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/falloc.h>
29#include <linux/backing-dev.h>
30#include <linux/mman.h>
31#include <linux/fadvise.h>
32#include <linux/mount.h>
33
34static const struct vm_operations_struct xfs_file_vm_ops;
35
36/*
37 * Decide if the given file range is aligned to the size of the fundamental
38 * allocation unit for the file.
39 */
40static bool
41xfs_is_falloc_aligned(
42 struct xfs_inode *ip,
43 loff_t pos,
44 long long int len)
45{
46 struct xfs_mount *mp = ip->i_mount;
47 uint64_t mask;
48
49 if (XFS_IS_REALTIME_INODE(ip)) {
50 if (!is_power_of_2(mp->m_sb.sb_rextsize)) {
51 u64 rextbytes;
52 u32 mod;
53
54 rextbytes = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize);
55 div_u64_rem(pos, rextbytes, &mod);
56 if (mod)
57 return false;
58 div_u64_rem(len, rextbytes, &mod);
59 return mod == 0;
60 }
61 mask = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize) - 1;
62 } else {
63 mask = mp->m_sb.sb_blocksize - 1;
64 }
65
66 return !((pos | len) & mask);
67}
68
69int
70xfs_update_prealloc_flags(
71 struct xfs_inode *ip,
72 enum xfs_prealloc_flags flags)
73{
74 struct xfs_trans *tp;
75 int error;
76
77 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_writeid,
78 0, 0, 0, &tp);
79 if (error)
80 return error;
81
82 xfs_ilock(ip, XFS_ILOCK_EXCL);
83 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
84
85 if (!(flags & XFS_PREALLOC_INVISIBLE)) {
86 VFS_I(ip)->i_mode &= ~S_ISUID;
87 if (VFS_I(ip)->i_mode & S_IXGRP)
88 VFS_I(ip)->i_mode &= ~S_ISGID;
89 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
90 }
91
92 if (flags & XFS_PREALLOC_SET)
93 ip->i_diflags |= XFS_DIFLAG_PREALLOC;
94 if (flags & XFS_PREALLOC_CLEAR)
95 ip->i_diflags &= ~XFS_DIFLAG_PREALLOC;
96
97 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
98 if (flags & XFS_PREALLOC_SYNC)
99 xfs_trans_set_sync(tp);
100 return xfs_trans_commit(tp);
101}
102
103/*
104 * Fsync operations on directories are much simpler than on regular files,
105 * as there is no file data to flush, and thus also no need for explicit
106 * cache flush operations, and there are no non-transaction metadata updates
107 * on directories either.
108 */
109STATIC int
110xfs_dir_fsync(
111 struct file *file,
112 loff_t start,
113 loff_t end,
114 int datasync)
115{
116 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
117
118 trace_xfs_dir_fsync(ip);
119 return xfs_log_force_inode(ip);
120}
121
122static xfs_csn_t
123xfs_fsync_seq(
124 struct xfs_inode *ip,
125 bool datasync)
126{
127 if (!xfs_ipincount(ip))
128 return 0;
129 if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
130 return 0;
131 return ip->i_itemp->ili_commit_seq;
132}
133
134/*
135 * All metadata updates are logged, which means that we just have to flush the
136 * log up to the latest LSN that touched the inode.
137 *
138 * If we have concurrent fsync/fdatasync() calls, we need them to all block on
139 * the log force before we clear the ili_fsync_fields field. This ensures that
140 * we don't get a racing sync operation that does not wait for the metadata to
141 * hit the journal before returning. If we race with clearing ili_fsync_fields,
142 * then all that will happen is the log force will do nothing as the lsn will
143 * already be on disk. We can't race with setting ili_fsync_fields because that
144 * is done under XFS_ILOCK_EXCL, and that can't happen because we hold the lock
145 * shared until after the ili_fsync_fields is cleared.
146 */
147static int
148xfs_fsync_flush_log(
149 struct xfs_inode *ip,
150 bool datasync,
151 int *log_flushed)
152{
153 int error = 0;
154 xfs_csn_t seq;
155
156 xfs_ilock(ip, XFS_ILOCK_SHARED);
157 seq = xfs_fsync_seq(ip, datasync);
158 if (seq) {
159 error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
160 log_flushed);
161
162 spin_lock(&ip->i_itemp->ili_lock);
163 ip->i_itemp->ili_fsync_fields = 0;
164 spin_unlock(&ip->i_itemp->ili_lock);
165 }
166 xfs_iunlock(ip, XFS_ILOCK_SHARED);
167 return error;
168}
169
170STATIC int
171xfs_file_fsync(
172 struct file *file,
173 loff_t start,
174 loff_t end,
175 int datasync)
176{
177 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
178 struct xfs_mount *mp = ip->i_mount;
179 int error = 0;
180 int log_flushed = 0;
181
182 trace_xfs_file_fsync(ip);
183
184 error = file_write_and_wait_range(file, start, end);
185 if (error)
186 return error;
187
188 if (XFS_FORCED_SHUTDOWN(mp))
189 return -EIO;
190
191 xfs_iflags_clear(ip, XFS_ITRUNCATED);
192
193 /*
194 * If we have an RT and/or log subvolume we need to make sure to flush
195 * the write cache the device used for file data first. This is to
196 * ensure newly written file data make it to disk before logging the new
197 * inode size in case of an extending write.
198 */
199 if (XFS_IS_REALTIME_INODE(ip))
200 blkdev_issue_flush(mp->m_rtdev_targp->bt_bdev);
201 else if (mp->m_logdev_targp != mp->m_ddev_targp)
202 blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
203
204 /*
205 * Any inode that has dirty modifications in the log is pinned. The
206 * racy check here for a pinned inode while not catch modifications
207 * that happen concurrently to the fsync call, but fsync semantics
208 * only require to sync previously completed I/O.
209 */
210 if (xfs_ipincount(ip))
211 error = xfs_fsync_flush_log(ip, datasync, &log_flushed);
212
213 /*
214 * If we only have a single device, and the log force about was
215 * a no-op we might have to flush the data device cache here.
216 * This can only happen for fdatasync/O_DSYNC if we were overwriting
217 * an already allocated file and thus do not have any metadata to
218 * commit.
219 */
220 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
221 mp->m_logdev_targp == mp->m_ddev_targp)
222 blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
223
224 return error;
225}
226
227static int
228xfs_ilock_iocb(
229 struct kiocb *iocb,
230 unsigned int lock_mode)
231{
232 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
233
234 if (iocb->ki_flags & IOCB_NOWAIT) {
235 if (!xfs_ilock_nowait(ip, lock_mode))
236 return -EAGAIN;
237 } else {
238 xfs_ilock(ip, lock_mode);
239 }
240
241 return 0;
242}
243
244STATIC ssize_t
245xfs_file_dio_read(
246 struct kiocb *iocb,
247 struct iov_iter *to)
248{
249 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
250 ssize_t ret;
251
252 trace_xfs_file_direct_read(iocb, to);
253
254 if (!iov_iter_count(to))
255 return 0; /* skip atime */
256
257 file_accessed(iocb->ki_filp);
258
259 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
260 if (ret)
261 return ret;
262 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0);
263 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
264
265 return ret;
266}
267
268static noinline ssize_t
269xfs_file_dax_read(
270 struct kiocb *iocb,
271 struct iov_iter *to)
272{
273 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
274 ssize_t ret = 0;
275
276 trace_xfs_file_dax_read(iocb, to);
277
278 if (!iov_iter_count(to))
279 return 0; /* skip atime */
280
281 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
282 if (ret)
283 return ret;
284 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
285 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
286
287 file_accessed(iocb->ki_filp);
288 return ret;
289}
290
291STATIC ssize_t
292xfs_file_buffered_read(
293 struct kiocb *iocb,
294 struct iov_iter *to)
295{
296 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
297 ssize_t ret;
298
299 trace_xfs_file_buffered_read(iocb, to);
300
301 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
302 if (ret)
303 return ret;
304 ret = generic_file_read_iter(iocb, to);
305 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
306
307 return ret;
308}
309
310STATIC ssize_t
311xfs_file_read_iter(
312 struct kiocb *iocb,
313 struct iov_iter *to)
314{
315 struct inode *inode = file_inode(iocb->ki_filp);
316 struct xfs_mount *mp = XFS_I(inode)->i_mount;
317 ssize_t ret = 0;
318
319 XFS_STATS_INC(mp, xs_read_calls);
320
321 if (XFS_FORCED_SHUTDOWN(mp))
322 return -EIO;
323
324 if (IS_DAX(inode))
325 ret = xfs_file_dax_read(iocb, to);
326 else if (iocb->ki_flags & IOCB_DIRECT)
327 ret = xfs_file_dio_read(iocb, to);
328 else
329 ret = xfs_file_buffered_read(iocb, to);
330
331 if (ret > 0)
332 XFS_STATS_ADD(mp, xs_read_bytes, ret);
333 return ret;
334}
335
336/*
337 * Common pre-write limit and setup checks.
338 *
339 * Called with the iolocked held either shared and exclusive according to
340 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
341 * if called for a direct write beyond i_size.
342 */
343STATIC ssize_t
344xfs_file_write_checks(
345 struct kiocb *iocb,
346 struct iov_iter *from,
347 int *iolock)
348{
349 struct file *file = iocb->ki_filp;
350 struct inode *inode = file->f_mapping->host;
351 struct xfs_inode *ip = XFS_I(inode);
352 ssize_t error = 0;
353 size_t count = iov_iter_count(from);
354 bool drained_dio = false;
355 loff_t isize;
356
357restart:
358 error = generic_write_checks(iocb, from);
359 if (error <= 0)
360 return error;
361
362 if (iocb->ki_flags & IOCB_NOWAIT) {
363 error = break_layout(inode, false);
364 if (error == -EWOULDBLOCK)
365 error = -EAGAIN;
366 } else {
367 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
368 }
369
370 if (error)
371 return error;
372
373 /*
374 * For changing security info in file_remove_privs() we need i_rwsem
375 * exclusively.
376 */
377 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
378 xfs_iunlock(ip, *iolock);
379 *iolock = XFS_IOLOCK_EXCL;
380 error = xfs_ilock_iocb(iocb, *iolock);
381 if (error) {
382 *iolock = 0;
383 return error;
384 }
385 goto restart;
386 }
387
388 /*
389 * If the offset is beyond the size of the file, we need to zero any
390 * blocks that fall between the existing EOF and the start of this
391 * write. If zeroing is needed and we are currently holding the iolock
392 * shared, we need to update it to exclusive which implies having to
393 * redo all checks before.
394 *
395 * We need to serialise against EOF updates that occur in IO completions
396 * here. We want to make sure that nobody is changing the size while we
397 * do this check until we have placed an IO barrier (i.e. hold the
398 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
399 * spinlock effectively forms a memory barrier once we have the
400 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
401 * hence be able to correctly determine if we need to run zeroing.
402 *
403 * We can do an unlocked check here safely as IO completion can only
404 * extend EOF. Truncate is locked out at this point, so the EOF can
405 * not move backwards, only forwards. Hence we only need to take the
406 * slow path and spin locks when we are at or beyond the current EOF.
407 */
408 if (iocb->ki_pos <= i_size_read(inode))
409 goto out;
410
411 spin_lock(&ip->i_flags_lock);
412 isize = i_size_read(inode);
413 if (iocb->ki_pos > isize) {
414 spin_unlock(&ip->i_flags_lock);
415
416 if (iocb->ki_flags & IOCB_NOWAIT)
417 return -EAGAIN;
418
419 if (!drained_dio) {
420 if (*iolock == XFS_IOLOCK_SHARED) {
421 xfs_iunlock(ip, *iolock);
422 *iolock = XFS_IOLOCK_EXCL;
423 xfs_ilock(ip, *iolock);
424 iov_iter_reexpand(from, count);
425 }
426 /*
427 * We now have an IO submission barrier in place, but
428 * AIO can do EOF updates during IO completion and hence
429 * we now need to wait for all of them to drain. Non-AIO
430 * DIO will have drained before we are given the
431 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
432 * no-op.
433 */
434 inode_dio_wait(inode);
435 drained_dio = true;
436 goto restart;
437 }
438
439 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
440 error = iomap_zero_range(inode, isize, iocb->ki_pos - isize,
441 NULL, &xfs_buffered_write_iomap_ops);
442 if (error)
443 return error;
444 } else
445 spin_unlock(&ip->i_flags_lock);
446
447out:
448 return file_modified(file);
449}
450
451static int
452xfs_dio_write_end_io(
453 struct kiocb *iocb,
454 ssize_t size,
455 int error,
456 unsigned flags)
457{
458 struct inode *inode = file_inode(iocb->ki_filp);
459 struct xfs_inode *ip = XFS_I(inode);
460 loff_t offset = iocb->ki_pos;
461 unsigned int nofs_flag;
462
463 trace_xfs_end_io_direct_write(ip, offset, size);
464
465 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
466 return -EIO;
467
468 if (error)
469 return error;
470 if (!size)
471 return 0;
472
473 /*
474 * Capture amount written on completion as we can't reliably account
475 * for it on submission.
476 */
477 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
478
479 /*
480 * We can allocate memory here while doing writeback on behalf of
481 * memory reclaim. To avoid memory allocation deadlocks set the
482 * task-wide nofs context for the following operations.
483 */
484 nofs_flag = memalloc_nofs_save();
485
486 if (flags & IOMAP_DIO_COW) {
487 error = xfs_reflink_end_cow(ip, offset, size);
488 if (error)
489 goto out;
490 }
491
492 /*
493 * Unwritten conversion updates the in-core isize after extent
494 * conversion but before updating the on-disk size. Updating isize any
495 * earlier allows a racing dio read to find unwritten extents before
496 * they are converted.
497 */
498 if (flags & IOMAP_DIO_UNWRITTEN) {
499 error = xfs_iomap_write_unwritten(ip, offset, size, true);
500 goto out;
501 }
502
503 /*
504 * We need to update the in-core inode size here so that we don't end up
505 * with the on-disk inode size being outside the in-core inode size. We
506 * have no other method of updating EOF for AIO, so always do it here
507 * if necessary.
508 *
509 * We need to lock the test/set EOF update as we can be racing with
510 * other IO completions here to update the EOF. Failing to serialise
511 * here can result in EOF moving backwards and Bad Things Happen when
512 * that occurs.
513 *
514 * As IO completion only ever extends EOF, we can do an unlocked check
515 * here to avoid taking the spinlock. If we land within the current EOF,
516 * then we do not need to do an extending update at all, and we don't
517 * need to take the lock to check this. If we race with an update moving
518 * EOF, then we'll either still be beyond EOF and need to take the lock,
519 * or we'll be within EOF and we don't need to take it at all.
520 */
521 if (offset + size <= i_size_read(inode))
522 goto out;
523
524 spin_lock(&ip->i_flags_lock);
525 if (offset + size > i_size_read(inode)) {
526 i_size_write(inode, offset + size);
527 spin_unlock(&ip->i_flags_lock);
528 error = xfs_setfilesize(ip, offset, size);
529 } else {
530 spin_unlock(&ip->i_flags_lock);
531 }
532
533out:
534 memalloc_nofs_restore(nofs_flag);
535 return error;
536}
537
538static const struct iomap_dio_ops xfs_dio_write_ops = {
539 .end_io = xfs_dio_write_end_io,
540};
541
542/*
543 * Handle block aligned direct I/O writes
544 */
545static noinline ssize_t
546xfs_file_dio_write_aligned(
547 struct xfs_inode *ip,
548 struct kiocb *iocb,
549 struct iov_iter *from)
550{
551 int iolock = XFS_IOLOCK_SHARED;
552 ssize_t ret;
553
554 ret = xfs_ilock_iocb(iocb, iolock);
555 if (ret)
556 return ret;
557 ret = xfs_file_write_checks(iocb, from, &iolock);
558 if (ret)
559 goto out_unlock;
560
561 /*
562 * We don't need to hold the IOLOCK exclusively across the IO, so demote
563 * the iolock back to shared if we had to take the exclusive lock in
564 * xfs_file_write_checks() for other reasons.
565 */
566 if (iolock == XFS_IOLOCK_EXCL) {
567 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
568 iolock = XFS_IOLOCK_SHARED;
569 }
570 trace_xfs_file_direct_write(iocb, from);
571 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
572 &xfs_dio_write_ops, 0);
573out_unlock:
574 if (iolock)
575 xfs_iunlock(ip, iolock);
576 return ret;
577}
578
579/*
580 * Handle block unaligned direct I/O writes
581 *
582 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
583 * them to be done in parallel with reads and other direct I/O writes. However,
584 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
585 * to do sub-block zeroing and that requires serialisation against other direct
586 * I/O to the same block. In this case we need to serialise the submission of
587 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
588 * In the case where sub-block zeroing is not required, we can do concurrent
589 * sub-block dios to the same block successfully.
590 *
591 * Optimistically submit the I/O using the shared lock first, but use the
592 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
593 * if block allocation or partial block zeroing would be required. In that case
594 * we try again with the exclusive lock.
595 */
596static noinline ssize_t
597xfs_file_dio_write_unaligned(
598 struct xfs_inode *ip,
599 struct kiocb *iocb,
600 struct iov_iter *from)
601{
602 size_t isize = i_size_read(VFS_I(ip));
603 size_t count = iov_iter_count(from);
604 int iolock = XFS_IOLOCK_SHARED;
605 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
606 ssize_t ret;
607
608 /*
609 * Extending writes need exclusivity because of the sub-block zeroing
610 * that the DIO code always does for partial tail blocks beyond EOF, so
611 * don't even bother trying the fast path in this case.
612 */
613 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
614retry_exclusive:
615 if (iocb->ki_flags & IOCB_NOWAIT)
616 return -EAGAIN;
617 iolock = XFS_IOLOCK_EXCL;
618 flags = IOMAP_DIO_FORCE_WAIT;
619 }
620
621 ret = xfs_ilock_iocb(iocb, iolock);
622 if (ret)
623 return ret;
624
625 /*
626 * We can't properly handle unaligned direct I/O to reflink files yet,
627 * as we can't unshare a partial block.
628 */
629 if (xfs_is_cow_inode(ip)) {
630 trace_xfs_reflink_bounce_dio_write(iocb, from);
631 ret = -ENOTBLK;
632 goto out_unlock;
633 }
634
635 ret = xfs_file_write_checks(iocb, from, &iolock);
636 if (ret)
637 goto out_unlock;
638
639 /*
640 * If we are doing exclusive unaligned I/O, this must be the only I/O
641 * in-flight. Otherwise we risk data corruption due to unwritten extent
642 * conversions from the AIO end_io handler. Wait for all other I/O to
643 * drain first.
644 */
645 if (flags & IOMAP_DIO_FORCE_WAIT)
646 inode_dio_wait(VFS_I(ip));
647
648 trace_xfs_file_direct_write(iocb, from);
649 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
650 &xfs_dio_write_ops, flags);
651
652 /*
653 * Retry unaligned I/O with exclusive blocking semantics if the DIO
654 * layer rejected it for mapping or locking reasons. If we are doing
655 * nonblocking user I/O, propagate the error.
656 */
657 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
658 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
659 xfs_iunlock(ip, iolock);
660 goto retry_exclusive;
661 }
662
663out_unlock:
664 if (iolock)
665 xfs_iunlock(ip, iolock);
666 return ret;
667}
668
669static ssize_t
670xfs_file_dio_write(
671 struct kiocb *iocb,
672 struct iov_iter *from)
673{
674 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
675 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
676 size_t count = iov_iter_count(from);
677
678 /* direct I/O must be aligned to device logical sector size */
679 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
680 return -EINVAL;
681 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
682 return xfs_file_dio_write_unaligned(ip, iocb, from);
683 return xfs_file_dio_write_aligned(ip, iocb, from);
684}
685
686static noinline ssize_t
687xfs_file_dax_write(
688 struct kiocb *iocb,
689 struct iov_iter *from)
690{
691 struct inode *inode = iocb->ki_filp->f_mapping->host;
692 struct xfs_inode *ip = XFS_I(inode);
693 int iolock = XFS_IOLOCK_EXCL;
694 ssize_t ret, error = 0;
695 loff_t pos;
696
697 ret = xfs_ilock_iocb(iocb, iolock);
698 if (ret)
699 return ret;
700 ret = xfs_file_write_checks(iocb, from, &iolock);
701 if (ret)
702 goto out;
703
704 pos = iocb->ki_pos;
705
706 trace_xfs_file_dax_write(iocb, from);
707 ret = dax_iomap_rw(iocb, from, &xfs_direct_write_iomap_ops);
708 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
709 i_size_write(inode, iocb->ki_pos);
710 error = xfs_setfilesize(ip, pos, ret);
711 }
712out:
713 if (iolock)
714 xfs_iunlock(ip, iolock);
715 if (error)
716 return error;
717
718 if (ret > 0) {
719 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
720
721 /* Handle various SYNC-type writes */
722 ret = generic_write_sync(iocb, ret);
723 }
724 return ret;
725}
726
727STATIC ssize_t
728xfs_file_buffered_write(
729 struct kiocb *iocb,
730 struct iov_iter *from)
731{
732 struct file *file = iocb->ki_filp;
733 struct address_space *mapping = file->f_mapping;
734 struct inode *inode = mapping->host;
735 struct xfs_inode *ip = XFS_I(inode);
736 ssize_t ret;
737 bool cleared_space = false;
738 int iolock;
739
740 if (iocb->ki_flags & IOCB_NOWAIT)
741 return -EOPNOTSUPP;
742
743write_retry:
744 iolock = XFS_IOLOCK_EXCL;
745 xfs_ilock(ip, iolock);
746
747 ret = xfs_file_write_checks(iocb, from, &iolock);
748 if (ret)
749 goto out;
750
751 /* We can write back this queue in page reclaim */
752 current->backing_dev_info = inode_to_bdi(inode);
753
754 trace_xfs_file_buffered_write(iocb, from);
755 ret = iomap_file_buffered_write(iocb, from,
756 &xfs_buffered_write_iomap_ops);
757 if (likely(ret >= 0))
758 iocb->ki_pos += ret;
759
760 /*
761 * If we hit a space limit, try to free up some lingering preallocated
762 * space before returning an error. In the case of ENOSPC, first try to
763 * write back all dirty inodes to free up some of the excess reserved
764 * metadata space. This reduces the chances that the eofblocks scan
765 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
766 * also behaves as a filter to prevent too many eofblocks scans from
767 * running at the same time. Use a synchronous scan to increase the
768 * effectiveness of the scan.
769 */
770 if (ret == -EDQUOT && !cleared_space) {
771 xfs_iunlock(ip, iolock);
772 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
773 cleared_space = true;
774 goto write_retry;
775 } else if (ret == -ENOSPC && !cleared_space) {
776 struct xfs_icwalk icw = {0};
777
778 cleared_space = true;
779 xfs_flush_inodes(ip->i_mount);
780
781 xfs_iunlock(ip, iolock);
782 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
783 xfs_blockgc_free_space(ip->i_mount, &icw);
784 goto write_retry;
785 }
786
787 current->backing_dev_info = NULL;
788out:
789 if (iolock)
790 xfs_iunlock(ip, iolock);
791
792 if (ret > 0) {
793 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
794 /* Handle various SYNC-type writes */
795 ret = generic_write_sync(iocb, ret);
796 }
797 return ret;
798}
799
800STATIC ssize_t
801xfs_file_write_iter(
802 struct kiocb *iocb,
803 struct iov_iter *from)
804{
805 struct file *file = iocb->ki_filp;
806 struct address_space *mapping = file->f_mapping;
807 struct inode *inode = mapping->host;
808 struct xfs_inode *ip = XFS_I(inode);
809 ssize_t ret;
810 size_t ocount = iov_iter_count(from);
811
812 XFS_STATS_INC(ip->i_mount, xs_write_calls);
813
814 if (ocount == 0)
815 return 0;
816
817 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
818 return -EIO;
819
820 if (IS_DAX(inode))
821 return xfs_file_dax_write(iocb, from);
822
823 if (iocb->ki_flags & IOCB_DIRECT) {
824 /*
825 * Allow a directio write to fall back to a buffered
826 * write *only* in the case that we're doing a reflink
827 * CoW. In all other directio scenarios we do not
828 * allow an operation to fall back to buffered mode.
829 */
830 ret = xfs_file_dio_write(iocb, from);
831 if (ret != -ENOTBLK)
832 return ret;
833 }
834
835 return xfs_file_buffered_write(iocb, from);
836}
837
838static void
839xfs_wait_dax_page(
840 struct inode *inode)
841{
842 struct xfs_inode *ip = XFS_I(inode);
843
844 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
845 schedule();
846 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
847}
848
849static int
850xfs_break_dax_layouts(
851 struct inode *inode,
852 bool *retry)
853{
854 struct page *page;
855
856 ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
857
858 page = dax_layout_busy_page(inode->i_mapping);
859 if (!page)
860 return 0;
861
862 *retry = true;
863 return ___wait_var_event(&page->_refcount,
864 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
865 0, 0, xfs_wait_dax_page(inode));
866}
867
868int
869xfs_break_layouts(
870 struct inode *inode,
871 uint *iolock,
872 enum layout_break_reason reason)
873{
874 bool retry;
875 int error;
876
877 ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
878
879 do {
880 retry = false;
881 switch (reason) {
882 case BREAK_UNMAP:
883 error = xfs_break_dax_layouts(inode, &retry);
884 if (error || retry)
885 break;
886 fallthrough;
887 case BREAK_WRITE:
888 error = xfs_break_leased_layouts(inode, iolock, &retry);
889 break;
890 default:
891 WARN_ON_ONCE(1);
892 error = -EINVAL;
893 }
894 } while (error == 0 && retry);
895
896 return error;
897}
898
899#define XFS_FALLOC_FL_SUPPORTED \
900 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
901 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
902 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
903
904STATIC long
905xfs_file_fallocate(
906 struct file *file,
907 int mode,
908 loff_t offset,
909 loff_t len)
910{
911 struct inode *inode = file_inode(file);
912 struct xfs_inode *ip = XFS_I(inode);
913 long error;
914 enum xfs_prealloc_flags flags = 0;
915 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
916 loff_t new_size = 0;
917 bool do_file_insert = false;
918
919 if (!S_ISREG(inode->i_mode))
920 return -EINVAL;
921 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
922 return -EOPNOTSUPP;
923
924 xfs_ilock(ip, iolock);
925 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
926 if (error)
927 goto out_unlock;
928
929 /*
930 * Must wait for all AIO to complete before we continue as AIO can
931 * change the file size on completion without holding any locks we
932 * currently hold. We must do this first because AIO can update both
933 * the on disk and in memory inode sizes, and the operations that follow
934 * require the in-memory size to be fully up-to-date.
935 */
936 inode_dio_wait(inode);
937
938 /*
939 * Now AIO and DIO has drained we flush and (if necessary) invalidate
940 * the cached range over the first operation we are about to run.
941 *
942 * We care about zero and collapse here because they both run a hole
943 * punch over the range first. Because that can zero data, and the range
944 * of invalidation for the shift operations is much larger, we still do
945 * the required flush for collapse in xfs_prepare_shift().
946 *
947 * Insert has the same range requirements as collapse, and we extend the
948 * file first which can zero data. Hence insert has the same
949 * flush/invalidate requirements as collapse and so they are both
950 * handled at the right time by xfs_prepare_shift().
951 */
952 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
953 FALLOC_FL_COLLAPSE_RANGE)) {
954 error = xfs_flush_unmap_range(ip, offset, len);
955 if (error)
956 goto out_unlock;
957 }
958
959 if (mode & FALLOC_FL_PUNCH_HOLE) {
960 error = xfs_free_file_space(ip, offset, len);
961 if (error)
962 goto out_unlock;
963 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
964 if (!xfs_is_falloc_aligned(ip, offset, len)) {
965 error = -EINVAL;
966 goto out_unlock;
967 }
968
969 /*
970 * There is no need to overlap collapse range with EOF,
971 * in which case it is effectively a truncate operation
972 */
973 if (offset + len >= i_size_read(inode)) {
974 error = -EINVAL;
975 goto out_unlock;
976 }
977
978 new_size = i_size_read(inode) - len;
979
980 error = xfs_collapse_file_space(ip, offset, len);
981 if (error)
982 goto out_unlock;
983 } else if (mode & FALLOC_FL_INSERT_RANGE) {
984 loff_t isize = i_size_read(inode);
985
986 if (!xfs_is_falloc_aligned(ip, offset, len)) {
987 error = -EINVAL;
988 goto out_unlock;
989 }
990
991 /*
992 * New inode size must not exceed ->s_maxbytes, accounting for
993 * possible signed overflow.
994 */
995 if (inode->i_sb->s_maxbytes - isize < len) {
996 error = -EFBIG;
997 goto out_unlock;
998 }
999 new_size = isize + len;
1000
1001 /* Offset should be less than i_size */
1002 if (offset >= isize) {
1003 error = -EINVAL;
1004 goto out_unlock;
1005 }
1006 do_file_insert = true;
1007 } else {
1008 flags |= XFS_PREALLOC_SET;
1009
1010 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
1011 offset + len > i_size_read(inode)) {
1012 new_size = offset + len;
1013 error = inode_newsize_ok(inode, new_size);
1014 if (error)
1015 goto out_unlock;
1016 }
1017
1018 if (mode & FALLOC_FL_ZERO_RANGE) {
1019 /*
1020 * Punch a hole and prealloc the range. We use a hole
1021 * punch rather than unwritten extent conversion for two
1022 * reasons:
1023 *
1024 * 1.) Hole punch handles partial block zeroing for us.
1025 * 2.) If prealloc returns ENOSPC, the file range is
1026 * still zero-valued by virtue of the hole punch.
1027 */
1028 unsigned int blksize = i_blocksize(inode);
1029
1030 trace_xfs_zero_file_space(ip);
1031
1032 error = xfs_free_file_space(ip, offset, len);
1033 if (error)
1034 goto out_unlock;
1035
1036 len = round_up(offset + len, blksize) -
1037 round_down(offset, blksize);
1038 offset = round_down(offset, blksize);
1039 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
1040 error = xfs_reflink_unshare(ip, offset, len);
1041 if (error)
1042 goto out_unlock;
1043 } else {
1044 /*
1045 * If always_cow mode we can't use preallocations and
1046 * thus should not create them.
1047 */
1048 if (xfs_is_always_cow_inode(ip)) {
1049 error = -EOPNOTSUPP;
1050 goto out_unlock;
1051 }
1052 }
1053
1054 if (!xfs_is_always_cow_inode(ip)) {
1055 error = xfs_alloc_file_space(ip, offset, len,
1056 XFS_BMAPI_PREALLOC);
1057 if (error)
1058 goto out_unlock;
1059 }
1060 }
1061
1062 if (file->f_flags & O_DSYNC)
1063 flags |= XFS_PREALLOC_SYNC;
1064
1065 error = xfs_update_prealloc_flags(ip, flags);
1066 if (error)
1067 goto out_unlock;
1068
1069 /* Change file size if needed */
1070 if (new_size) {
1071 struct iattr iattr;
1072
1073 iattr.ia_valid = ATTR_SIZE;
1074 iattr.ia_size = new_size;
1075 error = xfs_vn_setattr_size(file_mnt_user_ns(file),
1076 file_dentry(file), &iattr);
1077 if (error)
1078 goto out_unlock;
1079 }
1080
1081 /*
1082 * Perform hole insertion now that the file size has been
1083 * updated so that if we crash during the operation we don't
1084 * leave shifted extents past EOF and hence losing access to
1085 * the data that is contained within them.
1086 */
1087 if (do_file_insert)
1088 error = xfs_insert_file_space(ip, offset, len);
1089
1090out_unlock:
1091 xfs_iunlock(ip, iolock);
1092 return error;
1093}
1094
1095STATIC int
1096xfs_file_fadvise(
1097 struct file *file,
1098 loff_t start,
1099 loff_t end,
1100 int advice)
1101{
1102 struct xfs_inode *ip = XFS_I(file_inode(file));
1103 int ret;
1104 int lockflags = 0;
1105
1106 /*
1107 * Operations creating pages in page cache need protection from hole
1108 * punching and similar ops
1109 */
1110 if (advice == POSIX_FADV_WILLNEED) {
1111 lockflags = XFS_IOLOCK_SHARED;
1112 xfs_ilock(ip, lockflags);
1113 }
1114 ret = generic_fadvise(file, start, end, advice);
1115 if (lockflags)
1116 xfs_iunlock(ip, lockflags);
1117 return ret;
1118}
1119
1120/* Does this file, inode, or mount want synchronous writes? */
1121static inline bool xfs_file_sync_writes(struct file *filp)
1122{
1123 struct xfs_inode *ip = XFS_I(file_inode(filp));
1124
1125 if (ip->i_mount->m_flags & XFS_MOUNT_WSYNC)
1126 return true;
1127 if (filp->f_flags & (__O_SYNC | O_DSYNC))
1128 return true;
1129 if (IS_SYNC(file_inode(filp)))
1130 return true;
1131
1132 return false;
1133}
1134
1135STATIC loff_t
1136xfs_file_remap_range(
1137 struct file *file_in,
1138 loff_t pos_in,
1139 struct file *file_out,
1140 loff_t pos_out,
1141 loff_t len,
1142 unsigned int remap_flags)
1143{
1144 struct inode *inode_in = file_inode(file_in);
1145 struct xfs_inode *src = XFS_I(inode_in);
1146 struct inode *inode_out = file_inode(file_out);
1147 struct xfs_inode *dest = XFS_I(inode_out);
1148 struct xfs_mount *mp = src->i_mount;
1149 loff_t remapped = 0;
1150 xfs_extlen_t cowextsize;
1151 int ret;
1152
1153 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1154 return -EINVAL;
1155
1156 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1157 return -EOPNOTSUPP;
1158
1159 if (XFS_FORCED_SHUTDOWN(mp))
1160 return -EIO;
1161
1162 /* Prepare and then clone file data. */
1163 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1164 &len, remap_flags);
1165 if (ret || len == 0)
1166 return ret;
1167
1168 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1169
1170 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1171 &remapped);
1172 if (ret)
1173 goto out_unlock;
1174
1175 /*
1176 * Carry the cowextsize hint from src to dest if we're sharing the
1177 * entire source file to the entire destination file, the source file
1178 * has a cowextsize hint, and the destination file does not.
1179 */
1180 cowextsize = 0;
1181 if (pos_in == 0 && len == i_size_read(inode_in) &&
1182 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1183 pos_out == 0 && len >= i_size_read(inode_out) &&
1184 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1185 cowextsize = src->i_cowextsize;
1186
1187 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1188 remap_flags);
1189 if (ret)
1190 goto out_unlock;
1191
1192 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1193 xfs_log_force_inode(dest);
1194out_unlock:
1195 xfs_iunlock2_io_mmap(src, dest);
1196 if (ret)
1197 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1198 return remapped > 0 ? remapped : ret;
1199}
1200
1201STATIC int
1202xfs_file_open(
1203 struct inode *inode,
1204 struct file *file)
1205{
1206 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
1207 return -EFBIG;
1208 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
1209 return -EIO;
1210 file->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
1211 return 0;
1212}
1213
1214STATIC int
1215xfs_dir_open(
1216 struct inode *inode,
1217 struct file *file)
1218{
1219 struct xfs_inode *ip = XFS_I(inode);
1220 int mode;
1221 int error;
1222
1223 error = xfs_file_open(inode, file);
1224 if (error)
1225 return error;
1226
1227 /*
1228 * If there are any blocks, read-ahead block 0 as we're almost
1229 * certain to have the next operation be a read there.
1230 */
1231 mode = xfs_ilock_data_map_shared(ip);
1232 if (ip->i_df.if_nextents > 0)
1233 error = xfs_dir3_data_readahead(ip, 0, 0);
1234 xfs_iunlock(ip, mode);
1235 return error;
1236}
1237
1238STATIC int
1239xfs_file_release(
1240 struct inode *inode,
1241 struct file *filp)
1242{
1243 return xfs_release(XFS_I(inode));
1244}
1245
1246STATIC int
1247xfs_file_readdir(
1248 struct file *file,
1249 struct dir_context *ctx)
1250{
1251 struct inode *inode = file_inode(file);
1252 xfs_inode_t *ip = XFS_I(inode);
1253 size_t bufsize;
1254
1255 /*
1256 * The Linux API doesn't pass down the total size of the buffer
1257 * we read into down to the filesystem. With the filldir concept
1258 * it's not needed for correct information, but the XFS dir2 leaf
1259 * code wants an estimate of the buffer size to calculate it's
1260 * readahead window and size the buffers used for mapping to
1261 * physical blocks.
1262 *
1263 * Try to give it an estimate that's good enough, maybe at some
1264 * point we can change the ->readdir prototype to include the
1265 * buffer size. For now we use the current glibc buffer size.
1266 */
1267 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1268
1269 return xfs_readdir(NULL, ip, ctx, bufsize);
1270}
1271
1272STATIC loff_t
1273xfs_file_llseek(
1274 struct file *file,
1275 loff_t offset,
1276 int whence)
1277{
1278 struct inode *inode = file->f_mapping->host;
1279
1280 if (XFS_FORCED_SHUTDOWN(XFS_I(inode)->i_mount))
1281 return -EIO;
1282
1283 switch (whence) {
1284 default:
1285 return generic_file_llseek(file, offset, whence);
1286 case SEEK_HOLE:
1287 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1288 break;
1289 case SEEK_DATA:
1290 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1291 break;
1292 }
1293
1294 if (offset < 0)
1295 return offset;
1296 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1297}
1298
1299/*
1300 * Locking for serialisation of IO during page faults. This results in a lock
1301 * ordering of:
1302 *
1303 * mmap_lock (MM)
1304 * sb_start_pagefault(vfs, freeze)
1305 * i_mmaplock (XFS - truncate serialisation)
1306 * page_lock (MM)
1307 * i_lock (XFS - extent map serialisation)
1308 */
1309static vm_fault_t
1310__xfs_filemap_fault(
1311 struct vm_fault *vmf,
1312 enum page_entry_size pe_size,
1313 bool write_fault)
1314{
1315 struct inode *inode = file_inode(vmf->vma->vm_file);
1316 struct xfs_inode *ip = XFS_I(inode);
1317 vm_fault_t ret;
1318
1319 trace_xfs_filemap_fault(ip, pe_size, write_fault);
1320
1321 if (write_fault) {
1322 sb_start_pagefault(inode->i_sb);
1323 file_update_time(vmf->vma->vm_file);
1324 }
1325
1326 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1327 if (IS_DAX(inode)) {
1328 pfn_t pfn;
1329
1330 ret = dax_iomap_fault(vmf, pe_size, &pfn, NULL,
1331 (write_fault && !vmf->cow_page) ?
1332 &xfs_direct_write_iomap_ops :
1333 &xfs_read_iomap_ops);
1334 if (ret & VM_FAULT_NEEDDSYNC)
1335 ret = dax_finish_sync_fault(vmf, pe_size, pfn);
1336 } else {
1337 if (write_fault)
1338 ret = iomap_page_mkwrite(vmf,
1339 &xfs_buffered_write_iomap_ops);
1340 else
1341 ret = filemap_fault(vmf);
1342 }
1343 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1344
1345 if (write_fault)
1346 sb_end_pagefault(inode->i_sb);
1347 return ret;
1348}
1349
1350static inline bool
1351xfs_is_write_fault(
1352 struct vm_fault *vmf)
1353{
1354 return (vmf->flags & FAULT_FLAG_WRITE) &&
1355 (vmf->vma->vm_flags & VM_SHARED);
1356}
1357
1358static vm_fault_t
1359xfs_filemap_fault(
1360 struct vm_fault *vmf)
1361{
1362 /* DAX can shortcut the normal fault path on write faults! */
1363 return __xfs_filemap_fault(vmf, PE_SIZE_PTE,
1364 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1365 xfs_is_write_fault(vmf));
1366}
1367
1368static vm_fault_t
1369xfs_filemap_huge_fault(
1370 struct vm_fault *vmf,
1371 enum page_entry_size pe_size)
1372{
1373 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1374 return VM_FAULT_FALLBACK;
1375
1376 /* DAX can shortcut the normal fault path on write faults! */
1377 return __xfs_filemap_fault(vmf, pe_size,
1378 xfs_is_write_fault(vmf));
1379}
1380
1381static vm_fault_t
1382xfs_filemap_page_mkwrite(
1383 struct vm_fault *vmf)
1384{
1385 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1386}
1387
1388/*
1389 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1390 * on write faults. In reality, it needs to serialise against truncate and
1391 * prepare memory for writing so handle is as standard write fault.
1392 */
1393static vm_fault_t
1394xfs_filemap_pfn_mkwrite(
1395 struct vm_fault *vmf)
1396{
1397
1398 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1399}
1400
1401static vm_fault_t
1402xfs_filemap_map_pages(
1403 struct vm_fault *vmf,
1404 pgoff_t start_pgoff,
1405 pgoff_t end_pgoff)
1406{
1407 struct inode *inode = file_inode(vmf->vma->vm_file);
1408 vm_fault_t ret;
1409
1410 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1411 ret = filemap_map_pages(vmf, start_pgoff, end_pgoff);
1412 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1413 return ret;
1414}
1415
1416static const struct vm_operations_struct xfs_file_vm_ops = {
1417 .fault = xfs_filemap_fault,
1418 .huge_fault = xfs_filemap_huge_fault,
1419 .map_pages = xfs_filemap_map_pages,
1420 .page_mkwrite = xfs_filemap_page_mkwrite,
1421 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1422};
1423
1424STATIC int
1425xfs_file_mmap(
1426 struct file *file,
1427 struct vm_area_struct *vma)
1428{
1429 struct inode *inode = file_inode(file);
1430 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1431
1432 /*
1433 * We don't support synchronous mappings for non-DAX files and
1434 * for DAX files if underneath dax_device is not synchronous.
1435 */
1436 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1437 return -EOPNOTSUPP;
1438
1439 file_accessed(file);
1440 vma->vm_ops = &xfs_file_vm_ops;
1441 if (IS_DAX(inode))
1442 vma->vm_flags |= VM_HUGEPAGE;
1443 return 0;
1444}
1445
1446const struct file_operations xfs_file_operations = {
1447 .llseek = xfs_file_llseek,
1448 .read_iter = xfs_file_read_iter,
1449 .write_iter = xfs_file_write_iter,
1450 .splice_read = generic_file_splice_read,
1451 .splice_write = iter_file_splice_write,
1452 .iopoll = iomap_dio_iopoll,
1453 .unlocked_ioctl = xfs_file_ioctl,
1454#ifdef CONFIG_COMPAT
1455 .compat_ioctl = xfs_file_compat_ioctl,
1456#endif
1457 .mmap = xfs_file_mmap,
1458 .mmap_supported_flags = MAP_SYNC,
1459 .open = xfs_file_open,
1460 .release = xfs_file_release,
1461 .fsync = xfs_file_fsync,
1462 .get_unmapped_area = thp_get_unmapped_area,
1463 .fallocate = xfs_file_fallocate,
1464 .fadvise = xfs_file_fadvise,
1465 .remap_file_range = xfs_file_remap_range,
1466};
1467
1468const struct file_operations xfs_dir_file_operations = {
1469 .open = xfs_dir_open,
1470 .read = generic_read_dir,
1471 .iterate_shared = xfs_file_readdir,
1472 .llseek = generic_file_llseek,
1473 .unlocked_ioctl = xfs_file_ioctl,
1474#ifdef CONFIG_COMPAT
1475 .compat_ioctl = xfs_file_compat_ioctl,
1476#endif
1477 .fsync = xfs_dir_fsync,
1478};
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#include "xfs_file.h"
28
29#include <linux/dax.h>
30#include <linux/falloc.h>
31#include <linux/backing-dev.h>
32#include <linux/mman.h>
33#include <linux/fadvise.h>
34#include <linux/mount.h>
35
36static const struct vm_operations_struct xfs_file_vm_ops;
37
38/*
39 * Decide if the given file range is aligned to the size of the fundamental
40 * allocation unit for the file.
41 */
42bool
43xfs_is_falloc_aligned(
44 struct xfs_inode *ip,
45 loff_t pos,
46 long long int len)
47{
48 unsigned int alloc_unit = xfs_inode_alloc_unitsize(ip);
49
50 if (!is_power_of_2(alloc_unit))
51 return isaligned_64(pos, alloc_unit) &&
52 isaligned_64(len, alloc_unit);
53
54 return !((pos | len) & (alloc_unit - 1));
55}
56
57/*
58 * Fsync operations on directories are much simpler than on regular files,
59 * as there is no file data to flush, and thus also no need for explicit
60 * cache flush operations, and there are no non-transaction metadata updates
61 * on directories either.
62 */
63STATIC int
64xfs_dir_fsync(
65 struct file *file,
66 loff_t start,
67 loff_t end,
68 int datasync)
69{
70 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
71
72 trace_xfs_dir_fsync(ip);
73 return xfs_log_force_inode(ip);
74}
75
76static xfs_csn_t
77xfs_fsync_seq(
78 struct xfs_inode *ip,
79 bool datasync)
80{
81 if (!xfs_ipincount(ip))
82 return 0;
83 if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
84 return 0;
85 return ip->i_itemp->ili_commit_seq;
86}
87
88/*
89 * All metadata updates are logged, which means that we just have to flush the
90 * log up to the latest LSN that touched the inode.
91 *
92 * If we have concurrent fsync/fdatasync() calls, we need them to all block on
93 * the log force before we clear the ili_fsync_fields field. This ensures that
94 * we don't get a racing sync operation that does not wait for the metadata to
95 * hit the journal before returning. If we race with clearing ili_fsync_fields,
96 * then all that will happen is the log force will do nothing as the lsn will
97 * already be on disk. We can't race with setting ili_fsync_fields because that
98 * is done under XFS_ILOCK_EXCL, and that can't happen because we hold the lock
99 * shared until after the ili_fsync_fields is cleared.
100 */
101static int
102xfs_fsync_flush_log(
103 struct xfs_inode *ip,
104 bool datasync,
105 int *log_flushed)
106{
107 int error = 0;
108 xfs_csn_t seq;
109
110 xfs_ilock(ip, XFS_ILOCK_SHARED);
111 seq = xfs_fsync_seq(ip, datasync);
112 if (seq) {
113 error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
114 log_flushed);
115
116 spin_lock(&ip->i_itemp->ili_lock);
117 ip->i_itemp->ili_fsync_fields = 0;
118 spin_unlock(&ip->i_itemp->ili_lock);
119 }
120 xfs_iunlock(ip, XFS_ILOCK_SHARED);
121 return error;
122}
123
124STATIC int
125xfs_file_fsync(
126 struct file *file,
127 loff_t start,
128 loff_t end,
129 int datasync)
130{
131 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
132 struct xfs_mount *mp = ip->i_mount;
133 int error, err2;
134 int log_flushed = 0;
135
136 trace_xfs_file_fsync(ip);
137
138 error = file_write_and_wait_range(file, start, end);
139 if (error)
140 return error;
141
142 if (xfs_is_shutdown(mp))
143 return -EIO;
144
145 xfs_iflags_clear(ip, XFS_ITRUNCATED);
146
147 /*
148 * If we have an RT and/or log subvolume we need to make sure to flush
149 * the write cache the device used for file data first. This is to
150 * ensure newly written file data make it to disk before logging the new
151 * inode size in case of an extending write.
152 */
153 if (XFS_IS_REALTIME_INODE(ip))
154 error = blkdev_issue_flush(mp->m_rtdev_targp->bt_bdev);
155 else if (mp->m_logdev_targp != mp->m_ddev_targp)
156 error = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
157
158 /*
159 * Any inode that has dirty modifications in the log is pinned. The
160 * racy check here for a pinned inode will not catch modifications
161 * that happen concurrently to the fsync call, but fsync semantics
162 * only require to sync previously completed I/O.
163 */
164 if (xfs_ipincount(ip)) {
165 err2 = xfs_fsync_flush_log(ip, datasync, &log_flushed);
166 if (err2 && !error)
167 error = err2;
168 }
169
170 /*
171 * If we only have a single device, and the log force about was
172 * a no-op we might have to flush the data device cache here.
173 * This can only happen for fdatasync/O_DSYNC if we were overwriting
174 * an already allocated file and thus do not have any metadata to
175 * commit.
176 */
177 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
178 mp->m_logdev_targp == mp->m_ddev_targp) {
179 err2 = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
180 if (err2 && !error)
181 error = err2;
182 }
183
184 return error;
185}
186
187static int
188xfs_ilock_iocb(
189 struct kiocb *iocb,
190 unsigned int lock_mode)
191{
192 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
193
194 if (iocb->ki_flags & IOCB_NOWAIT) {
195 if (!xfs_ilock_nowait(ip, lock_mode))
196 return -EAGAIN;
197 } else {
198 xfs_ilock(ip, lock_mode);
199 }
200
201 return 0;
202}
203
204static int
205xfs_ilock_iocb_for_write(
206 struct kiocb *iocb,
207 unsigned int *lock_mode)
208{
209 ssize_t ret;
210 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
211
212 ret = xfs_ilock_iocb(iocb, *lock_mode);
213 if (ret)
214 return ret;
215
216 /*
217 * If a reflink remap is in progress we always need to take the iolock
218 * exclusively to wait for it to finish.
219 */
220 if (*lock_mode == XFS_IOLOCK_SHARED &&
221 xfs_iflags_test(ip, XFS_IREMAPPING)) {
222 xfs_iunlock(ip, *lock_mode);
223 *lock_mode = XFS_IOLOCK_EXCL;
224 return xfs_ilock_iocb(iocb, *lock_mode);
225 }
226
227 return 0;
228}
229
230STATIC ssize_t
231xfs_file_dio_read(
232 struct kiocb *iocb,
233 struct iov_iter *to)
234{
235 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
236 ssize_t ret;
237
238 trace_xfs_file_direct_read(iocb, to);
239
240 if (!iov_iter_count(to))
241 return 0; /* skip atime */
242
243 file_accessed(iocb->ki_filp);
244
245 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
246 if (ret)
247 return ret;
248 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
249 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
250
251 return ret;
252}
253
254static noinline ssize_t
255xfs_file_dax_read(
256 struct kiocb *iocb,
257 struct iov_iter *to)
258{
259 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
260 ssize_t ret = 0;
261
262 trace_xfs_file_dax_read(iocb, to);
263
264 if (!iov_iter_count(to))
265 return 0; /* skip atime */
266
267 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
268 if (ret)
269 return ret;
270 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
271 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
272
273 file_accessed(iocb->ki_filp);
274 return ret;
275}
276
277STATIC ssize_t
278xfs_file_buffered_read(
279 struct kiocb *iocb,
280 struct iov_iter *to)
281{
282 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
283 ssize_t ret;
284
285 trace_xfs_file_buffered_read(iocb, to);
286
287 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
288 if (ret)
289 return ret;
290 ret = generic_file_read_iter(iocb, to);
291 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
292
293 return ret;
294}
295
296STATIC ssize_t
297xfs_file_read_iter(
298 struct kiocb *iocb,
299 struct iov_iter *to)
300{
301 struct inode *inode = file_inode(iocb->ki_filp);
302 struct xfs_mount *mp = XFS_I(inode)->i_mount;
303 ssize_t ret = 0;
304
305 XFS_STATS_INC(mp, xs_read_calls);
306
307 if (xfs_is_shutdown(mp))
308 return -EIO;
309
310 if (IS_DAX(inode))
311 ret = xfs_file_dax_read(iocb, to);
312 else if (iocb->ki_flags & IOCB_DIRECT)
313 ret = xfs_file_dio_read(iocb, to);
314 else
315 ret = xfs_file_buffered_read(iocb, to);
316
317 if (ret > 0)
318 XFS_STATS_ADD(mp, xs_read_bytes, ret);
319 return ret;
320}
321
322STATIC ssize_t
323xfs_file_splice_read(
324 struct file *in,
325 loff_t *ppos,
326 struct pipe_inode_info *pipe,
327 size_t len,
328 unsigned int flags)
329{
330 struct inode *inode = file_inode(in);
331 struct xfs_inode *ip = XFS_I(inode);
332 struct xfs_mount *mp = ip->i_mount;
333 ssize_t ret = 0;
334
335 XFS_STATS_INC(mp, xs_read_calls);
336
337 if (xfs_is_shutdown(mp))
338 return -EIO;
339
340 trace_xfs_file_splice_read(ip, *ppos, len);
341
342 xfs_ilock(ip, XFS_IOLOCK_SHARED);
343 ret = filemap_splice_read(in, ppos, pipe, len, flags);
344 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
345 if (ret > 0)
346 XFS_STATS_ADD(mp, xs_read_bytes, ret);
347 return ret;
348}
349
350/*
351 * Take care of zeroing post-EOF blocks when they might exist.
352 *
353 * Returns 0 if successfully, a negative error for a failure, or 1 if this
354 * function dropped the iolock and reacquired it exclusively and the caller
355 * needs to restart the write sanity checks.
356 */
357static ssize_t
358xfs_file_write_zero_eof(
359 struct kiocb *iocb,
360 struct iov_iter *from,
361 unsigned int *iolock,
362 size_t count,
363 bool *drained_dio)
364{
365 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
366 loff_t isize;
367 int error;
368
369 /*
370 * We need to serialise against EOF updates that occur in IO completions
371 * here. We want to make sure that nobody is changing the size while
372 * we do this check until we have placed an IO barrier (i.e. hold
373 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
374 * spinlock effectively forms a memory barrier once we have
375 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
376 * hence be able to correctly determine if we need to run zeroing.
377 */
378 spin_lock(&ip->i_flags_lock);
379 isize = i_size_read(VFS_I(ip));
380 if (iocb->ki_pos <= isize) {
381 spin_unlock(&ip->i_flags_lock);
382 return 0;
383 }
384 spin_unlock(&ip->i_flags_lock);
385
386 if (iocb->ki_flags & IOCB_NOWAIT)
387 return -EAGAIN;
388
389 if (!*drained_dio) {
390 /*
391 * If zeroing is needed and we are currently holding the iolock
392 * shared, we need to update it to exclusive which implies
393 * having to redo all checks before.
394 */
395 if (*iolock == XFS_IOLOCK_SHARED) {
396 xfs_iunlock(ip, *iolock);
397 *iolock = XFS_IOLOCK_EXCL;
398 xfs_ilock(ip, *iolock);
399 iov_iter_reexpand(from, count);
400 }
401
402 /*
403 * We now have an IO submission barrier in place, but AIO can do
404 * EOF updates during IO completion and hence we now need to
405 * wait for all of them to drain. Non-AIO DIO will have drained
406 * before we are given the XFS_IOLOCK_EXCL, and so for most
407 * cases this wait is a no-op.
408 */
409 inode_dio_wait(VFS_I(ip));
410 *drained_dio = true;
411 return 1;
412 }
413
414 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
415
416 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
417 error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
418 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
419
420 return error;
421}
422
423/*
424 * Common pre-write limit and setup checks.
425 *
426 * Called with the iolock held either shared and exclusive according to
427 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
428 * if called for a direct write beyond i_size.
429 */
430STATIC ssize_t
431xfs_file_write_checks(
432 struct kiocb *iocb,
433 struct iov_iter *from,
434 unsigned int *iolock)
435{
436 struct inode *inode = iocb->ki_filp->f_mapping->host;
437 size_t count = iov_iter_count(from);
438 bool drained_dio = false;
439 ssize_t error;
440
441restart:
442 error = generic_write_checks(iocb, from);
443 if (error <= 0)
444 return error;
445
446 if (iocb->ki_flags & IOCB_NOWAIT) {
447 error = break_layout(inode, false);
448 if (error == -EWOULDBLOCK)
449 error = -EAGAIN;
450 } else {
451 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
452 }
453
454 if (error)
455 return error;
456
457 /*
458 * For changing security info in file_remove_privs() we need i_rwsem
459 * exclusively.
460 */
461 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
462 xfs_iunlock(XFS_I(inode), *iolock);
463 *iolock = XFS_IOLOCK_EXCL;
464 error = xfs_ilock_iocb(iocb, *iolock);
465 if (error) {
466 *iolock = 0;
467 return error;
468 }
469 goto restart;
470 }
471
472 /*
473 * If the offset is beyond the size of the file, we need to zero all
474 * blocks that fall between the existing EOF and the start of this
475 * write.
476 *
477 * We can do an unlocked check for i_size here safely as I/O completion
478 * can only extend EOF. Truncate is locked out at this point, so the
479 * EOF can not move backwards, only forwards. Hence we only need to take
480 * the slow path when we are at or beyond the current EOF.
481 */
482 if (iocb->ki_pos > i_size_read(inode)) {
483 error = xfs_file_write_zero_eof(iocb, from, iolock, count,
484 &drained_dio);
485 if (error == 1)
486 goto restart;
487 if (error)
488 return error;
489 }
490
491 return kiocb_modified(iocb);
492}
493
494static int
495xfs_dio_write_end_io(
496 struct kiocb *iocb,
497 ssize_t size,
498 int error,
499 unsigned flags)
500{
501 struct inode *inode = file_inode(iocb->ki_filp);
502 struct xfs_inode *ip = XFS_I(inode);
503 loff_t offset = iocb->ki_pos;
504 unsigned int nofs_flag;
505
506 trace_xfs_end_io_direct_write(ip, offset, size);
507
508 if (xfs_is_shutdown(ip->i_mount))
509 return -EIO;
510
511 if (error)
512 return error;
513 if (!size)
514 return 0;
515
516 /*
517 * Capture amount written on completion as we can't reliably account
518 * for it on submission.
519 */
520 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
521
522 /*
523 * We can allocate memory here while doing writeback on behalf of
524 * memory reclaim. To avoid memory allocation deadlocks set the
525 * task-wide nofs context for the following operations.
526 */
527 nofs_flag = memalloc_nofs_save();
528
529 if (flags & IOMAP_DIO_COW) {
530 error = xfs_reflink_end_cow(ip, offset, size);
531 if (error)
532 goto out;
533 }
534
535 /*
536 * Unwritten conversion updates the in-core isize after extent
537 * conversion but before updating the on-disk size. Updating isize any
538 * earlier allows a racing dio read to find unwritten extents before
539 * they are converted.
540 */
541 if (flags & IOMAP_DIO_UNWRITTEN) {
542 error = xfs_iomap_write_unwritten(ip, offset, size, true);
543 goto out;
544 }
545
546 /*
547 * We need to update the in-core inode size here so that we don't end up
548 * with the on-disk inode size being outside the in-core inode size. We
549 * have no other method of updating EOF for AIO, so always do it here
550 * if necessary.
551 *
552 * We need to lock the test/set EOF update as we can be racing with
553 * other IO completions here to update the EOF. Failing to serialise
554 * here can result in EOF moving backwards and Bad Things Happen when
555 * that occurs.
556 *
557 * As IO completion only ever extends EOF, we can do an unlocked check
558 * here to avoid taking the spinlock. If we land within the current EOF,
559 * then we do not need to do an extending update at all, and we don't
560 * need to take the lock to check this. If we race with an update moving
561 * EOF, then we'll either still be beyond EOF and need to take the lock,
562 * or we'll be within EOF and we don't need to take it at all.
563 */
564 if (offset + size <= i_size_read(inode))
565 goto out;
566
567 spin_lock(&ip->i_flags_lock);
568 if (offset + size > i_size_read(inode)) {
569 i_size_write(inode, offset + size);
570 spin_unlock(&ip->i_flags_lock);
571 error = xfs_setfilesize(ip, offset, size);
572 } else {
573 spin_unlock(&ip->i_flags_lock);
574 }
575
576out:
577 memalloc_nofs_restore(nofs_flag);
578 return error;
579}
580
581static const struct iomap_dio_ops xfs_dio_write_ops = {
582 .end_io = xfs_dio_write_end_io,
583};
584
585/*
586 * Handle block aligned direct I/O writes
587 */
588static noinline ssize_t
589xfs_file_dio_write_aligned(
590 struct xfs_inode *ip,
591 struct kiocb *iocb,
592 struct iov_iter *from)
593{
594 unsigned int iolock = XFS_IOLOCK_SHARED;
595 ssize_t ret;
596
597 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
598 if (ret)
599 return ret;
600 ret = xfs_file_write_checks(iocb, from, &iolock);
601 if (ret)
602 goto out_unlock;
603
604 /*
605 * We don't need to hold the IOLOCK exclusively across the IO, so demote
606 * the iolock back to shared if we had to take the exclusive lock in
607 * xfs_file_write_checks() for other reasons.
608 */
609 if (iolock == XFS_IOLOCK_EXCL) {
610 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
611 iolock = XFS_IOLOCK_SHARED;
612 }
613 trace_xfs_file_direct_write(iocb, from);
614 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
615 &xfs_dio_write_ops, 0, NULL, 0);
616out_unlock:
617 if (iolock)
618 xfs_iunlock(ip, iolock);
619 return ret;
620}
621
622/*
623 * Handle block unaligned direct I/O writes
624 *
625 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
626 * them to be done in parallel with reads and other direct I/O writes. However,
627 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
628 * to do sub-block zeroing and that requires serialisation against other direct
629 * I/O to the same block. In this case we need to serialise the submission of
630 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
631 * In the case where sub-block zeroing is not required, we can do concurrent
632 * sub-block dios to the same block successfully.
633 *
634 * Optimistically submit the I/O using the shared lock first, but use the
635 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
636 * if block allocation or partial block zeroing would be required. In that case
637 * we try again with the exclusive lock.
638 */
639static noinline ssize_t
640xfs_file_dio_write_unaligned(
641 struct xfs_inode *ip,
642 struct kiocb *iocb,
643 struct iov_iter *from)
644{
645 size_t isize = i_size_read(VFS_I(ip));
646 size_t count = iov_iter_count(from);
647 unsigned int iolock = XFS_IOLOCK_SHARED;
648 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
649 ssize_t ret;
650
651 /*
652 * Extending writes need exclusivity because of the sub-block zeroing
653 * that the DIO code always does for partial tail blocks beyond EOF, so
654 * don't even bother trying the fast path in this case.
655 */
656 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
657 if (iocb->ki_flags & IOCB_NOWAIT)
658 return -EAGAIN;
659retry_exclusive:
660 iolock = XFS_IOLOCK_EXCL;
661 flags = IOMAP_DIO_FORCE_WAIT;
662 }
663
664 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
665 if (ret)
666 return ret;
667
668 /*
669 * We can't properly handle unaligned direct I/O to reflink files yet,
670 * as we can't unshare a partial block.
671 */
672 if (xfs_is_cow_inode(ip)) {
673 trace_xfs_reflink_bounce_dio_write(iocb, from);
674 ret = -ENOTBLK;
675 goto out_unlock;
676 }
677
678 ret = xfs_file_write_checks(iocb, from, &iolock);
679 if (ret)
680 goto out_unlock;
681
682 /*
683 * If we are doing exclusive unaligned I/O, this must be the only I/O
684 * in-flight. Otherwise we risk data corruption due to unwritten extent
685 * conversions from the AIO end_io handler. Wait for all other I/O to
686 * drain first.
687 */
688 if (flags & IOMAP_DIO_FORCE_WAIT)
689 inode_dio_wait(VFS_I(ip));
690
691 trace_xfs_file_direct_write(iocb, from);
692 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
693 &xfs_dio_write_ops, flags, NULL, 0);
694
695 /*
696 * Retry unaligned I/O with exclusive blocking semantics if the DIO
697 * layer rejected it for mapping or locking reasons. If we are doing
698 * nonblocking user I/O, propagate the error.
699 */
700 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
701 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
702 xfs_iunlock(ip, iolock);
703 goto retry_exclusive;
704 }
705
706out_unlock:
707 if (iolock)
708 xfs_iunlock(ip, iolock);
709 return ret;
710}
711
712static ssize_t
713xfs_file_dio_write(
714 struct kiocb *iocb,
715 struct iov_iter *from)
716{
717 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
718 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
719 size_t count = iov_iter_count(from);
720
721 /* direct I/O must be aligned to device logical sector size */
722 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
723 return -EINVAL;
724 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
725 return xfs_file_dio_write_unaligned(ip, iocb, from);
726 return xfs_file_dio_write_aligned(ip, iocb, from);
727}
728
729static noinline ssize_t
730xfs_file_dax_write(
731 struct kiocb *iocb,
732 struct iov_iter *from)
733{
734 struct inode *inode = iocb->ki_filp->f_mapping->host;
735 struct xfs_inode *ip = XFS_I(inode);
736 unsigned int iolock = XFS_IOLOCK_EXCL;
737 ssize_t ret, error = 0;
738 loff_t pos;
739
740 ret = xfs_ilock_iocb(iocb, iolock);
741 if (ret)
742 return ret;
743 ret = xfs_file_write_checks(iocb, from, &iolock);
744 if (ret)
745 goto out;
746
747 pos = iocb->ki_pos;
748
749 trace_xfs_file_dax_write(iocb, from);
750 ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
751 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
752 i_size_write(inode, iocb->ki_pos);
753 error = xfs_setfilesize(ip, pos, ret);
754 }
755out:
756 if (iolock)
757 xfs_iunlock(ip, iolock);
758 if (error)
759 return error;
760
761 if (ret > 0) {
762 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
763
764 /* Handle various SYNC-type writes */
765 ret = generic_write_sync(iocb, ret);
766 }
767 return ret;
768}
769
770STATIC ssize_t
771xfs_file_buffered_write(
772 struct kiocb *iocb,
773 struct iov_iter *from)
774{
775 struct inode *inode = iocb->ki_filp->f_mapping->host;
776 struct xfs_inode *ip = XFS_I(inode);
777 ssize_t ret;
778 bool cleared_space = false;
779 unsigned int iolock;
780
781write_retry:
782 iolock = XFS_IOLOCK_EXCL;
783 ret = xfs_ilock_iocb(iocb, iolock);
784 if (ret)
785 return ret;
786
787 ret = xfs_file_write_checks(iocb, from, &iolock);
788 if (ret)
789 goto out;
790
791 trace_xfs_file_buffered_write(iocb, from);
792 ret = iomap_file_buffered_write(iocb, from,
793 &xfs_buffered_write_iomap_ops, NULL);
794
795 /*
796 * If we hit a space limit, try to free up some lingering preallocated
797 * space before returning an error. In the case of ENOSPC, first try to
798 * write back all dirty inodes to free up some of the excess reserved
799 * metadata space. This reduces the chances that the eofblocks scan
800 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
801 * also behaves as a filter to prevent too many eofblocks scans from
802 * running at the same time. Use a synchronous scan to increase the
803 * effectiveness of the scan.
804 */
805 if (ret == -EDQUOT && !cleared_space) {
806 xfs_iunlock(ip, iolock);
807 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
808 cleared_space = true;
809 goto write_retry;
810 } else if (ret == -ENOSPC && !cleared_space) {
811 struct xfs_icwalk icw = {0};
812
813 cleared_space = true;
814 xfs_flush_inodes(ip->i_mount);
815
816 xfs_iunlock(ip, iolock);
817 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
818 xfs_blockgc_free_space(ip->i_mount, &icw);
819 goto write_retry;
820 }
821
822out:
823 if (iolock)
824 xfs_iunlock(ip, iolock);
825
826 if (ret > 0) {
827 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
828 /* Handle various SYNC-type writes */
829 ret = generic_write_sync(iocb, ret);
830 }
831 return ret;
832}
833
834STATIC ssize_t
835xfs_file_write_iter(
836 struct kiocb *iocb,
837 struct iov_iter *from)
838{
839 struct inode *inode = iocb->ki_filp->f_mapping->host;
840 struct xfs_inode *ip = XFS_I(inode);
841 ssize_t ret;
842 size_t ocount = iov_iter_count(from);
843
844 XFS_STATS_INC(ip->i_mount, xs_write_calls);
845
846 if (ocount == 0)
847 return 0;
848
849 if (xfs_is_shutdown(ip->i_mount))
850 return -EIO;
851
852 if (IS_DAX(inode))
853 return xfs_file_dax_write(iocb, from);
854
855 if (iocb->ki_flags & IOCB_ATOMIC) {
856 /*
857 * Currently only atomic writing of a single FS block is
858 * supported. It would be possible to atomic write smaller than
859 * a FS block, but there is no requirement to support this.
860 * Note that iomap also does not support this yet.
861 */
862 if (ocount != ip->i_mount->m_sb.sb_blocksize)
863 return -EINVAL;
864 ret = generic_atomic_write_valid(iocb, from);
865 if (ret)
866 return ret;
867 }
868
869 if (iocb->ki_flags & IOCB_DIRECT) {
870 /*
871 * Allow a directio write to fall back to a buffered
872 * write *only* in the case that we're doing a reflink
873 * CoW. In all other directio scenarios we do not
874 * allow an operation to fall back to buffered mode.
875 */
876 ret = xfs_file_dio_write(iocb, from);
877 if (ret != -ENOTBLK)
878 return ret;
879 }
880
881 return xfs_file_buffered_write(iocb, from);
882}
883
884/* Does this file, inode, or mount want synchronous writes? */
885static inline bool xfs_file_sync_writes(struct file *filp)
886{
887 struct xfs_inode *ip = XFS_I(file_inode(filp));
888
889 if (xfs_has_wsync(ip->i_mount))
890 return true;
891 if (filp->f_flags & (__O_SYNC | O_DSYNC))
892 return true;
893 if (IS_SYNC(file_inode(filp)))
894 return true;
895
896 return false;
897}
898
899static int
900xfs_falloc_newsize(
901 struct file *file,
902 int mode,
903 loff_t offset,
904 loff_t len,
905 loff_t *new_size)
906{
907 struct inode *inode = file_inode(file);
908
909 if ((mode & FALLOC_FL_KEEP_SIZE) || offset + len <= i_size_read(inode))
910 return 0;
911 *new_size = offset + len;
912 return inode_newsize_ok(inode, *new_size);
913}
914
915static int
916xfs_falloc_setsize(
917 struct file *file,
918 loff_t new_size)
919{
920 struct iattr iattr = {
921 .ia_valid = ATTR_SIZE,
922 .ia_size = new_size,
923 };
924
925 if (!new_size)
926 return 0;
927 return xfs_vn_setattr_size(file_mnt_idmap(file), file_dentry(file),
928 &iattr);
929}
930
931static int
932xfs_falloc_collapse_range(
933 struct file *file,
934 loff_t offset,
935 loff_t len)
936{
937 struct inode *inode = file_inode(file);
938 loff_t new_size = i_size_read(inode) - len;
939 int error;
940
941 if (!xfs_is_falloc_aligned(XFS_I(inode), offset, len))
942 return -EINVAL;
943
944 /*
945 * There is no need to overlap collapse range with EOF, in which case it
946 * is effectively a truncate operation
947 */
948 if (offset + len >= i_size_read(inode))
949 return -EINVAL;
950
951 error = xfs_collapse_file_space(XFS_I(inode), offset, len);
952 if (error)
953 return error;
954 return xfs_falloc_setsize(file, new_size);
955}
956
957static int
958xfs_falloc_insert_range(
959 struct file *file,
960 loff_t offset,
961 loff_t len)
962{
963 struct inode *inode = file_inode(file);
964 loff_t isize = i_size_read(inode);
965 int error;
966
967 if (!xfs_is_falloc_aligned(XFS_I(inode), offset, len))
968 return -EINVAL;
969
970 /*
971 * New inode size must not exceed ->s_maxbytes, accounting for
972 * possible signed overflow.
973 */
974 if (inode->i_sb->s_maxbytes - isize < len)
975 return -EFBIG;
976
977 /* Offset should be less than i_size */
978 if (offset >= isize)
979 return -EINVAL;
980
981 error = xfs_falloc_setsize(file, isize + len);
982 if (error)
983 return error;
984
985 /*
986 * Perform hole insertion now that the file size has been updated so
987 * that if we crash during the operation we don't leave shifted extents
988 * past EOF and hence losing access to the data that is contained within
989 * them.
990 */
991 return xfs_insert_file_space(XFS_I(inode), offset, len);
992}
993
994/*
995 * Punch a hole and prealloc the range. We use a hole punch rather than
996 * unwritten extent conversion for two reasons:
997 *
998 * 1.) Hole punch handles partial block zeroing for us.
999 * 2.) If prealloc returns ENOSPC, the file range is still zero-valued by
1000 * virtue of the hole punch.
1001 */
1002static int
1003xfs_falloc_zero_range(
1004 struct file *file,
1005 int mode,
1006 loff_t offset,
1007 loff_t len)
1008{
1009 struct inode *inode = file_inode(file);
1010 unsigned int blksize = i_blocksize(inode);
1011 loff_t new_size = 0;
1012 int error;
1013
1014 trace_xfs_zero_file_space(XFS_I(inode));
1015
1016 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1017 if (error)
1018 return error;
1019
1020 error = xfs_free_file_space(XFS_I(inode), offset, len);
1021 if (error)
1022 return error;
1023
1024 len = round_up(offset + len, blksize) - round_down(offset, blksize);
1025 offset = round_down(offset, blksize);
1026 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1027 if (error)
1028 return error;
1029 return xfs_falloc_setsize(file, new_size);
1030}
1031
1032static int
1033xfs_falloc_unshare_range(
1034 struct file *file,
1035 int mode,
1036 loff_t offset,
1037 loff_t len)
1038{
1039 struct inode *inode = file_inode(file);
1040 loff_t new_size = 0;
1041 int error;
1042
1043 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1044 if (error)
1045 return error;
1046
1047 error = xfs_reflink_unshare(XFS_I(inode), offset, len);
1048 if (error)
1049 return error;
1050
1051 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1052 if (error)
1053 return error;
1054 return xfs_falloc_setsize(file, new_size);
1055}
1056
1057static int
1058xfs_falloc_allocate_range(
1059 struct file *file,
1060 int mode,
1061 loff_t offset,
1062 loff_t len)
1063{
1064 struct inode *inode = file_inode(file);
1065 loff_t new_size = 0;
1066 int error;
1067
1068 /*
1069 * If always_cow mode we can't use preallocations and thus should not
1070 * create them.
1071 */
1072 if (xfs_is_always_cow_inode(XFS_I(inode)))
1073 return -EOPNOTSUPP;
1074
1075 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1076 if (error)
1077 return error;
1078
1079 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1080 if (error)
1081 return error;
1082 return xfs_falloc_setsize(file, new_size);
1083}
1084
1085#define XFS_FALLOC_FL_SUPPORTED \
1086 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
1087 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
1088 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
1089
1090STATIC long
1091xfs_file_fallocate(
1092 struct file *file,
1093 int mode,
1094 loff_t offset,
1095 loff_t len)
1096{
1097 struct inode *inode = file_inode(file);
1098 struct xfs_inode *ip = XFS_I(inode);
1099 long error;
1100 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
1101
1102 if (!S_ISREG(inode->i_mode))
1103 return -EINVAL;
1104 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
1105 return -EOPNOTSUPP;
1106
1107 xfs_ilock(ip, iolock);
1108 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
1109 if (error)
1110 goto out_unlock;
1111
1112 /*
1113 * Must wait for all AIO to complete before we continue as AIO can
1114 * change the file size on completion without holding any locks we
1115 * currently hold. We must do this first because AIO can update both
1116 * the on disk and in memory inode sizes, and the operations that follow
1117 * require the in-memory size to be fully up-to-date.
1118 */
1119 inode_dio_wait(inode);
1120
1121 error = file_modified(file);
1122 if (error)
1123 goto out_unlock;
1124
1125 switch (mode & FALLOC_FL_MODE_MASK) {
1126 case FALLOC_FL_PUNCH_HOLE:
1127 error = xfs_free_file_space(ip, offset, len);
1128 break;
1129 case FALLOC_FL_COLLAPSE_RANGE:
1130 error = xfs_falloc_collapse_range(file, offset, len);
1131 break;
1132 case FALLOC_FL_INSERT_RANGE:
1133 error = xfs_falloc_insert_range(file, offset, len);
1134 break;
1135 case FALLOC_FL_ZERO_RANGE:
1136 error = xfs_falloc_zero_range(file, mode, offset, len);
1137 break;
1138 case FALLOC_FL_UNSHARE_RANGE:
1139 error = xfs_falloc_unshare_range(file, mode, offset, len);
1140 break;
1141 case FALLOC_FL_ALLOCATE_RANGE:
1142 error = xfs_falloc_allocate_range(file, mode, offset, len);
1143 break;
1144 default:
1145 error = -EOPNOTSUPP;
1146 break;
1147 }
1148
1149 if (!error && xfs_file_sync_writes(file))
1150 error = xfs_log_force_inode(ip);
1151
1152out_unlock:
1153 xfs_iunlock(ip, iolock);
1154 return error;
1155}
1156
1157STATIC int
1158xfs_file_fadvise(
1159 struct file *file,
1160 loff_t start,
1161 loff_t end,
1162 int advice)
1163{
1164 struct xfs_inode *ip = XFS_I(file_inode(file));
1165 int ret;
1166 int lockflags = 0;
1167
1168 /*
1169 * Operations creating pages in page cache need protection from hole
1170 * punching and similar ops
1171 */
1172 if (advice == POSIX_FADV_WILLNEED) {
1173 lockflags = XFS_IOLOCK_SHARED;
1174 xfs_ilock(ip, lockflags);
1175 }
1176 ret = generic_fadvise(file, start, end, advice);
1177 if (lockflags)
1178 xfs_iunlock(ip, lockflags);
1179 return ret;
1180}
1181
1182STATIC loff_t
1183xfs_file_remap_range(
1184 struct file *file_in,
1185 loff_t pos_in,
1186 struct file *file_out,
1187 loff_t pos_out,
1188 loff_t len,
1189 unsigned int remap_flags)
1190{
1191 struct inode *inode_in = file_inode(file_in);
1192 struct xfs_inode *src = XFS_I(inode_in);
1193 struct inode *inode_out = file_inode(file_out);
1194 struct xfs_inode *dest = XFS_I(inode_out);
1195 struct xfs_mount *mp = src->i_mount;
1196 loff_t remapped = 0;
1197 xfs_extlen_t cowextsize;
1198 int ret;
1199
1200 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1201 return -EINVAL;
1202
1203 if (!xfs_has_reflink(mp))
1204 return -EOPNOTSUPP;
1205
1206 if (xfs_is_shutdown(mp))
1207 return -EIO;
1208
1209 /* Prepare and then clone file data. */
1210 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1211 &len, remap_flags);
1212 if (ret || len == 0)
1213 return ret;
1214
1215 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1216
1217 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1218 &remapped);
1219 if (ret)
1220 goto out_unlock;
1221
1222 /*
1223 * Carry the cowextsize hint from src to dest if we're sharing the
1224 * entire source file to the entire destination file, the source file
1225 * has a cowextsize hint, and the destination file does not.
1226 */
1227 cowextsize = 0;
1228 if (pos_in == 0 && len == i_size_read(inode_in) &&
1229 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1230 pos_out == 0 && len >= i_size_read(inode_out) &&
1231 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1232 cowextsize = src->i_cowextsize;
1233
1234 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1235 remap_flags);
1236 if (ret)
1237 goto out_unlock;
1238
1239 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1240 xfs_log_force_inode(dest);
1241out_unlock:
1242 xfs_iunlock2_remapping(src, dest);
1243 if (ret)
1244 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1245 /*
1246 * If the caller did not set CAN_SHORTEN, then it is not prepared to
1247 * handle partial results -- either the whole remap succeeds, or we
1248 * must say why it did not. In this case, any error should be returned
1249 * to the caller.
1250 */
1251 if (ret && remapped < len && !(remap_flags & REMAP_FILE_CAN_SHORTEN))
1252 return ret;
1253 return remapped > 0 ? remapped : ret;
1254}
1255
1256STATIC int
1257xfs_file_open(
1258 struct inode *inode,
1259 struct file *file)
1260{
1261 if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1262 return -EIO;
1263 file->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
1264 if (xfs_inode_can_atomicwrite(XFS_I(inode)))
1265 file->f_mode |= FMODE_CAN_ATOMIC_WRITE;
1266 return generic_file_open(inode, file);
1267}
1268
1269STATIC int
1270xfs_dir_open(
1271 struct inode *inode,
1272 struct file *file)
1273{
1274 struct xfs_inode *ip = XFS_I(inode);
1275 unsigned int mode;
1276 int error;
1277
1278 if (xfs_is_shutdown(ip->i_mount))
1279 return -EIO;
1280 error = generic_file_open(inode, file);
1281 if (error)
1282 return error;
1283
1284 /*
1285 * If there are any blocks, read-ahead block 0 as we're almost
1286 * certain to have the next operation be a read there.
1287 */
1288 mode = xfs_ilock_data_map_shared(ip);
1289 if (ip->i_df.if_nextents > 0)
1290 error = xfs_dir3_data_readahead(ip, 0, 0);
1291 xfs_iunlock(ip, mode);
1292 return error;
1293}
1294
1295/*
1296 * Don't bother propagating errors. We're just doing cleanup, and the caller
1297 * ignores the return value anyway.
1298 */
1299STATIC int
1300xfs_file_release(
1301 struct inode *inode,
1302 struct file *file)
1303{
1304 struct xfs_inode *ip = XFS_I(inode);
1305 struct xfs_mount *mp = ip->i_mount;
1306
1307 /*
1308 * If this is a read-only mount or the file system has been shut down,
1309 * don't generate I/O.
1310 */
1311 if (xfs_is_readonly(mp) || xfs_is_shutdown(mp))
1312 return 0;
1313
1314 /*
1315 * If we previously truncated this file and removed old data in the
1316 * process, we want to initiate "early" writeout on the last close.
1317 * This is an attempt to combat the notorious NULL files problem which
1318 * is particularly noticeable from a truncate down, buffered (re-)write
1319 * (delalloc), followed by a crash. What we are effectively doing here
1320 * is significantly reducing the time window where we'd otherwise be
1321 * exposed to that problem.
1322 */
1323 if (xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED)) {
1324 xfs_iflags_clear(ip, XFS_EOFBLOCKS_RELEASED);
1325 if (ip->i_delayed_blks > 0)
1326 filemap_flush(inode->i_mapping);
1327 }
1328
1329 /*
1330 * XFS aggressively preallocates post-EOF space to generate contiguous
1331 * allocations for writers that append to the end of the file.
1332 *
1333 * To support workloads that close and reopen the file frequently, these
1334 * preallocations usually persist after a close unless it is the first
1335 * close for the inode. This is a tradeoff to generate tightly packed
1336 * data layouts for unpacking tarballs or similar archives that write
1337 * one file after another without going back to it while keeping the
1338 * preallocation for files that have recurring open/write/close cycles.
1339 *
1340 * This heuristic is skipped for inodes with the append-only flag as
1341 * that flag is rather pointless for inodes written only once.
1342 *
1343 * There is no point in freeing blocks here for open but unlinked files
1344 * as they will be taken care of by the inactivation path soon.
1345 *
1346 * When releasing a read-only context, don't flush data or trim post-EOF
1347 * blocks. This avoids open/read/close workloads from removing EOF
1348 * blocks that other writers depend upon to reduce fragmentation.
1349 *
1350 * If we can't get the iolock just skip truncating the blocks past EOF
1351 * because we could deadlock with the mmap_lock otherwise. We'll get
1352 * another chance to drop them once the last reference to the inode is
1353 * dropped, so we'll never leak blocks permanently.
1354 */
1355 if (inode->i_nlink &&
1356 (file->f_mode & FMODE_WRITE) &&
1357 !(ip->i_diflags & XFS_DIFLAG_APPEND) &&
1358 !xfs_iflags_test(ip, XFS_EOFBLOCKS_RELEASED) &&
1359 xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1360 if (xfs_can_free_eofblocks(ip) &&
1361 !xfs_iflags_test_and_set(ip, XFS_EOFBLOCKS_RELEASED))
1362 xfs_free_eofblocks(ip);
1363 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1364 }
1365
1366 return 0;
1367}
1368
1369STATIC int
1370xfs_file_readdir(
1371 struct file *file,
1372 struct dir_context *ctx)
1373{
1374 struct inode *inode = file_inode(file);
1375 xfs_inode_t *ip = XFS_I(inode);
1376 size_t bufsize;
1377
1378 /*
1379 * The Linux API doesn't pass down the total size of the buffer
1380 * we read into down to the filesystem. With the filldir concept
1381 * it's not needed for correct information, but the XFS dir2 leaf
1382 * code wants an estimate of the buffer size to calculate it's
1383 * readahead window and size the buffers used for mapping to
1384 * physical blocks.
1385 *
1386 * Try to give it an estimate that's good enough, maybe at some
1387 * point we can change the ->readdir prototype to include the
1388 * buffer size. For now we use the current glibc buffer size.
1389 */
1390 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1391
1392 return xfs_readdir(NULL, ip, ctx, bufsize);
1393}
1394
1395STATIC loff_t
1396xfs_file_llseek(
1397 struct file *file,
1398 loff_t offset,
1399 int whence)
1400{
1401 struct inode *inode = file->f_mapping->host;
1402
1403 if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1404 return -EIO;
1405
1406 switch (whence) {
1407 default:
1408 return generic_file_llseek(file, offset, whence);
1409 case SEEK_HOLE:
1410 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1411 break;
1412 case SEEK_DATA:
1413 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1414 break;
1415 }
1416
1417 if (offset < 0)
1418 return offset;
1419 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1420}
1421
1422static inline vm_fault_t
1423xfs_dax_fault_locked(
1424 struct vm_fault *vmf,
1425 unsigned int order,
1426 bool write_fault)
1427{
1428 vm_fault_t ret;
1429 pfn_t pfn;
1430
1431 if (!IS_ENABLED(CONFIG_FS_DAX)) {
1432 ASSERT(0);
1433 return VM_FAULT_SIGBUS;
1434 }
1435 ret = dax_iomap_fault(vmf, order, &pfn, NULL,
1436 (write_fault && !vmf->cow_page) ?
1437 &xfs_dax_write_iomap_ops :
1438 &xfs_read_iomap_ops);
1439 if (ret & VM_FAULT_NEEDDSYNC)
1440 ret = dax_finish_sync_fault(vmf, order, pfn);
1441 return ret;
1442}
1443
1444static vm_fault_t
1445xfs_dax_read_fault(
1446 struct vm_fault *vmf,
1447 unsigned int order)
1448{
1449 struct xfs_inode *ip = XFS_I(file_inode(vmf->vma->vm_file));
1450 vm_fault_t ret;
1451
1452 trace_xfs_read_fault(ip, order);
1453
1454 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1455 ret = xfs_dax_fault_locked(vmf, order, false);
1456 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1457
1458 return ret;
1459}
1460
1461/*
1462 * Locking for serialisation of IO during page faults. This results in a lock
1463 * ordering of:
1464 *
1465 * mmap_lock (MM)
1466 * sb_start_pagefault(vfs, freeze)
1467 * invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1468 * page_lock (MM)
1469 * i_lock (XFS - extent map serialisation)
1470 */
1471static vm_fault_t
1472xfs_write_fault(
1473 struct vm_fault *vmf,
1474 unsigned int order)
1475{
1476 struct inode *inode = file_inode(vmf->vma->vm_file);
1477 struct xfs_inode *ip = XFS_I(inode);
1478 unsigned int lock_mode = XFS_MMAPLOCK_SHARED;
1479 vm_fault_t ret;
1480
1481 trace_xfs_write_fault(ip, order);
1482
1483 sb_start_pagefault(inode->i_sb);
1484 file_update_time(vmf->vma->vm_file);
1485
1486 /*
1487 * Normally we only need the shared mmaplock, but if a reflink remap is
1488 * in progress we take the exclusive lock to wait for the remap to
1489 * finish before taking a write fault.
1490 */
1491 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1492 if (xfs_iflags_test(ip, XFS_IREMAPPING)) {
1493 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1494 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1495 lock_mode = XFS_MMAPLOCK_EXCL;
1496 }
1497
1498 if (IS_DAX(inode))
1499 ret = xfs_dax_fault_locked(vmf, order, true);
1500 else
1501 ret = iomap_page_mkwrite(vmf, &xfs_buffered_write_iomap_ops);
1502 xfs_iunlock(ip, lock_mode);
1503
1504 sb_end_pagefault(inode->i_sb);
1505 return ret;
1506}
1507
1508static inline bool
1509xfs_is_write_fault(
1510 struct vm_fault *vmf)
1511{
1512 return (vmf->flags & FAULT_FLAG_WRITE) &&
1513 (vmf->vma->vm_flags & VM_SHARED);
1514}
1515
1516static vm_fault_t
1517xfs_filemap_fault(
1518 struct vm_fault *vmf)
1519{
1520 struct inode *inode = file_inode(vmf->vma->vm_file);
1521
1522 /* DAX can shortcut the normal fault path on write faults! */
1523 if (IS_DAX(inode)) {
1524 if (xfs_is_write_fault(vmf))
1525 return xfs_write_fault(vmf, 0);
1526 return xfs_dax_read_fault(vmf, 0);
1527 }
1528
1529 trace_xfs_read_fault(XFS_I(inode), 0);
1530 return filemap_fault(vmf);
1531}
1532
1533static vm_fault_t
1534xfs_filemap_huge_fault(
1535 struct vm_fault *vmf,
1536 unsigned int order)
1537{
1538 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1539 return VM_FAULT_FALLBACK;
1540
1541 /* DAX can shortcut the normal fault path on write faults! */
1542 if (xfs_is_write_fault(vmf))
1543 return xfs_write_fault(vmf, order);
1544 return xfs_dax_read_fault(vmf, order);
1545}
1546
1547static vm_fault_t
1548xfs_filemap_page_mkwrite(
1549 struct vm_fault *vmf)
1550{
1551 return xfs_write_fault(vmf, 0);
1552}
1553
1554/*
1555 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1556 * on write faults. In reality, it needs to serialise against truncate and
1557 * prepare memory for writing so handle is as standard write fault.
1558 */
1559static vm_fault_t
1560xfs_filemap_pfn_mkwrite(
1561 struct vm_fault *vmf)
1562{
1563 return xfs_write_fault(vmf, 0);
1564}
1565
1566static const struct vm_operations_struct xfs_file_vm_ops = {
1567 .fault = xfs_filemap_fault,
1568 .huge_fault = xfs_filemap_huge_fault,
1569 .map_pages = filemap_map_pages,
1570 .page_mkwrite = xfs_filemap_page_mkwrite,
1571 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1572};
1573
1574STATIC int
1575xfs_file_mmap(
1576 struct file *file,
1577 struct vm_area_struct *vma)
1578{
1579 struct inode *inode = file_inode(file);
1580 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1581
1582 /*
1583 * We don't support synchronous mappings for non-DAX files and
1584 * for DAX files if underneath dax_device is not synchronous.
1585 */
1586 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1587 return -EOPNOTSUPP;
1588
1589 file_accessed(file);
1590 vma->vm_ops = &xfs_file_vm_ops;
1591 if (IS_DAX(inode))
1592 vm_flags_set(vma, VM_HUGEPAGE);
1593 return 0;
1594}
1595
1596const struct file_operations xfs_file_operations = {
1597 .llseek = xfs_file_llseek,
1598 .read_iter = xfs_file_read_iter,
1599 .write_iter = xfs_file_write_iter,
1600 .splice_read = xfs_file_splice_read,
1601 .splice_write = iter_file_splice_write,
1602 .iopoll = iocb_bio_iopoll,
1603 .unlocked_ioctl = xfs_file_ioctl,
1604#ifdef CONFIG_COMPAT
1605 .compat_ioctl = xfs_file_compat_ioctl,
1606#endif
1607 .mmap = xfs_file_mmap,
1608 .open = xfs_file_open,
1609 .release = xfs_file_release,
1610 .fsync = xfs_file_fsync,
1611 .get_unmapped_area = thp_get_unmapped_area,
1612 .fallocate = xfs_file_fallocate,
1613 .fadvise = xfs_file_fadvise,
1614 .remap_file_range = xfs_file_remap_range,
1615 .fop_flags = FOP_MMAP_SYNC | FOP_BUFFER_RASYNC |
1616 FOP_BUFFER_WASYNC | FOP_DIO_PARALLEL_WRITE,
1617};
1618
1619const struct file_operations xfs_dir_file_operations = {
1620 .open = xfs_dir_open,
1621 .read = generic_read_dir,
1622 .iterate_shared = xfs_file_readdir,
1623 .llseek = generic_file_llseek,
1624 .unlocked_ioctl = xfs_file_ioctl,
1625#ifdef CONFIG_COMPAT
1626 .compat_ioctl = xfs_file_compat_ioctl,
1627#endif
1628 .fsync = xfs_dir_fsync,
1629};