xref: /openbmc/linux/fs/xfs/xfs_file.c (revision b34e08d5)
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_shared.h"
21 #include "xfs_format.h"
22 #include "xfs_log_format.h"
23 #include "xfs_trans_resv.h"
24 #include "xfs_sb.h"
25 #include "xfs_ag.h"
26 #include "xfs_mount.h"
27 #include "xfs_da_format.h"
28 #include "xfs_da_btree.h"
29 #include "xfs_inode.h"
30 #include "xfs_trans.h"
31 #include "xfs_inode_item.h"
32 #include "xfs_bmap.h"
33 #include "xfs_bmap_util.h"
34 #include "xfs_error.h"
35 #include "xfs_dir2.h"
36 #include "xfs_dir2_priv.h"
37 #include "xfs_ioctl.h"
38 #include "xfs_trace.h"
39 #include "xfs_log.h"
40 #include "xfs_dinode.h"
41 
42 #include <linux/aio.h>
43 #include <linux/dcache.h>
44 #include <linux/falloc.h>
45 #include <linux/pagevec.h>
46 
47 static const struct vm_operations_struct xfs_file_vm_ops;
48 
49 /*
50  * Locking primitives for read and write IO paths to ensure we consistently use
51  * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
52  */
53 static inline void
54 xfs_rw_ilock(
55 	struct xfs_inode	*ip,
56 	int			type)
57 {
58 	if (type & XFS_IOLOCK_EXCL)
59 		mutex_lock(&VFS_I(ip)->i_mutex);
60 	xfs_ilock(ip, type);
61 }
62 
63 static inline void
64 xfs_rw_iunlock(
65 	struct xfs_inode	*ip,
66 	int			type)
67 {
68 	xfs_iunlock(ip, type);
69 	if (type & XFS_IOLOCK_EXCL)
70 		mutex_unlock(&VFS_I(ip)->i_mutex);
71 }
72 
73 static inline void
74 xfs_rw_ilock_demote(
75 	struct xfs_inode	*ip,
76 	int			type)
77 {
78 	xfs_ilock_demote(ip, type);
79 	if (type & XFS_IOLOCK_EXCL)
80 		mutex_unlock(&VFS_I(ip)->i_mutex);
81 }
82 
83 /*
84  *	xfs_iozero
85  *
86  *	xfs_iozero clears the specified range of buffer supplied,
87  *	and marks all the affected blocks as valid and modified.  If
88  *	an affected block is not allocated, it will be allocated.  If
89  *	an affected block is not completely overwritten, and is not
90  *	valid before the operation, it will be read from disk before
91  *	being partially zeroed.
92  */
93 int
94 xfs_iozero(
95 	struct xfs_inode	*ip,	/* inode			*/
96 	loff_t			pos,	/* offset in file		*/
97 	size_t			count)	/* size of data to zero		*/
98 {
99 	struct page		*page;
100 	struct address_space	*mapping;
101 	int			status;
102 
103 	mapping = VFS_I(ip)->i_mapping;
104 	do {
105 		unsigned offset, bytes;
106 		void *fsdata;
107 
108 		offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
109 		bytes = PAGE_CACHE_SIZE - offset;
110 		if (bytes > count)
111 			bytes = count;
112 
113 		status = pagecache_write_begin(NULL, mapping, pos, bytes,
114 					AOP_FLAG_UNINTERRUPTIBLE,
115 					&page, &fsdata);
116 		if (status)
117 			break;
118 
119 		zero_user(page, offset, bytes);
120 
121 		status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
122 					page, fsdata);
123 		WARN_ON(status <= 0); /* can't return less than zero! */
124 		pos += bytes;
125 		count -= bytes;
126 		status = 0;
127 	} while (count);
128 
129 	return (-status);
130 }
131 
132 /*
133  * Fsync operations on directories are much simpler than on regular files,
134  * as there is no file data to flush, and thus also no need for explicit
135  * cache flush operations, and there are no non-transaction metadata updates
136  * on directories either.
137  */
138 STATIC int
139 xfs_dir_fsync(
140 	struct file		*file,
141 	loff_t			start,
142 	loff_t			end,
143 	int			datasync)
144 {
145 	struct xfs_inode	*ip = XFS_I(file->f_mapping->host);
146 	struct xfs_mount	*mp = ip->i_mount;
147 	xfs_lsn_t		lsn = 0;
148 
149 	trace_xfs_dir_fsync(ip);
150 
151 	xfs_ilock(ip, XFS_ILOCK_SHARED);
152 	if (xfs_ipincount(ip))
153 		lsn = ip->i_itemp->ili_last_lsn;
154 	xfs_iunlock(ip, XFS_ILOCK_SHARED);
155 
156 	if (!lsn)
157 		return 0;
158 	return _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, NULL);
159 }
160 
161 STATIC int
162 xfs_file_fsync(
163 	struct file		*file,
164 	loff_t			start,
165 	loff_t			end,
166 	int			datasync)
167 {
168 	struct inode		*inode = file->f_mapping->host;
169 	struct xfs_inode	*ip = XFS_I(inode);
170 	struct xfs_mount	*mp = ip->i_mount;
171 	int			error = 0;
172 	int			log_flushed = 0;
173 	xfs_lsn_t		lsn = 0;
174 
175 	trace_xfs_file_fsync(ip);
176 
177 	error = filemap_write_and_wait_range(inode->i_mapping, start, end);
178 	if (error)
179 		return error;
180 
181 	if (XFS_FORCED_SHUTDOWN(mp))
182 		return -XFS_ERROR(EIO);
183 
184 	xfs_iflags_clear(ip, XFS_ITRUNCATED);
185 
186 	if (mp->m_flags & XFS_MOUNT_BARRIER) {
187 		/*
188 		 * If we have an RT and/or log subvolume we need to make sure
189 		 * to flush the write cache the device used for file data
190 		 * first.  This is to ensure newly written file data make
191 		 * it to disk before logging the new inode size in case of
192 		 * an extending write.
193 		 */
194 		if (XFS_IS_REALTIME_INODE(ip))
195 			xfs_blkdev_issue_flush(mp->m_rtdev_targp);
196 		else if (mp->m_logdev_targp != mp->m_ddev_targp)
197 			xfs_blkdev_issue_flush(mp->m_ddev_targp);
198 	}
199 
200 	/*
201 	 * All metadata updates are logged, which means that we just have
202 	 * to flush the log up to the latest LSN that touched the inode.
203 	 */
204 	xfs_ilock(ip, XFS_ILOCK_SHARED);
205 	if (xfs_ipincount(ip)) {
206 		if (!datasync ||
207 		    (ip->i_itemp->ili_fields & ~XFS_ILOG_TIMESTAMP))
208 			lsn = ip->i_itemp->ili_last_lsn;
209 	}
210 	xfs_iunlock(ip, XFS_ILOCK_SHARED);
211 
212 	if (lsn)
213 		error = _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
214 
215 	/*
216 	 * If we only have a single device, and the log force about was
217 	 * a no-op we might have to flush the data device cache here.
218 	 * This can only happen for fdatasync/O_DSYNC if we were overwriting
219 	 * an already allocated file and thus do not have any metadata to
220 	 * commit.
221 	 */
222 	if ((mp->m_flags & XFS_MOUNT_BARRIER) &&
223 	    mp->m_logdev_targp == mp->m_ddev_targp &&
224 	    !XFS_IS_REALTIME_INODE(ip) &&
225 	    !log_flushed)
226 		xfs_blkdev_issue_flush(mp->m_ddev_targp);
227 
228 	return -error;
229 }
230 
231 STATIC ssize_t
232 xfs_file_aio_read(
233 	struct kiocb		*iocb,
234 	const struct iovec	*iovp,
235 	unsigned long		nr_segs,
236 	loff_t			pos)
237 {
238 	struct file		*file = iocb->ki_filp;
239 	struct inode		*inode = file->f_mapping->host;
240 	struct xfs_inode	*ip = XFS_I(inode);
241 	struct xfs_mount	*mp = ip->i_mount;
242 	size_t			size = 0;
243 	ssize_t			ret = 0;
244 	int			ioflags = 0;
245 	xfs_fsize_t		n;
246 
247 	XFS_STATS_INC(xs_read_calls);
248 
249 	BUG_ON(iocb->ki_pos != pos);
250 
251 	if (unlikely(file->f_flags & O_DIRECT))
252 		ioflags |= IO_ISDIRECT;
253 	if (file->f_mode & FMODE_NOCMTIME)
254 		ioflags |= IO_INVIS;
255 
256 	ret = generic_segment_checks(iovp, &nr_segs, &size, VERIFY_WRITE);
257 	if (ret < 0)
258 		return ret;
259 
260 	if (unlikely(ioflags & IO_ISDIRECT)) {
261 		xfs_buftarg_t	*target =
262 			XFS_IS_REALTIME_INODE(ip) ?
263 				mp->m_rtdev_targp : mp->m_ddev_targp;
264 		/* DIO must be aligned to device logical sector size */
265 		if ((pos | size) & target->bt_logical_sectormask) {
266 			if (pos == i_size_read(inode))
267 				return 0;
268 			return -XFS_ERROR(EINVAL);
269 		}
270 	}
271 
272 	n = mp->m_super->s_maxbytes - pos;
273 	if (n <= 0 || size == 0)
274 		return 0;
275 
276 	if (n < size)
277 		size = n;
278 
279 	if (XFS_FORCED_SHUTDOWN(mp))
280 		return -EIO;
281 
282 	/*
283 	 * Locking is a bit tricky here. If we take an exclusive lock
284 	 * for direct IO, we effectively serialise all new concurrent
285 	 * read IO to this file and block it behind IO that is currently in
286 	 * progress because IO in progress holds the IO lock shared. We only
287 	 * need to hold the lock exclusive to blow away the page cache, so
288 	 * only take lock exclusively if the page cache needs invalidation.
289 	 * This allows the normal direct IO case of no page cache pages to
290 	 * proceeed concurrently without serialisation.
291 	 */
292 	xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
293 	if ((ioflags & IO_ISDIRECT) && inode->i_mapping->nrpages) {
294 		xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
295 		xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
296 
297 		if (inode->i_mapping->nrpages) {
298 			ret = -filemap_write_and_wait_range(
299 							VFS_I(ip)->i_mapping,
300 							pos, -1);
301 			if (ret) {
302 				xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
303 				return ret;
304 			}
305 			truncate_pagecache_range(VFS_I(ip), pos, -1);
306 		}
307 		xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
308 	}
309 
310 	trace_xfs_file_read(ip, size, pos, ioflags);
311 
312 	ret = generic_file_aio_read(iocb, iovp, nr_segs, pos);
313 	if (ret > 0)
314 		XFS_STATS_ADD(xs_read_bytes, ret);
315 
316 	xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
317 	return ret;
318 }
319 
320 STATIC ssize_t
321 xfs_file_splice_read(
322 	struct file		*infilp,
323 	loff_t			*ppos,
324 	struct pipe_inode_info	*pipe,
325 	size_t			count,
326 	unsigned int		flags)
327 {
328 	struct xfs_inode	*ip = XFS_I(infilp->f_mapping->host);
329 	int			ioflags = 0;
330 	ssize_t			ret;
331 
332 	XFS_STATS_INC(xs_read_calls);
333 
334 	if (infilp->f_mode & FMODE_NOCMTIME)
335 		ioflags |= IO_INVIS;
336 
337 	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
338 		return -EIO;
339 
340 	xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
341 
342 	trace_xfs_file_splice_read(ip, count, *ppos, ioflags);
343 
344 	ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
345 	if (ret > 0)
346 		XFS_STATS_ADD(xs_read_bytes, ret);
347 
348 	xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
349 	return ret;
350 }
351 
352 /*
353  * xfs_file_splice_write() does not use xfs_rw_ilock() because
354  * generic_file_splice_write() takes the i_mutex itself. This, in theory,
355  * couuld cause lock inversions between the aio_write path and the splice path
356  * if someone is doing concurrent splice(2) based writes and write(2) based
357  * writes to the same inode. The only real way to fix this is to re-implement
358  * the generic code here with correct locking orders.
359  */
360 STATIC ssize_t
361 xfs_file_splice_write(
362 	struct pipe_inode_info	*pipe,
363 	struct file		*outfilp,
364 	loff_t			*ppos,
365 	size_t			count,
366 	unsigned int		flags)
367 {
368 	struct inode		*inode = outfilp->f_mapping->host;
369 	struct xfs_inode	*ip = XFS_I(inode);
370 	int			ioflags = 0;
371 	ssize_t			ret;
372 
373 	XFS_STATS_INC(xs_write_calls);
374 
375 	if (outfilp->f_mode & FMODE_NOCMTIME)
376 		ioflags |= IO_INVIS;
377 
378 	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
379 		return -EIO;
380 
381 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
382 
383 	trace_xfs_file_splice_write(ip, count, *ppos, ioflags);
384 
385 	ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
386 	if (ret > 0)
387 		XFS_STATS_ADD(xs_write_bytes, ret);
388 
389 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
390 	return ret;
391 }
392 
393 /*
394  * This routine is called to handle zeroing any space in the last block of the
395  * file that is beyond the EOF.  We do this since the size is being increased
396  * without writing anything to that block and we don't want to read the
397  * garbage on the disk.
398  */
399 STATIC int				/* error (positive) */
400 xfs_zero_last_block(
401 	struct xfs_inode	*ip,
402 	xfs_fsize_t		offset,
403 	xfs_fsize_t		isize)
404 {
405 	struct xfs_mount	*mp = ip->i_mount;
406 	xfs_fileoff_t		last_fsb = XFS_B_TO_FSBT(mp, isize);
407 	int			zero_offset = XFS_B_FSB_OFFSET(mp, isize);
408 	int			zero_len;
409 	int			nimaps = 1;
410 	int			error = 0;
411 	struct xfs_bmbt_irec	imap;
412 
413 	xfs_ilock(ip, XFS_ILOCK_EXCL);
414 	error = xfs_bmapi_read(ip, last_fsb, 1, &imap, &nimaps, 0);
415 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
416 	if (error)
417 		return error;
418 
419 	ASSERT(nimaps > 0);
420 
421 	/*
422 	 * If the block underlying isize is just a hole, then there
423 	 * is nothing to zero.
424 	 */
425 	if (imap.br_startblock == HOLESTARTBLOCK)
426 		return 0;
427 
428 	zero_len = mp->m_sb.sb_blocksize - zero_offset;
429 	if (isize + zero_len > offset)
430 		zero_len = offset - isize;
431 	return xfs_iozero(ip, isize, zero_len);
432 }
433 
434 /*
435  * Zero any on disk space between the current EOF and the new, larger EOF.
436  *
437  * This handles the normal case of zeroing the remainder of the last block in
438  * the file and the unusual case of zeroing blocks out beyond the size of the
439  * file.  This second case only happens with fixed size extents and when the
440  * system crashes before the inode size was updated but after blocks were
441  * allocated.
442  *
443  * Expects the iolock to be held exclusive, and will take the ilock internally.
444  */
445 int					/* error (positive) */
446 xfs_zero_eof(
447 	struct xfs_inode	*ip,
448 	xfs_off_t		offset,		/* starting I/O offset */
449 	xfs_fsize_t		isize)		/* current inode size */
450 {
451 	struct xfs_mount	*mp = ip->i_mount;
452 	xfs_fileoff_t		start_zero_fsb;
453 	xfs_fileoff_t		end_zero_fsb;
454 	xfs_fileoff_t		zero_count_fsb;
455 	xfs_fileoff_t		last_fsb;
456 	xfs_fileoff_t		zero_off;
457 	xfs_fsize_t		zero_len;
458 	int			nimaps;
459 	int			error = 0;
460 	struct xfs_bmbt_irec	imap;
461 
462 	ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
463 	ASSERT(offset > isize);
464 
465 	/*
466 	 * First handle zeroing the block on which isize resides.
467 	 *
468 	 * We only zero a part of that block so it is handled specially.
469 	 */
470 	if (XFS_B_FSB_OFFSET(mp, isize) != 0) {
471 		error = xfs_zero_last_block(ip, offset, isize);
472 		if (error)
473 			return error;
474 	}
475 
476 	/*
477 	 * Calculate the range between the new size and the old where blocks
478 	 * needing to be zeroed may exist.
479 	 *
480 	 * To get the block where the last byte in the file currently resides,
481 	 * we need to subtract one from the size and truncate back to a block
482 	 * boundary.  We subtract 1 in case the size is exactly on a block
483 	 * boundary.
484 	 */
485 	last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
486 	start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
487 	end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
488 	ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
489 	if (last_fsb == end_zero_fsb) {
490 		/*
491 		 * The size was only incremented on its last block.
492 		 * We took care of that above, so just return.
493 		 */
494 		return 0;
495 	}
496 
497 	ASSERT(start_zero_fsb <= end_zero_fsb);
498 	while (start_zero_fsb <= end_zero_fsb) {
499 		nimaps = 1;
500 		zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
501 
502 		xfs_ilock(ip, XFS_ILOCK_EXCL);
503 		error = xfs_bmapi_read(ip, start_zero_fsb, zero_count_fsb,
504 					  &imap, &nimaps, 0);
505 		xfs_iunlock(ip, XFS_ILOCK_EXCL);
506 		if (error)
507 			return error;
508 
509 		ASSERT(nimaps > 0);
510 
511 		if (imap.br_state == XFS_EXT_UNWRITTEN ||
512 		    imap.br_startblock == HOLESTARTBLOCK) {
513 			start_zero_fsb = imap.br_startoff + imap.br_blockcount;
514 			ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
515 			continue;
516 		}
517 
518 		/*
519 		 * There are blocks we need to zero.
520 		 */
521 		zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
522 		zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
523 
524 		if ((zero_off + zero_len) > offset)
525 			zero_len = offset - zero_off;
526 
527 		error = xfs_iozero(ip, zero_off, zero_len);
528 		if (error)
529 			return error;
530 
531 		start_zero_fsb = imap.br_startoff + imap.br_blockcount;
532 		ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
533 	}
534 
535 	return 0;
536 }
537 
538 /*
539  * Common pre-write limit and setup checks.
540  *
541  * Called with the iolocked held either shared and exclusive according to
542  * @iolock, and returns with it held.  Might upgrade the iolock to exclusive
543  * if called for a direct write beyond i_size.
544  */
545 STATIC ssize_t
546 xfs_file_aio_write_checks(
547 	struct file		*file,
548 	loff_t			*pos,
549 	size_t			*count,
550 	int			*iolock)
551 {
552 	struct inode		*inode = file->f_mapping->host;
553 	struct xfs_inode	*ip = XFS_I(inode);
554 	int			error = 0;
555 
556 restart:
557 	error = generic_write_checks(file, pos, count, S_ISBLK(inode->i_mode));
558 	if (error)
559 		return error;
560 
561 	/*
562 	 * If the offset is beyond the size of the file, we need to zero any
563 	 * blocks that fall between the existing EOF and the start of this
564 	 * write.  If zeroing is needed and we are currently holding the
565 	 * iolock shared, we need to update it to exclusive which implies
566 	 * having to redo all checks before.
567 	 */
568 	if (*pos > i_size_read(inode)) {
569 		if (*iolock == XFS_IOLOCK_SHARED) {
570 			xfs_rw_iunlock(ip, *iolock);
571 			*iolock = XFS_IOLOCK_EXCL;
572 			xfs_rw_ilock(ip, *iolock);
573 			goto restart;
574 		}
575 		error = -xfs_zero_eof(ip, *pos, i_size_read(inode));
576 		if (error)
577 			return error;
578 	}
579 
580 	/*
581 	 * Updating the timestamps will grab the ilock again from
582 	 * xfs_fs_dirty_inode, so we have to call it after dropping the
583 	 * lock above.  Eventually we should look into a way to avoid
584 	 * the pointless lock roundtrip.
585 	 */
586 	if (likely(!(file->f_mode & FMODE_NOCMTIME))) {
587 		error = file_update_time(file);
588 		if (error)
589 			return error;
590 	}
591 
592 	/*
593 	 * If we're writing the file then make sure to clear the setuid and
594 	 * setgid bits if the process is not being run by root.  This keeps
595 	 * people from modifying setuid and setgid binaries.
596 	 */
597 	return file_remove_suid(file);
598 }
599 
600 /*
601  * xfs_file_dio_aio_write - handle direct IO writes
602  *
603  * Lock the inode appropriately to prepare for and issue a direct IO write.
604  * By separating it from the buffered write path we remove all the tricky to
605  * follow locking changes and looping.
606  *
607  * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
608  * until we're sure the bytes at the new EOF have been zeroed and/or the cached
609  * pages are flushed out.
610  *
611  * In most cases the direct IO writes will be done holding IOLOCK_SHARED
612  * allowing them to be done in parallel with reads and other direct IO writes.
613  * However, if the IO is not aligned to filesystem blocks, the direct IO layer
614  * needs to do sub-block zeroing and that requires serialisation against other
615  * direct IOs to the same block. In this case we need to serialise the
616  * submission of the unaligned IOs so that we don't get racing block zeroing in
617  * the dio layer.  To avoid the problem with aio, we also need to wait for
618  * outstanding IOs to complete so that unwritten extent conversion is completed
619  * before we try to map the overlapping block. This is currently implemented by
620  * hitting it with a big hammer (i.e. inode_dio_wait()).
621  *
622  * Returns with locks held indicated by @iolock and errors indicated by
623  * negative return values.
624  */
625 STATIC ssize_t
626 xfs_file_dio_aio_write(
627 	struct kiocb		*iocb,
628 	const struct iovec	*iovp,
629 	unsigned long		nr_segs,
630 	loff_t			pos,
631 	size_t			ocount)
632 {
633 	struct file		*file = iocb->ki_filp;
634 	struct address_space	*mapping = file->f_mapping;
635 	struct inode		*inode = mapping->host;
636 	struct xfs_inode	*ip = XFS_I(inode);
637 	struct xfs_mount	*mp = ip->i_mount;
638 	ssize_t			ret = 0;
639 	size_t			count = ocount;
640 	int			unaligned_io = 0;
641 	int			iolock;
642 	struct xfs_buftarg	*target = XFS_IS_REALTIME_INODE(ip) ?
643 					mp->m_rtdev_targp : mp->m_ddev_targp;
644 
645 	/* DIO must be aligned to device logical sector size */
646 	if ((pos | count) & target->bt_logical_sectormask)
647 		return -XFS_ERROR(EINVAL);
648 
649 	/* "unaligned" here means not aligned to a filesystem block */
650 	if ((pos & mp->m_blockmask) || ((pos + count) & mp->m_blockmask))
651 		unaligned_io = 1;
652 
653 	/*
654 	 * We don't need to take an exclusive lock unless there page cache needs
655 	 * to be invalidated or unaligned IO is being executed. We don't need to
656 	 * consider the EOF extension case here because
657 	 * xfs_file_aio_write_checks() will relock the inode as necessary for
658 	 * EOF zeroing cases and fill out the new inode size as appropriate.
659 	 */
660 	if (unaligned_io || mapping->nrpages)
661 		iolock = XFS_IOLOCK_EXCL;
662 	else
663 		iolock = XFS_IOLOCK_SHARED;
664 	xfs_rw_ilock(ip, iolock);
665 
666 	/*
667 	 * Recheck if there are cached pages that need invalidate after we got
668 	 * the iolock to protect against other threads adding new pages while
669 	 * we were waiting for the iolock.
670 	 */
671 	if (mapping->nrpages && iolock == XFS_IOLOCK_SHARED) {
672 		xfs_rw_iunlock(ip, iolock);
673 		iolock = XFS_IOLOCK_EXCL;
674 		xfs_rw_ilock(ip, iolock);
675 	}
676 
677 	ret = xfs_file_aio_write_checks(file, &pos, &count, &iolock);
678 	if (ret)
679 		goto out;
680 
681 	if (mapping->nrpages) {
682 		ret = -filemap_write_and_wait_range(VFS_I(ip)->i_mapping,
683 						    pos, -1);
684 		if (ret)
685 			goto out;
686 		truncate_pagecache_range(VFS_I(ip), pos, -1);
687 	}
688 
689 	/*
690 	 * If we are doing unaligned IO, wait for all other IO to drain,
691 	 * otherwise demote the lock if we had to flush cached pages
692 	 */
693 	if (unaligned_io)
694 		inode_dio_wait(inode);
695 	else if (iolock == XFS_IOLOCK_EXCL) {
696 		xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
697 		iolock = XFS_IOLOCK_SHARED;
698 	}
699 
700 	trace_xfs_file_direct_write(ip, count, iocb->ki_pos, 0);
701 	ret = generic_file_direct_write(iocb, iovp,
702 			&nr_segs, pos, count, ocount);
703 
704 out:
705 	xfs_rw_iunlock(ip, iolock);
706 
707 	/* No fallback to buffered IO on errors for XFS. */
708 	ASSERT(ret < 0 || ret == count);
709 	return ret;
710 }
711 
712 STATIC ssize_t
713 xfs_file_buffered_aio_write(
714 	struct kiocb		*iocb,
715 	const struct iovec	*iovp,
716 	unsigned long		nr_segs,
717 	loff_t			pos,
718 	size_t			count)
719 {
720 	struct file		*file = iocb->ki_filp;
721 	struct address_space	*mapping = file->f_mapping;
722 	struct inode		*inode = mapping->host;
723 	struct xfs_inode	*ip = XFS_I(inode);
724 	ssize_t			ret;
725 	int			enospc = 0;
726 	int			iolock = XFS_IOLOCK_EXCL;
727 	struct iov_iter		from;
728 
729 	xfs_rw_ilock(ip, iolock);
730 
731 	ret = xfs_file_aio_write_checks(file, &pos, &count, &iolock);
732 	if (ret)
733 		goto out;
734 
735 	iov_iter_init(&from, iovp, nr_segs, count, 0);
736 	/* We can write back this queue in page reclaim */
737 	current->backing_dev_info = mapping->backing_dev_info;
738 
739 write_retry:
740 	trace_xfs_file_buffered_write(ip, count, iocb->ki_pos, 0);
741 	ret = generic_perform_write(file, &from, pos);
742 	if (likely(ret >= 0))
743 		iocb->ki_pos = pos + ret;
744 	/*
745 	 * If we just got an ENOSPC, try to write back all dirty inodes to
746 	 * convert delalloc space to free up some of the excess reserved
747 	 * metadata space.
748 	 */
749 	if (ret == -ENOSPC && !enospc) {
750 		enospc = 1;
751 		xfs_flush_inodes(ip->i_mount);
752 		goto write_retry;
753 	}
754 
755 	current->backing_dev_info = NULL;
756 out:
757 	xfs_rw_iunlock(ip, iolock);
758 	return ret;
759 }
760 
761 STATIC ssize_t
762 xfs_file_aio_write(
763 	struct kiocb		*iocb,
764 	const struct iovec	*iovp,
765 	unsigned long		nr_segs,
766 	loff_t			pos)
767 {
768 	struct file		*file = iocb->ki_filp;
769 	struct address_space	*mapping = file->f_mapping;
770 	struct inode		*inode = mapping->host;
771 	struct xfs_inode	*ip = XFS_I(inode);
772 	ssize_t			ret;
773 	size_t			ocount = 0;
774 
775 	XFS_STATS_INC(xs_write_calls);
776 
777 	BUG_ON(iocb->ki_pos != pos);
778 
779 	ret = generic_segment_checks(iovp, &nr_segs, &ocount, VERIFY_READ);
780 	if (ret)
781 		return ret;
782 
783 	if (ocount == 0)
784 		return 0;
785 
786 	if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
787 		ret = -EIO;
788 		goto out;
789 	}
790 
791 	if (unlikely(file->f_flags & O_DIRECT))
792 		ret = xfs_file_dio_aio_write(iocb, iovp, nr_segs, pos, ocount);
793 	else
794 		ret = xfs_file_buffered_aio_write(iocb, iovp, nr_segs, pos,
795 						  ocount);
796 
797 	if (ret > 0) {
798 		ssize_t err;
799 
800 		XFS_STATS_ADD(xs_write_bytes, ret);
801 
802 		/* Handle various SYNC-type writes */
803 		err = generic_write_sync(file, iocb->ki_pos - ret, ret);
804 		if (err < 0)
805 			ret = err;
806 	}
807 
808 out:
809 	return ret;
810 }
811 
812 STATIC long
813 xfs_file_fallocate(
814 	struct file		*file,
815 	int			mode,
816 	loff_t			offset,
817 	loff_t			len)
818 {
819 	struct inode		*inode = file_inode(file);
820 	struct xfs_inode	*ip = XFS_I(inode);
821 	struct xfs_trans	*tp;
822 	long			error;
823 	loff_t			new_size = 0;
824 
825 	if (!S_ISREG(inode->i_mode))
826 		return -EINVAL;
827 	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
828 		     FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
829 		return -EOPNOTSUPP;
830 
831 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
832 	if (mode & FALLOC_FL_PUNCH_HOLE) {
833 		error = xfs_free_file_space(ip, offset, len);
834 		if (error)
835 			goto out_unlock;
836 	} else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
837 		unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
838 
839 		if (offset & blksize_mask || len & blksize_mask) {
840 			error = -EINVAL;
841 			goto out_unlock;
842 		}
843 
844 		ASSERT(offset + len < i_size_read(inode));
845 		new_size = i_size_read(inode) - len;
846 
847 		error = xfs_collapse_file_space(ip, offset, len);
848 		if (error)
849 			goto out_unlock;
850 	} else {
851 		if (!(mode & FALLOC_FL_KEEP_SIZE) &&
852 		    offset + len > i_size_read(inode)) {
853 			new_size = offset + len;
854 			error = -inode_newsize_ok(inode, new_size);
855 			if (error)
856 				goto out_unlock;
857 		}
858 
859 		if (mode & FALLOC_FL_ZERO_RANGE)
860 			error = xfs_zero_file_space(ip, offset, len);
861 		else
862 			error = xfs_alloc_file_space(ip, offset, len,
863 						     XFS_BMAPI_PREALLOC);
864 		if (error)
865 			goto out_unlock;
866 	}
867 
868 	tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_WRITEID);
869 	error = xfs_trans_reserve(tp, &M_RES(ip->i_mount)->tr_writeid, 0, 0);
870 	if (error) {
871 		xfs_trans_cancel(tp, 0);
872 		goto out_unlock;
873 	}
874 
875 	xfs_ilock(ip, XFS_ILOCK_EXCL);
876 	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
877 	ip->i_d.di_mode &= ~S_ISUID;
878 	if (ip->i_d.di_mode & S_IXGRP)
879 		ip->i_d.di_mode &= ~S_ISGID;
880 
881 	if (!(mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_COLLAPSE_RANGE)))
882 		ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
883 
884 	xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
885 	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
886 
887 	if (file->f_flags & O_DSYNC)
888 		xfs_trans_set_sync(tp);
889 	error = xfs_trans_commit(tp, 0);
890 	if (error)
891 		goto out_unlock;
892 
893 	/* Change file size if needed */
894 	if (new_size) {
895 		struct iattr iattr;
896 
897 		iattr.ia_valid = ATTR_SIZE;
898 		iattr.ia_size = new_size;
899 		error = xfs_setattr_size(ip, &iattr);
900 	}
901 
902 out_unlock:
903 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
904 	return -error;
905 }
906 
907 
908 STATIC int
909 xfs_file_open(
910 	struct inode	*inode,
911 	struct file	*file)
912 {
913 	if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
914 		return -EFBIG;
915 	if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
916 		return -EIO;
917 	return 0;
918 }
919 
920 STATIC int
921 xfs_dir_open(
922 	struct inode	*inode,
923 	struct file	*file)
924 {
925 	struct xfs_inode *ip = XFS_I(inode);
926 	int		mode;
927 	int		error;
928 
929 	error = xfs_file_open(inode, file);
930 	if (error)
931 		return error;
932 
933 	/*
934 	 * If there are any blocks, read-ahead block 0 as we're almost
935 	 * certain to have the next operation be a read there.
936 	 */
937 	mode = xfs_ilock_data_map_shared(ip);
938 	if (ip->i_d.di_nextents > 0)
939 		xfs_dir3_data_readahead(NULL, ip, 0, -1);
940 	xfs_iunlock(ip, mode);
941 	return 0;
942 }
943 
944 STATIC int
945 xfs_file_release(
946 	struct inode	*inode,
947 	struct file	*filp)
948 {
949 	return -xfs_release(XFS_I(inode));
950 }
951 
952 STATIC int
953 xfs_file_readdir(
954 	struct file	*file,
955 	struct dir_context *ctx)
956 {
957 	struct inode	*inode = file_inode(file);
958 	xfs_inode_t	*ip = XFS_I(inode);
959 	int		error;
960 	size_t		bufsize;
961 
962 	/*
963 	 * The Linux API doesn't pass down the total size of the buffer
964 	 * we read into down to the filesystem.  With the filldir concept
965 	 * it's not needed for correct information, but the XFS dir2 leaf
966 	 * code wants an estimate of the buffer size to calculate it's
967 	 * readahead window and size the buffers used for mapping to
968 	 * physical blocks.
969 	 *
970 	 * Try to give it an estimate that's good enough, maybe at some
971 	 * point we can change the ->readdir prototype to include the
972 	 * buffer size.  For now we use the current glibc buffer size.
973 	 */
974 	bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
975 
976 	error = xfs_readdir(ip, ctx, bufsize);
977 	if (error)
978 		return -error;
979 	return 0;
980 }
981 
982 STATIC int
983 xfs_file_mmap(
984 	struct file	*filp,
985 	struct vm_area_struct *vma)
986 {
987 	vma->vm_ops = &xfs_file_vm_ops;
988 
989 	file_accessed(filp);
990 	return 0;
991 }
992 
993 /*
994  * mmap()d file has taken write protection fault and is being made
995  * writable. We can set the page state up correctly for a writable
996  * page, which means we can do correct delalloc accounting (ENOSPC
997  * checking!) and unwritten extent mapping.
998  */
999 STATIC int
1000 xfs_vm_page_mkwrite(
1001 	struct vm_area_struct	*vma,
1002 	struct vm_fault		*vmf)
1003 {
1004 	return block_page_mkwrite(vma, vmf, xfs_get_blocks);
1005 }
1006 
1007 /*
1008  * This type is designed to indicate the type of offset we would like
1009  * to search from page cache for either xfs_seek_data() or xfs_seek_hole().
1010  */
1011 enum {
1012 	HOLE_OFF = 0,
1013 	DATA_OFF,
1014 };
1015 
1016 /*
1017  * Lookup the desired type of offset from the given page.
1018  *
1019  * On success, return true and the offset argument will point to the
1020  * start of the region that was found.  Otherwise this function will
1021  * return false and keep the offset argument unchanged.
1022  */
1023 STATIC bool
1024 xfs_lookup_buffer_offset(
1025 	struct page		*page,
1026 	loff_t			*offset,
1027 	unsigned int		type)
1028 {
1029 	loff_t			lastoff = page_offset(page);
1030 	bool			found = false;
1031 	struct buffer_head	*bh, *head;
1032 
1033 	bh = head = page_buffers(page);
1034 	do {
1035 		/*
1036 		 * Unwritten extents that have data in the page
1037 		 * cache covering them can be identified by the
1038 		 * BH_Unwritten state flag.  Pages with multiple
1039 		 * buffers might have a mix of holes, data and
1040 		 * unwritten extents - any buffer with valid
1041 		 * data in it should have BH_Uptodate flag set
1042 		 * on it.
1043 		 */
1044 		if (buffer_unwritten(bh) ||
1045 		    buffer_uptodate(bh)) {
1046 			if (type == DATA_OFF)
1047 				found = true;
1048 		} else {
1049 			if (type == HOLE_OFF)
1050 				found = true;
1051 		}
1052 
1053 		if (found) {
1054 			*offset = lastoff;
1055 			break;
1056 		}
1057 		lastoff += bh->b_size;
1058 	} while ((bh = bh->b_this_page) != head);
1059 
1060 	return found;
1061 }
1062 
1063 /*
1064  * This routine is called to find out and return a data or hole offset
1065  * from the page cache for unwritten extents according to the desired
1066  * type for xfs_seek_data() or xfs_seek_hole().
1067  *
1068  * The argument offset is used to tell where we start to search from the
1069  * page cache.  Map is used to figure out the end points of the range to
1070  * lookup pages.
1071  *
1072  * Return true if the desired type of offset was found, and the argument
1073  * offset is filled with that address.  Otherwise, return false and keep
1074  * offset unchanged.
1075  */
1076 STATIC bool
1077 xfs_find_get_desired_pgoff(
1078 	struct inode		*inode,
1079 	struct xfs_bmbt_irec	*map,
1080 	unsigned int		type,
1081 	loff_t			*offset)
1082 {
1083 	struct xfs_inode	*ip = XFS_I(inode);
1084 	struct xfs_mount	*mp = ip->i_mount;
1085 	struct pagevec		pvec;
1086 	pgoff_t			index;
1087 	pgoff_t			end;
1088 	loff_t			endoff;
1089 	loff_t			startoff = *offset;
1090 	loff_t			lastoff = startoff;
1091 	bool			found = false;
1092 
1093 	pagevec_init(&pvec, 0);
1094 
1095 	index = startoff >> PAGE_CACHE_SHIFT;
1096 	endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1097 	end = endoff >> PAGE_CACHE_SHIFT;
1098 	do {
1099 		int		want;
1100 		unsigned	nr_pages;
1101 		unsigned int	i;
1102 
1103 		want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1104 		nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1105 					  want);
1106 		/*
1107 		 * No page mapped into given range.  If we are searching holes
1108 		 * and if this is the first time we got into the loop, it means
1109 		 * that the given offset is landed in a hole, return it.
1110 		 *
1111 		 * If we have already stepped through some block buffers to find
1112 		 * holes but they all contains data.  In this case, the last
1113 		 * offset is already updated and pointed to the end of the last
1114 		 * mapped page, if it does not reach the endpoint to search,
1115 		 * that means there should be a hole between them.
1116 		 */
1117 		if (nr_pages == 0) {
1118 			/* Data search found nothing */
1119 			if (type == DATA_OFF)
1120 				break;
1121 
1122 			ASSERT(type == HOLE_OFF);
1123 			if (lastoff == startoff || lastoff < endoff) {
1124 				found = true;
1125 				*offset = lastoff;
1126 			}
1127 			break;
1128 		}
1129 
1130 		/*
1131 		 * At lease we found one page.  If this is the first time we
1132 		 * step into the loop, and if the first page index offset is
1133 		 * greater than the given search offset, a hole was found.
1134 		 */
1135 		if (type == HOLE_OFF && lastoff == startoff &&
1136 		    lastoff < page_offset(pvec.pages[0])) {
1137 			found = true;
1138 			break;
1139 		}
1140 
1141 		for (i = 0; i < nr_pages; i++) {
1142 			struct page	*page = pvec.pages[i];
1143 			loff_t		b_offset;
1144 
1145 			/*
1146 			 * At this point, the page may be truncated or
1147 			 * invalidated (changing page->mapping to NULL),
1148 			 * or even swizzled back from swapper_space to tmpfs
1149 			 * file mapping. However, page->index will not change
1150 			 * because we have a reference on the page.
1151 			 *
1152 			 * Searching done if the page index is out of range.
1153 			 * If the current offset is not reaches the end of
1154 			 * the specified search range, there should be a hole
1155 			 * between them.
1156 			 */
1157 			if (page->index > end) {
1158 				if (type == HOLE_OFF && lastoff < endoff) {
1159 					*offset = lastoff;
1160 					found = true;
1161 				}
1162 				goto out;
1163 			}
1164 
1165 			lock_page(page);
1166 			/*
1167 			 * Page truncated or invalidated(page->mapping == NULL).
1168 			 * We can freely skip it and proceed to check the next
1169 			 * page.
1170 			 */
1171 			if (unlikely(page->mapping != inode->i_mapping)) {
1172 				unlock_page(page);
1173 				continue;
1174 			}
1175 
1176 			if (!page_has_buffers(page)) {
1177 				unlock_page(page);
1178 				continue;
1179 			}
1180 
1181 			found = xfs_lookup_buffer_offset(page, &b_offset, type);
1182 			if (found) {
1183 				/*
1184 				 * The found offset may be less than the start
1185 				 * point to search if this is the first time to
1186 				 * come here.
1187 				 */
1188 				*offset = max_t(loff_t, startoff, b_offset);
1189 				unlock_page(page);
1190 				goto out;
1191 			}
1192 
1193 			/*
1194 			 * We either searching data but nothing was found, or
1195 			 * searching hole but found a data buffer.  In either
1196 			 * case, probably the next page contains the desired
1197 			 * things, update the last offset to it so.
1198 			 */
1199 			lastoff = page_offset(page) + PAGE_SIZE;
1200 			unlock_page(page);
1201 		}
1202 
1203 		/*
1204 		 * The number of returned pages less than our desired, search
1205 		 * done.  In this case, nothing was found for searching data,
1206 		 * but we found a hole behind the last offset.
1207 		 */
1208 		if (nr_pages < want) {
1209 			if (type == HOLE_OFF) {
1210 				*offset = lastoff;
1211 				found = true;
1212 			}
1213 			break;
1214 		}
1215 
1216 		index = pvec.pages[i - 1]->index + 1;
1217 		pagevec_release(&pvec);
1218 	} while (index <= end);
1219 
1220 out:
1221 	pagevec_release(&pvec);
1222 	return found;
1223 }
1224 
1225 STATIC loff_t
1226 xfs_seek_data(
1227 	struct file		*file,
1228 	loff_t			start)
1229 {
1230 	struct inode		*inode = file->f_mapping->host;
1231 	struct xfs_inode	*ip = XFS_I(inode);
1232 	struct xfs_mount	*mp = ip->i_mount;
1233 	loff_t			uninitialized_var(offset);
1234 	xfs_fsize_t		isize;
1235 	xfs_fileoff_t		fsbno;
1236 	xfs_filblks_t		end;
1237 	uint			lock;
1238 	int			error;
1239 
1240 	lock = xfs_ilock_data_map_shared(ip);
1241 
1242 	isize = i_size_read(inode);
1243 	if (start >= isize) {
1244 		error = ENXIO;
1245 		goto out_unlock;
1246 	}
1247 
1248 	/*
1249 	 * Try to read extents from the first block indicated
1250 	 * by fsbno to the end block of the file.
1251 	 */
1252 	fsbno = XFS_B_TO_FSBT(mp, start);
1253 	end = XFS_B_TO_FSB(mp, isize);
1254 	for (;;) {
1255 		struct xfs_bmbt_irec	map[2];
1256 		int			nmap = 2;
1257 		unsigned int		i;
1258 
1259 		error = xfs_bmapi_read(ip, fsbno, end - fsbno, map, &nmap,
1260 				       XFS_BMAPI_ENTIRE);
1261 		if (error)
1262 			goto out_unlock;
1263 
1264 		/* No extents at given offset, must be beyond EOF */
1265 		if (nmap == 0) {
1266 			error = ENXIO;
1267 			goto out_unlock;
1268 		}
1269 
1270 		for (i = 0; i < nmap; i++) {
1271 			offset = max_t(loff_t, start,
1272 				       XFS_FSB_TO_B(mp, map[i].br_startoff));
1273 
1274 			/* Landed in a data extent */
1275 			if (map[i].br_startblock == DELAYSTARTBLOCK ||
1276 			    (map[i].br_state == XFS_EXT_NORM &&
1277 			     !isnullstartblock(map[i].br_startblock)))
1278 				goto out;
1279 
1280 			/*
1281 			 * Landed in an unwritten extent, try to search data
1282 			 * from page cache.
1283 			 */
1284 			if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1285 				if (xfs_find_get_desired_pgoff(inode, &map[i],
1286 							DATA_OFF, &offset))
1287 					goto out;
1288 			}
1289 		}
1290 
1291 		/*
1292 		 * map[0] is hole or its an unwritten extent but
1293 		 * without data in page cache.  Probably means that
1294 		 * we are reading after EOF if nothing in map[1].
1295 		 */
1296 		if (nmap == 1) {
1297 			error = ENXIO;
1298 			goto out_unlock;
1299 		}
1300 
1301 		ASSERT(i > 1);
1302 
1303 		/*
1304 		 * Nothing was found, proceed to the next round of search
1305 		 * if reading offset not beyond or hit EOF.
1306 		 */
1307 		fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1308 		start = XFS_FSB_TO_B(mp, fsbno);
1309 		if (start >= isize) {
1310 			error = ENXIO;
1311 			goto out_unlock;
1312 		}
1313 	}
1314 
1315 out:
1316 	offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1317 
1318 out_unlock:
1319 	xfs_iunlock(ip, lock);
1320 
1321 	if (error)
1322 		return -error;
1323 	return offset;
1324 }
1325 
1326 STATIC loff_t
1327 xfs_seek_hole(
1328 	struct file		*file,
1329 	loff_t			start)
1330 {
1331 	struct inode		*inode = file->f_mapping->host;
1332 	struct xfs_inode	*ip = XFS_I(inode);
1333 	struct xfs_mount	*mp = ip->i_mount;
1334 	loff_t			uninitialized_var(offset);
1335 	xfs_fsize_t		isize;
1336 	xfs_fileoff_t		fsbno;
1337 	xfs_filblks_t		end;
1338 	uint			lock;
1339 	int			error;
1340 
1341 	if (XFS_FORCED_SHUTDOWN(mp))
1342 		return -XFS_ERROR(EIO);
1343 
1344 	lock = xfs_ilock_data_map_shared(ip);
1345 
1346 	isize = i_size_read(inode);
1347 	if (start >= isize) {
1348 		error = ENXIO;
1349 		goto out_unlock;
1350 	}
1351 
1352 	fsbno = XFS_B_TO_FSBT(mp, start);
1353 	end = XFS_B_TO_FSB(mp, isize);
1354 
1355 	for (;;) {
1356 		struct xfs_bmbt_irec	map[2];
1357 		int			nmap = 2;
1358 		unsigned int		i;
1359 
1360 		error = xfs_bmapi_read(ip, fsbno, end - fsbno, map, &nmap,
1361 				       XFS_BMAPI_ENTIRE);
1362 		if (error)
1363 			goto out_unlock;
1364 
1365 		/* No extents at given offset, must be beyond EOF */
1366 		if (nmap == 0) {
1367 			error = ENXIO;
1368 			goto out_unlock;
1369 		}
1370 
1371 		for (i = 0; i < nmap; i++) {
1372 			offset = max_t(loff_t, start,
1373 				       XFS_FSB_TO_B(mp, map[i].br_startoff));
1374 
1375 			/* Landed in a hole */
1376 			if (map[i].br_startblock == HOLESTARTBLOCK)
1377 				goto out;
1378 
1379 			/*
1380 			 * Landed in an unwritten extent, try to search hole
1381 			 * from page cache.
1382 			 */
1383 			if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1384 				if (xfs_find_get_desired_pgoff(inode, &map[i],
1385 							HOLE_OFF, &offset))
1386 					goto out;
1387 			}
1388 		}
1389 
1390 		/*
1391 		 * map[0] contains data or its unwritten but contains
1392 		 * data in page cache, probably means that we are
1393 		 * reading after EOF.  We should fix offset to point
1394 		 * to the end of the file(i.e., there is an implicit
1395 		 * hole at the end of any file).
1396 		 */
1397 		if (nmap == 1) {
1398 			offset = isize;
1399 			break;
1400 		}
1401 
1402 		ASSERT(i > 1);
1403 
1404 		/*
1405 		 * Both mappings contains data, proceed to the next round of
1406 		 * search if the current reading offset not beyond or hit EOF.
1407 		 */
1408 		fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1409 		start = XFS_FSB_TO_B(mp, fsbno);
1410 		if (start >= isize) {
1411 			offset = isize;
1412 			break;
1413 		}
1414 	}
1415 
1416 out:
1417 	/*
1418 	 * At this point, we must have found a hole.  However, the returned
1419 	 * offset may be bigger than the file size as it may be aligned to
1420 	 * page boundary for unwritten extents, we need to deal with this
1421 	 * situation in particular.
1422 	 */
1423 	offset = min_t(loff_t, offset, isize);
1424 	offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1425 
1426 out_unlock:
1427 	xfs_iunlock(ip, lock);
1428 
1429 	if (error)
1430 		return -error;
1431 	return offset;
1432 }
1433 
1434 STATIC loff_t
1435 xfs_file_llseek(
1436 	struct file	*file,
1437 	loff_t		offset,
1438 	int		origin)
1439 {
1440 	switch (origin) {
1441 	case SEEK_END:
1442 	case SEEK_CUR:
1443 	case SEEK_SET:
1444 		return generic_file_llseek(file, offset, origin);
1445 	case SEEK_DATA:
1446 		return xfs_seek_data(file, offset);
1447 	case SEEK_HOLE:
1448 		return xfs_seek_hole(file, offset);
1449 	default:
1450 		return -EINVAL;
1451 	}
1452 }
1453 
1454 const struct file_operations xfs_file_operations = {
1455 	.llseek		= xfs_file_llseek,
1456 	.read		= do_sync_read,
1457 	.write		= do_sync_write,
1458 	.aio_read	= xfs_file_aio_read,
1459 	.aio_write	= xfs_file_aio_write,
1460 	.splice_read	= xfs_file_splice_read,
1461 	.splice_write	= xfs_file_splice_write,
1462 	.unlocked_ioctl	= xfs_file_ioctl,
1463 #ifdef CONFIG_COMPAT
1464 	.compat_ioctl	= xfs_file_compat_ioctl,
1465 #endif
1466 	.mmap		= xfs_file_mmap,
1467 	.open		= xfs_file_open,
1468 	.release	= xfs_file_release,
1469 	.fsync		= xfs_file_fsync,
1470 	.fallocate	= xfs_file_fallocate,
1471 };
1472 
1473 const struct file_operations xfs_dir_file_operations = {
1474 	.open		= xfs_dir_open,
1475 	.read		= generic_read_dir,
1476 	.iterate	= xfs_file_readdir,
1477 	.llseek		= generic_file_llseek,
1478 	.unlocked_ioctl	= xfs_file_ioctl,
1479 #ifdef CONFIG_COMPAT
1480 	.compat_ioctl	= xfs_file_compat_ioctl,
1481 #endif
1482 	.fsync		= xfs_dir_fsync,
1483 };
1484 
1485 static const struct vm_operations_struct xfs_file_vm_ops = {
1486 	.fault		= filemap_fault,
1487 	.map_pages	= filemap_map_pages,
1488 	.page_mkwrite	= xfs_vm_page_mkwrite,
1489 	.remap_pages	= generic_file_remap_pages,
1490 };
1491