Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_inode.h"
14#include "xfs_trans.h"
15#include "xfs_inode_item.h"
16#include "xfs_bmap.h"
17#include "xfs_bmap_util.h"
18#include "xfs_dir2.h"
19#include "xfs_dir2_priv.h"
20#include "xfs_ioctl.h"
21#include "xfs_trace.h"
22#include "xfs_log.h"
23#include "xfs_icache.h"
24#include "xfs_pnfs.h"
25#include "xfs_iomap.h"
26#include "xfs_reflink.h"
27
28#include <linux/dax.h>
29#include <linux/falloc.h>
30#include <linux/backing-dev.h>
31#include <linux/mman.h>
32#include <linux/fadvise.h>
33#include <linux/mount.h>
34
35static const struct vm_operations_struct xfs_file_vm_ops;
36
37/*
38 * Decide if the given file range is aligned to the size of the fundamental
39 * allocation unit for the file.
40 */
41static bool
42xfs_is_falloc_aligned(
43 struct xfs_inode *ip,
44 loff_t pos,
45 long long int len)
46{
47 struct xfs_mount *mp = ip->i_mount;
48 uint64_t mask;
49
50 if (XFS_IS_REALTIME_INODE(ip)) {
51 if (!is_power_of_2(mp->m_sb.sb_rextsize)) {
52 u64 rextbytes;
53 u32 mod;
54
55 rextbytes = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize);
56 div_u64_rem(pos, rextbytes, &mod);
57 if (mod)
58 return false;
59 div_u64_rem(len, rextbytes, &mod);
60 return mod == 0;
61 }
62 mask = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize) - 1;
63 } else {
64 mask = mp->m_sb.sb_blocksize - 1;
65 }
66
67 return !((pos | len) & mask);
68}
69
70/*
71 * Fsync operations on directories are much simpler than on regular files,
72 * as there is no file data to flush, and thus also no need for explicit
73 * cache flush operations, and there are no non-transaction metadata updates
74 * on directories either.
75 */
76STATIC int
77xfs_dir_fsync(
78 struct file *file,
79 loff_t start,
80 loff_t end,
81 int datasync)
82{
83 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
84
85 trace_xfs_dir_fsync(ip);
86 return xfs_log_force_inode(ip);
87}
88
89static xfs_csn_t
90xfs_fsync_seq(
91 struct xfs_inode *ip,
92 bool datasync)
93{
94 if (!xfs_ipincount(ip))
95 return 0;
96 if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
97 return 0;
98 return ip->i_itemp->ili_commit_seq;
99}
100
101/*
102 * All metadata updates are logged, which means that we just have to flush the
103 * log up to the latest LSN that touched the inode.
104 *
105 * If we have concurrent fsync/fdatasync() calls, we need them to all block on
106 * the log force before we clear the ili_fsync_fields field. This ensures that
107 * we don't get a racing sync operation that does not wait for the metadata to
108 * hit the journal before returning. If we race with clearing ili_fsync_fields,
109 * then all that will happen is the log force will do nothing as the lsn will
110 * already be on disk. We can't race with setting ili_fsync_fields because that
111 * is done under XFS_ILOCK_EXCL, and that can't happen because we hold the lock
112 * shared until after the ili_fsync_fields is cleared.
113 */
114static int
115xfs_fsync_flush_log(
116 struct xfs_inode *ip,
117 bool datasync,
118 int *log_flushed)
119{
120 int error = 0;
121 xfs_csn_t seq;
122
123 xfs_ilock(ip, XFS_ILOCK_SHARED);
124 seq = xfs_fsync_seq(ip, datasync);
125 if (seq) {
126 error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
127 log_flushed);
128
129 spin_lock(&ip->i_itemp->ili_lock);
130 ip->i_itemp->ili_fsync_fields = 0;
131 spin_unlock(&ip->i_itemp->ili_lock);
132 }
133 xfs_iunlock(ip, XFS_ILOCK_SHARED);
134 return error;
135}
136
137STATIC int
138xfs_file_fsync(
139 struct file *file,
140 loff_t start,
141 loff_t end,
142 int datasync)
143{
144 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
145 struct xfs_mount *mp = ip->i_mount;
146 int error, err2;
147 int log_flushed = 0;
148
149 trace_xfs_file_fsync(ip);
150
151 error = file_write_and_wait_range(file, start, end);
152 if (error)
153 return error;
154
155 if (xfs_is_shutdown(mp))
156 return -EIO;
157
158 xfs_iflags_clear(ip, XFS_ITRUNCATED);
159
160 /*
161 * If we have an RT and/or log subvolume we need to make sure to flush
162 * the write cache the device used for file data first. This is to
163 * ensure newly written file data make it to disk before logging the new
164 * inode size in case of an extending write.
165 */
166 if (XFS_IS_REALTIME_INODE(ip))
167 error = blkdev_issue_flush(mp->m_rtdev_targp->bt_bdev);
168 else if (mp->m_logdev_targp != mp->m_ddev_targp)
169 error = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
170
171 /*
172 * Any inode that has dirty modifications in the log is pinned. The
173 * racy check here for a pinned inode will not catch modifications
174 * that happen concurrently to the fsync call, but fsync semantics
175 * only require to sync previously completed I/O.
176 */
177 if (xfs_ipincount(ip)) {
178 err2 = xfs_fsync_flush_log(ip, datasync, &log_flushed);
179 if (err2 && !error)
180 error = err2;
181 }
182
183 /*
184 * If we only have a single device, and the log force about was
185 * a no-op we might have to flush the data device cache here.
186 * This can only happen for fdatasync/O_DSYNC if we were overwriting
187 * an already allocated file and thus do not have any metadata to
188 * commit.
189 */
190 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
191 mp->m_logdev_targp == mp->m_ddev_targp) {
192 err2 = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
193 if (err2 && !error)
194 error = err2;
195 }
196
197 return error;
198}
199
200static int
201xfs_ilock_iocb(
202 struct kiocb *iocb,
203 unsigned int lock_mode)
204{
205 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
206
207 if (iocb->ki_flags & IOCB_NOWAIT) {
208 if (!xfs_ilock_nowait(ip, lock_mode))
209 return -EAGAIN;
210 } else {
211 xfs_ilock(ip, lock_mode);
212 }
213
214 return 0;
215}
216
217STATIC ssize_t
218xfs_file_dio_read(
219 struct kiocb *iocb,
220 struct iov_iter *to)
221{
222 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
223 ssize_t ret;
224
225 trace_xfs_file_direct_read(iocb, to);
226
227 if (!iov_iter_count(to))
228 return 0; /* skip atime */
229
230 file_accessed(iocb->ki_filp);
231
232 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
233 if (ret)
234 return ret;
235 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
236 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
237
238 return ret;
239}
240
241static noinline ssize_t
242xfs_file_dax_read(
243 struct kiocb *iocb,
244 struct iov_iter *to)
245{
246 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
247 ssize_t ret = 0;
248
249 trace_xfs_file_dax_read(iocb, to);
250
251 if (!iov_iter_count(to))
252 return 0; /* skip atime */
253
254 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
255 if (ret)
256 return ret;
257 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
258 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
259
260 file_accessed(iocb->ki_filp);
261 return ret;
262}
263
264STATIC ssize_t
265xfs_file_buffered_read(
266 struct kiocb *iocb,
267 struct iov_iter *to)
268{
269 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
270 ssize_t ret;
271
272 trace_xfs_file_buffered_read(iocb, to);
273
274 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
275 if (ret)
276 return ret;
277 ret = generic_file_read_iter(iocb, to);
278 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
279
280 return ret;
281}
282
283STATIC ssize_t
284xfs_file_read_iter(
285 struct kiocb *iocb,
286 struct iov_iter *to)
287{
288 struct inode *inode = file_inode(iocb->ki_filp);
289 struct xfs_mount *mp = XFS_I(inode)->i_mount;
290 ssize_t ret = 0;
291
292 XFS_STATS_INC(mp, xs_read_calls);
293
294 if (xfs_is_shutdown(mp))
295 return -EIO;
296
297 if (IS_DAX(inode))
298 ret = xfs_file_dax_read(iocb, to);
299 else if (iocb->ki_flags & IOCB_DIRECT)
300 ret = xfs_file_dio_read(iocb, to);
301 else
302 ret = xfs_file_buffered_read(iocb, to);
303
304 if (ret > 0)
305 XFS_STATS_ADD(mp, xs_read_bytes, ret);
306 return ret;
307}
308
309/*
310 * Common pre-write limit and setup checks.
311 *
312 * Called with the iolocked held either shared and exclusive according to
313 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
314 * if called for a direct write beyond i_size.
315 */
316STATIC ssize_t
317xfs_file_write_checks(
318 struct kiocb *iocb,
319 struct iov_iter *from,
320 unsigned int *iolock)
321{
322 struct file *file = iocb->ki_filp;
323 struct inode *inode = file->f_mapping->host;
324 struct xfs_inode *ip = XFS_I(inode);
325 ssize_t error = 0;
326 size_t count = iov_iter_count(from);
327 bool drained_dio = false;
328 loff_t isize;
329
330restart:
331 error = generic_write_checks(iocb, from);
332 if (error <= 0)
333 return error;
334
335 if (iocb->ki_flags & IOCB_NOWAIT) {
336 error = break_layout(inode, false);
337 if (error == -EWOULDBLOCK)
338 error = -EAGAIN;
339 } else {
340 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
341 }
342
343 if (error)
344 return error;
345
346 /*
347 * For changing security info in file_remove_privs() we need i_rwsem
348 * exclusively.
349 */
350 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
351 xfs_iunlock(ip, *iolock);
352 *iolock = XFS_IOLOCK_EXCL;
353 error = xfs_ilock_iocb(iocb, *iolock);
354 if (error) {
355 *iolock = 0;
356 return error;
357 }
358 goto restart;
359 }
360
361 /*
362 * If the offset is beyond the size of the file, we need to zero any
363 * blocks that fall between the existing EOF and the start of this
364 * write. If zeroing is needed and we are currently holding the iolock
365 * shared, we need to update it to exclusive which implies having to
366 * redo all checks before.
367 *
368 * We need to serialise against EOF updates that occur in IO completions
369 * here. We want to make sure that nobody is changing the size while we
370 * do this check until we have placed an IO barrier (i.e. hold the
371 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
372 * spinlock effectively forms a memory barrier once we have the
373 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
374 * hence be able to correctly determine if we need to run zeroing.
375 *
376 * We can do an unlocked check here safely as IO completion can only
377 * extend EOF. Truncate is locked out at this point, so the EOF can
378 * not move backwards, only forwards. Hence we only need to take the
379 * slow path and spin locks when we are at or beyond the current EOF.
380 */
381 if (iocb->ki_pos <= i_size_read(inode))
382 goto out;
383
384 spin_lock(&ip->i_flags_lock);
385 isize = i_size_read(inode);
386 if (iocb->ki_pos > isize) {
387 spin_unlock(&ip->i_flags_lock);
388
389 if (iocb->ki_flags & IOCB_NOWAIT)
390 return -EAGAIN;
391
392 if (!drained_dio) {
393 if (*iolock == XFS_IOLOCK_SHARED) {
394 xfs_iunlock(ip, *iolock);
395 *iolock = XFS_IOLOCK_EXCL;
396 xfs_ilock(ip, *iolock);
397 iov_iter_reexpand(from, count);
398 }
399 /*
400 * We now have an IO submission barrier in place, but
401 * AIO can do EOF updates during IO completion and hence
402 * we now need to wait for all of them to drain. Non-AIO
403 * DIO will have drained before we are given the
404 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
405 * no-op.
406 */
407 inode_dio_wait(inode);
408 drained_dio = true;
409 goto restart;
410 }
411
412 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
413 error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
414 if (error)
415 return error;
416 } else
417 spin_unlock(&ip->i_flags_lock);
418
419out:
420 return kiocb_modified(iocb);
421}
422
423static int
424xfs_dio_write_end_io(
425 struct kiocb *iocb,
426 ssize_t size,
427 int error,
428 unsigned flags)
429{
430 struct inode *inode = file_inode(iocb->ki_filp);
431 struct xfs_inode *ip = XFS_I(inode);
432 loff_t offset = iocb->ki_pos;
433 unsigned int nofs_flag;
434
435 trace_xfs_end_io_direct_write(ip, offset, size);
436
437 if (xfs_is_shutdown(ip->i_mount))
438 return -EIO;
439
440 if (error)
441 return error;
442 if (!size)
443 return 0;
444
445 /*
446 * Capture amount written on completion as we can't reliably account
447 * for it on submission.
448 */
449 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
450
451 /*
452 * We can allocate memory here while doing writeback on behalf of
453 * memory reclaim. To avoid memory allocation deadlocks set the
454 * task-wide nofs context for the following operations.
455 */
456 nofs_flag = memalloc_nofs_save();
457
458 if (flags & IOMAP_DIO_COW) {
459 error = xfs_reflink_end_cow(ip, offset, size);
460 if (error)
461 goto out;
462 }
463
464 /*
465 * Unwritten conversion updates the in-core isize after extent
466 * conversion but before updating the on-disk size. Updating isize any
467 * earlier allows a racing dio read to find unwritten extents before
468 * they are converted.
469 */
470 if (flags & IOMAP_DIO_UNWRITTEN) {
471 error = xfs_iomap_write_unwritten(ip, offset, size, true);
472 goto out;
473 }
474
475 /*
476 * We need to update the in-core inode size here so that we don't end up
477 * with the on-disk inode size being outside the in-core inode size. We
478 * have no other method of updating EOF for AIO, so always do it here
479 * if necessary.
480 *
481 * We need to lock the test/set EOF update as we can be racing with
482 * other IO completions here to update the EOF. Failing to serialise
483 * here can result in EOF moving backwards and Bad Things Happen when
484 * that occurs.
485 *
486 * As IO completion only ever extends EOF, we can do an unlocked check
487 * here to avoid taking the spinlock. If we land within the current EOF,
488 * then we do not need to do an extending update at all, and we don't
489 * need to take the lock to check this. If we race with an update moving
490 * EOF, then we'll either still be beyond EOF and need to take the lock,
491 * or we'll be within EOF and we don't need to take it at all.
492 */
493 if (offset + size <= i_size_read(inode))
494 goto out;
495
496 spin_lock(&ip->i_flags_lock);
497 if (offset + size > i_size_read(inode)) {
498 i_size_write(inode, offset + size);
499 spin_unlock(&ip->i_flags_lock);
500 error = xfs_setfilesize(ip, offset, size);
501 } else {
502 spin_unlock(&ip->i_flags_lock);
503 }
504
505out:
506 memalloc_nofs_restore(nofs_flag);
507 return error;
508}
509
510static const struct iomap_dio_ops xfs_dio_write_ops = {
511 .end_io = xfs_dio_write_end_io,
512};
513
514/*
515 * Handle block aligned direct I/O writes
516 */
517static noinline ssize_t
518xfs_file_dio_write_aligned(
519 struct xfs_inode *ip,
520 struct kiocb *iocb,
521 struct iov_iter *from)
522{
523 unsigned int iolock = XFS_IOLOCK_SHARED;
524 ssize_t ret;
525
526 ret = xfs_ilock_iocb(iocb, iolock);
527 if (ret)
528 return ret;
529 ret = xfs_file_write_checks(iocb, from, &iolock);
530 if (ret)
531 goto out_unlock;
532
533 /*
534 * We don't need to hold the IOLOCK exclusively across the IO, so demote
535 * the iolock back to shared if we had to take the exclusive lock in
536 * xfs_file_write_checks() for other reasons.
537 */
538 if (iolock == XFS_IOLOCK_EXCL) {
539 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
540 iolock = XFS_IOLOCK_SHARED;
541 }
542 trace_xfs_file_direct_write(iocb, from);
543 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
544 &xfs_dio_write_ops, 0, NULL, 0);
545out_unlock:
546 if (iolock)
547 xfs_iunlock(ip, iolock);
548 return ret;
549}
550
551/*
552 * Handle block unaligned direct I/O writes
553 *
554 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
555 * them to be done in parallel with reads and other direct I/O writes. However,
556 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
557 * to do sub-block zeroing and that requires serialisation against other direct
558 * I/O to the same block. In this case we need to serialise the submission of
559 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
560 * In the case where sub-block zeroing is not required, we can do concurrent
561 * sub-block dios to the same block successfully.
562 *
563 * Optimistically submit the I/O using the shared lock first, but use the
564 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
565 * if block allocation or partial block zeroing would be required. In that case
566 * we try again with the exclusive lock.
567 */
568static noinline ssize_t
569xfs_file_dio_write_unaligned(
570 struct xfs_inode *ip,
571 struct kiocb *iocb,
572 struct iov_iter *from)
573{
574 size_t isize = i_size_read(VFS_I(ip));
575 size_t count = iov_iter_count(from);
576 unsigned int iolock = XFS_IOLOCK_SHARED;
577 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
578 ssize_t ret;
579
580 /*
581 * Extending writes need exclusivity because of the sub-block zeroing
582 * that the DIO code always does for partial tail blocks beyond EOF, so
583 * don't even bother trying the fast path in this case.
584 */
585 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
586 if (iocb->ki_flags & IOCB_NOWAIT)
587 return -EAGAIN;
588retry_exclusive:
589 iolock = XFS_IOLOCK_EXCL;
590 flags = IOMAP_DIO_FORCE_WAIT;
591 }
592
593 ret = xfs_ilock_iocb(iocb, iolock);
594 if (ret)
595 return ret;
596
597 /*
598 * We can't properly handle unaligned direct I/O to reflink files yet,
599 * as we can't unshare a partial block.
600 */
601 if (xfs_is_cow_inode(ip)) {
602 trace_xfs_reflink_bounce_dio_write(iocb, from);
603 ret = -ENOTBLK;
604 goto out_unlock;
605 }
606
607 ret = xfs_file_write_checks(iocb, from, &iolock);
608 if (ret)
609 goto out_unlock;
610
611 /*
612 * If we are doing exclusive unaligned I/O, this must be the only I/O
613 * in-flight. Otherwise we risk data corruption due to unwritten extent
614 * conversions from the AIO end_io handler. Wait for all other I/O to
615 * drain first.
616 */
617 if (flags & IOMAP_DIO_FORCE_WAIT)
618 inode_dio_wait(VFS_I(ip));
619
620 trace_xfs_file_direct_write(iocb, from);
621 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
622 &xfs_dio_write_ops, flags, NULL, 0);
623
624 /*
625 * Retry unaligned I/O with exclusive blocking semantics if the DIO
626 * layer rejected it for mapping or locking reasons. If we are doing
627 * nonblocking user I/O, propagate the error.
628 */
629 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
630 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
631 xfs_iunlock(ip, iolock);
632 goto retry_exclusive;
633 }
634
635out_unlock:
636 if (iolock)
637 xfs_iunlock(ip, iolock);
638 return ret;
639}
640
641static ssize_t
642xfs_file_dio_write(
643 struct kiocb *iocb,
644 struct iov_iter *from)
645{
646 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
647 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
648 size_t count = iov_iter_count(from);
649
650 /* direct I/O must be aligned to device logical sector size */
651 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
652 return -EINVAL;
653 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
654 return xfs_file_dio_write_unaligned(ip, iocb, from);
655 return xfs_file_dio_write_aligned(ip, iocb, from);
656}
657
658static noinline ssize_t
659xfs_file_dax_write(
660 struct kiocb *iocb,
661 struct iov_iter *from)
662{
663 struct inode *inode = iocb->ki_filp->f_mapping->host;
664 struct xfs_inode *ip = XFS_I(inode);
665 unsigned int iolock = XFS_IOLOCK_EXCL;
666 ssize_t ret, error = 0;
667 loff_t pos;
668
669 ret = xfs_ilock_iocb(iocb, iolock);
670 if (ret)
671 return ret;
672 ret = xfs_file_write_checks(iocb, from, &iolock);
673 if (ret)
674 goto out;
675
676 pos = iocb->ki_pos;
677
678 trace_xfs_file_dax_write(iocb, from);
679 ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
680 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
681 i_size_write(inode, iocb->ki_pos);
682 error = xfs_setfilesize(ip, pos, ret);
683 }
684out:
685 if (iolock)
686 xfs_iunlock(ip, iolock);
687 if (error)
688 return error;
689
690 if (ret > 0) {
691 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
692
693 /* Handle various SYNC-type writes */
694 ret = generic_write_sync(iocb, ret);
695 }
696 return ret;
697}
698
699STATIC ssize_t
700xfs_file_buffered_write(
701 struct kiocb *iocb,
702 struct iov_iter *from)
703{
704 struct inode *inode = iocb->ki_filp->f_mapping->host;
705 struct xfs_inode *ip = XFS_I(inode);
706 ssize_t ret;
707 bool cleared_space = false;
708 unsigned int iolock;
709
710write_retry:
711 iolock = XFS_IOLOCK_EXCL;
712 ret = xfs_ilock_iocb(iocb, iolock);
713 if (ret)
714 return ret;
715
716 ret = xfs_file_write_checks(iocb, from, &iolock);
717 if (ret)
718 goto out;
719
720 /* We can write back this queue in page reclaim */
721 current->backing_dev_info = inode_to_bdi(inode);
722
723 trace_xfs_file_buffered_write(iocb, from);
724 ret = iomap_file_buffered_write(iocb, from,
725 &xfs_buffered_write_iomap_ops);
726 if (likely(ret >= 0))
727 iocb->ki_pos += ret;
728
729 /*
730 * If we hit a space limit, try to free up some lingering preallocated
731 * space before returning an error. In the case of ENOSPC, first try to
732 * write back all dirty inodes to free up some of the excess reserved
733 * metadata space. This reduces the chances that the eofblocks scan
734 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
735 * also behaves as a filter to prevent too many eofblocks scans from
736 * running at the same time. Use a synchronous scan to increase the
737 * effectiveness of the scan.
738 */
739 if (ret == -EDQUOT && !cleared_space) {
740 xfs_iunlock(ip, iolock);
741 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
742 cleared_space = true;
743 goto write_retry;
744 } else if (ret == -ENOSPC && !cleared_space) {
745 struct xfs_icwalk icw = {0};
746
747 cleared_space = true;
748 xfs_flush_inodes(ip->i_mount);
749
750 xfs_iunlock(ip, iolock);
751 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
752 xfs_blockgc_free_space(ip->i_mount, &icw);
753 goto write_retry;
754 }
755
756 current->backing_dev_info = NULL;
757out:
758 if (iolock)
759 xfs_iunlock(ip, iolock);
760
761 if (ret > 0) {
762 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
763 /* Handle various SYNC-type writes */
764 ret = generic_write_sync(iocb, ret);
765 }
766 return ret;
767}
768
769STATIC ssize_t
770xfs_file_write_iter(
771 struct kiocb *iocb,
772 struct iov_iter *from)
773{
774 struct inode *inode = iocb->ki_filp->f_mapping->host;
775 struct xfs_inode *ip = XFS_I(inode);
776 ssize_t ret;
777 size_t ocount = iov_iter_count(from);
778
779 XFS_STATS_INC(ip->i_mount, xs_write_calls);
780
781 if (ocount == 0)
782 return 0;
783
784 if (xfs_is_shutdown(ip->i_mount))
785 return -EIO;
786
787 if (IS_DAX(inode))
788 return xfs_file_dax_write(iocb, from);
789
790 if (iocb->ki_flags & IOCB_DIRECT) {
791 /*
792 * Allow a directio write to fall back to a buffered
793 * write *only* in the case that we're doing a reflink
794 * CoW. In all other directio scenarios we do not
795 * allow an operation to fall back to buffered mode.
796 */
797 ret = xfs_file_dio_write(iocb, from);
798 if (ret != -ENOTBLK)
799 return ret;
800 }
801
802 return xfs_file_buffered_write(iocb, from);
803}
804
805static void
806xfs_wait_dax_page(
807 struct inode *inode)
808{
809 struct xfs_inode *ip = XFS_I(inode);
810
811 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
812 schedule();
813 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
814}
815
816int
817xfs_break_dax_layouts(
818 struct inode *inode,
819 bool *retry)
820{
821 struct page *page;
822
823 ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
824
825 page = dax_layout_busy_page(inode->i_mapping);
826 if (!page)
827 return 0;
828
829 *retry = true;
830 return ___wait_var_event(&page->_refcount,
831 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
832 0, 0, xfs_wait_dax_page(inode));
833}
834
835int
836xfs_break_layouts(
837 struct inode *inode,
838 uint *iolock,
839 enum layout_break_reason reason)
840{
841 bool retry;
842 int error;
843
844 ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
845
846 do {
847 retry = false;
848 switch (reason) {
849 case BREAK_UNMAP:
850 error = xfs_break_dax_layouts(inode, &retry);
851 if (error || retry)
852 break;
853 fallthrough;
854 case BREAK_WRITE:
855 error = xfs_break_leased_layouts(inode, iolock, &retry);
856 break;
857 default:
858 WARN_ON_ONCE(1);
859 error = -EINVAL;
860 }
861 } while (error == 0 && retry);
862
863 return error;
864}
865
866/* Does this file, inode, or mount want synchronous writes? */
867static inline bool xfs_file_sync_writes(struct file *filp)
868{
869 struct xfs_inode *ip = XFS_I(file_inode(filp));
870
871 if (xfs_has_wsync(ip->i_mount))
872 return true;
873 if (filp->f_flags & (__O_SYNC | O_DSYNC))
874 return true;
875 if (IS_SYNC(file_inode(filp)))
876 return true;
877
878 return false;
879}
880
881#define XFS_FALLOC_FL_SUPPORTED \
882 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
883 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
884 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
885
886STATIC long
887xfs_file_fallocate(
888 struct file *file,
889 int mode,
890 loff_t offset,
891 loff_t len)
892{
893 struct inode *inode = file_inode(file);
894 struct xfs_inode *ip = XFS_I(inode);
895 long error;
896 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
897 loff_t new_size = 0;
898 bool do_file_insert = false;
899
900 if (!S_ISREG(inode->i_mode))
901 return -EINVAL;
902 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
903 return -EOPNOTSUPP;
904
905 xfs_ilock(ip, iolock);
906 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
907 if (error)
908 goto out_unlock;
909
910 /*
911 * Must wait for all AIO to complete before we continue as AIO can
912 * change the file size on completion without holding any locks we
913 * currently hold. We must do this first because AIO can update both
914 * the on disk and in memory inode sizes, and the operations that follow
915 * require the in-memory size to be fully up-to-date.
916 */
917 inode_dio_wait(inode);
918
919 /*
920 * Now AIO and DIO has drained we flush and (if necessary) invalidate
921 * the cached range over the first operation we are about to run.
922 *
923 * We care about zero and collapse here because they both run a hole
924 * punch over the range first. Because that can zero data, and the range
925 * of invalidation for the shift operations is much larger, we still do
926 * the required flush for collapse in xfs_prepare_shift().
927 *
928 * Insert has the same range requirements as collapse, and we extend the
929 * file first which can zero data. Hence insert has the same
930 * flush/invalidate requirements as collapse and so they are both
931 * handled at the right time by xfs_prepare_shift().
932 */
933 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
934 FALLOC_FL_COLLAPSE_RANGE)) {
935 error = xfs_flush_unmap_range(ip, offset, len);
936 if (error)
937 goto out_unlock;
938 }
939
940 error = file_modified(file);
941 if (error)
942 goto out_unlock;
943
944 if (mode & FALLOC_FL_PUNCH_HOLE) {
945 error = xfs_free_file_space(ip, offset, len);
946 if (error)
947 goto out_unlock;
948 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
949 if (!xfs_is_falloc_aligned(ip, offset, len)) {
950 error = -EINVAL;
951 goto out_unlock;
952 }
953
954 /*
955 * There is no need to overlap collapse range with EOF,
956 * in which case it is effectively a truncate operation
957 */
958 if (offset + len >= i_size_read(inode)) {
959 error = -EINVAL;
960 goto out_unlock;
961 }
962
963 new_size = i_size_read(inode) - len;
964
965 error = xfs_collapse_file_space(ip, offset, len);
966 if (error)
967 goto out_unlock;
968 } else if (mode & FALLOC_FL_INSERT_RANGE) {
969 loff_t isize = i_size_read(inode);
970
971 if (!xfs_is_falloc_aligned(ip, offset, len)) {
972 error = -EINVAL;
973 goto out_unlock;
974 }
975
976 /*
977 * New inode size must not exceed ->s_maxbytes, accounting for
978 * possible signed overflow.
979 */
980 if (inode->i_sb->s_maxbytes - isize < len) {
981 error = -EFBIG;
982 goto out_unlock;
983 }
984 new_size = isize + len;
985
986 /* Offset should be less than i_size */
987 if (offset >= isize) {
988 error = -EINVAL;
989 goto out_unlock;
990 }
991 do_file_insert = true;
992 } else {
993 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
994 offset + len > i_size_read(inode)) {
995 new_size = offset + len;
996 error = inode_newsize_ok(inode, new_size);
997 if (error)
998 goto out_unlock;
999 }
1000
1001 if (mode & FALLOC_FL_ZERO_RANGE) {
1002 /*
1003 * Punch a hole and prealloc the range. We use a hole
1004 * punch rather than unwritten extent conversion for two
1005 * reasons:
1006 *
1007 * 1.) Hole punch handles partial block zeroing for us.
1008 * 2.) If prealloc returns ENOSPC, the file range is
1009 * still zero-valued by virtue of the hole punch.
1010 */
1011 unsigned int blksize = i_blocksize(inode);
1012
1013 trace_xfs_zero_file_space(ip);
1014
1015 error = xfs_free_file_space(ip, offset, len);
1016 if (error)
1017 goto out_unlock;
1018
1019 len = round_up(offset + len, blksize) -
1020 round_down(offset, blksize);
1021 offset = round_down(offset, blksize);
1022 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
1023 error = xfs_reflink_unshare(ip, offset, len);
1024 if (error)
1025 goto out_unlock;
1026 } else {
1027 /*
1028 * If always_cow mode we can't use preallocations and
1029 * thus should not create them.
1030 */
1031 if (xfs_is_always_cow_inode(ip)) {
1032 error = -EOPNOTSUPP;
1033 goto out_unlock;
1034 }
1035 }
1036
1037 if (!xfs_is_always_cow_inode(ip)) {
1038 error = xfs_alloc_file_space(ip, offset, len);
1039 if (error)
1040 goto out_unlock;
1041 }
1042 }
1043
1044 /* Change file size if needed */
1045 if (new_size) {
1046 struct iattr iattr;
1047
1048 iattr.ia_valid = ATTR_SIZE;
1049 iattr.ia_size = new_size;
1050 error = xfs_vn_setattr_size(file_mnt_user_ns(file),
1051 file_dentry(file), &iattr);
1052 if (error)
1053 goto out_unlock;
1054 }
1055
1056 /*
1057 * Perform hole insertion now that the file size has been
1058 * updated so that if we crash during the operation we don't
1059 * leave shifted extents past EOF and hence losing access to
1060 * the data that is contained within them.
1061 */
1062 if (do_file_insert) {
1063 error = xfs_insert_file_space(ip, offset, len);
1064 if (error)
1065 goto out_unlock;
1066 }
1067
1068 if (xfs_file_sync_writes(file))
1069 error = xfs_log_force_inode(ip);
1070
1071out_unlock:
1072 xfs_iunlock(ip, iolock);
1073 return error;
1074}
1075
1076STATIC int
1077xfs_file_fadvise(
1078 struct file *file,
1079 loff_t start,
1080 loff_t end,
1081 int advice)
1082{
1083 struct xfs_inode *ip = XFS_I(file_inode(file));
1084 int ret;
1085 int lockflags = 0;
1086
1087 /*
1088 * Operations creating pages in page cache need protection from hole
1089 * punching and similar ops
1090 */
1091 if (advice == POSIX_FADV_WILLNEED) {
1092 lockflags = XFS_IOLOCK_SHARED;
1093 xfs_ilock(ip, lockflags);
1094 }
1095 ret = generic_fadvise(file, start, end, advice);
1096 if (lockflags)
1097 xfs_iunlock(ip, lockflags);
1098 return ret;
1099}
1100
1101STATIC loff_t
1102xfs_file_remap_range(
1103 struct file *file_in,
1104 loff_t pos_in,
1105 struct file *file_out,
1106 loff_t pos_out,
1107 loff_t len,
1108 unsigned int remap_flags)
1109{
1110 struct inode *inode_in = file_inode(file_in);
1111 struct xfs_inode *src = XFS_I(inode_in);
1112 struct inode *inode_out = file_inode(file_out);
1113 struct xfs_inode *dest = XFS_I(inode_out);
1114 struct xfs_mount *mp = src->i_mount;
1115 loff_t remapped = 0;
1116 xfs_extlen_t cowextsize;
1117 int ret;
1118
1119 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1120 return -EINVAL;
1121
1122 if (!xfs_has_reflink(mp))
1123 return -EOPNOTSUPP;
1124
1125 if (xfs_is_shutdown(mp))
1126 return -EIO;
1127
1128 /* Prepare and then clone file data. */
1129 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1130 &len, remap_flags);
1131 if (ret || len == 0)
1132 return ret;
1133
1134 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1135
1136 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1137 &remapped);
1138 if (ret)
1139 goto out_unlock;
1140
1141 /*
1142 * Carry the cowextsize hint from src to dest if we're sharing the
1143 * entire source file to the entire destination file, the source file
1144 * has a cowextsize hint, and the destination file does not.
1145 */
1146 cowextsize = 0;
1147 if (pos_in == 0 && len == i_size_read(inode_in) &&
1148 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1149 pos_out == 0 && len >= i_size_read(inode_out) &&
1150 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1151 cowextsize = src->i_cowextsize;
1152
1153 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1154 remap_flags);
1155 if (ret)
1156 goto out_unlock;
1157
1158 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1159 xfs_log_force_inode(dest);
1160out_unlock:
1161 xfs_iunlock2_io_mmap(src, dest);
1162 if (ret)
1163 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1164 return remapped > 0 ? remapped : ret;
1165}
1166
1167STATIC int
1168xfs_file_open(
1169 struct inode *inode,
1170 struct file *file)
1171{
1172 if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1173 return -EIO;
1174 file->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC;
1175 return generic_file_open(inode, file);
1176}
1177
1178STATIC int
1179xfs_dir_open(
1180 struct inode *inode,
1181 struct file *file)
1182{
1183 struct xfs_inode *ip = XFS_I(inode);
1184 unsigned int mode;
1185 int error;
1186
1187 error = xfs_file_open(inode, file);
1188 if (error)
1189 return error;
1190
1191 /*
1192 * If there are any blocks, read-ahead block 0 as we're almost
1193 * certain to have the next operation be a read there.
1194 */
1195 mode = xfs_ilock_data_map_shared(ip);
1196 if (ip->i_df.if_nextents > 0)
1197 error = xfs_dir3_data_readahead(ip, 0, 0);
1198 xfs_iunlock(ip, mode);
1199 return error;
1200}
1201
1202STATIC int
1203xfs_file_release(
1204 struct inode *inode,
1205 struct file *filp)
1206{
1207 return xfs_release(XFS_I(inode));
1208}
1209
1210STATIC int
1211xfs_file_readdir(
1212 struct file *file,
1213 struct dir_context *ctx)
1214{
1215 struct inode *inode = file_inode(file);
1216 xfs_inode_t *ip = XFS_I(inode);
1217 size_t bufsize;
1218
1219 /*
1220 * The Linux API doesn't pass down the total size of the buffer
1221 * we read into down to the filesystem. With the filldir concept
1222 * it's not needed for correct information, but the XFS dir2 leaf
1223 * code wants an estimate of the buffer size to calculate it's
1224 * readahead window and size the buffers used for mapping to
1225 * physical blocks.
1226 *
1227 * Try to give it an estimate that's good enough, maybe at some
1228 * point we can change the ->readdir prototype to include the
1229 * buffer size. For now we use the current glibc buffer size.
1230 */
1231 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1232
1233 return xfs_readdir(NULL, ip, ctx, bufsize);
1234}
1235
1236STATIC loff_t
1237xfs_file_llseek(
1238 struct file *file,
1239 loff_t offset,
1240 int whence)
1241{
1242 struct inode *inode = file->f_mapping->host;
1243
1244 if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1245 return -EIO;
1246
1247 switch (whence) {
1248 default:
1249 return generic_file_llseek(file, offset, whence);
1250 case SEEK_HOLE:
1251 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1252 break;
1253 case SEEK_DATA:
1254 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1255 break;
1256 }
1257
1258 if (offset < 0)
1259 return offset;
1260 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1261}
1262
1263#ifdef CONFIG_FS_DAX
1264static inline vm_fault_t
1265xfs_dax_fault(
1266 struct vm_fault *vmf,
1267 enum page_entry_size pe_size,
1268 bool write_fault,
1269 pfn_t *pfn)
1270{
1271 return dax_iomap_fault(vmf, pe_size, pfn, NULL,
1272 (write_fault && !vmf->cow_page) ?
1273 &xfs_dax_write_iomap_ops :
1274 &xfs_read_iomap_ops);
1275}
1276#else
1277static inline vm_fault_t
1278xfs_dax_fault(
1279 struct vm_fault *vmf,
1280 enum page_entry_size pe_size,
1281 bool write_fault,
1282 pfn_t *pfn)
1283{
1284 ASSERT(0);
1285 return VM_FAULT_SIGBUS;
1286}
1287#endif
1288
1289/*
1290 * Locking for serialisation of IO during page faults. This results in a lock
1291 * ordering of:
1292 *
1293 * mmap_lock (MM)
1294 * sb_start_pagefault(vfs, freeze)
1295 * invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1296 * page_lock (MM)
1297 * i_lock (XFS - extent map serialisation)
1298 */
1299static vm_fault_t
1300__xfs_filemap_fault(
1301 struct vm_fault *vmf,
1302 enum page_entry_size pe_size,
1303 bool write_fault)
1304{
1305 struct inode *inode = file_inode(vmf->vma->vm_file);
1306 struct xfs_inode *ip = XFS_I(inode);
1307 vm_fault_t ret;
1308
1309 trace_xfs_filemap_fault(ip, pe_size, write_fault);
1310
1311 if (write_fault) {
1312 sb_start_pagefault(inode->i_sb);
1313 file_update_time(vmf->vma->vm_file);
1314 }
1315
1316 if (IS_DAX(inode)) {
1317 pfn_t pfn;
1318
1319 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1320 ret = xfs_dax_fault(vmf, pe_size, write_fault, &pfn);
1321 if (ret & VM_FAULT_NEEDDSYNC)
1322 ret = dax_finish_sync_fault(vmf, pe_size, pfn);
1323 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1324 } else {
1325 if (write_fault) {
1326 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1327 ret = iomap_page_mkwrite(vmf,
1328 &xfs_page_mkwrite_iomap_ops);
1329 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1330 } else {
1331 ret = filemap_fault(vmf);
1332 }
1333 }
1334
1335 if (write_fault)
1336 sb_end_pagefault(inode->i_sb);
1337 return ret;
1338}
1339
1340static inline bool
1341xfs_is_write_fault(
1342 struct vm_fault *vmf)
1343{
1344 return (vmf->flags & FAULT_FLAG_WRITE) &&
1345 (vmf->vma->vm_flags & VM_SHARED);
1346}
1347
1348static vm_fault_t
1349xfs_filemap_fault(
1350 struct vm_fault *vmf)
1351{
1352 /* DAX can shortcut the normal fault path on write faults! */
1353 return __xfs_filemap_fault(vmf, PE_SIZE_PTE,
1354 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1355 xfs_is_write_fault(vmf));
1356}
1357
1358static vm_fault_t
1359xfs_filemap_huge_fault(
1360 struct vm_fault *vmf,
1361 enum page_entry_size pe_size)
1362{
1363 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1364 return VM_FAULT_FALLBACK;
1365
1366 /* DAX can shortcut the normal fault path on write faults! */
1367 return __xfs_filemap_fault(vmf, pe_size,
1368 xfs_is_write_fault(vmf));
1369}
1370
1371static vm_fault_t
1372xfs_filemap_page_mkwrite(
1373 struct vm_fault *vmf)
1374{
1375 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1376}
1377
1378/*
1379 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1380 * on write faults. In reality, it needs to serialise against truncate and
1381 * prepare memory for writing so handle is as standard write fault.
1382 */
1383static vm_fault_t
1384xfs_filemap_pfn_mkwrite(
1385 struct vm_fault *vmf)
1386{
1387
1388 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1389}
1390
1391static vm_fault_t
1392xfs_filemap_map_pages(
1393 struct vm_fault *vmf,
1394 pgoff_t start_pgoff,
1395 pgoff_t end_pgoff)
1396{
1397 struct inode *inode = file_inode(vmf->vma->vm_file);
1398 vm_fault_t ret;
1399
1400 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1401 ret = filemap_map_pages(vmf, start_pgoff, end_pgoff);
1402 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1403 return ret;
1404}
1405
1406static const struct vm_operations_struct xfs_file_vm_ops = {
1407 .fault = xfs_filemap_fault,
1408 .huge_fault = xfs_filemap_huge_fault,
1409 .map_pages = xfs_filemap_map_pages,
1410 .page_mkwrite = xfs_filemap_page_mkwrite,
1411 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1412};
1413
1414STATIC int
1415xfs_file_mmap(
1416 struct file *file,
1417 struct vm_area_struct *vma)
1418{
1419 struct inode *inode = file_inode(file);
1420 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1421
1422 /*
1423 * We don't support synchronous mappings for non-DAX files and
1424 * for DAX files if underneath dax_device is not synchronous.
1425 */
1426 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1427 return -EOPNOTSUPP;
1428
1429 file_accessed(file);
1430 vma->vm_ops = &xfs_file_vm_ops;
1431 if (IS_DAX(inode))
1432 vma->vm_flags |= VM_HUGEPAGE;
1433 return 0;
1434}
1435
1436const struct file_operations xfs_file_operations = {
1437 .llseek = xfs_file_llseek,
1438 .read_iter = xfs_file_read_iter,
1439 .write_iter = xfs_file_write_iter,
1440 .splice_read = generic_file_splice_read,
1441 .splice_write = iter_file_splice_write,
1442 .iopoll = iocb_bio_iopoll,
1443 .unlocked_ioctl = xfs_file_ioctl,
1444#ifdef CONFIG_COMPAT
1445 .compat_ioctl = xfs_file_compat_ioctl,
1446#endif
1447 .mmap = xfs_file_mmap,
1448 .mmap_supported_flags = MAP_SYNC,
1449 .open = xfs_file_open,
1450 .release = xfs_file_release,
1451 .fsync = xfs_file_fsync,
1452 .get_unmapped_area = thp_get_unmapped_area,
1453 .fallocate = xfs_file_fallocate,
1454 .fadvise = xfs_file_fadvise,
1455 .remap_file_range = xfs_file_remap_range,
1456};
1457
1458const struct file_operations xfs_dir_file_operations = {
1459 .open = xfs_dir_open,
1460 .read = generic_read_dir,
1461 .iterate_shared = xfs_file_readdir,
1462 .llseek = generic_file_llseek,
1463 .unlocked_ioctl = xfs_file_ioctl,
1464#ifdef CONFIG_COMPAT
1465 .compat_ioctl = xfs_file_compat_ioctl,
1466#endif
1467 .fsync = xfs_dir_fsync,
1468};
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
47static 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 */
53static inline void
54xfs_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
63static inline void
64xfs_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
73static inline void
74xfs_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 */
93int
94xfs_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 */
138STATIC int
139xfs_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
161STATIC int
162xfs_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
231STATIC ssize_t
232xfs_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
320STATIC ssize_t
321xfs_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 */
360STATIC ssize_t
361xfs_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 */
399STATIC int /* error (positive) */
400xfs_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 */
445int /* error (positive) */
446xfs_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 */
545STATIC ssize_t
546xfs_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
556restart:
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 */
625STATIC ssize_t
626xfs_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
704out:
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
712STATIC ssize_t
713xfs_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
739write_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;
756out:
757 xfs_rw_iunlock(ip, iolock);
758 return ret;
759}
760
761STATIC ssize_t
762xfs_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
808out:
809 return ret;
810}
811
812STATIC long
813xfs_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 /*
845 * There is no need to overlap collapse range with EOF,
846 * in which case it is effectively a truncate operation
847 */
848 if (offset + len >= i_size_read(inode)) {
849 error = EINVAL;
850 goto out_unlock;
851 }
852
853 new_size = i_size_read(inode) - len;
854
855 error = xfs_collapse_file_space(ip, offset, len);
856 if (error)
857 goto out_unlock;
858 } else {
859 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
860 offset + len > i_size_read(inode)) {
861 new_size = offset + len;
862 error = -inode_newsize_ok(inode, new_size);
863 if (error)
864 goto out_unlock;
865 }
866
867 if (mode & FALLOC_FL_ZERO_RANGE)
868 error = xfs_zero_file_space(ip, offset, len);
869 else
870 error = xfs_alloc_file_space(ip, offset, len,
871 XFS_BMAPI_PREALLOC);
872 if (error)
873 goto out_unlock;
874 }
875
876 tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_WRITEID);
877 error = xfs_trans_reserve(tp, &M_RES(ip->i_mount)->tr_writeid, 0, 0);
878 if (error) {
879 xfs_trans_cancel(tp, 0);
880 goto out_unlock;
881 }
882
883 xfs_ilock(ip, XFS_ILOCK_EXCL);
884 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
885 ip->i_d.di_mode &= ~S_ISUID;
886 if (ip->i_d.di_mode & S_IXGRP)
887 ip->i_d.di_mode &= ~S_ISGID;
888
889 if (!(mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_COLLAPSE_RANGE)))
890 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
891
892 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
893 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
894
895 if (file->f_flags & O_DSYNC)
896 xfs_trans_set_sync(tp);
897 error = xfs_trans_commit(tp, 0);
898 if (error)
899 goto out_unlock;
900
901 /* Change file size if needed */
902 if (new_size) {
903 struct iattr iattr;
904
905 iattr.ia_valid = ATTR_SIZE;
906 iattr.ia_size = new_size;
907 error = xfs_setattr_size(ip, &iattr);
908 }
909
910out_unlock:
911 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
912 return -error;
913}
914
915
916STATIC int
917xfs_file_open(
918 struct inode *inode,
919 struct file *file)
920{
921 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
922 return -EFBIG;
923 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
924 return -EIO;
925 return 0;
926}
927
928STATIC int
929xfs_dir_open(
930 struct inode *inode,
931 struct file *file)
932{
933 struct xfs_inode *ip = XFS_I(inode);
934 int mode;
935 int error;
936
937 error = xfs_file_open(inode, file);
938 if (error)
939 return error;
940
941 /*
942 * If there are any blocks, read-ahead block 0 as we're almost
943 * certain to have the next operation be a read there.
944 */
945 mode = xfs_ilock_data_map_shared(ip);
946 if (ip->i_d.di_nextents > 0)
947 xfs_dir3_data_readahead(NULL, ip, 0, -1);
948 xfs_iunlock(ip, mode);
949 return 0;
950}
951
952STATIC int
953xfs_file_release(
954 struct inode *inode,
955 struct file *filp)
956{
957 return -xfs_release(XFS_I(inode));
958}
959
960STATIC int
961xfs_file_readdir(
962 struct file *file,
963 struct dir_context *ctx)
964{
965 struct inode *inode = file_inode(file);
966 xfs_inode_t *ip = XFS_I(inode);
967 int error;
968 size_t bufsize;
969
970 /*
971 * The Linux API doesn't pass down the total size of the buffer
972 * we read into down to the filesystem. With the filldir concept
973 * it's not needed for correct information, but the XFS dir2 leaf
974 * code wants an estimate of the buffer size to calculate it's
975 * readahead window and size the buffers used for mapping to
976 * physical blocks.
977 *
978 * Try to give it an estimate that's good enough, maybe at some
979 * point we can change the ->readdir prototype to include the
980 * buffer size. For now we use the current glibc buffer size.
981 */
982 bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
983
984 error = xfs_readdir(ip, ctx, bufsize);
985 if (error)
986 return -error;
987 return 0;
988}
989
990STATIC int
991xfs_file_mmap(
992 struct file *filp,
993 struct vm_area_struct *vma)
994{
995 vma->vm_ops = &xfs_file_vm_ops;
996
997 file_accessed(filp);
998 return 0;
999}
1000
1001/*
1002 * mmap()d file has taken write protection fault and is being made
1003 * writable. We can set the page state up correctly for a writable
1004 * page, which means we can do correct delalloc accounting (ENOSPC
1005 * checking!) and unwritten extent mapping.
1006 */
1007STATIC int
1008xfs_vm_page_mkwrite(
1009 struct vm_area_struct *vma,
1010 struct vm_fault *vmf)
1011{
1012 return block_page_mkwrite(vma, vmf, xfs_get_blocks);
1013}
1014
1015/*
1016 * This type is designed to indicate the type of offset we would like
1017 * to search from page cache for either xfs_seek_data() or xfs_seek_hole().
1018 */
1019enum {
1020 HOLE_OFF = 0,
1021 DATA_OFF,
1022};
1023
1024/*
1025 * Lookup the desired type of offset from the given page.
1026 *
1027 * On success, return true and the offset argument will point to the
1028 * start of the region that was found. Otherwise this function will
1029 * return false and keep the offset argument unchanged.
1030 */
1031STATIC bool
1032xfs_lookup_buffer_offset(
1033 struct page *page,
1034 loff_t *offset,
1035 unsigned int type)
1036{
1037 loff_t lastoff = page_offset(page);
1038 bool found = false;
1039 struct buffer_head *bh, *head;
1040
1041 bh = head = page_buffers(page);
1042 do {
1043 /*
1044 * Unwritten extents that have data in the page
1045 * cache covering them can be identified by the
1046 * BH_Unwritten state flag. Pages with multiple
1047 * buffers might have a mix of holes, data and
1048 * unwritten extents - any buffer with valid
1049 * data in it should have BH_Uptodate flag set
1050 * on it.
1051 */
1052 if (buffer_unwritten(bh) ||
1053 buffer_uptodate(bh)) {
1054 if (type == DATA_OFF)
1055 found = true;
1056 } else {
1057 if (type == HOLE_OFF)
1058 found = true;
1059 }
1060
1061 if (found) {
1062 *offset = lastoff;
1063 break;
1064 }
1065 lastoff += bh->b_size;
1066 } while ((bh = bh->b_this_page) != head);
1067
1068 return found;
1069}
1070
1071/*
1072 * This routine is called to find out and return a data or hole offset
1073 * from the page cache for unwritten extents according to the desired
1074 * type for xfs_seek_data() or xfs_seek_hole().
1075 *
1076 * The argument offset is used to tell where we start to search from the
1077 * page cache. Map is used to figure out the end points of the range to
1078 * lookup pages.
1079 *
1080 * Return true if the desired type of offset was found, and the argument
1081 * offset is filled with that address. Otherwise, return false and keep
1082 * offset unchanged.
1083 */
1084STATIC bool
1085xfs_find_get_desired_pgoff(
1086 struct inode *inode,
1087 struct xfs_bmbt_irec *map,
1088 unsigned int type,
1089 loff_t *offset)
1090{
1091 struct xfs_inode *ip = XFS_I(inode);
1092 struct xfs_mount *mp = ip->i_mount;
1093 struct pagevec pvec;
1094 pgoff_t index;
1095 pgoff_t end;
1096 loff_t endoff;
1097 loff_t startoff = *offset;
1098 loff_t lastoff = startoff;
1099 bool found = false;
1100
1101 pagevec_init(&pvec, 0);
1102
1103 index = startoff >> PAGE_CACHE_SHIFT;
1104 endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1105 end = endoff >> PAGE_CACHE_SHIFT;
1106 do {
1107 int want;
1108 unsigned nr_pages;
1109 unsigned int i;
1110
1111 want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1112 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1113 want);
1114 /*
1115 * No page mapped into given range. If we are searching holes
1116 * and if this is the first time we got into the loop, it means
1117 * that the given offset is landed in a hole, return it.
1118 *
1119 * If we have already stepped through some block buffers to find
1120 * holes but they all contains data. In this case, the last
1121 * offset is already updated and pointed to the end of the last
1122 * mapped page, if it does not reach the endpoint to search,
1123 * that means there should be a hole between them.
1124 */
1125 if (nr_pages == 0) {
1126 /* Data search found nothing */
1127 if (type == DATA_OFF)
1128 break;
1129
1130 ASSERT(type == HOLE_OFF);
1131 if (lastoff == startoff || lastoff < endoff) {
1132 found = true;
1133 *offset = lastoff;
1134 }
1135 break;
1136 }
1137
1138 /*
1139 * At lease we found one page. If this is the first time we
1140 * step into the loop, and if the first page index offset is
1141 * greater than the given search offset, a hole was found.
1142 */
1143 if (type == HOLE_OFF && lastoff == startoff &&
1144 lastoff < page_offset(pvec.pages[0])) {
1145 found = true;
1146 break;
1147 }
1148
1149 for (i = 0; i < nr_pages; i++) {
1150 struct page *page = pvec.pages[i];
1151 loff_t b_offset;
1152
1153 /*
1154 * At this point, the page may be truncated or
1155 * invalidated (changing page->mapping to NULL),
1156 * or even swizzled back from swapper_space to tmpfs
1157 * file mapping. However, page->index will not change
1158 * because we have a reference on the page.
1159 *
1160 * Searching done if the page index is out of range.
1161 * If the current offset is not reaches the end of
1162 * the specified search range, there should be a hole
1163 * between them.
1164 */
1165 if (page->index > end) {
1166 if (type == HOLE_OFF && lastoff < endoff) {
1167 *offset = lastoff;
1168 found = true;
1169 }
1170 goto out;
1171 }
1172
1173 lock_page(page);
1174 /*
1175 * Page truncated or invalidated(page->mapping == NULL).
1176 * We can freely skip it and proceed to check the next
1177 * page.
1178 */
1179 if (unlikely(page->mapping != inode->i_mapping)) {
1180 unlock_page(page);
1181 continue;
1182 }
1183
1184 if (!page_has_buffers(page)) {
1185 unlock_page(page);
1186 continue;
1187 }
1188
1189 found = xfs_lookup_buffer_offset(page, &b_offset, type);
1190 if (found) {
1191 /*
1192 * The found offset may be less than the start
1193 * point to search if this is the first time to
1194 * come here.
1195 */
1196 *offset = max_t(loff_t, startoff, b_offset);
1197 unlock_page(page);
1198 goto out;
1199 }
1200
1201 /*
1202 * We either searching data but nothing was found, or
1203 * searching hole but found a data buffer. In either
1204 * case, probably the next page contains the desired
1205 * things, update the last offset to it so.
1206 */
1207 lastoff = page_offset(page) + PAGE_SIZE;
1208 unlock_page(page);
1209 }
1210
1211 /*
1212 * The number of returned pages less than our desired, search
1213 * done. In this case, nothing was found for searching data,
1214 * but we found a hole behind the last offset.
1215 */
1216 if (nr_pages < want) {
1217 if (type == HOLE_OFF) {
1218 *offset = lastoff;
1219 found = true;
1220 }
1221 break;
1222 }
1223
1224 index = pvec.pages[i - 1]->index + 1;
1225 pagevec_release(&pvec);
1226 } while (index <= end);
1227
1228out:
1229 pagevec_release(&pvec);
1230 return found;
1231}
1232
1233STATIC loff_t
1234xfs_seek_data(
1235 struct file *file,
1236 loff_t start)
1237{
1238 struct inode *inode = file->f_mapping->host;
1239 struct xfs_inode *ip = XFS_I(inode);
1240 struct xfs_mount *mp = ip->i_mount;
1241 loff_t uninitialized_var(offset);
1242 xfs_fsize_t isize;
1243 xfs_fileoff_t fsbno;
1244 xfs_filblks_t end;
1245 uint lock;
1246 int error;
1247
1248 lock = xfs_ilock_data_map_shared(ip);
1249
1250 isize = i_size_read(inode);
1251 if (start >= isize) {
1252 error = ENXIO;
1253 goto out_unlock;
1254 }
1255
1256 /*
1257 * Try to read extents from the first block indicated
1258 * by fsbno to the end block of the file.
1259 */
1260 fsbno = XFS_B_TO_FSBT(mp, start);
1261 end = XFS_B_TO_FSB(mp, isize);
1262 for (;;) {
1263 struct xfs_bmbt_irec map[2];
1264 int nmap = 2;
1265 unsigned int i;
1266
1267 error = xfs_bmapi_read(ip, fsbno, end - fsbno, map, &nmap,
1268 XFS_BMAPI_ENTIRE);
1269 if (error)
1270 goto out_unlock;
1271
1272 /* No extents at given offset, must be beyond EOF */
1273 if (nmap == 0) {
1274 error = ENXIO;
1275 goto out_unlock;
1276 }
1277
1278 for (i = 0; i < nmap; i++) {
1279 offset = max_t(loff_t, start,
1280 XFS_FSB_TO_B(mp, map[i].br_startoff));
1281
1282 /* Landed in a data extent */
1283 if (map[i].br_startblock == DELAYSTARTBLOCK ||
1284 (map[i].br_state == XFS_EXT_NORM &&
1285 !isnullstartblock(map[i].br_startblock)))
1286 goto out;
1287
1288 /*
1289 * Landed in an unwritten extent, try to search data
1290 * from page cache.
1291 */
1292 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1293 if (xfs_find_get_desired_pgoff(inode, &map[i],
1294 DATA_OFF, &offset))
1295 goto out;
1296 }
1297 }
1298
1299 /*
1300 * map[0] is hole or its an unwritten extent but
1301 * without data in page cache. Probably means that
1302 * we are reading after EOF if nothing in map[1].
1303 */
1304 if (nmap == 1) {
1305 error = ENXIO;
1306 goto out_unlock;
1307 }
1308
1309 ASSERT(i > 1);
1310
1311 /*
1312 * Nothing was found, proceed to the next round of search
1313 * if reading offset not beyond or hit EOF.
1314 */
1315 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1316 start = XFS_FSB_TO_B(mp, fsbno);
1317 if (start >= isize) {
1318 error = ENXIO;
1319 goto out_unlock;
1320 }
1321 }
1322
1323out:
1324 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1325
1326out_unlock:
1327 xfs_iunlock(ip, lock);
1328
1329 if (error)
1330 return -error;
1331 return offset;
1332}
1333
1334STATIC loff_t
1335xfs_seek_hole(
1336 struct file *file,
1337 loff_t start)
1338{
1339 struct inode *inode = file->f_mapping->host;
1340 struct xfs_inode *ip = XFS_I(inode);
1341 struct xfs_mount *mp = ip->i_mount;
1342 loff_t uninitialized_var(offset);
1343 xfs_fsize_t isize;
1344 xfs_fileoff_t fsbno;
1345 xfs_filblks_t end;
1346 uint lock;
1347 int error;
1348
1349 if (XFS_FORCED_SHUTDOWN(mp))
1350 return -XFS_ERROR(EIO);
1351
1352 lock = xfs_ilock_data_map_shared(ip);
1353
1354 isize = i_size_read(inode);
1355 if (start >= isize) {
1356 error = ENXIO;
1357 goto out_unlock;
1358 }
1359
1360 fsbno = XFS_B_TO_FSBT(mp, start);
1361 end = XFS_B_TO_FSB(mp, isize);
1362
1363 for (;;) {
1364 struct xfs_bmbt_irec map[2];
1365 int nmap = 2;
1366 unsigned int i;
1367
1368 error = xfs_bmapi_read(ip, fsbno, end - fsbno, map, &nmap,
1369 XFS_BMAPI_ENTIRE);
1370 if (error)
1371 goto out_unlock;
1372
1373 /* No extents at given offset, must be beyond EOF */
1374 if (nmap == 0) {
1375 error = ENXIO;
1376 goto out_unlock;
1377 }
1378
1379 for (i = 0; i < nmap; i++) {
1380 offset = max_t(loff_t, start,
1381 XFS_FSB_TO_B(mp, map[i].br_startoff));
1382
1383 /* Landed in a hole */
1384 if (map[i].br_startblock == HOLESTARTBLOCK)
1385 goto out;
1386
1387 /*
1388 * Landed in an unwritten extent, try to search hole
1389 * from page cache.
1390 */
1391 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1392 if (xfs_find_get_desired_pgoff(inode, &map[i],
1393 HOLE_OFF, &offset))
1394 goto out;
1395 }
1396 }
1397
1398 /*
1399 * map[0] contains data or its unwritten but contains
1400 * data in page cache, probably means that we are
1401 * reading after EOF. We should fix offset to point
1402 * to the end of the file(i.e., there is an implicit
1403 * hole at the end of any file).
1404 */
1405 if (nmap == 1) {
1406 offset = isize;
1407 break;
1408 }
1409
1410 ASSERT(i > 1);
1411
1412 /*
1413 * Both mappings contains data, proceed to the next round of
1414 * search if the current reading offset not beyond or hit EOF.
1415 */
1416 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1417 start = XFS_FSB_TO_B(mp, fsbno);
1418 if (start >= isize) {
1419 offset = isize;
1420 break;
1421 }
1422 }
1423
1424out:
1425 /*
1426 * At this point, we must have found a hole. However, the returned
1427 * offset may be bigger than the file size as it may be aligned to
1428 * page boundary for unwritten extents, we need to deal with this
1429 * situation in particular.
1430 */
1431 offset = min_t(loff_t, offset, isize);
1432 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1433
1434out_unlock:
1435 xfs_iunlock(ip, lock);
1436
1437 if (error)
1438 return -error;
1439 return offset;
1440}
1441
1442STATIC loff_t
1443xfs_file_llseek(
1444 struct file *file,
1445 loff_t offset,
1446 int origin)
1447{
1448 switch (origin) {
1449 case SEEK_END:
1450 case SEEK_CUR:
1451 case SEEK_SET:
1452 return generic_file_llseek(file, offset, origin);
1453 case SEEK_DATA:
1454 return xfs_seek_data(file, offset);
1455 case SEEK_HOLE:
1456 return xfs_seek_hole(file, offset);
1457 default:
1458 return -EINVAL;
1459 }
1460}
1461
1462const struct file_operations xfs_file_operations = {
1463 .llseek = xfs_file_llseek,
1464 .read = do_sync_read,
1465 .write = do_sync_write,
1466 .aio_read = xfs_file_aio_read,
1467 .aio_write = xfs_file_aio_write,
1468 .splice_read = xfs_file_splice_read,
1469 .splice_write = xfs_file_splice_write,
1470 .unlocked_ioctl = xfs_file_ioctl,
1471#ifdef CONFIG_COMPAT
1472 .compat_ioctl = xfs_file_compat_ioctl,
1473#endif
1474 .mmap = xfs_file_mmap,
1475 .open = xfs_file_open,
1476 .release = xfs_file_release,
1477 .fsync = xfs_file_fsync,
1478 .fallocate = xfs_file_fallocate,
1479};
1480
1481const struct file_operations xfs_dir_file_operations = {
1482 .open = xfs_dir_open,
1483 .read = generic_read_dir,
1484 .iterate = xfs_file_readdir,
1485 .llseek = generic_file_llseek,
1486 .unlocked_ioctl = xfs_file_ioctl,
1487#ifdef CONFIG_COMPAT
1488 .compat_ioctl = xfs_file_compat_ioctl,
1489#endif
1490 .fsync = xfs_dir_fsync,
1491};
1492
1493static const struct vm_operations_struct xfs_file_vm_ops = {
1494 .fault = filemap_fault,
1495 .map_pages = filemap_map_pages,
1496 .page_mkwrite = xfs_vm_page_mkwrite,
1497 .remap_pages = generic_file_remap_pages,
1498};