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