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 int
218xfs_ilock_iocb_for_write(
219 struct kiocb *iocb,
220 unsigned int *lock_mode)
221{
222 ssize_t ret;
223 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
224
225 ret = xfs_ilock_iocb(iocb, *lock_mode);
226 if (ret)
227 return ret;
228
229 if (*lock_mode == XFS_IOLOCK_EXCL)
230 return 0;
231 if (!xfs_iflags_test(ip, XFS_IREMAPPING))
232 return 0;
233
234 xfs_iunlock(ip, *lock_mode);
235 *lock_mode = XFS_IOLOCK_EXCL;
236 return xfs_ilock_iocb(iocb, *lock_mode);
237}
238
239static unsigned int
240xfs_ilock_for_write_fault(
241 struct xfs_inode *ip)
242{
243 /* get a shared lock if no remapping in progress */
244 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
245 if (!xfs_iflags_test(ip, XFS_IREMAPPING))
246 return XFS_MMAPLOCK_SHARED;
247
248 /* wait for remapping to complete */
249 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
250 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
251 return XFS_MMAPLOCK_EXCL;
252}
253
254STATIC ssize_t
255xfs_file_dio_read(
256 struct kiocb *iocb,
257 struct iov_iter *to)
258{
259 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
260 ssize_t ret;
261
262 trace_xfs_file_direct_read(iocb, to);
263
264 if (!iov_iter_count(to))
265 return 0; /* skip atime */
266
267 file_accessed(iocb->ki_filp);
268
269 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
270 if (ret)
271 return ret;
272 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
273 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
274
275 return ret;
276}
277
278static noinline ssize_t
279xfs_file_dax_read(
280 struct kiocb *iocb,
281 struct iov_iter *to)
282{
283 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
284 ssize_t ret = 0;
285
286 trace_xfs_file_dax_read(iocb, to);
287
288 if (!iov_iter_count(to))
289 return 0; /* skip atime */
290
291 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
292 if (ret)
293 return ret;
294 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
295 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
296
297 file_accessed(iocb->ki_filp);
298 return ret;
299}
300
301STATIC ssize_t
302xfs_file_buffered_read(
303 struct kiocb *iocb,
304 struct iov_iter *to)
305{
306 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
307 ssize_t ret;
308
309 trace_xfs_file_buffered_read(iocb, to);
310
311 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
312 if (ret)
313 return ret;
314 ret = generic_file_read_iter(iocb, to);
315 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
316
317 return ret;
318}
319
320STATIC ssize_t
321xfs_file_read_iter(
322 struct kiocb *iocb,
323 struct iov_iter *to)
324{
325 struct inode *inode = file_inode(iocb->ki_filp);
326 struct xfs_mount *mp = XFS_I(inode)->i_mount;
327 ssize_t ret = 0;
328
329 XFS_STATS_INC(mp, xs_read_calls);
330
331 if (xfs_is_shutdown(mp))
332 return -EIO;
333
334 if (IS_DAX(inode))
335 ret = xfs_file_dax_read(iocb, to);
336 else if (iocb->ki_flags & IOCB_DIRECT)
337 ret = xfs_file_dio_read(iocb, to);
338 else
339 ret = xfs_file_buffered_read(iocb, to);
340
341 if (ret > 0)
342 XFS_STATS_ADD(mp, xs_read_bytes, ret);
343 return ret;
344}
345
346STATIC ssize_t
347xfs_file_splice_read(
348 struct file *in,
349 loff_t *ppos,
350 struct pipe_inode_info *pipe,
351 size_t len,
352 unsigned int flags)
353{
354 struct inode *inode = file_inode(in);
355 struct xfs_inode *ip = XFS_I(inode);
356 struct xfs_mount *mp = ip->i_mount;
357 ssize_t ret = 0;
358
359 XFS_STATS_INC(mp, xs_read_calls);
360
361 if (xfs_is_shutdown(mp))
362 return -EIO;
363
364 trace_xfs_file_splice_read(ip, *ppos, len);
365
366 xfs_ilock(ip, XFS_IOLOCK_SHARED);
367 ret = filemap_splice_read(in, ppos, pipe, len, flags);
368 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
369 if (ret > 0)
370 XFS_STATS_ADD(mp, xs_read_bytes, ret);
371 return ret;
372}
373
374/*
375 * Common pre-write limit and setup checks.
376 *
377 * Called with the iolocked held either shared and exclusive according to
378 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
379 * if called for a direct write beyond i_size.
380 */
381STATIC ssize_t
382xfs_file_write_checks(
383 struct kiocb *iocb,
384 struct iov_iter *from,
385 unsigned int *iolock)
386{
387 struct file *file = iocb->ki_filp;
388 struct inode *inode = file->f_mapping->host;
389 struct xfs_inode *ip = XFS_I(inode);
390 ssize_t error = 0;
391 size_t count = iov_iter_count(from);
392 bool drained_dio = false;
393 loff_t isize;
394
395restart:
396 error = generic_write_checks(iocb, from);
397 if (error <= 0)
398 return error;
399
400 if (iocb->ki_flags & IOCB_NOWAIT) {
401 error = break_layout(inode, false);
402 if (error == -EWOULDBLOCK)
403 error = -EAGAIN;
404 } else {
405 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
406 }
407
408 if (error)
409 return error;
410
411 /*
412 * For changing security info in file_remove_privs() we need i_rwsem
413 * exclusively.
414 */
415 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
416 xfs_iunlock(ip, *iolock);
417 *iolock = XFS_IOLOCK_EXCL;
418 error = xfs_ilock_iocb(iocb, *iolock);
419 if (error) {
420 *iolock = 0;
421 return error;
422 }
423 goto restart;
424 }
425
426 /*
427 * If the offset is beyond the size of the file, we need to zero any
428 * blocks that fall between the existing EOF and the start of this
429 * write. If zeroing is needed and we are currently holding the iolock
430 * shared, we need to update it to exclusive which implies having to
431 * redo all checks before.
432 *
433 * We need to serialise against EOF updates that occur in IO completions
434 * here. We want to make sure that nobody is changing the size while we
435 * do this check until we have placed an IO barrier (i.e. hold the
436 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
437 * spinlock effectively forms a memory barrier once we have the
438 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
439 * hence be able to correctly determine if we need to run zeroing.
440 *
441 * We can do an unlocked check here safely as IO completion can only
442 * extend EOF. Truncate is locked out at this point, so the EOF can
443 * not move backwards, only forwards. Hence we only need to take the
444 * slow path and spin locks when we are at or beyond the current EOF.
445 */
446 if (iocb->ki_pos <= i_size_read(inode))
447 goto out;
448
449 spin_lock(&ip->i_flags_lock);
450 isize = i_size_read(inode);
451 if (iocb->ki_pos > isize) {
452 spin_unlock(&ip->i_flags_lock);
453
454 if (iocb->ki_flags & IOCB_NOWAIT)
455 return -EAGAIN;
456
457 if (!drained_dio) {
458 if (*iolock == XFS_IOLOCK_SHARED) {
459 xfs_iunlock(ip, *iolock);
460 *iolock = XFS_IOLOCK_EXCL;
461 xfs_ilock(ip, *iolock);
462 iov_iter_reexpand(from, count);
463 }
464 /*
465 * We now have an IO submission barrier in place, but
466 * AIO can do EOF updates during IO completion and hence
467 * we now need to wait for all of them to drain. Non-AIO
468 * DIO will have drained before we are given the
469 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
470 * no-op.
471 */
472 inode_dio_wait(inode);
473 drained_dio = true;
474 goto restart;
475 }
476
477 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
478 error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
479 if (error)
480 return error;
481 } else
482 spin_unlock(&ip->i_flags_lock);
483
484out:
485 return kiocb_modified(iocb);
486}
487
488static int
489xfs_dio_write_end_io(
490 struct kiocb *iocb,
491 ssize_t size,
492 int error,
493 unsigned flags)
494{
495 struct inode *inode = file_inode(iocb->ki_filp);
496 struct xfs_inode *ip = XFS_I(inode);
497 loff_t offset = iocb->ki_pos;
498 unsigned int nofs_flag;
499
500 trace_xfs_end_io_direct_write(ip, offset, size);
501
502 if (xfs_is_shutdown(ip->i_mount))
503 return -EIO;
504
505 if (error)
506 return error;
507 if (!size)
508 return 0;
509
510 /*
511 * Capture amount written on completion as we can't reliably account
512 * for it on submission.
513 */
514 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
515
516 /*
517 * We can allocate memory here while doing writeback on behalf of
518 * memory reclaim. To avoid memory allocation deadlocks set the
519 * task-wide nofs context for the following operations.
520 */
521 nofs_flag = memalloc_nofs_save();
522
523 if (flags & IOMAP_DIO_COW) {
524 error = xfs_reflink_end_cow(ip, offset, size);
525 if (error)
526 goto out;
527 }
528
529 /*
530 * Unwritten conversion updates the in-core isize after extent
531 * conversion but before updating the on-disk size. Updating isize any
532 * earlier allows a racing dio read to find unwritten extents before
533 * they are converted.
534 */
535 if (flags & IOMAP_DIO_UNWRITTEN) {
536 error = xfs_iomap_write_unwritten(ip, offset, size, true);
537 goto out;
538 }
539
540 /*
541 * We need to update the in-core inode size here so that we don't end up
542 * with the on-disk inode size being outside the in-core inode size. We
543 * have no other method of updating EOF for AIO, so always do it here
544 * if necessary.
545 *
546 * We need to lock the test/set EOF update as we can be racing with
547 * other IO completions here to update the EOF. Failing to serialise
548 * here can result in EOF moving backwards and Bad Things Happen when
549 * that occurs.
550 *
551 * As IO completion only ever extends EOF, we can do an unlocked check
552 * here to avoid taking the spinlock. If we land within the current EOF,
553 * then we do not need to do an extending update at all, and we don't
554 * need to take the lock to check this. If we race with an update moving
555 * EOF, then we'll either still be beyond EOF and need to take the lock,
556 * or we'll be within EOF and we don't need to take it at all.
557 */
558 if (offset + size <= i_size_read(inode))
559 goto out;
560
561 spin_lock(&ip->i_flags_lock);
562 if (offset + size > i_size_read(inode)) {
563 i_size_write(inode, offset + size);
564 spin_unlock(&ip->i_flags_lock);
565 error = xfs_setfilesize(ip, offset, size);
566 } else {
567 spin_unlock(&ip->i_flags_lock);
568 }
569
570out:
571 memalloc_nofs_restore(nofs_flag);
572 return error;
573}
574
575static const struct iomap_dio_ops xfs_dio_write_ops = {
576 .end_io = xfs_dio_write_end_io,
577};
578
579/*
580 * Handle block aligned direct I/O writes
581 */
582static noinline ssize_t
583xfs_file_dio_write_aligned(
584 struct xfs_inode *ip,
585 struct kiocb *iocb,
586 struct iov_iter *from)
587{
588 unsigned int iolock = XFS_IOLOCK_SHARED;
589 ssize_t ret;
590
591 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
592 if (ret)
593 return ret;
594 ret = xfs_file_write_checks(iocb, from, &iolock);
595 if (ret)
596 goto out_unlock;
597
598 /*
599 * We don't need to hold the IOLOCK exclusively across the IO, so demote
600 * the iolock back to shared if we had to take the exclusive lock in
601 * xfs_file_write_checks() for other reasons.
602 */
603 if (iolock == XFS_IOLOCK_EXCL) {
604 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
605 iolock = XFS_IOLOCK_SHARED;
606 }
607 trace_xfs_file_direct_write(iocb, from);
608 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
609 &xfs_dio_write_ops, 0, NULL, 0);
610out_unlock:
611 if (iolock)
612 xfs_iunlock(ip, iolock);
613 return ret;
614}
615
616/*
617 * Handle block unaligned direct I/O writes
618 *
619 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
620 * them to be done in parallel with reads and other direct I/O writes. However,
621 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
622 * to do sub-block zeroing and that requires serialisation against other direct
623 * I/O to the same block. In this case we need to serialise the submission of
624 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
625 * In the case where sub-block zeroing is not required, we can do concurrent
626 * sub-block dios to the same block successfully.
627 *
628 * Optimistically submit the I/O using the shared lock first, but use the
629 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
630 * if block allocation or partial block zeroing would be required. In that case
631 * we try again with the exclusive lock.
632 */
633static noinline ssize_t
634xfs_file_dio_write_unaligned(
635 struct xfs_inode *ip,
636 struct kiocb *iocb,
637 struct iov_iter *from)
638{
639 size_t isize = i_size_read(VFS_I(ip));
640 size_t count = iov_iter_count(from);
641 unsigned int iolock = XFS_IOLOCK_SHARED;
642 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
643 ssize_t ret;
644
645 /*
646 * Extending writes need exclusivity because of the sub-block zeroing
647 * that the DIO code always does for partial tail blocks beyond EOF, so
648 * don't even bother trying the fast path in this case.
649 */
650 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
651 if (iocb->ki_flags & IOCB_NOWAIT)
652 return -EAGAIN;
653retry_exclusive:
654 iolock = XFS_IOLOCK_EXCL;
655 flags = IOMAP_DIO_FORCE_WAIT;
656 }
657
658 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
659 if (ret)
660 return ret;
661
662 /*
663 * We can't properly handle unaligned direct I/O to reflink files yet,
664 * as we can't unshare a partial block.
665 */
666 if (xfs_is_cow_inode(ip)) {
667 trace_xfs_reflink_bounce_dio_write(iocb, from);
668 ret = -ENOTBLK;
669 goto out_unlock;
670 }
671
672 ret = xfs_file_write_checks(iocb, from, &iolock);
673 if (ret)
674 goto out_unlock;
675
676 /*
677 * If we are doing exclusive unaligned I/O, this must be the only I/O
678 * in-flight. Otherwise we risk data corruption due to unwritten extent
679 * conversions from the AIO end_io handler. Wait for all other I/O to
680 * drain first.
681 */
682 if (flags & IOMAP_DIO_FORCE_WAIT)
683 inode_dio_wait(VFS_I(ip));
684
685 trace_xfs_file_direct_write(iocb, from);
686 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
687 &xfs_dio_write_ops, flags, NULL, 0);
688
689 /*
690 * Retry unaligned I/O with exclusive blocking semantics if the DIO
691 * layer rejected it for mapping or locking reasons. If we are doing
692 * nonblocking user I/O, propagate the error.
693 */
694 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
695 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
696 xfs_iunlock(ip, iolock);
697 goto retry_exclusive;
698 }
699
700out_unlock:
701 if (iolock)
702 xfs_iunlock(ip, iolock);
703 return ret;
704}
705
706static ssize_t
707xfs_file_dio_write(
708 struct kiocb *iocb,
709 struct iov_iter *from)
710{
711 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
712 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
713 size_t count = iov_iter_count(from);
714
715 /* direct I/O must be aligned to device logical sector size */
716 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
717 return -EINVAL;
718 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
719 return xfs_file_dio_write_unaligned(ip, iocb, from);
720 return xfs_file_dio_write_aligned(ip, iocb, from);
721}
722
723static noinline ssize_t
724xfs_file_dax_write(
725 struct kiocb *iocb,
726 struct iov_iter *from)
727{
728 struct inode *inode = iocb->ki_filp->f_mapping->host;
729 struct xfs_inode *ip = XFS_I(inode);
730 unsigned int iolock = XFS_IOLOCK_EXCL;
731 ssize_t ret, error = 0;
732 loff_t pos;
733
734 ret = xfs_ilock_iocb(iocb, iolock);
735 if (ret)
736 return ret;
737 ret = xfs_file_write_checks(iocb, from, &iolock);
738 if (ret)
739 goto out;
740
741 pos = iocb->ki_pos;
742
743 trace_xfs_file_dax_write(iocb, from);
744 ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
745 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
746 i_size_write(inode, iocb->ki_pos);
747 error = xfs_setfilesize(ip, pos, ret);
748 }
749out:
750 if (iolock)
751 xfs_iunlock(ip, iolock);
752 if (error)
753 return error;
754
755 if (ret > 0) {
756 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
757
758 /* Handle various SYNC-type writes */
759 ret = generic_write_sync(iocb, ret);
760 }
761 return ret;
762}
763
764STATIC ssize_t
765xfs_file_buffered_write(
766 struct kiocb *iocb,
767 struct iov_iter *from)
768{
769 struct inode *inode = iocb->ki_filp->f_mapping->host;
770 struct xfs_inode *ip = XFS_I(inode);
771 ssize_t ret;
772 bool cleared_space = false;
773 unsigned int iolock;
774
775write_retry:
776 iolock = XFS_IOLOCK_EXCL;
777 ret = xfs_ilock_iocb(iocb, iolock);
778 if (ret)
779 return ret;
780
781 ret = xfs_file_write_checks(iocb, from, &iolock);
782 if (ret)
783 goto out;
784
785 trace_xfs_file_buffered_write(iocb, from);
786 ret = iomap_file_buffered_write(iocb, from,
787 &xfs_buffered_write_iomap_ops);
788
789 /*
790 * If we hit a space limit, try to free up some lingering preallocated
791 * space before returning an error. In the case of ENOSPC, first try to
792 * write back all dirty inodes to free up some of the excess reserved
793 * metadata space. This reduces the chances that the eofblocks scan
794 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
795 * also behaves as a filter to prevent too many eofblocks scans from
796 * running at the same time. Use a synchronous scan to increase the
797 * effectiveness of the scan.
798 */
799 if (ret == -EDQUOT && !cleared_space) {
800 xfs_iunlock(ip, iolock);
801 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
802 cleared_space = true;
803 goto write_retry;
804 } else if (ret == -ENOSPC && !cleared_space) {
805 struct xfs_icwalk icw = {0};
806
807 cleared_space = true;
808 xfs_flush_inodes(ip->i_mount);
809
810 xfs_iunlock(ip, iolock);
811 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
812 xfs_blockgc_free_space(ip->i_mount, &icw);
813 goto write_retry;
814 }
815
816out:
817 if (iolock)
818 xfs_iunlock(ip, iolock);
819
820 if (ret > 0) {
821 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
822 /* Handle various SYNC-type writes */
823 ret = generic_write_sync(iocb, ret);
824 }
825 return ret;
826}
827
828STATIC ssize_t
829xfs_file_write_iter(
830 struct kiocb *iocb,
831 struct iov_iter *from)
832{
833 struct inode *inode = iocb->ki_filp->f_mapping->host;
834 struct xfs_inode *ip = XFS_I(inode);
835 ssize_t ret;
836 size_t ocount = iov_iter_count(from);
837
838 XFS_STATS_INC(ip->i_mount, xs_write_calls);
839
840 if (ocount == 0)
841 return 0;
842
843 if (xfs_is_shutdown(ip->i_mount))
844 return -EIO;
845
846 if (IS_DAX(inode))
847 return xfs_file_dax_write(iocb, from);
848
849 if (iocb->ki_flags & IOCB_DIRECT) {
850 /*
851 * Allow a directio write to fall back to a buffered
852 * write *only* in the case that we're doing a reflink
853 * CoW. In all other directio scenarios we do not
854 * allow an operation to fall back to buffered mode.
855 */
856 ret = xfs_file_dio_write(iocb, from);
857 if (ret != -ENOTBLK)
858 return ret;
859 }
860
861 return xfs_file_buffered_write(iocb, from);
862}
863
864static void
865xfs_wait_dax_page(
866 struct inode *inode)
867{
868 struct xfs_inode *ip = XFS_I(inode);
869
870 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
871 schedule();
872 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
873}
874
875int
876xfs_break_dax_layouts(
877 struct inode *inode,
878 bool *retry)
879{
880 struct page *page;
881
882 ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
883
884 page = dax_layout_busy_page(inode->i_mapping);
885 if (!page)
886 return 0;
887
888 *retry = true;
889 return ___wait_var_event(&page->_refcount,
890 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
891 0, 0, xfs_wait_dax_page(inode));
892}
893
894int
895xfs_break_layouts(
896 struct inode *inode,
897 uint *iolock,
898 enum layout_break_reason reason)
899{
900 bool retry;
901 int error;
902
903 ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
904
905 do {
906 retry = false;
907 switch (reason) {
908 case BREAK_UNMAP:
909 error = xfs_break_dax_layouts(inode, &retry);
910 if (error || retry)
911 break;
912 fallthrough;
913 case BREAK_WRITE:
914 error = xfs_break_leased_layouts(inode, iolock, &retry);
915 break;
916 default:
917 WARN_ON_ONCE(1);
918 error = -EINVAL;
919 }
920 } while (error == 0 && retry);
921
922 return error;
923}
924
925/* Does this file, inode, or mount want synchronous writes? */
926static inline bool xfs_file_sync_writes(struct file *filp)
927{
928 struct xfs_inode *ip = XFS_I(file_inode(filp));
929
930 if (xfs_has_wsync(ip->i_mount))
931 return true;
932 if (filp->f_flags & (__O_SYNC | O_DSYNC))
933 return true;
934 if (IS_SYNC(file_inode(filp)))
935 return true;
936
937 return false;
938}
939
940#define XFS_FALLOC_FL_SUPPORTED \
941 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
942 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
943 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
944
945STATIC long
946xfs_file_fallocate(
947 struct file *file,
948 int mode,
949 loff_t offset,
950 loff_t len)
951{
952 struct inode *inode = file_inode(file);
953 struct xfs_inode *ip = XFS_I(inode);
954 long error;
955 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
956 loff_t new_size = 0;
957 bool do_file_insert = false;
958
959 if (!S_ISREG(inode->i_mode))
960 return -EINVAL;
961 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
962 return -EOPNOTSUPP;
963
964 xfs_ilock(ip, iolock);
965 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
966 if (error)
967 goto out_unlock;
968
969 /*
970 * Must wait for all AIO to complete before we continue as AIO can
971 * change the file size on completion without holding any locks we
972 * currently hold. We must do this first because AIO can update both
973 * the on disk and in memory inode sizes, and the operations that follow
974 * require the in-memory size to be fully up-to-date.
975 */
976 inode_dio_wait(inode);
977
978 /*
979 * Now AIO and DIO has drained we flush and (if necessary) invalidate
980 * the cached range over the first operation we are about to run.
981 *
982 * We care about zero and collapse here because they both run a hole
983 * punch over the range first. Because that can zero data, and the range
984 * of invalidation for the shift operations is much larger, we still do
985 * the required flush for collapse in xfs_prepare_shift().
986 *
987 * Insert has the same range requirements as collapse, and we extend the
988 * file first which can zero data. Hence insert has the same
989 * flush/invalidate requirements as collapse and so they are both
990 * handled at the right time by xfs_prepare_shift().
991 */
992 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
993 FALLOC_FL_COLLAPSE_RANGE)) {
994 error = xfs_flush_unmap_range(ip, offset, len);
995 if (error)
996 goto out_unlock;
997 }
998
999 error = file_modified(file);
1000 if (error)
1001 goto out_unlock;
1002
1003 if (mode & FALLOC_FL_PUNCH_HOLE) {
1004 error = xfs_free_file_space(ip, offset, len);
1005 if (error)
1006 goto out_unlock;
1007 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
1008 if (!xfs_is_falloc_aligned(ip, offset, len)) {
1009 error = -EINVAL;
1010 goto out_unlock;
1011 }
1012
1013 /*
1014 * There is no need to overlap collapse range with EOF,
1015 * in which case it is effectively a truncate operation
1016 */
1017 if (offset + len >= i_size_read(inode)) {
1018 error = -EINVAL;
1019 goto out_unlock;
1020 }
1021
1022 new_size = i_size_read(inode) - len;
1023
1024 error = xfs_collapse_file_space(ip, offset, len);
1025 if (error)
1026 goto out_unlock;
1027 } else if (mode & FALLOC_FL_INSERT_RANGE) {
1028 loff_t isize = i_size_read(inode);
1029
1030 if (!xfs_is_falloc_aligned(ip, offset, len)) {
1031 error = -EINVAL;
1032 goto out_unlock;
1033 }
1034
1035 /*
1036 * New inode size must not exceed ->s_maxbytes, accounting for
1037 * possible signed overflow.
1038 */
1039 if (inode->i_sb->s_maxbytes - isize < len) {
1040 error = -EFBIG;
1041 goto out_unlock;
1042 }
1043 new_size = isize + len;
1044
1045 /* Offset should be less than i_size */
1046 if (offset >= isize) {
1047 error = -EINVAL;
1048 goto out_unlock;
1049 }
1050 do_file_insert = true;
1051 } else {
1052 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
1053 offset + len > i_size_read(inode)) {
1054 new_size = offset + len;
1055 error = inode_newsize_ok(inode, new_size);
1056 if (error)
1057 goto out_unlock;
1058 }
1059
1060 if (mode & FALLOC_FL_ZERO_RANGE) {
1061 /*
1062 * Punch a hole and prealloc the range. We use a hole
1063 * punch rather than unwritten extent conversion for two
1064 * reasons:
1065 *
1066 * 1.) Hole punch handles partial block zeroing for us.
1067 * 2.) If prealloc returns ENOSPC, the file range is
1068 * still zero-valued by virtue of the hole punch.
1069 */
1070 unsigned int blksize = i_blocksize(inode);
1071
1072 trace_xfs_zero_file_space(ip);
1073
1074 error = xfs_free_file_space(ip, offset, len);
1075 if (error)
1076 goto out_unlock;
1077
1078 len = round_up(offset + len, blksize) -
1079 round_down(offset, blksize);
1080 offset = round_down(offset, blksize);
1081 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
1082 error = xfs_reflink_unshare(ip, offset, len);
1083 if (error)
1084 goto out_unlock;
1085 } else {
1086 /*
1087 * If always_cow mode we can't use preallocations and
1088 * thus should not create them.
1089 */
1090 if (xfs_is_always_cow_inode(ip)) {
1091 error = -EOPNOTSUPP;
1092 goto out_unlock;
1093 }
1094 }
1095
1096 if (!xfs_is_always_cow_inode(ip)) {
1097 error = xfs_alloc_file_space(ip, offset, len);
1098 if (error)
1099 goto out_unlock;
1100 }
1101 }
1102
1103 /* Change file size if needed */
1104 if (new_size) {
1105 struct iattr iattr;
1106
1107 iattr.ia_valid = ATTR_SIZE;
1108 iattr.ia_size = new_size;
1109 error = xfs_vn_setattr_size(file_mnt_idmap(file),
1110 file_dentry(file), &iattr);
1111 if (error)
1112 goto out_unlock;
1113 }
1114
1115 /*
1116 * Perform hole insertion now that the file size has been
1117 * updated so that if we crash during the operation we don't
1118 * leave shifted extents past EOF and hence losing access to
1119 * the data that is contained within them.
1120 */
1121 if (do_file_insert) {
1122 error = xfs_insert_file_space(ip, offset, len);
1123 if (error)
1124 goto out_unlock;
1125 }
1126
1127 if (xfs_file_sync_writes(file))
1128 error = xfs_log_force_inode(ip);
1129
1130out_unlock:
1131 xfs_iunlock(ip, iolock);
1132 return error;
1133}
1134
1135STATIC int
1136xfs_file_fadvise(
1137 struct file *file,
1138 loff_t start,
1139 loff_t end,
1140 int advice)
1141{
1142 struct xfs_inode *ip = XFS_I(file_inode(file));
1143 int ret;
1144 int lockflags = 0;
1145
1146 /*
1147 * Operations creating pages in page cache need protection from hole
1148 * punching and similar ops
1149 */
1150 if (advice == POSIX_FADV_WILLNEED) {
1151 lockflags = XFS_IOLOCK_SHARED;
1152 xfs_ilock(ip, lockflags);
1153 }
1154 ret = generic_fadvise(file, start, end, advice);
1155 if (lockflags)
1156 xfs_iunlock(ip, lockflags);
1157 return ret;
1158}
1159
1160STATIC loff_t
1161xfs_file_remap_range(
1162 struct file *file_in,
1163 loff_t pos_in,
1164 struct file *file_out,
1165 loff_t pos_out,
1166 loff_t len,
1167 unsigned int remap_flags)
1168{
1169 struct inode *inode_in = file_inode(file_in);
1170 struct xfs_inode *src = XFS_I(inode_in);
1171 struct inode *inode_out = file_inode(file_out);
1172 struct xfs_inode *dest = XFS_I(inode_out);
1173 struct xfs_mount *mp = src->i_mount;
1174 loff_t remapped = 0;
1175 xfs_extlen_t cowextsize;
1176 int ret;
1177
1178 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1179 return -EINVAL;
1180
1181 if (!xfs_has_reflink(mp))
1182 return -EOPNOTSUPP;
1183
1184 if (xfs_is_shutdown(mp))
1185 return -EIO;
1186
1187 /* Prepare and then clone file data. */
1188 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1189 &len, remap_flags);
1190 if (ret || len == 0)
1191 return ret;
1192
1193 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1194
1195 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1196 &remapped);
1197 if (ret)
1198 goto out_unlock;
1199
1200 /*
1201 * Carry the cowextsize hint from src to dest if we're sharing the
1202 * entire source file to the entire destination file, the source file
1203 * has a cowextsize hint, and the destination file does not.
1204 */
1205 cowextsize = 0;
1206 if (pos_in == 0 && len == i_size_read(inode_in) &&
1207 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1208 pos_out == 0 && len >= i_size_read(inode_out) &&
1209 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1210 cowextsize = src->i_cowextsize;
1211
1212 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1213 remap_flags);
1214 if (ret)
1215 goto out_unlock;
1216
1217 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1218 xfs_log_force_inode(dest);
1219out_unlock:
1220 xfs_iunlock2_remapping(src, dest);
1221 if (ret)
1222 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1223 return remapped > 0 ? remapped : ret;
1224}
1225
1226STATIC int
1227xfs_file_open(
1228 struct inode *inode,
1229 struct file *file)
1230{
1231 if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1232 return -EIO;
1233 file->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC |
1234 FMODE_DIO_PARALLEL_WRITE | FMODE_CAN_ODIRECT;
1235 return generic_file_open(inode, file);
1236}
1237
1238STATIC int
1239xfs_dir_open(
1240 struct inode *inode,
1241 struct file *file)
1242{
1243 struct xfs_inode *ip = XFS_I(inode);
1244 unsigned int mode;
1245 int error;
1246
1247 error = xfs_file_open(inode, file);
1248 if (error)
1249 return error;
1250
1251 /*
1252 * If there are any blocks, read-ahead block 0 as we're almost
1253 * certain to have the next operation be a read there.
1254 */
1255 mode = xfs_ilock_data_map_shared(ip);
1256 if (ip->i_df.if_nextents > 0)
1257 error = xfs_dir3_data_readahead(ip, 0, 0);
1258 xfs_iunlock(ip, mode);
1259 return error;
1260}
1261
1262STATIC int
1263xfs_file_release(
1264 struct inode *inode,
1265 struct file *filp)
1266{
1267 return xfs_release(XFS_I(inode));
1268}
1269
1270STATIC int
1271xfs_file_readdir(
1272 struct file *file,
1273 struct dir_context *ctx)
1274{
1275 struct inode *inode = file_inode(file);
1276 xfs_inode_t *ip = XFS_I(inode);
1277 size_t bufsize;
1278
1279 /*
1280 * The Linux API doesn't pass down the total size of the buffer
1281 * we read into down to the filesystem. With the filldir concept
1282 * it's not needed for correct information, but the XFS dir2 leaf
1283 * code wants an estimate of the buffer size to calculate it's
1284 * readahead window and size the buffers used for mapping to
1285 * physical blocks.
1286 *
1287 * Try to give it an estimate that's good enough, maybe at some
1288 * point we can change the ->readdir prototype to include the
1289 * buffer size. For now we use the current glibc buffer size.
1290 */
1291 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1292
1293 return xfs_readdir(NULL, ip, ctx, bufsize);
1294}
1295
1296STATIC loff_t
1297xfs_file_llseek(
1298 struct file *file,
1299 loff_t offset,
1300 int whence)
1301{
1302 struct inode *inode = file->f_mapping->host;
1303
1304 if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1305 return -EIO;
1306
1307 switch (whence) {
1308 default:
1309 return generic_file_llseek(file, offset, whence);
1310 case SEEK_HOLE:
1311 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1312 break;
1313 case SEEK_DATA:
1314 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1315 break;
1316 }
1317
1318 if (offset < 0)
1319 return offset;
1320 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1321}
1322
1323#ifdef CONFIG_FS_DAX
1324static inline vm_fault_t
1325xfs_dax_fault(
1326 struct vm_fault *vmf,
1327 unsigned int order,
1328 bool write_fault,
1329 pfn_t *pfn)
1330{
1331 return dax_iomap_fault(vmf, order, pfn, NULL,
1332 (write_fault && !vmf->cow_page) ?
1333 &xfs_dax_write_iomap_ops :
1334 &xfs_read_iomap_ops);
1335}
1336#else
1337static inline vm_fault_t
1338xfs_dax_fault(
1339 struct vm_fault *vmf,
1340 unsigned int order,
1341 bool write_fault,
1342 pfn_t *pfn)
1343{
1344 ASSERT(0);
1345 return VM_FAULT_SIGBUS;
1346}
1347#endif
1348
1349/*
1350 * Locking for serialisation of IO during page faults. This results in a lock
1351 * ordering of:
1352 *
1353 * mmap_lock (MM)
1354 * sb_start_pagefault(vfs, freeze)
1355 * invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1356 * page_lock (MM)
1357 * i_lock (XFS - extent map serialisation)
1358 */
1359static vm_fault_t
1360__xfs_filemap_fault(
1361 struct vm_fault *vmf,
1362 unsigned int order,
1363 bool write_fault)
1364{
1365 struct inode *inode = file_inode(vmf->vma->vm_file);
1366 struct xfs_inode *ip = XFS_I(inode);
1367 vm_fault_t ret;
1368 unsigned int lock_mode = 0;
1369
1370 trace_xfs_filemap_fault(ip, order, write_fault);
1371
1372 if (write_fault) {
1373 sb_start_pagefault(inode->i_sb);
1374 file_update_time(vmf->vma->vm_file);
1375 }
1376
1377 if (IS_DAX(inode) || write_fault)
1378 lock_mode = xfs_ilock_for_write_fault(XFS_I(inode));
1379
1380 if (IS_DAX(inode)) {
1381 pfn_t pfn;
1382
1383 ret = xfs_dax_fault(vmf, order, write_fault, &pfn);
1384 if (ret & VM_FAULT_NEEDDSYNC)
1385 ret = dax_finish_sync_fault(vmf, order, pfn);
1386 } else if (write_fault) {
1387 ret = iomap_page_mkwrite(vmf, &xfs_page_mkwrite_iomap_ops);
1388 } else {
1389 ret = filemap_fault(vmf);
1390 }
1391
1392 if (lock_mode)
1393 xfs_iunlock(XFS_I(inode), lock_mode);
1394
1395 if (write_fault)
1396 sb_end_pagefault(inode->i_sb);
1397 return ret;
1398}
1399
1400static inline bool
1401xfs_is_write_fault(
1402 struct vm_fault *vmf)
1403{
1404 return (vmf->flags & FAULT_FLAG_WRITE) &&
1405 (vmf->vma->vm_flags & VM_SHARED);
1406}
1407
1408static vm_fault_t
1409xfs_filemap_fault(
1410 struct vm_fault *vmf)
1411{
1412 /* DAX can shortcut the normal fault path on write faults! */
1413 return __xfs_filemap_fault(vmf, 0,
1414 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1415 xfs_is_write_fault(vmf));
1416}
1417
1418static vm_fault_t
1419xfs_filemap_huge_fault(
1420 struct vm_fault *vmf,
1421 unsigned int order)
1422{
1423 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1424 return VM_FAULT_FALLBACK;
1425
1426 /* DAX can shortcut the normal fault path on write faults! */
1427 return __xfs_filemap_fault(vmf, order,
1428 xfs_is_write_fault(vmf));
1429}
1430
1431static vm_fault_t
1432xfs_filemap_page_mkwrite(
1433 struct vm_fault *vmf)
1434{
1435 return __xfs_filemap_fault(vmf, 0, true);
1436}
1437
1438/*
1439 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1440 * on write faults. In reality, it needs to serialise against truncate and
1441 * prepare memory for writing so handle is as standard write fault.
1442 */
1443static vm_fault_t
1444xfs_filemap_pfn_mkwrite(
1445 struct vm_fault *vmf)
1446{
1447
1448 return __xfs_filemap_fault(vmf, 0, true);
1449}
1450
1451static const struct vm_operations_struct xfs_file_vm_ops = {
1452 .fault = xfs_filemap_fault,
1453 .huge_fault = xfs_filemap_huge_fault,
1454 .map_pages = filemap_map_pages,
1455 .page_mkwrite = xfs_filemap_page_mkwrite,
1456 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1457};
1458
1459STATIC int
1460xfs_file_mmap(
1461 struct file *file,
1462 struct vm_area_struct *vma)
1463{
1464 struct inode *inode = file_inode(file);
1465 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1466
1467 /*
1468 * We don't support synchronous mappings for non-DAX files and
1469 * for DAX files if underneath dax_device is not synchronous.
1470 */
1471 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1472 return -EOPNOTSUPP;
1473
1474 file_accessed(file);
1475 vma->vm_ops = &xfs_file_vm_ops;
1476 if (IS_DAX(inode))
1477 vm_flags_set(vma, VM_HUGEPAGE);
1478 return 0;
1479}
1480
1481const struct file_operations xfs_file_operations = {
1482 .llseek = xfs_file_llseek,
1483 .read_iter = xfs_file_read_iter,
1484 .write_iter = xfs_file_write_iter,
1485 .splice_read = xfs_file_splice_read,
1486 .splice_write = iter_file_splice_write,
1487 .iopoll = iocb_bio_iopoll,
1488 .unlocked_ioctl = xfs_file_ioctl,
1489#ifdef CONFIG_COMPAT
1490 .compat_ioctl = xfs_file_compat_ioctl,
1491#endif
1492 .mmap = xfs_file_mmap,
1493 .mmap_supported_flags = MAP_SYNC,
1494 .open = xfs_file_open,
1495 .release = xfs_file_release,
1496 .fsync = xfs_file_fsync,
1497 .get_unmapped_area = thp_get_unmapped_area,
1498 .fallocate = xfs_file_fallocate,
1499 .fadvise = xfs_file_fadvise,
1500 .remap_file_range = xfs_file_remap_range,
1501};
1502
1503const struct file_operations xfs_dir_file_operations = {
1504 .open = xfs_dir_open,
1505 .read = generic_read_dir,
1506 .iterate_shared = xfs_file_readdir,
1507 .llseek = generic_file_llseek,
1508 .unlocked_ioctl = xfs_file_ioctl,
1509#ifdef CONFIG_COMPAT
1510 .compat_ioctl = xfs_file_compat_ioctl,
1511#endif
1512 .fsync = xfs_dir_fsync,
1513};
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_mount.h"
25#include "xfs_da_format.h"
26#include "xfs_da_btree.h"
27#include "xfs_inode.h"
28#include "xfs_trans.h"
29#include "xfs_inode_item.h"
30#include "xfs_bmap.h"
31#include "xfs_bmap_util.h"
32#include "xfs_error.h"
33#include "xfs_dir2.h"
34#include "xfs_dir2_priv.h"
35#include "xfs_ioctl.h"
36#include "xfs_trace.h"
37#include "xfs_log.h"
38#include "xfs_icache.h"
39#include "xfs_pnfs.h"
40
41#include <linux/dcache.h>
42#include <linux/falloc.h>
43#include <linux/pagevec.h>
44#include <linux/backing-dev.h>
45
46static const struct vm_operations_struct xfs_file_vm_ops;
47
48/*
49 * Locking primitives for read and write IO paths to ensure we consistently use
50 * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
51 */
52static inline void
53xfs_rw_ilock(
54 struct xfs_inode *ip,
55 int type)
56{
57 if (type & XFS_IOLOCK_EXCL)
58 inode_lock(VFS_I(ip));
59 xfs_ilock(ip, type);
60}
61
62static inline void
63xfs_rw_iunlock(
64 struct xfs_inode *ip,
65 int type)
66{
67 xfs_iunlock(ip, type);
68 if (type & XFS_IOLOCK_EXCL)
69 inode_unlock(VFS_I(ip));
70}
71
72static inline void
73xfs_rw_ilock_demote(
74 struct xfs_inode *ip,
75 int type)
76{
77 xfs_ilock_demote(ip, type);
78 if (type & XFS_IOLOCK_EXCL)
79 inode_unlock(VFS_I(ip));
80}
81
82/*
83 * xfs_iozero clears the specified range supplied via the page cache (except in
84 * the DAX case). Writes through the page cache will allocate blocks over holes,
85 * though the callers usually map the holes first and avoid them. If a block is
86 * not completely zeroed, then it will be read from disk before being partially
87 * zeroed.
88 *
89 * In the DAX case, we can just directly write to the underlying pages. This
90 * will not allocate blocks, but will avoid holes and unwritten extents and so
91 * not do unnecessary work.
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 = 0;
102
103
104 mapping = VFS_I(ip)->i_mapping;
105 do {
106 unsigned offset, bytes;
107 void *fsdata;
108
109 offset = (pos & (PAGE_SIZE -1)); /* Within page */
110 bytes = PAGE_SIZE - offset;
111 if (bytes > count)
112 bytes = count;
113
114 if (IS_DAX(VFS_I(ip))) {
115 status = dax_zero_page_range(VFS_I(ip), pos, bytes,
116 xfs_get_blocks_direct);
117 if (status)
118 break;
119 } else {
120 status = pagecache_write_begin(NULL, mapping, pos, bytes,
121 AOP_FLAG_UNINTERRUPTIBLE,
122 &page, &fsdata);
123 if (status)
124 break;
125
126 zero_user(page, offset, bytes);
127
128 status = pagecache_write_end(NULL, mapping, pos, bytes,
129 bytes, page, fsdata);
130 WARN_ON(status <= 0); /* can't return less than zero! */
131 status = 0;
132 }
133 pos += bytes;
134 count -= bytes;
135 } while (count);
136
137 return status;
138}
139
140int
141xfs_update_prealloc_flags(
142 struct xfs_inode *ip,
143 enum xfs_prealloc_flags flags)
144{
145 struct xfs_trans *tp;
146 int error;
147
148 tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_WRITEID);
149 error = xfs_trans_reserve(tp, &M_RES(ip->i_mount)->tr_writeid, 0, 0);
150 if (error) {
151 xfs_trans_cancel(tp);
152 return error;
153 }
154
155 xfs_ilock(ip, XFS_ILOCK_EXCL);
156 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
157
158 if (!(flags & XFS_PREALLOC_INVISIBLE)) {
159 VFS_I(ip)->i_mode &= ~S_ISUID;
160 if (VFS_I(ip)->i_mode & S_IXGRP)
161 VFS_I(ip)->i_mode &= ~S_ISGID;
162 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
163 }
164
165 if (flags & XFS_PREALLOC_SET)
166 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
167 if (flags & XFS_PREALLOC_CLEAR)
168 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
169
170 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
171 if (flags & XFS_PREALLOC_SYNC)
172 xfs_trans_set_sync(tp);
173 return xfs_trans_commit(tp);
174}
175
176/*
177 * Fsync operations on directories are much simpler than on regular files,
178 * as there is no file data to flush, and thus also no need for explicit
179 * cache flush operations, and there are no non-transaction metadata updates
180 * on directories either.
181 */
182STATIC int
183xfs_dir_fsync(
184 struct file *file,
185 loff_t start,
186 loff_t end,
187 int datasync)
188{
189 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
190 struct xfs_mount *mp = ip->i_mount;
191 xfs_lsn_t lsn = 0;
192
193 trace_xfs_dir_fsync(ip);
194
195 xfs_ilock(ip, XFS_ILOCK_SHARED);
196 if (xfs_ipincount(ip))
197 lsn = ip->i_itemp->ili_last_lsn;
198 xfs_iunlock(ip, XFS_ILOCK_SHARED);
199
200 if (!lsn)
201 return 0;
202 return _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, NULL);
203}
204
205STATIC int
206xfs_file_fsync(
207 struct file *file,
208 loff_t start,
209 loff_t end,
210 int datasync)
211{
212 struct inode *inode = file->f_mapping->host;
213 struct xfs_inode *ip = XFS_I(inode);
214 struct xfs_mount *mp = ip->i_mount;
215 int error = 0;
216 int log_flushed = 0;
217 xfs_lsn_t lsn = 0;
218
219 trace_xfs_file_fsync(ip);
220
221 error = filemap_write_and_wait_range(inode->i_mapping, start, end);
222 if (error)
223 return error;
224
225 if (XFS_FORCED_SHUTDOWN(mp))
226 return -EIO;
227
228 xfs_iflags_clear(ip, XFS_ITRUNCATED);
229
230 if (mp->m_flags & XFS_MOUNT_BARRIER) {
231 /*
232 * If we have an RT and/or log subvolume we need to make sure
233 * to flush the write cache the device used for file data
234 * first. This is to ensure newly written file data make
235 * it to disk before logging the new inode size in case of
236 * an extending write.
237 */
238 if (XFS_IS_REALTIME_INODE(ip))
239 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
240 else if (mp->m_logdev_targp != mp->m_ddev_targp)
241 xfs_blkdev_issue_flush(mp->m_ddev_targp);
242 }
243
244 /*
245 * All metadata updates are logged, which means that we just have to
246 * flush the log up to the latest LSN that touched the inode. If we have
247 * concurrent fsync/fdatasync() calls, we need them to all block on the
248 * log force before we clear the ili_fsync_fields field. This ensures
249 * that we don't get a racing sync operation that does not wait for the
250 * metadata to hit the journal before returning. If we race with
251 * clearing the ili_fsync_fields, then all that will happen is the log
252 * force will do nothing as the lsn will already be on disk. We can't
253 * race with setting ili_fsync_fields because that is done under
254 * XFS_ILOCK_EXCL, and that can't happen because we hold the lock shared
255 * until after the ili_fsync_fields is cleared.
256 */
257 xfs_ilock(ip, XFS_ILOCK_SHARED);
258 if (xfs_ipincount(ip)) {
259 if (!datasync ||
260 (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
261 lsn = ip->i_itemp->ili_last_lsn;
262 }
263
264 if (lsn) {
265 error = _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
266 ip->i_itemp->ili_fsync_fields = 0;
267 }
268 xfs_iunlock(ip, XFS_ILOCK_SHARED);
269
270 /*
271 * If we only have a single device, and the log force about was
272 * a no-op we might have to flush the data device cache here.
273 * This can only happen for fdatasync/O_DSYNC if we were overwriting
274 * an already allocated file and thus do not have any metadata to
275 * commit.
276 */
277 if ((mp->m_flags & XFS_MOUNT_BARRIER) &&
278 mp->m_logdev_targp == mp->m_ddev_targp &&
279 !XFS_IS_REALTIME_INODE(ip) &&
280 !log_flushed)
281 xfs_blkdev_issue_flush(mp->m_ddev_targp);
282
283 return error;
284}
285
286STATIC ssize_t
287xfs_file_read_iter(
288 struct kiocb *iocb,
289 struct iov_iter *to)
290{
291 struct file *file = iocb->ki_filp;
292 struct inode *inode = file->f_mapping->host;
293 struct xfs_inode *ip = XFS_I(inode);
294 struct xfs_mount *mp = ip->i_mount;
295 size_t size = iov_iter_count(to);
296 ssize_t ret = 0;
297 int ioflags = 0;
298 xfs_fsize_t n;
299 loff_t pos = iocb->ki_pos;
300
301 XFS_STATS_INC(mp, xs_read_calls);
302
303 if (unlikely(iocb->ki_flags & IOCB_DIRECT))
304 ioflags |= XFS_IO_ISDIRECT;
305 if (file->f_mode & FMODE_NOCMTIME)
306 ioflags |= XFS_IO_INVIS;
307
308 if ((ioflags & XFS_IO_ISDIRECT) && !IS_DAX(inode)) {
309 xfs_buftarg_t *target =
310 XFS_IS_REALTIME_INODE(ip) ?
311 mp->m_rtdev_targp : mp->m_ddev_targp;
312 /* DIO must be aligned to device logical sector size */
313 if ((pos | size) & target->bt_logical_sectormask) {
314 if (pos == i_size_read(inode))
315 return 0;
316 return -EINVAL;
317 }
318 }
319
320 n = mp->m_super->s_maxbytes - pos;
321 if (n <= 0 || size == 0)
322 return 0;
323
324 if (n < size)
325 size = n;
326
327 if (XFS_FORCED_SHUTDOWN(mp))
328 return -EIO;
329
330 /*
331 * Locking is a bit tricky here. If we take an exclusive lock for direct
332 * IO, we effectively serialise all new concurrent read IO to this file
333 * and block it behind IO that is currently in progress because IO in
334 * progress holds the IO lock shared. We only need to hold the lock
335 * exclusive to blow away the page cache, so only take lock exclusively
336 * if the page cache needs invalidation. This allows the normal direct
337 * IO case of no page cache pages to proceeed concurrently without
338 * serialisation.
339 */
340 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
341 if ((ioflags & XFS_IO_ISDIRECT) && inode->i_mapping->nrpages) {
342 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
343 xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
344
345 /*
346 * The generic dio code only flushes the range of the particular
347 * I/O. Because we take an exclusive lock here, this whole
348 * sequence is considerably more expensive for us. This has a
349 * noticeable performance impact for any file with cached pages,
350 * even when outside of the range of the particular I/O.
351 *
352 * Hence, amortize the cost of the lock against a full file
353 * flush and reduce the chances of repeated iolock cycles going
354 * forward.
355 */
356 if (inode->i_mapping->nrpages) {
357 ret = filemap_write_and_wait(VFS_I(ip)->i_mapping);
358 if (ret) {
359 xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
360 return ret;
361 }
362
363 /*
364 * Invalidate whole pages. This can return an error if
365 * we fail to invalidate a page, but this should never
366 * happen on XFS. Warn if it does fail.
367 */
368 ret = invalidate_inode_pages2(VFS_I(ip)->i_mapping);
369 WARN_ON_ONCE(ret);
370 ret = 0;
371 }
372 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
373 }
374
375 trace_xfs_file_read(ip, size, pos, ioflags);
376
377 ret = generic_file_read_iter(iocb, to);
378 if (ret > 0)
379 XFS_STATS_ADD(mp, xs_read_bytes, ret);
380
381 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
382 return ret;
383}
384
385STATIC ssize_t
386xfs_file_splice_read(
387 struct file *infilp,
388 loff_t *ppos,
389 struct pipe_inode_info *pipe,
390 size_t count,
391 unsigned int flags)
392{
393 struct xfs_inode *ip = XFS_I(infilp->f_mapping->host);
394 int ioflags = 0;
395 ssize_t ret;
396
397 XFS_STATS_INC(ip->i_mount, xs_read_calls);
398
399 if (infilp->f_mode & FMODE_NOCMTIME)
400 ioflags |= XFS_IO_INVIS;
401
402 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
403 return -EIO;
404
405 trace_xfs_file_splice_read(ip, count, *ppos, ioflags);
406
407 /*
408 * DAX inodes cannot ues the page cache for splice, so we have to push
409 * them through the VFS IO path. This means it goes through
410 * ->read_iter, which for us takes the XFS_IOLOCK_SHARED. Hence we
411 * cannot lock the splice operation at this level for DAX inodes.
412 */
413 if (IS_DAX(VFS_I(ip))) {
414 ret = default_file_splice_read(infilp, ppos, pipe, count,
415 flags);
416 goto out;
417 }
418
419 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
420 ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
421 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
422out:
423 if (ret > 0)
424 XFS_STATS_ADD(ip->i_mount, xs_read_bytes, ret);
425 return ret;
426}
427
428/*
429 * This routine is called to handle zeroing any space in the last block of the
430 * file that is beyond the EOF. We do this since the size is being increased
431 * without writing anything to that block and we don't want to read the
432 * garbage on the disk.
433 */
434STATIC int /* error (positive) */
435xfs_zero_last_block(
436 struct xfs_inode *ip,
437 xfs_fsize_t offset,
438 xfs_fsize_t isize,
439 bool *did_zeroing)
440{
441 struct xfs_mount *mp = ip->i_mount;
442 xfs_fileoff_t last_fsb = XFS_B_TO_FSBT(mp, isize);
443 int zero_offset = XFS_B_FSB_OFFSET(mp, isize);
444 int zero_len;
445 int nimaps = 1;
446 int error = 0;
447 struct xfs_bmbt_irec imap;
448
449 xfs_ilock(ip, XFS_ILOCK_EXCL);
450 error = xfs_bmapi_read(ip, last_fsb, 1, &imap, &nimaps, 0);
451 xfs_iunlock(ip, XFS_ILOCK_EXCL);
452 if (error)
453 return error;
454
455 ASSERT(nimaps > 0);
456
457 /*
458 * If the block underlying isize is just a hole, then there
459 * is nothing to zero.
460 */
461 if (imap.br_startblock == HOLESTARTBLOCK)
462 return 0;
463
464 zero_len = mp->m_sb.sb_blocksize - zero_offset;
465 if (isize + zero_len > offset)
466 zero_len = offset - isize;
467 *did_zeroing = true;
468 return xfs_iozero(ip, isize, zero_len);
469}
470
471/*
472 * Zero any on disk space between the current EOF and the new, larger EOF.
473 *
474 * This handles the normal case of zeroing the remainder of the last block in
475 * the file and the unusual case of zeroing blocks out beyond the size of the
476 * file. This second case only happens with fixed size extents and when the
477 * system crashes before the inode size was updated but after blocks were
478 * allocated.
479 *
480 * Expects the iolock to be held exclusive, and will take the ilock internally.
481 */
482int /* error (positive) */
483xfs_zero_eof(
484 struct xfs_inode *ip,
485 xfs_off_t offset, /* starting I/O offset */
486 xfs_fsize_t isize, /* current inode size */
487 bool *did_zeroing)
488{
489 struct xfs_mount *mp = ip->i_mount;
490 xfs_fileoff_t start_zero_fsb;
491 xfs_fileoff_t end_zero_fsb;
492 xfs_fileoff_t zero_count_fsb;
493 xfs_fileoff_t last_fsb;
494 xfs_fileoff_t zero_off;
495 xfs_fsize_t zero_len;
496 int nimaps;
497 int error = 0;
498 struct xfs_bmbt_irec imap;
499
500 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
501 ASSERT(offset > isize);
502
503 trace_xfs_zero_eof(ip, isize, offset - isize);
504
505 /*
506 * First handle zeroing the block on which isize resides.
507 *
508 * We only zero a part of that block so it is handled specially.
509 */
510 if (XFS_B_FSB_OFFSET(mp, isize) != 0) {
511 error = xfs_zero_last_block(ip, offset, isize, did_zeroing);
512 if (error)
513 return error;
514 }
515
516 /*
517 * Calculate the range between the new size and the old where blocks
518 * needing to be zeroed may exist.
519 *
520 * To get the block where the last byte in the file currently resides,
521 * we need to subtract one from the size and truncate back to a block
522 * boundary. We subtract 1 in case the size is exactly on a block
523 * boundary.
524 */
525 last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
526 start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
527 end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
528 ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
529 if (last_fsb == end_zero_fsb) {
530 /*
531 * The size was only incremented on its last block.
532 * We took care of that above, so just return.
533 */
534 return 0;
535 }
536
537 ASSERT(start_zero_fsb <= end_zero_fsb);
538 while (start_zero_fsb <= end_zero_fsb) {
539 nimaps = 1;
540 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
541
542 xfs_ilock(ip, XFS_ILOCK_EXCL);
543 error = xfs_bmapi_read(ip, start_zero_fsb, zero_count_fsb,
544 &imap, &nimaps, 0);
545 xfs_iunlock(ip, XFS_ILOCK_EXCL);
546 if (error)
547 return error;
548
549 ASSERT(nimaps > 0);
550
551 if (imap.br_state == XFS_EXT_UNWRITTEN ||
552 imap.br_startblock == HOLESTARTBLOCK) {
553 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
554 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
555 continue;
556 }
557
558 /*
559 * There are blocks we need to zero.
560 */
561 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
562 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
563
564 if ((zero_off + zero_len) > offset)
565 zero_len = offset - zero_off;
566
567 error = xfs_iozero(ip, zero_off, zero_len);
568 if (error)
569 return error;
570
571 *did_zeroing = true;
572 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
573 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
574 }
575
576 return 0;
577}
578
579/*
580 * Common pre-write limit and setup checks.
581 *
582 * Called with the iolocked held either shared and exclusive according to
583 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
584 * if called for a direct write beyond i_size.
585 */
586STATIC ssize_t
587xfs_file_aio_write_checks(
588 struct kiocb *iocb,
589 struct iov_iter *from,
590 int *iolock)
591{
592 struct file *file = iocb->ki_filp;
593 struct inode *inode = file->f_mapping->host;
594 struct xfs_inode *ip = XFS_I(inode);
595 ssize_t error = 0;
596 size_t count = iov_iter_count(from);
597 bool drained_dio = false;
598
599restart:
600 error = generic_write_checks(iocb, from);
601 if (error <= 0)
602 return error;
603
604 error = xfs_break_layouts(inode, iolock, true);
605 if (error)
606 return error;
607
608 /* For changing security info in file_remove_privs() we need i_mutex */
609 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
610 xfs_rw_iunlock(ip, *iolock);
611 *iolock = XFS_IOLOCK_EXCL;
612 xfs_rw_ilock(ip, *iolock);
613 goto restart;
614 }
615 /*
616 * If the offset is beyond the size of the file, we need to zero any
617 * blocks that fall between the existing EOF and the start of this
618 * write. If zeroing is needed and we are currently holding the
619 * iolock shared, we need to update it to exclusive which implies
620 * having to redo all checks before.
621 *
622 * We need to serialise against EOF updates that occur in IO
623 * completions here. We want to make sure that nobody is changing the
624 * size while we do this check until we have placed an IO barrier (i.e.
625 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
626 * The spinlock effectively forms a memory barrier once we have the
627 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
628 * and hence be able to correctly determine if we need to run zeroing.
629 */
630 spin_lock(&ip->i_flags_lock);
631 if (iocb->ki_pos > i_size_read(inode)) {
632 bool zero = false;
633
634 spin_unlock(&ip->i_flags_lock);
635 if (!drained_dio) {
636 if (*iolock == XFS_IOLOCK_SHARED) {
637 xfs_rw_iunlock(ip, *iolock);
638 *iolock = XFS_IOLOCK_EXCL;
639 xfs_rw_ilock(ip, *iolock);
640 iov_iter_reexpand(from, count);
641 }
642 /*
643 * We now have an IO submission barrier in place, but
644 * AIO can do EOF updates during IO completion and hence
645 * we now need to wait for all of them to drain. Non-AIO
646 * DIO will have drained before we are given the
647 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
648 * no-op.
649 */
650 inode_dio_wait(inode);
651 drained_dio = true;
652 goto restart;
653 }
654 error = xfs_zero_eof(ip, iocb->ki_pos, i_size_read(inode), &zero);
655 if (error)
656 return error;
657 } else
658 spin_unlock(&ip->i_flags_lock);
659
660 /*
661 * Updating the timestamps will grab the ilock again from
662 * xfs_fs_dirty_inode, so we have to call it after dropping the
663 * lock above. Eventually we should look into a way to avoid
664 * the pointless lock roundtrip.
665 */
666 if (likely(!(file->f_mode & FMODE_NOCMTIME))) {
667 error = file_update_time(file);
668 if (error)
669 return error;
670 }
671
672 /*
673 * If we're writing the file then make sure to clear the setuid and
674 * setgid bits if the process is not being run by root. This keeps
675 * people from modifying setuid and setgid binaries.
676 */
677 if (!IS_NOSEC(inode))
678 return file_remove_privs(file);
679 return 0;
680}
681
682/*
683 * xfs_file_dio_aio_write - handle direct IO writes
684 *
685 * Lock the inode appropriately to prepare for and issue a direct IO write.
686 * By separating it from the buffered write path we remove all the tricky to
687 * follow locking changes and looping.
688 *
689 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
690 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
691 * pages are flushed out.
692 *
693 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
694 * allowing them to be done in parallel with reads and other direct IO writes.
695 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
696 * needs to do sub-block zeroing and that requires serialisation against other
697 * direct IOs to the same block. In this case we need to serialise the
698 * submission of the unaligned IOs so that we don't get racing block zeroing in
699 * the dio layer. To avoid the problem with aio, we also need to wait for
700 * outstanding IOs to complete so that unwritten extent conversion is completed
701 * before we try to map the overlapping block. This is currently implemented by
702 * hitting it with a big hammer (i.e. inode_dio_wait()).
703 *
704 * Returns with locks held indicated by @iolock and errors indicated by
705 * negative return values.
706 */
707STATIC ssize_t
708xfs_file_dio_aio_write(
709 struct kiocb *iocb,
710 struct iov_iter *from)
711{
712 struct file *file = iocb->ki_filp;
713 struct address_space *mapping = file->f_mapping;
714 struct inode *inode = mapping->host;
715 struct xfs_inode *ip = XFS_I(inode);
716 struct xfs_mount *mp = ip->i_mount;
717 ssize_t ret = 0;
718 int unaligned_io = 0;
719 int iolock;
720 size_t count = iov_iter_count(from);
721 loff_t pos = iocb->ki_pos;
722 loff_t end;
723 struct iov_iter data;
724 struct xfs_buftarg *target = XFS_IS_REALTIME_INODE(ip) ?
725 mp->m_rtdev_targp : mp->m_ddev_targp;
726
727 /* DIO must be aligned to device logical sector size */
728 if (!IS_DAX(inode) && ((pos | count) & target->bt_logical_sectormask))
729 return -EINVAL;
730
731 /* "unaligned" here means not aligned to a filesystem block */
732 if ((pos & mp->m_blockmask) || ((pos + count) & mp->m_blockmask))
733 unaligned_io = 1;
734
735 /*
736 * We don't need to take an exclusive lock unless there page cache needs
737 * to be invalidated or unaligned IO is being executed. We don't need to
738 * consider the EOF extension case here because
739 * xfs_file_aio_write_checks() will relock the inode as necessary for
740 * EOF zeroing cases and fill out the new inode size as appropriate.
741 */
742 if (unaligned_io || mapping->nrpages)
743 iolock = XFS_IOLOCK_EXCL;
744 else
745 iolock = XFS_IOLOCK_SHARED;
746 xfs_rw_ilock(ip, iolock);
747
748 /*
749 * Recheck if there are cached pages that need invalidate after we got
750 * the iolock to protect against other threads adding new pages while
751 * we were waiting for the iolock.
752 */
753 if (mapping->nrpages && iolock == XFS_IOLOCK_SHARED) {
754 xfs_rw_iunlock(ip, iolock);
755 iolock = XFS_IOLOCK_EXCL;
756 xfs_rw_ilock(ip, iolock);
757 }
758
759 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
760 if (ret)
761 goto out;
762 count = iov_iter_count(from);
763 pos = iocb->ki_pos;
764 end = pos + count - 1;
765
766 /*
767 * See xfs_file_read_iter() for why we do a full-file flush here.
768 */
769 if (mapping->nrpages) {
770 ret = filemap_write_and_wait(VFS_I(ip)->i_mapping);
771 if (ret)
772 goto out;
773 /*
774 * Invalidate whole pages. This can return an error if we fail
775 * to invalidate a page, but this should never happen on XFS.
776 * Warn if it does fail.
777 */
778 ret = invalidate_inode_pages2(VFS_I(ip)->i_mapping);
779 WARN_ON_ONCE(ret);
780 ret = 0;
781 }
782
783 /*
784 * If we are doing unaligned IO, wait for all other IO to drain,
785 * otherwise demote the lock if we had to flush cached pages
786 */
787 if (unaligned_io)
788 inode_dio_wait(inode);
789 else if (iolock == XFS_IOLOCK_EXCL) {
790 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
791 iolock = XFS_IOLOCK_SHARED;
792 }
793
794 trace_xfs_file_direct_write(ip, count, iocb->ki_pos, 0);
795
796 data = *from;
797 ret = mapping->a_ops->direct_IO(iocb, &data, pos);
798
799 /* see generic_file_direct_write() for why this is necessary */
800 if (mapping->nrpages) {
801 invalidate_inode_pages2_range(mapping,
802 pos >> PAGE_SHIFT,
803 end >> PAGE_SHIFT);
804 }
805
806 if (ret > 0) {
807 pos += ret;
808 iov_iter_advance(from, ret);
809 iocb->ki_pos = pos;
810 }
811out:
812 xfs_rw_iunlock(ip, iolock);
813
814 /*
815 * No fallback to buffered IO on errors for XFS. DAX can result in
816 * partial writes, but direct IO will either complete fully or fail.
817 */
818 ASSERT(ret < 0 || ret == count || IS_DAX(VFS_I(ip)));
819 return ret;
820}
821
822STATIC ssize_t
823xfs_file_buffered_aio_write(
824 struct kiocb *iocb,
825 struct iov_iter *from)
826{
827 struct file *file = iocb->ki_filp;
828 struct address_space *mapping = file->f_mapping;
829 struct inode *inode = mapping->host;
830 struct xfs_inode *ip = XFS_I(inode);
831 ssize_t ret;
832 int enospc = 0;
833 int iolock = XFS_IOLOCK_EXCL;
834
835 xfs_rw_ilock(ip, iolock);
836
837 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
838 if (ret)
839 goto out;
840
841 /* We can write back this queue in page reclaim */
842 current->backing_dev_info = inode_to_bdi(inode);
843
844write_retry:
845 trace_xfs_file_buffered_write(ip, iov_iter_count(from),
846 iocb->ki_pos, 0);
847 ret = generic_perform_write(file, from, iocb->ki_pos);
848 if (likely(ret >= 0))
849 iocb->ki_pos += ret;
850
851 /*
852 * If we hit a space limit, try to free up some lingering preallocated
853 * space before returning an error. In the case of ENOSPC, first try to
854 * write back all dirty inodes to free up some of the excess reserved
855 * metadata space. This reduces the chances that the eofblocks scan
856 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
857 * also behaves as a filter to prevent too many eofblocks scans from
858 * running at the same time.
859 */
860 if (ret == -EDQUOT && !enospc) {
861 enospc = xfs_inode_free_quota_eofblocks(ip);
862 if (enospc)
863 goto write_retry;
864 } else if (ret == -ENOSPC && !enospc) {
865 struct xfs_eofblocks eofb = {0};
866
867 enospc = 1;
868 xfs_flush_inodes(ip->i_mount);
869 eofb.eof_scan_owner = ip->i_ino; /* for locking */
870 eofb.eof_flags = XFS_EOF_FLAGS_SYNC;
871 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
872 goto write_retry;
873 }
874
875 current->backing_dev_info = NULL;
876out:
877 xfs_rw_iunlock(ip, iolock);
878 return ret;
879}
880
881STATIC ssize_t
882xfs_file_write_iter(
883 struct kiocb *iocb,
884 struct iov_iter *from)
885{
886 struct file *file = iocb->ki_filp;
887 struct address_space *mapping = file->f_mapping;
888 struct inode *inode = mapping->host;
889 struct xfs_inode *ip = XFS_I(inode);
890 ssize_t ret;
891 size_t ocount = iov_iter_count(from);
892
893 XFS_STATS_INC(ip->i_mount, xs_write_calls);
894
895 if (ocount == 0)
896 return 0;
897
898 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
899 return -EIO;
900
901 if ((iocb->ki_flags & IOCB_DIRECT) || IS_DAX(inode))
902 ret = xfs_file_dio_aio_write(iocb, from);
903 else
904 ret = xfs_file_buffered_aio_write(iocb, from);
905
906 if (ret > 0) {
907 ssize_t err;
908
909 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
910
911 /* Handle various SYNC-type writes */
912 err = generic_write_sync(file, iocb->ki_pos - ret, ret);
913 if (err < 0)
914 ret = err;
915 }
916 return ret;
917}
918
919#define XFS_FALLOC_FL_SUPPORTED \
920 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
921 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
922 FALLOC_FL_INSERT_RANGE)
923
924STATIC long
925xfs_file_fallocate(
926 struct file *file,
927 int mode,
928 loff_t offset,
929 loff_t len)
930{
931 struct inode *inode = file_inode(file);
932 struct xfs_inode *ip = XFS_I(inode);
933 long error;
934 enum xfs_prealloc_flags flags = 0;
935 uint iolock = XFS_IOLOCK_EXCL;
936 loff_t new_size = 0;
937 bool do_file_insert = 0;
938
939 if (!S_ISREG(inode->i_mode))
940 return -EINVAL;
941 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
942 return -EOPNOTSUPP;
943
944 xfs_ilock(ip, iolock);
945 error = xfs_break_layouts(inode, &iolock, false);
946 if (error)
947 goto out_unlock;
948
949 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
950 iolock |= XFS_MMAPLOCK_EXCL;
951
952 if (mode & FALLOC_FL_PUNCH_HOLE) {
953 error = xfs_free_file_space(ip, offset, len);
954 if (error)
955 goto out_unlock;
956 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
957 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
958
959 if (offset & blksize_mask || len & blksize_mask) {
960 error = -EINVAL;
961 goto out_unlock;
962 }
963
964 /*
965 * There is no need to overlap collapse range with EOF,
966 * in which case it is effectively a truncate operation
967 */
968 if (offset + len >= i_size_read(inode)) {
969 error = -EINVAL;
970 goto out_unlock;
971 }
972
973 new_size = i_size_read(inode) - len;
974
975 error = xfs_collapse_file_space(ip, offset, len);
976 if (error)
977 goto out_unlock;
978 } else if (mode & FALLOC_FL_INSERT_RANGE) {
979 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
980
981 new_size = i_size_read(inode) + len;
982 if (offset & blksize_mask || len & blksize_mask) {
983 error = -EINVAL;
984 goto out_unlock;
985 }
986
987 /* check the new inode size does not wrap through zero */
988 if (new_size > inode->i_sb->s_maxbytes) {
989 error = -EFBIG;
990 goto out_unlock;
991 }
992
993 /* Offset should be less than i_size */
994 if (offset >= i_size_read(inode)) {
995 error = -EINVAL;
996 goto out_unlock;
997 }
998 do_file_insert = 1;
999 } else {
1000 flags |= XFS_PREALLOC_SET;
1001
1002 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
1003 offset + len > i_size_read(inode)) {
1004 new_size = offset + len;
1005 error = inode_newsize_ok(inode, new_size);
1006 if (error)
1007 goto out_unlock;
1008 }
1009
1010 if (mode & FALLOC_FL_ZERO_RANGE)
1011 error = xfs_zero_file_space(ip, offset, len);
1012 else
1013 error = xfs_alloc_file_space(ip, offset, len,
1014 XFS_BMAPI_PREALLOC);
1015 if (error)
1016 goto out_unlock;
1017 }
1018
1019 if (file->f_flags & O_DSYNC)
1020 flags |= XFS_PREALLOC_SYNC;
1021
1022 error = xfs_update_prealloc_flags(ip, flags);
1023 if (error)
1024 goto out_unlock;
1025
1026 /* Change file size if needed */
1027 if (new_size) {
1028 struct iattr iattr;
1029
1030 iattr.ia_valid = ATTR_SIZE;
1031 iattr.ia_size = new_size;
1032 error = xfs_setattr_size(ip, &iattr);
1033 if (error)
1034 goto out_unlock;
1035 }
1036
1037 /*
1038 * Perform hole insertion now that the file size has been
1039 * updated so that if we crash during the operation we don't
1040 * leave shifted extents past EOF and hence losing access to
1041 * the data that is contained within them.
1042 */
1043 if (do_file_insert)
1044 error = xfs_insert_file_space(ip, offset, len);
1045
1046out_unlock:
1047 xfs_iunlock(ip, iolock);
1048 return error;
1049}
1050
1051
1052STATIC int
1053xfs_file_open(
1054 struct inode *inode,
1055 struct file *file)
1056{
1057 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
1058 return -EFBIG;
1059 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
1060 return -EIO;
1061 return 0;
1062}
1063
1064STATIC int
1065xfs_dir_open(
1066 struct inode *inode,
1067 struct file *file)
1068{
1069 struct xfs_inode *ip = XFS_I(inode);
1070 int mode;
1071 int error;
1072
1073 error = xfs_file_open(inode, file);
1074 if (error)
1075 return error;
1076
1077 /*
1078 * If there are any blocks, read-ahead block 0 as we're almost
1079 * certain to have the next operation be a read there.
1080 */
1081 mode = xfs_ilock_data_map_shared(ip);
1082 if (ip->i_d.di_nextents > 0)
1083 xfs_dir3_data_readahead(ip, 0, -1);
1084 xfs_iunlock(ip, mode);
1085 return 0;
1086}
1087
1088STATIC int
1089xfs_file_release(
1090 struct inode *inode,
1091 struct file *filp)
1092{
1093 return xfs_release(XFS_I(inode));
1094}
1095
1096STATIC int
1097xfs_file_readdir(
1098 struct file *file,
1099 struct dir_context *ctx)
1100{
1101 struct inode *inode = file_inode(file);
1102 xfs_inode_t *ip = XFS_I(inode);
1103 size_t bufsize;
1104
1105 /*
1106 * The Linux API doesn't pass down the total size of the buffer
1107 * we read into down to the filesystem. With the filldir concept
1108 * it's not needed for correct information, but the XFS dir2 leaf
1109 * code wants an estimate of the buffer size to calculate it's
1110 * readahead window and size the buffers used for mapping to
1111 * physical blocks.
1112 *
1113 * Try to give it an estimate that's good enough, maybe at some
1114 * point we can change the ->readdir prototype to include the
1115 * buffer size. For now we use the current glibc buffer size.
1116 */
1117 bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
1118
1119 return xfs_readdir(ip, ctx, bufsize);
1120}
1121
1122/*
1123 * This type is designed to indicate the type of offset we would like
1124 * to search from page cache for xfs_seek_hole_data().
1125 */
1126enum {
1127 HOLE_OFF = 0,
1128 DATA_OFF,
1129};
1130
1131/*
1132 * Lookup the desired type of offset from the given page.
1133 *
1134 * On success, return true and the offset argument will point to the
1135 * start of the region that was found. Otherwise this function will
1136 * return false and keep the offset argument unchanged.
1137 */
1138STATIC bool
1139xfs_lookup_buffer_offset(
1140 struct page *page,
1141 loff_t *offset,
1142 unsigned int type)
1143{
1144 loff_t lastoff = page_offset(page);
1145 bool found = false;
1146 struct buffer_head *bh, *head;
1147
1148 bh = head = page_buffers(page);
1149 do {
1150 /*
1151 * Unwritten extents that have data in the page
1152 * cache covering them can be identified by the
1153 * BH_Unwritten state flag. Pages with multiple
1154 * buffers might have a mix of holes, data and
1155 * unwritten extents - any buffer with valid
1156 * data in it should have BH_Uptodate flag set
1157 * on it.
1158 */
1159 if (buffer_unwritten(bh) ||
1160 buffer_uptodate(bh)) {
1161 if (type == DATA_OFF)
1162 found = true;
1163 } else {
1164 if (type == HOLE_OFF)
1165 found = true;
1166 }
1167
1168 if (found) {
1169 *offset = lastoff;
1170 break;
1171 }
1172 lastoff += bh->b_size;
1173 } while ((bh = bh->b_this_page) != head);
1174
1175 return found;
1176}
1177
1178/*
1179 * This routine is called to find out and return a data or hole offset
1180 * from the page cache for unwritten extents according to the desired
1181 * type for xfs_seek_hole_data().
1182 *
1183 * The argument offset is used to tell where we start to search from the
1184 * page cache. Map is used to figure out the end points of the range to
1185 * lookup pages.
1186 *
1187 * Return true if the desired type of offset was found, and the argument
1188 * offset is filled with that address. Otherwise, return false and keep
1189 * offset unchanged.
1190 */
1191STATIC bool
1192xfs_find_get_desired_pgoff(
1193 struct inode *inode,
1194 struct xfs_bmbt_irec *map,
1195 unsigned int type,
1196 loff_t *offset)
1197{
1198 struct xfs_inode *ip = XFS_I(inode);
1199 struct xfs_mount *mp = ip->i_mount;
1200 struct pagevec pvec;
1201 pgoff_t index;
1202 pgoff_t end;
1203 loff_t endoff;
1204 loff_t startoff = *offset;
1205 loff_t lastoff = startoff;
1206 bool found = false;
1207
1208 pagevec_init(&pvec, 0);
1209
1210 index = startoff >> PAGE_SHIFT;
1211 endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1212 end = endoff >> PAGE_SHIFT;
1213 do {
1214 int want;
1215 unsigned nr_pages;
1216 unsigned int i;
1217
1218 want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1219 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1220 want);
1221 /*
1222 * No page mapped into given range. If we are searching holes
1223 * and if this is the first time we got into the loop, it means
1224 * that the given offset is landed in a hole, return it.
1225 *
1226 * If we have already stepped through some block buffers to find
1227 * holes but they all contains data. In this case, the last
1228 * offset is already updated and pointed to the end of the last
1229 * mapped page, if it does not reach the endpoint to search,
1230 * that means there should be a hole between them.
1231 */
1232 if (nr_pages == 0) {
1233 /* Data search found nothing */
1234 if (type == DATA_OFF)
1235 break;
1236
1237 ASSERT(type == HOLE_OFF);
1238 if (lastoff == startoff || lastoff < endoff) {
1239 found = true;
1240 *offset = lastoff;
1241 }
1242 break;
1243 }
1244
1245 /*
1246 * At lease we found one page. If this is the first time we
1247 * step into the loop, and if the first page index offset is
1248 * greater than the given search offset, a hole was found.
1249 */
1250 if (type == HOLE_OFF && lastoff == startoff &&
1251 lastoff < page_offset(pvec.pages[0])) {
1252 found = true;
1253 break;
1254 }
1255
1256 for (i = 0; i < nr_pages; i++) {
1257 struct page *page = pvec.pages[i];
1258 loff_t b_offset;
1259
1260 /*
1261 * At this point, the page may be truncated or
1262 * invalidated (changing page->mapping to NULL),
1263 * or even swizzled back from swapper_space to tmpfs
1264 * file mapping. However, page->index will not change
1265 * because we have a reference on the page.
1266 *
1267 * Searching done if the page index is out of range.
1268 * If the current offset is not reaches the end of
1269 * the specified search range, there should be a hole
1270 * between them.
1271 */
1272 if (page->index > end) {
1273 if (type == HOLE_OFF && lastoff < endoff) {
1274 *offset = lastoff;
1275 found = true;
1276 }
1277 goto out;
1278 }
1279
1280 lock_page(page);
1281 /*
1282 * Page truncated or invalidated(page->mapping == NULL).
1283 * We can freely skip it and proceed to check the next
1284 * page.
1285 */
1286 if (unlikely(page->mapping != inode->i_mapping)) {
1287 unlock_page(page);
1288 continue;
1289 }
1290
1291 if (!page_has_buffers(page)) {
1292 unlock_page(page);
1293 continue;
1294 }
1295
1296 found = xfs_lookup_buffer_offset(page, &b_offset, type);
1297 if (found) {
1298 /*
1299 * The found offset may be less than the start
1300 * point to search if this is the first time to
1301 * come here.
1302 */
1303 *offset = max_t(loff_t, startoff, b_offset);
1304 unlock_page(page);
1305 goto out;
1306 }
1307
1308 /*
1309 * We either searching data but nothing was found, or
1310 * searching hole but found a data buffer. In either
1311 * case, probably the next page contains the desired
1312 * things, update the last offset to it so.
1313 */
1314 lastoff = page_offset(page) + PAGE_SIZE;
1315 unlock_page(page);
1316 }
1317
1318 /*
1319 * The number of returned pages less than our desired, search
1320 * done. In this case, nothing was found for searching data,
1321 * but we found a hole behind the last offset.
1322 */
1323 if (nr_pages < want) {
1324 if (type == HOLE_OFF) {
1325 *offset = lastoff;
1326 found = true;
1327 }
1328 break;
1329 }
1330
1331 index = pvec.pages[i - 1]->index + 1;
1332 pagevec_release(&pvec);
1333 } while (index <= end);
1334
1335out:
1336 pagevec_release(&pvec);
1337 return found;
1338}
1339
1340/*
1341 * caller must lock inode with xfs_ilock_data_map_shared,
1342 * can we craft an appropriate ASSERT?
1343 *
1344 * end is because the VFS-level lseek interface is defined such that any
1345 * offset past i_size shall return -ENXIO, but we use this for quota code
1346 * which does not maintain i_size, and we want to SEEK_DATA past i_size.
1347 */
1348loff_t
1349__xfs_seek_hole_data(
1350 struct inode *inode,
1351 loff_t start,
1352 loff_t end,
1353 int whence)
1354{
1355 struct xfs_inode *ip = XFS_I(inode);
1356 struct xfs_mount *mp = ip->i_mount;
1357 loff_t uninitialized_var(offset);
1358 xfs_fileoff_t fsbno;
1359 xfs_filblks_t lastbno;
1360 int error;
1361
1362 if (start >= end) {
1363 error = -ENXIO;
1364 goto out_error;
1365 }
1366
1367 /*
1368 * Try to read extents from the first block indicated
1369 * by fsbno to the end block of the file.
1370 */
1371 fsbno = XFS_B_TO_FSBT(mp, start);
1372 lastbno = XFS_B_TO_FSB(mp, end);
1373
1374 for (;;) {
1375 struct xfs_bmbt_irec map[2];
1376 int nmap = 2;
1377 unsigned int i;
1378
1379 error = xfs_bmapi_read(ip, fsbno, lastbno - fsbno, map, &nmap,
1380 XFS_BMAPI_ENTIRE);
1381 if (error)
1382 goto out_error;
1383
1384 /* No extents at given offset, must be beyond EOF */
1385 if (nmap == 0) {
1386 error = -ENXIO;
1387 goto out_error;
1388 }
1389
1390 for (i = 0; i < nmap; i++) {
1391 offset = max_t(loff_t, start,
1392 XFS_FSB_TO_B(mp, map[i].br_startoff));
1393
1394 /* Landed in the hole we wanted? */
1395 if (whence == SEEK_HOLE &&
1396 map[i].br_startblock == HOLESTARTBLOCK)
1397 goto out;
1398
1399 /* Landed in the data extent we wanted? */
1400 if (whence == SEEK_DATA &&
1401 (map[i].br_startblock == DELAYSTARTBLOCK ||
1402 (map[i].br_state == XFS_EXT_NORM &&
1403 !isnullstartblock(map[i].br_startblock))))
1404 goto out;
1405
1406 /*
1407 * Landed in an unwritten extent, try to search
1408 * for hole or data from page cache.
1409 */
1410 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1411 if (xfs_find_get_desired_pgoff(inode, &map[i],
1412 whence == SEEK_HOLE ? HOLE_OFF : DATA_OFF,
1413 &offset))
1414 goto out;
1415 }
1416 }
1417
1418 /*
1419 * We only received one extent out of the two requested. This
1420 * means we've hit EOF and didn't find what we are looking for.
1421 */
1422 if (nmap == 1) {
1423 /*
1424 * If we were looking for a hole, set offset to
1425 * the end of the file (i.e., there is an implicit
1426 * hole at the end of any file).
1427 */
1428 if (whence == SEEK_HOLE) {
1429 offset = end;
1430 break;
1431 }
1432 /*
1433 * If we were looking for data, it's nowhere to be found
1434 */
1435 ASSERT(whence == SEEK_DATA);
1436 error = -ENXIO;
1437 goto out_error;
1438 }
1439
1440 ASSERT(i > 1);
1441
1442 /*
1443 * Nothing was found, proceed to the next round of search
1444 * if the next reading offset is not at or beyond EOF.
1445 */
1446 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1447 start = XFS_FSB_TO_B(mp, fsbno);
1448 if (start >= end) {
1449 if (whence == SEEK_HOLE) {
1450 offset = end;
1451 break;
1452 }
1453 ASSERT(whence == SEEK_DATA);
1454 error = -ENXIO;
1455 goto out_error;
1456 }
1457 }
1458
1459out:
1460 /*
1461 * If at this point we have found the hole we wanted, the returned
1462 * offset may be bigger than the file size as it may be aligned to
1463 * page boundary for unwritten extents. We need to deal with this
1464 * situation in particular.
1465 */
1466 if (whence == SEEK_HOLE)
1467 offset = min_t(loff_t, offset, end);
1468
1469 return offset;
1470
1471out_error:
1472 return error;
1473}
1474
1475STATIC loff_t
1476xfs_seek_hole_data(
1477 struct file *file,
1478 loff_t start,
1479 int whence)
1480{
1481 struct inode *inode = file->f_mapping->host;
1482 struct xfs_inode *ip = XFS_I(inode);
1483 struct xfs_mount *mp = ip->i_mount;
1484 uint lock;
1485 loff_t offset, end;
1486 int error = 0;
1487
1488 if (XFS_FORCED_SHUTDOWN(mp))
1489 return -EIO;
1490
1491 lock = xfs_ilock_data_map_shared(ip);
1492
1493 end = i_size_read(inode);
1494 offset = __xfs_seek_hole_data(inode, start, end, whence);
1495 if (offset < 0) {
1496 error = offset;
1497 goto out_unlock;
1498 }
1499
1500 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1501
1502out_unlock:
1503 xfs_iunlock(ip, lock);
1504
1505 if (error)
1506 return error;
1507 return offset;
1508}
1509
1510STATIC loff_t
1511xfs_file_llseek(
1512 struct file *file,
1513 loff_t offset,
1514 int whence)
1515{
1516 switch (whence) {
1517 case SEEK_END:
1518 case SEEK_CUR:
1519 case SEEK_SET:
1520 return generic_file_llseek(file, offset, whence);
1521 case SEEK_HOLE:
1522 case SEEK_DATA:
1523 return xfs_seek_hole_data(file, offset, whence);
1524 default:
1525 return -EINVAL;
1526 }
1527}
1528
1529/*
1530 * Locking for serialisation of IO during page faults. This results in a lock
1531 * ordering of:
1532 *
1533 * mmap_sem (MM)
1534 * sb_start_pagefault(vfs, freeze)
1535 * i_mmaplock (XFS - truncate serialisation)
1536 * page_lock (MM)
1537 * i_lock (XFS - extent map serialisation)
1538 */
1539
1540/*
1541 * mmap()d file has taken write protection fault and is being made writable. We
1542 * can set the page state up correctly for a writable page, which means we can
1543 * do correct delalloc accounting (ENOSPC checking!) and unwritten extent
1544 * mapping.
1545 */
1546STATIC int
1547xfs_filemap_page_mkwrite(
1548 struct vm_area_struct *vma,
1549 struct vm_fault *vmf)
1550{
1551 struct inode *inode = file_inode(vma->vm_file);
1552 int ret;
1553
1554 trace_xfs_filemap_page_mkwrite(XFS_I(inode));
1555
1556 sb_start_pagefault(inode->i_sb);
1557 file_update_time(vma->vm_file);
1558 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1559
1560 if (IS_DAX(inode)) {
1561 ret = __dax_mkwrite(vma, vmf, xfs_get_blocks_dax_fault, NULL);
1562 } else {
1563 ret = block_page_mkwrite(vma, vmf, xfs_get_blocks);
1564 ret = block_page_mkwrite_return(ret);
1565 }
1566
1567 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1568 sb_end_pagefault(inode->i_sb);
1569
1570 return ret;
1571}
1572
1573STATIC int
1574xfs_filemap_fault(
1575 struct vm_area_struct *vma,
1576 struct vm_fault *vmf)
1577{
1578 struct inode *inode = file_inode(vma->vm_file);
1579 int ret;
1580
1581 trace_xfs_filemap_fault(XFS_I(inode));
1582
1583 /* DAX can shortcut the normal fault path on write faults! */
1584 if ((vmf->flags & FAULT_FLAG_WRITE) && IS_DAX(inode))
1585 return xfs_filemap_page_mkwrite(vma, vmf);
1586
1587 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1588 if (IS_DAX(inode)) {
1589 /*
1590 * we do not want to trigger unwritten extent conversion on read
1591 * faults - that is unnecessary overhead and would also require
1592 * changes to xfs_get_blocks_direct() to map unwritten extent
1593 * ioend for conversion on read-only mappings.
1594 */
1595 ret = __dax_fault(vma, vmf, xfs_get_blocks_dax_fault, NULL);
1596 } else
1597 ret = filemap_fault(vma, vmf);
1598 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1599
1600 return ret;
1601}
1602
1603/*
1604 * Similar to xfs_filemap_fault(), the DAX fault path can call into here on
1605 * both read and write faults. Hence we need to handle both cases. There is no
1606 * ->pmd_mkwrite callout for huge pages, so we have a single function here to
1607 * handle both cases here. @flags carries the information on the type of fault
1608 * occuring.
1609 */
1610STATIC int
1611xfs_filemap_pmd_fault(
1612 struct vm_area_struct *vma,
1613 unsigned long addr,
1614 pmd_t *pmd,
1615 unsigned int flags)
1616{
1617 struct inode *inode = file_inode(vma->vm_file);
1618 struct xfs_inode *ip = XFS_I(inode);
1619 int ret;
1620
1621 if (!IS_DAX(inode))
1622 return VM_FAULT_FALLBACK;
1623
1624 trace_xfs_filemap_pmd_fault(ip);
1625
1626 if (flags & FAULT_FLAG_WRITE) {
1627 sb_start_pagefault(inode->i_sb);
1628 file_update_time(vma->vm_file);
1629 }
1630
1631 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1632 ret = __dax_pmd_fault(vma, addr, pmd, flags, xfs_get_blocks_dax_fault,
1633 NULL);
1634 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1635
1636 if (flags & FAULT_FLAG_WRITE)
1637 sb_end_pagefault(inode->i_sb);
1638
1639 return ret;
1640}
1641
1642/*
1643 * pfn_mkwrite was originally inteneded to ensure we capture time stamp
1644 * updates on write faults. In reality, it's need to serialise against
1645 * truncate similar to page_mkwrite. Hence we cycle the XFS_MMAPLOCK_SHARED
1646 * to ensure we serialise the fault barrier in place.
1647 */
1648static int
1649xfs_filemap_pfn_mkwrite(
1650 struct vm_area_struct *vma,
1651 struct vm_fault *vmf)
1652{
1653
1654 struct inode *inode = file_inode(vma->vm_file);
1655 struct xfs_inode *ip = XFS_I(inode);
1656 int ret = VM_FAULT_NOPAGE;
1657 loff_t size;
1658
1659 trace_xfs_filemap_pfn_mkwrite(ip);
1660
1661 sb_start_pagefault(inode->i_sb);
1662 file_update_time(vma->vm_file);
1663
1664 /* check if the faulting page hasn't raced with truncate */
1665 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1666 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1667 if (vmf->pgoff >= size)
1668 ret = VM_FAULT_SIGBUS;
1669 else if (IS_DAX(inode))
1670 ret = dax_pfn_mkwrite(vma, vmf);
1671 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1672 sb_end_pagefault(inode->i_sb);
1673 return ret;
1674
1675}
1676
1677static const struct vm_operations_struct xfs_file_vm_ops = {
1678 .fault = xfs_filemap_fault,
1679 .pmd_fault = xfs_filemap_pmd_fault,
1680 .map_pages = filemap_map_pages,
1681 .page_mkwrite = xfs_filemap_page_mkwrite,
1682 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1683};
1684
1685STATIC int
1686xfs_file_mmap(
1687 struct file *filp,
1688 struct vm_area_struct *vma)
1689{
1690 file_accessed(filp);
1691 vma->vm_ops = &xfs_file_vm_ops;
1692 if (IS_DAX(file_inode(filp)))
1693 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
1694 return 0;
1695}
1696
1697const struct file_operations xfs_file_operations = {
1698 .llseek = xfs_file_llseek,
1699 .read_iter = xfs_file_read_iter,
1700 .write_iter = xfs_file_write_iter,
1701 .splice_read = xfs_file_splice_read,
1702 .splice_write = iter_file_splice_write,
1703 .unlocked_ioctl = xfs_file_ioctl,
1704#ifdef CONFIG_COMPAT
1705 .compat_ioctl = xfs_file_compat_ioctl,
1706#endif
1707 .mmap = xfs_file_mmap,
1708 .open = xfs_file_open,
1709 .release = xfs_file_release,
1710 .fsync = xfs_file_fsync,
1711 .fallocate = xfs_file_fallocate,
1712};
1713
1714const struct file_operations xfs_dir_file_operations = {
1715 .open = xfs_dir_open,
1716 .read = generic_read_dir,
1717 .iterate = xfs_file_readdir,
1718 .llseek = generic_file_llseek,
1719 .unlocked_ioctl = xfs_file_ioctl,
1720#ifdef CONFIG_COMPAT
1721 .compat_ioctl = xfs_file_compat_ioctl,
1722#endif
1723 .fsync = xfs_dir_fsync,
1724};