138 {
139 	struct xfs_inode	*ip = XFS_I(file->f_mapping->host);
140 	struct xfs_mount	*mp = ip->i_mount;
141 	xfs_lsn_t		lsn = 0;
142 
143 	trace_xfs_dir_fsync(ip);
144 
145 	xfs_ilock(ip, XFS_ILOCK_SHARED);
146 	if (xfs_ipincount(ip))
147 		lsn = ip->i_itemp->ili_last_lsn;
148 	xfs_iunlock(ip, XFS_ILOCK_SHARED);
149 
150 	if (!lsn)
151 		return 0;
152 	return _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, NULL);
153 }
154 
155 STATIC int
156 xfs_file_fsync(
157 	struct file		*file,
158 	loff_t			start,
159 	loff_t			end,
160 	int			datasync)
161 {
162 	struct inode		*inode = file->f_mapping->host;
163 	struct xfs_inode	*ip = XFS_I(inode);
164 	struct xfs_mount	*mp = ip->i_mount;
165 	int			error = 0;
166 	int			log_flushed = 0;
167 	xfs_lsn_t		lsn = 0;
168 
169 	trace_xfs_file_fsync(ip);
170 
171 	error = filemap_write_and_wait_range(inode->i_mapping, start, end);
172 	if (error)
173 		return error;
174 
175 	if (XFS_FORCED_SHUTDOWN(mp))
176 		return -XFS_ERROR(EIO);
177 
178 	xfs_iflags_clear(ip, XFS_ITRUNCATED);
179 
180 	if (mp->m_flags & XFS_MOUNT_BARRIER) {
181 		/*
182 		 * If we have an RT and/or log subvolume we need to make sure
183 		 * to flush the write cache the device used for file data
184 		 * first.  This is to ensure newly written file data make
185 		 * it to disk before logging the new inode size in case of
186 		 * an extending write.
187 		 */
188 		if (XFS_IS_REALTIME_INODE(ip))
189 			xfs_blkdev_issue_flush(mp->m_rtdev_targp);
190 		else if (mp->m_logdev_targp != mp->m_ddev_targp)
191 			xfs_blkdev_issue_flush(mp->m_ddev_targp);
192 	}
193 
194 	/*
195 	 * All metadata updates are logged, which means that we just have
196 	 * to flush the log up to the latest LSN that touched the inode.
197 	 */
198 	xfs_ilock(ip, XFS_ILOCK_SHARED);
199 	if (xfs_ipincount(ip)) {
200 		if (!datasync ||
201 		    (ip->i_itemp->ili_fields & ~XFS_ILOG_TIMESTAMP))
202 			lsn = ip->i_itemp->ili_last_lsn;
203 	}
204 	xfs_iunlock(ip, XFS_ILOCK_SHARED);
205 
206 	if (lsn)
207 		error = _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
208 
209 	/*
210 	 * If we only have a single device, and the log force about was
211 	 * a no-op we might have to flush the data device cache here.
212 	 * This can only happen for fdatasync/O_DSYNC if we were overwriting
213 	 * an already allocated file and thus do not have any metadata to
214 	 * commit.
215 	 */
216 	if ((mp->m_flags & XFS_MOUNT_BARRIER) &&
217 	    mp->m_logdev_targp == mp->m_ddev_targp &&
218 	    !XFS_IS_REALTIME_INODE(ip) &&
219 	    !log_flushed)
220 		xfs_blkdev_issue_flush(mp->m_ddev_targp);
221 
222 	return -error;
223 }
224 
225 STATIC ssize_t
226 xfs_file_aio_read(
227 	struct kiocb		*iocb,
228 	const struct iovec	*iovp,
229 	unsigned long		nr_segs,
230 	loff_t			pos)
231 {
232 	struct file		*file = iocb->ki_filp;
233 	struct inode		*inode = file->f_mapping->host;
234 	struct xfs_inode	*ip = XFS_I(inode);
235 	struct xfs_mount	*mp = ip->i_mount;
236 	size_t			size = 0;
237 	ssize_t			ret = 0;
238 	int			ioflags = 0;
239 	xfs_fsize_t		n;
240 
241 	XFS_STATS_INC(xs_read_calls);
242 
243 	BUG_ON(iocb->ki_pos != pos);
244 
245 	if (unlikely(file->f_flags & O_DIRECT))
246 		ioflags |= IO_ISDIRECT;
247 	if (file->f_mode & FMODE_NOCMTIME)
248 		ioflags |= IO_INVIS;
249 
250 	ret = generic_segment_checks(iovp, &nr_segs, &size, VERIFY_WRITE);
251 	if (ret < 0)
252 		return ret;
253 
254 	if (unlikely(ioflags & IO_ISDIRECT)) {
255 		xfs_buftarg_t	*target =
256 			XFS_IS_REALTIME_INODE(ip) ?
257 				mp->m_rtdev_targp : mp->m_ddev_targp;
258 		if ((iocb->ki_pos & target->bt_smask) ||
259 		    (size & target->bt_smask)) {
260 			if (iocb->ki_pos == i_size_read(inode))
261 				return 0;
262 			return -XFS_ERROR(EINVAL);
263 		}
264 	}
265 
266 	n = mp->m_super->s_maxbytes - iocb->ki_pos;
267 	if (n <= 0 || size == 0)
268 		return 0;
269 
270 	if (n < size)
271 		size = n;
272 
273 	if (XFS_FORCED_SHUTDOWN(mp))
274 		return -EIO;
275 
276 	/*
277 	 * Locking is a bit tricky here. If we take an exclusive lock
278 	 * for direct IO, we effectively serialise all new concurrent
279 	 * read IO to this file and block it behind IO that is currently in
280 	 * progress because IO in progress holds the IO lock shared. We only
281 	 * need to hold the lock exclusive to blow away the page cache, so
282 	 * only take lock exclusively if the page cache needs invalidation.
283 	 * This allows the normal direct IO case of no page cache pages to
284 	 * proceeed concurrently without serialisation.
285 	 */
286 	xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
287 	if ((ioflags & IO_ISDIRECT) && inode->i_mapping->nrpages) {
288 		xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
289 		xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
290 
291 		if (inode->i_mapping->nrpages) {
292 			ret = -xfs_flushinval_pages(ip,
293 					(iocb->ki_pos & PAGE_CACHE_MASK),
294 					-1, FI_REMAPF_LOCKED);
295 			if (ret) {
296 				xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
297 				return ret;
298 			}
299 		}
300 		xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
301 	}
302 
303 	trace_xfs_file_read(ip, size, iocb->ki_pos, ioflags);
304 
305 	ret = generic_file_aio_read(iocb, iovp, nr_segs, iocb->ki_pos);
306 	if (ret > 0)
307 		XFS_STATS_ADD(xs_read_bytes, ret);
308 
309 	xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
310 	return ret;
311 }
312 
313 STATIC ssize_t
314 xfs_file_splice_read(
315 	struct file		*infilp,
316 	loff_t			*ppos,
317 	struct pipe_inode_info	*pipe,
318 	size_t			count,
319 	unsigned int		flags)
320 {
321 	struct xfs_inode	*ip = XFS_I(infilp->f_mapping->host);
322 	int			ioflags = 0;
323 	ssize_t			ret;
324 
325 	XFS_STATS_INC(xs_read_calls);
326 
327 	if (infilp->f_mode & FMODE_NOCMTIME)
328 		ioflags |= IO_INVIS;
329 
330 	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
331 		return -EIO;
332 
333 	xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
334 
335 	trace_xfs_file_splice_read(ip, count, *ppos, ioflags);
336 
337 	ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
338 	if (ret > 0)
339 		XFS_STATS_ADD(xs_read_bytes, ret);
340 
341 	xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
342 	return ret;
343 }
344 
345 /*
346  * xfs_file_splice_write() does not use xfs_rw_ilock() because
347  * generic_file_splice_write() takes the i_mutex itself. This, in theory,
348  * couuld cause lock inversions between the aio_write path and the splice path
349  * if someone is doing concurrent splice(2) based writes and write(2) based
350  * writes to the same inode. The only real way to fix this is to re-implement
351  * the generic code here with correct locking orders.
352  */
353 STATIC ssize_t
354 xfs_file_splice_write(
355 	struct pipe_inode_info	*pipe,
356 	struct file		*outfilp,
357 	loff_t			*ppos,
358 	size_t			count,
359 	unsigned int		flags)
360 {
361 	struct inode		*inode = outfilp->f_mapping->host;
362 	struct xfs_inode	*ip = XFS_I(inode);
363 	int			ioflags = 0;
364 	ssize_t			ret;
365 
366 	XFS_STATS_INC(xs_write_calls);
367 
368 	if (outfilp->f_mode & FMODE_NOCMTIME)
369 		ioflags |= IO_INVIS;
370 
371 	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
372 		return -EIO;
373 
374 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
375 
376 	trace_xfs_file_splice_write(ip, count, *ppos, ioflags);
377 
378 	ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
379 	if (ret > 0)
380 		XFS_STATS_ADD(xs_write_bytes, ret);
381 
382 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
383 	return ret;
384 }
385 
386 /*
387  * This routine is called to handle zeroing any space in the last block of the
388  * file that is beyond the EOF.  We do this since the size is being increased
389  * without writing anything to that block and we don't want to read the
390  * garbage on the disk.
391  */
392 STATIC int				/* error (positive) */
393 xfs_zero_last_block(
394 	struct xfs_inode	*ip,
395 	xfs_fsize_t		offset,
396 	xfs_fsize_t		isize)
397 {
398 	struct xfs_mount	*mp = ip->i_mount;
399 	xfs_fileoff_t		last_fsb = XFS_B_TO_FSBT(mp, isize);
400 	int			zero_offset = XFS_B_FSB_OFFSET(mp, isize);
401 	int			zero_len;
402 	int			nimaps = 1;
403 	int			error = 0;
404 	struct xfs_bmbt_irec	imap;
405 
406 	xfs_ilock(ip, XFS_ILOCK_EXCL);
407 	error = xfs_bmapi_read(ip, last_fsb, 1, &imap, &nimaps, 0);
408 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
409 	if (error)
410 		return error;
411 
412 	ASSERT(nimaps > 0);
413 
414 	/*
415 	 * If the block underlying isize is just a hole, then there
416 	 * is nothing to zero.
417 	 */
418 	if (imap.br_startblock == HOLESTARTBLOCK)
419 		return 0;
420 
421 	zero_len = mp->m_sb.sb_blocksize - zero_offset;
422 	if (isize + zero_len > offset)
423 		zero_len = offset - isize;
424 	return xfs_iozero(ip, isize, zero_len);
425 }
426 
427 /*
428  * Zero any on disk space between the current EOF and the new, larger EOF.
429  *
430  * This handles the normal case of zeroing the remainder of the last block in
431  * the file and the unusual case of zeroing blocks out beyond the size of the
432  * file.  This second case only happens with fixed size extents and when the
433  * system crashes before the inode size was updated but after blocks were
434  * allocated.
435  *
436  * Expects the iolock to be held exclusive, and will take the ilock internally.
437  */
438 int					/* error (positive) */
439 xfs_zero_eof(
440 	struct xfs_inode	*ip,
441 	xfs_off_t		offset,		/* starting I/O offset */
442 	xfs_fsize_t		isize)		/* current inode size */
443 {
444 	struct xfs_mount	*mp = ip->i_mount;
445 	xfs_fileoff_t		start_zero_fsb;
446 	xfs_fileoff_t		end_zero_fsb;
447 	xfs_fileoff_t		zero_count_fsb;
448 	xfs_fileoff_t		last_fsb;
449 	xfs_fileoff_t		zero_off;
450 	xfs_fsize_t		zero_len;
451 	int			nimaps;
452 	int			error = 0;
453 	struct xfs_bmbt_irec	imap;
454 
455 	ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
456 	ASSERT(offset > isize);
457 
458 	/*
459 	 * First handle zeroing the block on which isize resides.
460 	 *
461 	 * We only zero a part of that block so it is handled specially.
462 	 */
463 	if (XFS_B_FSB_OFFSET(mp, isize) != 0) {
464 		error = xfs_zero_last_block(ip, offset, isize);
465 		if (error)
466 			return error;
467 	}
468 
469 	/*
470 	 * Calculate the range between the new size and the old where blocks
471 	 * needing to be zeroed may exist.
472 	 *
473 	 * To get the block where the last byte in the file currently resides,
474 	 * we need to subtract one from the size and truncate back to a block
475 	 * boundary.  We subtract 1 in case the size is exactly on a block
476 	 * boundary.
477 	 */
478 	last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
479 	start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
480 	end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
481 	ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
482 	if (last_fsb == end_zero_fsb) {
483 		/*
484 		 * The size was only incremented on its last block.
485 		 * We took care of that above, so just return.
486 		 */
487 		return 0;
488 	}
489 
490 	ASSERT(start_zero_fsb <= end_zero_fsb);
491 	while (start_zero_fsb <= end_zero_fsb) {
492 		nimaps = 1;
493 		zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
494 
495 		xfs_ilock(ip, XFS_ILOCK_EXCL);
496 		error = xfs_bmapi_read(ip, start_zero_fsb, zero_count_fsb,
497 					  &imap, &nimaps, 0);
498 		xfs_iunlock(ip, XFS_ILOCK_EXCL);
499 		if (error)
500 			return error;
501 
502 		ASSERT(nimaps > 0);
503 
504 		if (imap.br_state == XFS_EXT_UNWRITTEN ||
505 		    imap.br_startblock == HOLESTARTBLOCK) {
506 			start_zero_fsb = imap.br_startoff + imap.br_blockcount;
507 			ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
508 			continue;
509 		}
510 
511 		/*
512 		 * There are blocks we need to zero.
513 		 */
514 		zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
515 		zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
516 
517 		if ((zero_off + zero_len) > offset)
518 			zero_len = offset - zero_off;
519 
520 		error = xfs_iozero(ip, zero_off, zero_len);
521 		if (error)
522 			return error;
523 
524 		start_zero_fsb = imap.br_startoff + imap.br_blockcount;
525 		ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
526 	}
527 
528 	return 0;
529 }
530 
531 /*
532  * Common pre-write limit and setup checks.
533  *
534  * Called with the iolocked held either shared and exclusive according to
535  * @iolock, and returns with it held.  Might upgrade the iolock to exclusive
536  * if called for a direct write beyond i_size.
537  */
538 STATIC ssize_t
539 xfs_file_aio_write_checks(
540 	struct file		*file,
541 	loff_t			*pos,
542 	size_t			*count,
543 	int			*iolock)
544 {
545 	struct inode		*inode = file->f_mapping->host;
546 	struct xfs_inode	*ip = XFS_I(inode);
547 	int			error = 0;
548 
549 restart:
550 	error = generic_write_checks(file, pos, count, S_ISBLK(inode->i_mode));
551 	if (error)
552 		return error;
553 
554 	/*
555 	 * If the offset is beyond the size of the file, we need to zero any
556 	 * blocks that fall between the existing EOF and the start of this
557 	 * write.  If zeroing is needed and we are currently holding the
558 	 * iolock shared, we need to update it to exclusive which implies
559 	 * having to redo all checks before.
560 	 */
561 	if (*pos > i_size_read(inode)) {
562 		if (*iolock == XFS_IOLOCK_SHARED) {
563 			xfs_rw_iunlock(ip, *iolock);
564 			*iolock = XFS_IOLOCK_EXCL;
565 			xfs_rw_ilock(ip, *iolock);
566 			goto restart;
567 		}
568 		error = -xfs_zero_eof(ip, *pos, i_size_read(inode));
569 		if (error)
570 			return error;
571 	}
572 
573 	/*
574 	 * Updating the timestamps will grab the ilock again from
575 	 * xfs_fs_dirty_inode, so we have to call it after dropping the
576 	 * lock above.  Eventually we should look into a way to avoid
577 	 * the pointless lock roundtrip.
578 	 */
579 	if (likely(!(file->f_mode & FMODE_NOCMTIME))) {
580 		error = file_update_time(file);
581 		if (error)
582 			return error;
583 	}
584 
585 	/*
586 	 * If we're writing the file then make sure to clear the setuid and
587 	 * setgid bits if the process is not being run by root.  This keeps
588 	 * people from modifying setuid and setgid binaries.
589 	 */
590 	return file_remove_suid(file);
591 }
592 
593 /*
594  * xfs_file_dio_aio_write - handle direct IO writes
595  *
596  * Lock the inode appropriately to prepare for and issue a direct IO write.
597  * By separating it from the buffered write path we remove all the tricky to
598  * follow locking changes and looping.
599  *
600  * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
601  * until we're sure the bytes at the new EOF have been zeroed and/or the cached
602  * pages are flushed out.
603  *
604  * In most cases the direct IO writes will be done holding IOLOCK_SHARED
605  * allowing them to be done in parallel with reads and other direct IO writes.
606  * However, if the IO is not aligned to filesystem blocks, the direct IO layer
607  * needs to do sub-block zeroing and that requires serialisation against other
608  * direct IOs to the same block. In this case we need to serialise the
609  * submission of the unaligned IOs so that we don't get racing block zeroing in
610  * the dio layer.  To avoid the problem with aio, we also need to wait for
611  * outstanding IOs to complete so that unwritten extent conversion is completed
612  * before we try to map the overlapping block. This is currently implemented by
613  * hitting it with a big hammer (i.e. inode_dio_wait()).
614  *
615  * Returns with locks held indicated by @iolock and errors indicated by
616  * negative return values.
617  */
618 STATIC ssize_t
619 xfs_file_dio_aio_write(
620 	struct kiocb		*iocb,
621 	const struct iovec	*iovp,
622 	unsigned long		nr_segs,
623 	loff_t			pos,
624 	size_t			ocount)
625 {
626 	struct file		*file = iocb->ki_filp;
627 	struct address_space	*mapping = file->f_mapping;
628 	struct inode		*inode = mapping->host;
629 	struct xfs_inode	*ip = XFS_I(inode);
630 	struct xfs_mount	*mp = ip->i_mount;
631 	ssize_t			ret = 0;
632 	size_t			count = ocount;
633 	int			unaligned_io = 0;
634 	int			iolock;
635 	struct xfs_buftarg	*target = XFS_IS_REALTIME_INODE(ip) ?
636 					mp->m_rtdev_targp : mp->m_ddev_targp;
637 
638 	if ((pos & target->bt_smask) || (count & target->bt_smask))
639 		return -XFS_ERROR(EINVAL);
640 
641 	if ((pos & mp->m_blockmask) || ((pos + count) & mp->m_blockmask))
642 		unaligned_io = 1;
643 
644 	/*
645 	 * We don't need to take an exclusive lock unless there page cache needs
646 	 * to be invalidated or unaligned IO is being executed. We don't need to
647 	 * consider the EOF extension case here because
648 	 * xfs_file_aio_write_checks() will relock the inode as necessary for
649 	 * EOF zeroing cases and fill out the new inode size as appropriate.
650 	 */
651 	if (unaligned_io || mapping->nrpages)
652 		iolock = XFS_IOLOCK_EXCL;
653 	else
654 		iolock = XFS_IOLOCK_SHARED;
655 	xfs_rw_ilock(ip, iolock);
656 
657 	/*
658 	 * Recheck if there are cached pages that need invalidate after we got
659 	 * the iolock to protect against other threads adding new pages while
660 	 * we were waiting for the iolock.
661 	 */
662 	if (mapping->nrpages && iolock == XFS_IOLOCK_SHARED) {
663 		xfs_rw_iunlock(ip, iolock);
664 		iolock = XFS_IOLOCK_EXCL;
665 		xfs_rw_ilock(ip, iolock);
666 	}
667 
668 	ret = xfs_file_aio_write_checks(file, &pos, &count, &iolock);
669 	if (ret)
670 		goto out;
671 
672 	if (mapping->nrpages) {
673 		ret = -xfs_flushinval_pages(ip, (pos & PAGE_CACHE_MASK), -1,
674 							FI_REMAPF_LOCKED);
675 		if (ret)
676 			goto out;
677 	}
678 
679 	/*
680 	 * If we are doing unaligned IO, wait for all other IO to drain,
681 	 * otherwise demote the lock if we had to flush cached pages
682 	 */
683 	if (unaligned_io)
684 		inode_dio_wait(inode);
685 	else if (iolock == XFS_IOLOCK_EXCL) {
686 		xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
687 		iolock = XFS_IOLOCK_SHARED;
688 	}
689 
690 	trace_xfs_file_direct_write(ip, count, iocb->ki_pos, 0);
691 	ret = generic_file_direct_write(iocb, iovp,
692 			&nr_segs, pos, &iocb->ki_pos, count, ocount);
693 
694 out:
695 	xfs_rw_iunlock(ip, iolock);
696 
697 	/* No fallback to buffered IO on errors for XFS. */
698 	ASSERT(ret < 0 || ret == count);
699 	return ret;
700 }
701 
702 STATIC ssize_t
703 xfs_file_buffered_aio_write(
704 	struct kiocb		*iocb,
705 	const struct iovec	*iovp,
706 	unsigned long		nr_segs,
707 	loff_t			pos,
708 	size_t			ocount)
709 {
710 	struct file		*file = iocb->ki_filp;
711 	struct address_space	*mapping = file->f_mapping;
712 	struct inode		*inode = mapping->host;
713 	struct xfs_inode	*ip = XFS_I(inode);
714 	ssize_t			ret;
715 	int			enospc = 0;
716 	int			iolock = XFS_IOLOCK_EXCL;
717 	size_t			count = ocount;
718 
719 	xfs_rw_ilock(ip, iolock);
720 
721 	ret = xfs_file_aio_write_checks(file, &pos, &count, &iolock);
722 	if (ret)
723 		goto out;
724 
725 	/* We can write back this queue in page reclaim */
726 	current->backing_dev_info = mapping->backing_dev_info;
727 
728 write_retry:
729 	trace_xfs_file_buffered_write(ip, count, iocb->ki_pos, 0);
730 	ret = generic_file_buffered_write(iocb, iovp, nr_segs,
731 			pos, &iocb->ki_pos, count, ret);
732 	/*
733 	 * if we just got an ENOSPC, flush the inode now we aren't holding any
734 	 * page locks and retry *once*
735 	 */
736 	if (ret == -ENOSPC && !enospc) {
737 		enospc = 1;
738 		ret = -xfs_flush_pages(ip, 0, -1, 0, FI_NONE);
739 		if (!ret)
740 			goto write_retry;
741 	}
742 
743 	current->backing_dev_info = NULL;
744 out:
745 	xfs_rw_iunlock(ip, iolock);
746 	return ret;
747 }
748 
749 STATIC ssize_t
750 xfs_file_aio_write(
751 	struct kiocb		*iocb,
752 	const struct iovec	*iovp,
753 	unsigned long		nr_segs,
754 	loff_t			pos)
755 {
756 	struct file		*file = iocb->ki_filp;
757 	struct address_space	*mapping = file->f_mapping;
758 	struct inode		*inode = mapping->host;
759 	struct xfs_inode	*ip = XFS_I(inode);
760 	ssize_t			ret;
761 	size_t			ocount = 0;
762 
763 	XFS_STATS_INC(xs_write_calls);
764 
765 	BUG_ON(iocb->ki_pos != pos);
766 
767 	ret = generic_segment_checks(iovp, &nr_segs, &ocount, VERIFY_READ);
768 	if (ret)
769 		return ret;
770 
771 	if (ocount == 0)
772 		return 0;
773 
774 	sb_start_write(inode->i_sb);
775 
776 	if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
777 		ret = -EIO;
778 		goto out;
779 	}
780 
781 	if (unlikely(file->f_flags & O_DIRECT))
782 		ret = xfs_file_dio_aio_write(iocb, iovp, nr_segs, pos, ocount);
783 	else
784 		ret = xfs_file_buffered_aio_write(iocb, iovp, nr_segs, pos,
785 						  ocount);
786 
787 	if (ret > 0) {
788 		ssize_t err;
789 
790 		XFS_STATS_ADD(xs_write_bytes, ret);
791 
792 		/* Handle various SYNC-type writes */
793 		err = generic_write_sync(file, pos, ret);
794 		if (err < 0)
795 			ret = err;
796 	}
797 
798 out:
799 	sb_end_write(inode->i_sb);
800 	return ret;
801 }
802 
803 STATIC long
804 xfs_file_fallocate(
805 	struct file	*file,
806 	int		mode,
807 	loff_t		offset,
808 	loff_t		len)
809 {
810 	struct inode	*inode = file->f_path.dentry->d_inode;
811 	long		error;
812 	loff_t		new_size = 0;
813 	xfs_flock64_t	bf;
814 	xfs_inode_t	*ip = XFS_I(inode);
815 	int		cmd = XFS_IOC_RESVSP;
816 	int		attr_flags = XFS_ATTR_NOLOCK;
817 
818 	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
819 		return -EOPNOTSUPP;
820 
821 	bf.l_whence = 0;
822 	bf.l_start = offset;
823 	bf.l_len = len;
824 
825 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
826 
827 	if (mode & FALLOC_FL_PUNCH_HOLE)
828 		cmd = XFS_IOC_UNRESVSP;
829 
830 	/* check the new inode size is valid before allocating */
831 	if (!(mode & FALLOC_FL_KEEP_SIZE) &&
832 	    offset + len > i_size_read(inode)) {
833 		new_size = offset + len;
834 		error = inode_newsize_ok(inode, new_size);
835 		if (error)
836 			goto out_unlock;
837 	}
838 
839 	if (file->f_flags & O_DSYNC)
840 		attr_flags |= XFS_ATTR_SYNC;
841 
842 	error = -xfs_change_file_space(ip, cmd, &bf, 0, attr_flags);
843 	if (error)
844 		goto out_unlock;
845 
846 	/* Change file size if needed */
847 	if (new_size) {
848 		struct iattr iattr;
849 
850 		iattr.ia_valid = ATTR_SIZE;
851 		iattr.ia_size = new_size;
852 		error = -xfs_setattr_size(ip, &iattr, XFS_ATTR_NOLOCK);
853 	}
854 
855 out_unlock:
856 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
857 	return error;
858 }
859 
860 
861 STATIC int
862 xfs_file_open(
863 	struct inode	*inode,
864 	struct file	*file)
865 {
866 	if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
867 		return -EFBIG;
868 	if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
869 		return -EIO;
870 	return 0;
871 }
872 
873 STATIC int
874 xfs_dir_open(
875 	struct inode	*inode,
876 	struct file	*file)
877 {
878 	struct xfs_inode *ip = XFS_I(inode);
879 	int		mode;
880 	int		error;
881 
882 	error = xfs_file_open(inode, file);
883 	if (error)
884 		return error;
885 
886 	/*
887 	 * If there are any blocks, read-ahead block 0 as we're almost
888 	 * certain to have the next operation be a read there.
889 	 */
890 	mode = xfs_ilock_map_shared(ip);
891 	if (ip->i_d.di_nextents > 0)
892 		xfs_da_reada_buf(NULL, ip, 0, XFS_DATA_FORK);
893 	xfs_iunlock(ip, mode);
894 	return 0;
895 }
896 
897 STATIC int
898 xfs_file_release(
899 	struct inode	*inode,
900 	struct file	*filp)
901 {
902 	return -xfs_release(XFS_I(inode));
903 }
904 
905 STATIC int
906 xfs_file_readdir(
907 	struct file	*filp,
908 	void		*dirent,
909 	filldir_t	filldir)
910 {
911 	struct inode	*inode = filp->f_path.dentry->d_inode;
912 	xfs_inode_t	*ip = XFS_I(inode);
913 	int		error;
914 	size_t		bufsize;
915 
916 	/*
917 	 * The Linux API doesn't pass down the total size of the buffer
918 	 * we read into down to the filesystem.  With the filldir concept
919 	 * it's not needed for correct information, but the XFS dir2 leaf
920 	 * code wants an estimate of the buffer size to calculate it's
921 	 * readahead window and size the buffers used for mapping to
922 	 * physical blocks.
923 	 *
924 	 * Try to give it an estimate that's good enough, maybe at some
925 	 * point we can change the ->readdir prototype to include the
926 	 * buffer size.  For now we use the current glibc buffer size.
927 	 */
928 	bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
929 
930 	error = xfs_readdir(ip, dirent, bufsize,
931 				(xfs_off_t *)&filp->f_pos, filldir);
932 	if (error)
933 		return -error;
934 	return 0;
935 }
936 
937 STATIC int
938 xfs_file_mmap(
939 	struct file	*filp,
940 	struct vm_area_struct *vma)
941 {
942 	vma->vm_ops = &xfs_file_vm_ops;
943 
944 	file_accessed(filp);
945 	return 0;
946 }
947 
948 /*
949  * mmap()d file has taken write protection fault and is being made
950  * writable. We can set the page state up correctly for a writable
951  * page, which means we can do correct delalloc accounting (ENOSPC
952  * checking!) and unwritten extent mapping.
953  */
954 STATIC int
955 xfs_vm_page_mkwrite(
956 	struct vm_area_struct	*vma,
957 	struct vm_fault		*vmf)
958 {
959 	return block_page_mkwrite(vma, vmf, xfs_get_blocks);
960 }
961 
962 /*
963  * This type is designed to indicate the type of offset we would like
964  * to search from page cache for either xfs_seek_data() or xfs_seek_hole().
965  */
966 enum {
967 	HOLE_OFF = 0,
968 	DATA_OFF,
969 };
970 
971 /*
972  * Lookup the desired type of offset from the given page.
973  *
974  * On success, return true and the offset argument will point to the
975  * start of the region that was found.  Otherwise this function will
976  * return false and keep the offset argument unchanged.
977  */
978 STATIC bool
979 xfs_lookup_buffer_offset(
980 	struct page		*page,
981 	loff_t			*offset,
982 	unsigned int		type)
983 {
984 	loff_t			lastoff = page_offset(page);
985 	bool			found = false;
986 	struct buffer_head	*bh, *head;
987 
988 	bh = head = page_buffers(page);
989 	do {
990 		/*
991 		 * Unwritten extents that have data in the page
992 		 * cache covering them can be identified by the
993 		 * BH_Unwritten state flag.  Pages with multiple
994 		 * buffers might have a mix of holes, data and
995 		 * unwritten extents - any buffer with valid
996 		 * data in it should have BH_Uptodate flag set
997 		 * on it.
998 		 */
999 		if (buffer_unwritten(bh) ||
1000 		    buffer_uptodate(bh)) {
1001 			if (type == DATA_OFF)
1002 				found = true;
1003 		} else {
1004 			if (type == HOLE_OFF)
1005 				found = true;
1006 		}
1007 
1008 		if (found) {
1009 			*offset = lastoff;
1010 			break;
1011 		}
1012 		lastoff += bh->b_size;
1013 	} while ((bh = bh->b_this_page) != head);
1014 
1015 	return found;
1016 }
1017 
1018 /*
1019  * This routine is called to find out and return a data or hole offset
1020  * from the page cache for unwritten extents according to the desired
1021  * type for xfs_seek_data() or xfs_seek_hole().
1022  *
1023  * The argument offset is used to tell where we start to search from the
1024  * page cache.  Map is used to figure out the end points of the range to
1025  * lookup pages.
1026  *
1027  * Return true if the desired type of offset was found, and the argument
1028  * offset is filled with that address.  Otherwise, return false and keep
1029  * offset unchanged.
1030  */
1031 STATIC bool
1032 xfs_find_get_desired_pgoff(
1033 	struct inode		*inode,
1034 	struct xfs_bmbt_irec	*map,
1035 	unsigned int		type,
1036 	loff_t			*offset)
1037 {
1038 	struct xfs_inode	*ip = XFS_I(inode);
1039 	struct xfs_mount	*mp = ip->i_mount;
1040 	struct pagevec		pvec;
1041 	pgoff_t			index;
1042 	pgoff_t			end;
1043 	loff_t			endoff;
1044 	loff_t			startoff = *offset;
1045 	loff_t			lastoff = startoff;
1046 	bool			found = false;
1047 
1048 	pagevec_init(&pvec, 0);
1049 
1050 	index = startoff >> PAGE_CACHE_SHIFT;
1051 	endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1052 	end = endoff >> PAGE_CACHE_SHIFT;
1053 	do {
1054 		int		want;
1055 		unsigned	nr_pages;
1056 		unsigned int	i;
1057 
1058 		want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1059 		nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1060 					  want);
1061 		/*
1062 		 * No page mapped into given range.  If we are searching holes
1063 		 * and if this is the first time we got into the loop, it means
1064 		 * that the given offset is landed in a hole, return it.
1065 		 *
1066 		 * If we have already stepped through some block buffers to find
1067 		 * holes but they all contains data.  In this case, the last
1068 		 * offset is already updated and pointed to the end of the last
1069 		 * mapped page, if it does not reach the endpoint to search,
1070 		 * that means there should be a hole between them.
1071 		 */
1072 		if (nr_pages == 0) {
1073 			/* Data search found nothing */
1074 			if (type == DATA_OFF)
1075 				break;
1076 
1077 			ASSERT(type == HOLE_OFF);
1078 			if (lastoff == startoff || lastoff < endoff) {
1079 				found = true;
1080 				*offset = lastoff;
1081 			}
1082 			break;
1083 		}
1084 
1085 		/*
1086 		 * At lease we found one page.  If this is the first time we
1087 		 * step into the loop, and if the first page index offset is
1088 		 * greater than the given search offset, a hole was found.
1089 		 */
1090 		if (type == HOLE_OFF && lastoff == startoff &&
1091 		    lastoff < page_offset(pvec.pages[0])) {
1092 			found = true;
1093 			break;
1094 		}
1095 
1096 		for (i = 0; i < nr_pages; i++) {
1097 			struct page	*page = pvec.pages[i];
1098 			loff_t		b_offset;
1099 
1100 			/*
1101 			 * At this point, the page may be truncated or
1102 			 * invalidated (changing page->mapping to NULL),
1103 			 * or even swizzled back from swapper_space to tmpfs
1104 			 * file mapping. However, page->index will not change
1105 			 * because we have a reference on the page.
1106 			 *
1107 			 * Searching done if the page index is out of range.
1108 			 * If the current offset is not reaches the end of
1109 			 * the specified search range, there should be a hole
1110 			 * between them.
1111 			 */
1112 			if (page->index > end) {
1113 				if (type == HOLE_OFF && lastoff < endoff) {
1114 					*offset = lastoff;
1115 					found = true;
1116 				}
1117 				goto out;
1118 			}
1119 
1120 			lock_page(page);
1121 			/*
1122 			 * Page truncated or invalidated(page->mapping == NULL).
1123 			 * We can freely skip it and proceed to check the next
1124 			 * page.
1125 			 */
1126 			if (unlikely(page->mapping != inode->i_mapping)) {
1127 				unlock_page(page);
1128 				continue;
1129 			}
1130 
1131 			if (!page_has_buffers(page)) {
1132 				unlock_page(page);
1133 				continue;
1134 			}
1135 
1136 			found = xfs_lookup_buffer_offset(page, &b_offset, type);
1137 			if (found) {
1138 				/*
1139 				 * The found offset may be less than the start
1140 				 * point to search if this is the first time to
1141 				 * come here.
1142 				 */
1143 				*offset = max_t(loff_t, startoff, b_offset);
1144 				unlock_page(page);
1145 				goto out;
1146 			}
1147 
1148 			/*
1149 			 * We either searching data but nothing was found, or
1150 			 * searching hole but found a data buffer.  In either
1151 			 * case, probably the next page contains the desired
1152 			 * things, update the last offset to it so.
1153 			 */
1154 			lastoff = page_offset(page) + PAGE_SIZE;
1155 			unlock_page(page);
1156 		}
1157 
1158 		/*
1159 		 * The number of returned pages less than our desired, search
1160 		 * done.  In this case, nothing was found for searching data,
1161 		 * but we found a hole behind the last offset.
1162 		 */
1163 		if (nr_pages < want) {
1164 			if (type == HOLE_OFF) {
1165 				*offset = lastoff;
1166 				found = true;
1167 			}
1168 			break;
1169 		}
1170 
1171 		index = pvec.pages[i - 1]->index + 1;
1172 		pagevec_release(&pvec);
1173 	} while (index <= end);
1174 
1175 out:
1176 	pagevec_release(&pvec);
1177 	return found;
1178 }
1179 
1180 STATIC loff_t
1181 xfs_seek_data(
1182 	struct file		*file,
1183 	loff_t			start)
1184 {
1185 	struct inode		*inode = file->f_mapping->host;
1186 	struct xfs_inode	*ip = XFS_I(inode);
1187 	struct xfs_mount	*mp = ip->i_mount;
1188 	loff_t			uninitialized_var(offset);
1189 	xfs_fsize_t		isize;
1190 	xfs_fileoff_t		fsbno;
1191 	xfs_filblks_t		end;
1192 	uint			lock;
1193 	int			error;
1194 
1195 	lock = xfs_ilock_map_shared(ip);
1196 
1197 	isize = i_size_read(inode);
1198 	if (start >= isize) {
1199 		error = ENXIO;
1200 		goto out_unlock;
1201 	}
1202 
1203 	/*
1204 	 * Try to read extents from the first block indicated
1205 	 * by fsbno to the end block of the file.
1206 	 */
1207 	fsbno = XFS_B_TO_FSBT(mp, start);
1208 	end = XFS_B_TO_FSB(mp, isize);
1209 	for (;;) {
1210 		struct xfs_bmbt_irec	map[2];
1211 		int			nmap = 2;
1212 		unsigned int		i;
1213 
1214 		error = xfs_bmapi_read(ip, fsbno, end - fsbno, map, &nmap,
1215 				       XFS_BMAPI_ENTIRE);
1216 		if (error)
1217 			goto out_unlock;
1218 
1219 		/* No extents at given offset, must be beyond EOF */
1220 		if (nmap == 0) {
1221 			error = ENXIO;
1222 			goto out_unlock;
1223 		}
1224 
1225 		for (i = 0; i < nmap; i++) {
1226 			offset = max_t(loff_t, start,
1227 				       XFS_FSB_TO_B(mp, map[i].br_startoff));
1228 
1229 			/* Landed in a data extent */
1230 			if (map[i].br_startblock == DELAYSTARTBLOCK ||
1231 			    (map[i].br_state == XFS_EXT_NORM &&
1232 			     !isnullstartblock(map[i].br_startblock)))
1233 				goto out;
1234 
1235 			/*
1236 			 * Landed in an unwritten extent, try to search data
1237 			 * from page cache.
1238 			 */
1239 			if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1240 				if (xfs_find_get_desired_pgoff(inode, &map[i],
1241 							DATA_OFF, &offset))
1242 					goto out;
1243 			}
1244 		}
1245 
1246 		/*
1247 		 * map[0] is hole or its an unwritten extent but
1248 		 * without data in page cache.  Probably means that
1249 		 * we are reading after EOF if nothing in map[1].
1250 		 */
1251 		if (nmap == 1) {
1252 			error = ENXIO;
1253 			goto out_unlock;
1254 		}
1255 
1256 		ASSERT(i > 1);
1257 
1258 		/*
1259 		 * Nothing was found, proceed to the next round of search
1260 		 * if reading offset not beyond or hit EOF.
1261 		 */
1262 		fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1263 		start = XFS_FSB_TO_B(mp, fsbno);
1264 		if (start >= isize) {
1265 			error = ENXIO;
1266 			goto out_unlock;
1267 		}
1268 	}
1269 
1270 out:
1271 	if (offset != file->f_pos)
1272 		file->f_pos = offset;
1273 
1274 out_unlock:
1275 	xfs_iunlock_map_shared(ip, lock);
1276 
1277 	if (error)
1278 		return -error;
1279 	return offset;
1280 }
1281 
1282 STATIC loff_t
1283 xfs_seek_hole(
1284 	struct file		*file,
1285 	loff_t			start)
1286 {
1287 	struct inode		*inode = file->f_mapping->host;
1288 	struct xfs_inode	*ip = XFS_I(inode);
1289 	struct xfs_mount	*mp = ip->i_mount;
1290 	loff_t			uninitialized_var(offset);
1291 	xfs_fsize_t		isize;
1292 	xfs_fileoff_t		fsbno;
1293 	xfs_filblks_t		end;
1294 	uint			lock;
1295 	int			error;
1296 
1297 	if (XFS_FORCED_SHUTDOWN(mp))
1298 		return -XFS_ERROR(EIO);
1299 
1300 	lock = xfs_ilock_map_shared(ip);
1301 
1302 	isize = i_size_read(inode);
1303 	if (start >= isize) {
1304 		error = ENXIO;
1305 		goto out_unlock;
1306 	}
1307 
1308 	fsbno = XFS_B_TO_FSBT(mp, start);
1309 	end = XFS_B_TO_FSB(mp, isize);
1310 
1311 	for (;;) {
1312 		struct xfs_bmbt_irec	map[2];
1313 		int			nmap = 2;
1314 		unsigned int		i;
1315 
1316 		error = xfs_bmapi_read(ip, fsbno, end - fsbno, map, &nmap,
1317 				       XFS_BMAPI_ENTIRE);
1318 		if (error)
1319 			goto out_unlock;
1320 
1321 		/* No extents at given offset, must be beyond EOF */
1322 		if (nmap == 0) {
1323 			error = ENXIO;
1324 			goto out_unlock;
1325 		}
1326 
1327 		for (i = 0; i < nmap; i++) {
1328 			offset = max_t(loff_t, start,
1329 				       XFS_FSB_TO_B(mp, map[i].br_startoff));
1330 
1331 			/* Landed in a hole */
1332 			if (map[i].br_startblock == HOLESTARTBLOCK)
1333 				goto out;
1334 
1335 			/*
1336 			 * Landed in an unwritten extent, try to search hole
1337 			 * from page cache.
1338 			 */
1339 			if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1340 				if (xfs_find_get_desired_pgoff(inode, &map[i],
1341 							HOLE_OFF, &offset))
1342 					goto out;
1343 			}
1344 		}
1345 
1346 		/*
1347 		 * map[0] contains data or its unwritten but contains
1348 		 * data in page cache, probably means that we are
1349 		 * reading after EOF.  We should fix offset to point
1350 		 * to the end of the file(i.e., there is an implicit
1351 		 * hole at the end of any file).
1352 		 */
1353 		if (nmap == 1) {
1354 			offset = isize;
1355 			break;
1356 		}
1357 
1358 		ASSERT(i > 1);
1359 
1360 		/*
1361 		 * Both mappings contains data, proceed to the next round of
1362 		 * search if the current reading offset not beyond or hit EOF.
1363 		 */
1364 		fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1365 		start = XFS_FSB_TO_B(mp, fsbno);
1366 		if (start >= isize) {
1367 			offset = isize;
1368 			break;
1369 		}
1370 	}
1371 
1372 out:
1373 	/*
1374 	 * At this point, we must have found a hole.  However, the returned
1375 	 * offset may be bigger than the file size as it may be aligned to
1376 	 * page boundary for unwritten extents, we need to deal with this
1377 	 * situation in particular.
1378 	 */
1379 	offset = min_t(loff_t, offset, isize);
1380 	if (offset != file->f_pos)
1381 		file->f_pos = offset;
1382 
1383 out_unlock:
1384 	xfs_iunlock_map_shared(ip, lock);
1385 
1386 	if (error)
1387 		return -error;
1388 	return offset;
1389 }
1390 
1391 STATIC loff_t
1392 xfs_file_llseek(
1393 	struct file	*file,
1394 	loff_t		offset,
1395 	int		origin)
1396 {
1397 	switch (origin) {
1398 	case SEEK_END:
1399 	case SEEK_CUR:
1400 	case SEEK_SET:
1401 		return generic_file_llseek(file, offset, origin);
1402 	case SEEK_DATA:
1403 		return xfs_seek_data(file, offset);
1404 	case SEEK_HOLE:
1405 		return xfs_seek_hole(file, offset);
1406 	default:
1407 		return -EINVAL;
1408 	}
1409 }
1410 
1411 const struct file_operations xfs_file_operations = {
1412 	.llseek		= xfs_file_llseek,
1413 	.read		= do_sync_read,
1414 	.write		= do_sync_write,
1415 	.aio_read	= xfs_file_aio_read,
1416 	.aio_write	= xfs_file_aio_write,
1417 	.splice_read	= xfs_file_splice_read,
1418 	.splice_write	= xfs_file_splice_write,
1419 	.unlocked_ioctl	= xfs_file_ioctl,
1420 #ifdef CONFIG_COMPAT
1421 	.compat_ioctl	= xfs_file_compat_ioctl,
1422 #endif
1423 	.mmap		= xfs_file_mmap,
1424 	.open		= xfs_file_open,
1425 	.release	= xfs_file_release,
1426 	.fsync		= xfs_file_fsync,
1427 	.fallocate	= xfs_file_fallocate,
1428 };
1429 
1430 const struct file_operations xfs_dir_file_operations = {
1431 	.open		= xfs_dir_open,
1432 	.read		= generic_read_dir,
1433 	.readdir	= xfs_file_readdir,
1434 	.llseek		= generic_file_llseek,
1435 	.unlocked_ioctl	= xfs_file_ioctl,
1436 #ifdef CONFIG_COMPAT
1437 	.compat_ioctl	= xfs_file_compat_ioctl,
1438 #endif
1439 	.fsync		= xfs_dir_fsync,
1440 };
1441 
1442 static const struct vm_operations_struct xfs_file_vm_ops = {
1443 	.fault		= filemap_fault,
1444 	.page_mkwrite	= xfs_vm_page_mkwrite,
1445 	.remap_pages	= generic_file_remap_pages,
1446 };
1447