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/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};