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
33static const struct vm_operations_struct xfs_file_vm_ops;
34
35int
36xfs_update_prealloc_flags(
37 struct xfs_inode *ip,
38 enum xfs_prealloc_flags flags)
39{
40 struct xfs_trans *tp;
41 int error;
42
43 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_writeid,
44 0, 0, 0, &tp);
45 if (error)
46 return error;
47
48 xfs_ilock(ip, XFS_ILOCK_EXCL);
49 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
50
51 if (!(flags & XFS_PREALLOC_INVISIBLE)) {
52 VFS_I(ip)->i_mode &= ~S_ISUID;
53 if (VFS_I(ip)->i_mode & S_IXGRP)
54 VFS_I(ip)->i_mode &= ~S_ISGID;
55 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
56 }
57
58 if (flags & XFS_PREALLOC_SET)
59 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
60 if (flags & XFS_PREALLOC_CLEAR)
61 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
62
63 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
64 if (flags & XFS_PREALLOC_SYNC)
65 xfs_trans_set_sync(tp);
66 return xfs_trans_commit(tp);
67}
68
69/*
70 * Fsync operations on directories are much simpler than on regular files,
71 * as there is no file data to flush, and thus also no need for explicit
72 * cache flush operations, and there are no non-transaction metadata updates
73 * on directories either.
74 */
75STATIC int
76xfs_dir_fsync(
77 struct file *file,
78 loff_t start,
79 loff_t end,
80 int datasync)
81{
82 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
83 struct xfs_mount *mp = ip->i_mount;
84 xfs_lsn_t lsn = 0;
85
86 trace_xfs_dir_fsync(ip);
87
88 xfs_ilock(ip, XFS_ILOCK_SHARED);
89 if (xfs_ipincount(ip))
90 lsn = ip->i_itemp->ili_last_lsn;
91 xfs_iunlock(ip, XFS_ILOCK_SHARED);
92
93 if (!lsn)
94 return 0;
95 return xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, NULL);
96}
97
98STATIC int
99xfs_file_fsync(
100 struct file *file,
101 loff_t start,
102 loff_t end,
103 int datasync)
104{
105 struct inode *inode = file->f_mapping->host;
106 struct xfs_inode *ip = XFS_I(inode);
107 struct xfs_mount *mp = ip->i_mount;
108 int error = 0;
109 int log_flushed = 0;
110 xfs_lsn_t lsn = 0;
111
112 trace_xfs_file_fsync(ip);
113
114 error = file_write_and_wait_range(file, start, end);
115 if (error)
116 return error;
117
118 if (XFS_FORCED_SHUTDOWN(mp))
119 return -EIO;
120
121 xfs_iflags_clear(ip, XFS_ITRUNCATED);
122
123 /*
124 * If we have an RT and/or log subvolume we need to make sure to flush
125 * the write cache the device used for file data first. This is to
126 * ensure newly written file data make it to disk before logging the new
127 * inode size in case of an extending write.
128 */
129 if (XFS_IS_REALTIME_INODE(ip))
130 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
131 else if (mp->m_logdev_targp != mp->m_ddev_targp)
132 xfs_blkdev_issue_flush(mp->m_ddev_targp);
133
134 /*
135 * All metadata updates are logged, which means that we just have to
136 * flush the log up to the latest LSN that touched the inode. If we have
137 * concurrent fsync/fdatasync() calls, we need them to all block on the
138 * log force before we clear the ili_fsync_fields field. This ensures
139 * that we don't get a racing sync operation that does not wait for the
140 * metadata to hit the journal before returning. If we race with
141 * clearing the ili_fsync_fields, then all that will happen is the log
142 * force will do nothing as the lsn will already be on disk. We can't
143 * race with setting ili_fsync_fields because that is done under
144 * XFS_ILOCK_EXCL, and that can't happen because we hold the lock shared
145 * until after the ili_fsync_fields is cleared.
146 */
147 xfs_ilock(ip, XFS_ILOCK_SHARED);
148 if (xfs_ipincount(ip)) {
149 if (!datasync ||
150 (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
151 lsn = ip->i_itemp->ili_last_lsn;
152 }
153
154 if (lsn) {
155 error = xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
156 ip->i_itemp->ili_fsync_fields = 0;
157 }
158 xfs_iunlock(ip, XFS_ILOCK_SHARED);
159
160 /*
161 * If we only have a single device, and the log force about was
162 * a no-op we might have to flush the data device cache here.
163 * This can only happen for fdatasync/O_DSYNC if we were overwriting
164 * an already allocated file and thus do not have any metadata to
165 * commit.
166 */
167 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
168 mp->m_logdev_targp == mp->m_ddev_targp)
169 xfs_blkdev_issue_flush(mp->m_ddev_targp);
170
171 return error;
172}
173
174STATIC ssize_t
175xfs_file_dio_aio_read(
176 struct kiocb *iocb,
177 struct iov_iter *to)
178{
179 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
180 size_t count = iov_iter_count(to);
181 ssize_t ret;
182
183 trace_xfs_file_direct_read(ip, count, iocb->ki_pos);
184
185 if (!count)
186 return 0; /* skip atime */
187
188 file_accessed(iocb->ki_filp);
189
190 xfs_ilock(ip, XFS_IOLOCK_SHARED);
191 ret = iomap_dio_rw(iocb, to, &xfs_iomap_ops, NULL);
192 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
193
194 return ret;
195}
196
197static noinline ssize_t
198xfs_file_dax_read(
199 struct kiocb *iocb,
200 struct iov_iter *to)
201{
202 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
203 size_t count = iov_iter_count(to);
204 ssize_t ret = 0;
205
206 trace_xfs_file_dax_read(ip, count, iocb->ki_pos);
207
208 if (!count)
209 return 0; /* skip atime */
210
211 if (iocb->ki_flags & IOCB_NOWAIT) {
212 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
213 return -EAGAIN;
214 } else {
215 xfs_ilock(ip, XFS_IOLOCK_SHARED);
216 }
217
218 ret = dax_iomap_rw(iocb, to, &xfs_iomap_ops);
219 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
220
221 file_accessed(iocb->ki_filp);
222 return ret;
223}
224
225STATIC ssize_t
226xfs_file_buffered_aio_read(
227 struct kiocb *iocb,
228 struct iov_iter *to)
229{
230 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
231 ssize_t ret;
232
233 trace_xfs_file_buffered_read(ip, iov_iter_count(to), iocb->ki_pos);
234
235 if (iocb->ki_flags & IOCB_NOWAIT) {
236 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
237 return -EAGAIN;
238 } else {
239 xfs_ilock(ip, XFS_IOLOCK_SHARED);
240 }
241 ret = generic_file_read_iter(iocb, to);
242 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
243
244 return ret;
245}
246
247STATIC ssize_t
248xfs_file_read_iter(
249 struct kiocb *iocb,
250 struct iov_iter *to)
251{
252 struct inode *inode = file_inode(iocb->ki_filp);
253 struct xfs_mount *mp = XFS_I(inode)->i_mount;
254 ssize_t ret = 0;
255
256 XFS_STATS_INC(mp, xs_read_calls);
257
258 if (XFS_FORCED_SHUTDOWN(mp))
259 return -EIO;
260
261 if (IS_DAX(inode))
262 ret = xfs_file_dax_read(iocb, to);
263 else if (iocb->ki_flags & IOCB_DIRECT)
264 ret = xfs_file_dio_aio_read(iocb, to);
265 else
266 ret = xfs_file_buffered_aio_read(iocb, to);
267
268 if (ret > 0)
269 XFS_STATS_ADD(mp, xs_read_bytes, ret);
270 return ret;
271}
272
273/*
274 * Common pre-write limit and setup checks.
275 *
276 * Called with the iolocked held either shared and exclusive according to
277 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
278 * if called for a direct write beyond i_size.
279 */
280STATIC ssize_t
281xfs_file_aio_write_checks(
282 struct kiocb *iocb,
283 struct iov_iter *from,
284 int *iolock)
285{
286 struct file *file = iocb->ki_filp;
287 struct inode *inode = file->f_mapping->host;
288 struct xfs_inode *ip = XFS_I(inode);
289 ssize_t error = 0;
290 size_t count = iov_iter_count(from);
291 bool drained_dio = false;
292 loff_t isize;
293
294restart:
295 error = generic_write_checks(iocb, from);
296 if (error <= 0)
297 return error;
298
299 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
300 if (error)
301 return error;
302
303 /*
304 * For changing security info in file_remove_privs() we need i_rwsem
305 * exclusively.
306 */
307 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
308 xfs_iunlock(ip, *iolock);
309 *iolock = XFS_IOLOCK_EXCL;
310 xfs_ilock(ip, *iolock);
311 goto restart;
312 }
313 /*
314 * If the offset is beyond the size of the file, we need to zero any
315 * blocks that fall between the existing EOF and the start of this
316 * write. If zeroing is needed and we are currently holding the
317 * iolock shared, we need to update it to exclusive which implies
318 * having to redo all checks before.
319 *
320 * We need to serialise against EOF updates that occur in IO
321 * completions here. We want to make sure that nobody is changing the
322 * size while we do this check until we have placed an IO barrier (i.e.
323 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
324 * The spinlock effectively forms a memory barrier once we have the
325 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
326 * and hence be able to correctly determine if we need to run zeroing.
327 */
328 spin_lock(&ip->i_flags_lock);
329 isize = i_size_read(inode);
330 if (iocb->ki_pos > isize) {
331 spin_unlock(&ip->i_flags_lock);
332 if (!drained_dio) {
333 if (*iolock == XFS_IOLOCK_SHARED) {
334 xfs_iunlock(ip, *iolock);
335 *iolock = XFS_IOLOCK_EXCL;
336 xfs_ilock(ip, *iolock);
337 iov_iter_reexpand(from, count);
338 }
339 /*
340 * We now have an IO submission barrier in place, but
341 * AIO can do EOF updates during IO completion and hence
342 * we now need to wait for all of them to drain. Non-AIO
343 * DIO will have drained before we are given the
344 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
345 * no-op.
346 */
347 inode_dio_wait(inode);
348 drained_dio = true;
349 goto restart;
350 }
351
352 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
353 error = iomap_zero_range(inode, isize, iocb->ki_pos - isize,
354 NULL, &xfs_iomap_ops);
355 if (error)
356 return error;
357 } else
358 spin_unlock(&ip->i_flags_lock);
359
360 /*
361 * Updating the timestamps will grab the ilock again from
362 * xfs_fs_dirty_inode, so we have to call it after dropping the
363 * lock above. Eventually we should look into a way to avoid
364 * the pointless lock roundtrip.
365 */
366 return file_modified(file);
367}
368
369static int
370xfs_dio_write_end_io(
371 struct kiocb *iocb,
372 ssize_t size,
373 int error,
374 unsigned flags)
375{
376 struct inode *inode = file_inode(iocb->ki_filp);
377 struct xfs_inode *ip = XFS_I(inode);
378 loff_t offset = iocb->ki_pos;
379 unsigned int nofs_flag;
380
381 trace_xfs_end_io_direct_write(ip, offset, size);
382
383 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
384 return -EIO;
385
386 if (error)
387 return error;
388 if (!size)
389 return 0;
390
391 /*
392 * Capture amount written on completion as we can't reliably account
393 * for it on submission.
394 */
395 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
396
397 /*
398 * We can allocate memory here while doing writeback on behalf of
399 * memory reclaim. To avoid memory allocation deadlocks set the
400 * task-wide nofs context for the following operations.
401 */
402 nofs_flag = memalloc_nofs_save();
403
404 if (flags & IOMAP_DIO_COW) {
405 error = xfs_reflink_end_cow(ip, offset, size);
406 if (error)
407 goto out;
408 }
409
410 /*
411 * Unwritten conversion updates the in-core isize after extent
412 * conversion but before updating the on-disk size. Updating isize any
413 * earlier allows a racing dio read to find unwritten extents before
414 * they are converted.
415 */
416 if (flags & IOMAP_DIO_UNWRITTEN) {
417 error = xfs_iomap_write_unwritten(ip, offset, size, true);
418 goto out;
419 }
420
421 /*
422 * We need to update the in-core inode size here so that we don't end up
423 * with the on-disk inode size being outside the in-core inode size. We
424 * have no other method of updating EOF for AIO, so always do it here
425 * if necessary.
426 *
427 * We need to lock the test/set EOF update as we can be racing with
428 * other IO completions here to update the EOF. Failing to serialise
429 * here can result in EOF moving backwards and Bad Things Happen when
430 * that occurs.
431 */
432 spin_lock(&ip->i_flags_lock);
433 if (offset + size > i_size_read(inode)) {
434 i_size_write(inode, offset + size);
435 spin_unlock(&ip->i_flags_lock);
436 error = xfs_setfilesize(ip, offset, size);
437 } else {
438 spin_unlock(&ip->i_flags_lock);
439 }
440
441out:
442 memalloc_nofs_restore(nofs_flag);
443 return error;
444}
445
446static const struct iomap_dio_ops xfs_dio_write_ops = {
447 .end_io = xfs_dio_write_end_io,
448};
449
450/*
451 * xfs_file_dio_aio_write - handle direct IO writes
452 *
453 * Lock the inode appropriately to prepare for and issue a direct IO write.
454 * By separating it from the buffered write path we remove all the tricky to
455 * follow locking changes and looping.
456 *
457 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
458 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
459 * pages are flushed out.
460 *
461 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
462 * allowing them to be done in parallel with reads and other direct IO writes.
463 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
464 * needs to do sub-block zeroing and that requires serialisation against other
465 * direct IOs to the same block. In this case we need to serialise the
466 * submission of the unaligned IOs so that we don't get racing block zeroing in
467 * the dio layer. To avoid the problem with aio, we also need to wait for
468 * outstanding IOs to complete so that unwritten extent conversion is completed
469 * before we try to map the overlapping block. This is currently implemented by
470 * hitting it with a big hammer (i.e. inode_dio_wait()).
471 *
472 * Returns with locks held indicated by @iolock and errors indicated by
473 * negative return values.
474 */
475STATIC ssize_t
476xfs_file_dio_aio_write(
477 struct kiocb *iocb,
478 struct iov_iter *from)
479{
480 struct file *file = iocb->ki_filp;
481 struct address_space *mapping = file->f_mapping;
482 struct inode *inode = mapping->host;
483 struct xfs_inode *ip = XFS_I(inode);
484 struct xfs_mount *mp = ip->i_mount;
485 ssize_t ret = 0;
486 int unaligned_io = 0;
487 int iolock;
488 size_t count = iov_iter_count(from);
489 struct xfs_buftarg *target = XFS_IS_REALTIME_INODE(ip) ?
490 mp->m_rtdev_targp : mp->m_ddev_targp;
491
492 /* DIO must be aligned to device logical sector size */
493 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
494 return -EINVAL;
495
496 /*
497 * Don't take the exclusive iolock here unless the I/O is unaligned to
498 * the file system block size. We don't need to consider the EOF
499 * extension case here because xfs_file_aio_write_checks() will relock
500 * the inode as necessary for EOF zeroing cases and fill out the new
501 * inode size as appropriate.
502 */
503 if ((iocb->ki_pos & mp->m_blockmask) ||
504 ((iocb->ki_pos + count) & mp->m_blockmask)) {
505 unaligned_io = 1;
506
507 /*
508 * We can't properly handle unaligned direct I/O to reflink
509 * files yet, as we can't unshare a partial block.
510 */
511 if (xfs_is_cow_inode(ip)) {
512 trace_xfs_reflink_bounce_dio_write(ip, iocb->ki_pos, count);
513 return -EREMCHG;
514 }
515 iolock = XFS_IOLOCK_EXCL;
516 } else {
517 iolock = XFS_IOLOCK_SHARED;
518 }
519
520 if (iocb->ki_flags & IOCB_NOWAIT) {
521 /* unaligned dio always waits, bail */
522 if (unaligned_io)
523 return -EAGAIN;
524 if (!xfs_ilock_nowait(ip, iolock))
525 return -EAGAIN;
526 } else {
527 xfs_ilock(ip, iolock);
528 }
529
530 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
531 if (ret)
532 goto out;
533 count = iov_iter_count(from);
534
535 /*
536 * If we are doing unaligned IO, we can't allow any other overlapping IO
537 * in-flight at the same time or we risk data corruption. Wait for all
538 * other IO to drain before we submit. If the IO is aligned, demote the
539 * iolock if we had to take the exclusive lock in
540 * xfs_file_aio_write_checks() for other reasons.
541 */
542 if (unaligned_io) {
543 inode_dio_wait(inode);
544 } else if (iolock == XFS_IOLOCK_EXCL) {
545 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
546 iolock = XFS_IOLOCK_SHARED;
547 }
548
549 trace_xfs_file_direct_write(ip, count, iocb->ki_pos);
550 ret = iomap_dio_rw(iocb, from, &xfs_iomap_ops, &xfs_dio_write_ops);
551
552 /*
553 * If unaligned, this is the only IO in-flight. If it has not yet
554 * completed, wait on it before we release the iolock to prevent
555 * subsequent overlapping IO.
556 */
557 if (ret == -EIOCBQUEUED && unaligned_io)
558 inode_dio_wait(inode);
559out:
560 xfs_iunlock(ip, iolock);
561
562 /*
563 * No fallback to buffered IO on errors for XFS, direct IO will either
564 * complete fully or fail.
565 */
566 ASSERT(ret < 0 || ret == count);
567 return ret;
568}
569
570static noinline ssize_t
571xfs_file_dax_write(
572 struct kiocb *iocb,
573 struct iov_iter *from)
574{
575 struct inode *inode = iocb->ki_filp->f_mapping->host;
576 struct xfs_inode *ip = XFS_I(inode);
577 int iolock = XFS_IOLOCK_EXCL;
578 ssize_t ret, error = 0;
579 size_t count;
580 loff_t pos;
581
582 if (iocb->ki_flags & IOCB_NOWAIT) {
583 if (!xfs_ilock_nowait(ip, iolock))
584 return -EAGAIN;
585 } else {
586 xfs_ilock(ip, iolock);
587 }
588
589 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
590 if (ret)
591 goto out;
592
593 pos = iocb->ki_pos;
594 count = iov_iter_count(from);
595
596 trace_xfs_file_dax_write(ip, count, pos);
597 ret = dax_iomap_rw(iocb, from, &xfs_iomap_ops);
598 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
599 i_size_write(inode, iocb->ki_pos);
600 error = xfs_setfilesize(ip, pos, ret);
601 }
602out:
603 xfs_iunlock(ip, iolock);
604 if (error)
605 return error;
606
607 if (ret > 0) {
608 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
609
610 /* Handle various SYNC-type writes */
611 ret = generic_write_sync(iocb, ret);
612 }
613 return ret;
614}
615
616STATIC ssize_t
617xfs_file_buffered_aio_write(
618 struct kiocb *iocb,
619 struct iov_iter *from)
620{
621 struct file *file = iocb->ki_filp;
622 struct address_space *mapping = file->f_mapping;
623 struct inode *inode = mapping->host;
624 struct xfs_inode *ip = XFS_I(inode);
625 ssize_t ret;
626 int enospc = 0;
627 int iolock;
628
629 if (iocb->ki_flags & IOCB_NOWAIT)
630 return -EOPNOTSUPP;
631
632write_retry:
633 iolock = XFS_IOLOCK_EXCL;
634 xfs_ilock(ip, iolock);
635
636 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
637 if (ret)
638 goto out;
639
640 /* We can write back this queue in page reclaim */
641 current->backing_dev_info = inode_to_bdi(inode);
642
643 trace_xfs_file_buffered_write(ip, iov_iter_count(from), iocb->ki_pos);
644 ret = iomap_file_buffered_write(iocb, from, &xfs_iomap_ops);
645 if (likely(ret >= 0))
646 iocb->ki_pos += ret;
647
648 /*
649 * If we hit a space limit, try to free up some lingering preallocated
650 * space before returning an error. In the case of ENOSPC, first try to
651 * write back all dirty inodes to free up some of the excess reserved
652 * metadata space. This reduces the chances that the eofblocks scan
653 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
654 * also behaves as a filter to prevent too many eofblocks scans from
655 * running at the same time.
656 */
657 if (ret == -EDQUOT && !enospc) {
658 xfs_iunlock(ip, iolock);
659 enospc = xfs_inode_free_quota_eofblocks(ip);
660 if (enospc)
661 goto write_retry;
662 enospc = xfs_inode_free_quota_cowblocks(ip);
663 if (enospc)
664 goto write_retry;
665 iolock = 0;
666 } else if (ret == -ENOSPC && !enospc) {
667 struct xfs_eofblocks eofb = {0};
668
669 enospc = 1;
670 xfs_flush_inodes(ip->i_mount);
671
672 xfs_iunlock(ip, iolock);
673 eofb.eof_flags = XFS_EOF_FLAGS_SYNC;
674 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
675 xfs_icache_free_cowblocks(ip->i_mount, &eofb);
676 goto write_retry;
677 }
678
679 current->backing_dev_info = NULL;
680out:
681 if (iolock)
682 xfs_iunlock(ip, iolock);
683
684 if (ret > 0) {
685 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
686 /* Handle various SYNC-type writes */
687 ret = generic_write_sync(iocb, ret);
688 }
689 return ret;
690}
691
692STATIC ssize_t
693xfs_file_write_iter(
694 struct kiocb *iocb,
695 struct iov_iter *from)
696{
697 struct file *file = iocb->ki_filp;
698 struct address_space *mapping = file->f_mapping;
699 struct inode *inode = mapping->host;
700 struct xfs_inode *ip = XFS_I(inode);
701 ssize_t ret;
702 size_t ocount = iov_iter_count(from);
703
704 XFS_STATS_INC(ip->i_mount, xs_write_calls);
705
706 if (ocount == 0)
707 return 0;
708
709 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
710 return -EIO;
711
712 if (IS_DAX(inode))
713 return xfs_file_dax_write(iocb, from);
714
715 if (iocb->ki_flags & IOCB_DIRECT) {
716 /*
717 * Allow a directio write to fall back to a buffered
718 * write *only* in the case that we're doing a reflink
719 * CoW. In all other directio scenarios we do not
720 * allow an operation to fall back to buffered mode.
721 */
722 ret = xfs_file_dio_aio_write(iocb, from);
723 if (ret != -EREMCHG)
724 return ret;
725 }
726
727 return xfs_file_buffered_aio_write(iocb, from);
728}
729
730static void
731xfs_wait_dax_page(
732 struct inode *inode)
733{
734 struct xfs_inode *ip = XFS_I(inode);
735
736 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
737 schedule();
738 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
739}
740
741static int
742xfs_break_dax_layouts(
743 struct inode *inode,
744 bool *retry)
745{
746 struct page *page;
747
748 ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
749
750 page = dax_layout_busy_page(inode->i_mapping);
751 if (!page)
752 return 0;
753
754 *retry = true;
755 return ___wait_var_event(&page->_refcount,
756 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
757 0, 0, xfs_wait_dax_page(inode));
758}
759
760int
761xfs_break_layouts(
762 struct inode *inode,
763 uint *iolock,
764 enum layout_break_reason reason)
765{
766 bool retry;
767 int error;
768
769 ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
770
771 do {
772 retry = false;
773 switch (reason) {
774 case BREAK_UNMAP:
775 error = xfs_break_dax_layouts(inode, &retry);
776 if (error || retry)
777 break;
778 /* fall through */
779 case BREAK_WRITE:
780 error = xfs_break_leased_layouts(inode, iolock, &retry);
781 break;
782 default:
783 WARN_ON_ONCE(1);
784 error = -EINVAL;
785 }
786 } while (error == 0 && retry);
787
788 return error;
789}
790
791#define XFS_FALLOC_FL_SUPPORTED \
792 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
793 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
794 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
795
796STATIC long
797xfs_file_fallocate(
798 struct file *file,
799 int mode,
800 loff_t offset,
801 loff_t len)
802{
803 struct inode *inode = file_inode(file);
804 struct xfs_inode *ip = XFS_I(inode);
805 long error;
806 enum xfs_prealloc_flags flags = 0;
807 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
808 loff_t new_size = 0;
809 bool do_file_insert = false;
810
811 if (!S_ISREG(inode->i_mode))
812 return -EINVAL;
813 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
814 return -EOPNOTSUPP;
815
816 xfs_ilock(ip, iolock);
817 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
818 if (error)
819 goto out_unlock;
820
821 if (mode & FALLOC_FL_PUNCH_HOLE) {
822 error = xfs_free_file_space(ip, offset, len);
823 if (error)
824 goto out_unlock;
825 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
826 unsigned int blksize_mask = i_blocksize(inode) - 1;
827
828 if (offset & blksize_mask || len & blksize_mask) {
829 error = -EINVAL;
830 goto out_unlock;
831 }
832
833 /*
834 * There is no need to overlap collapse range with EOF,
835 * in which case it is effectively a truncate operation
836 */
837 if (offset + len >= i_size_read(inode)) {
838 error = -EINVAL;
839 goto out_unlock;
840 }
841
842 new_size = i_size_read(inode) - len;
843
844 error = xfs_collapse_file_space(ip, offset, len);
845 if (error)
846 goto out_unlock;
847 } else if (mode & FALLOC_FL_INSERT_RANGE) {
848 unsigned int blksize_mask = i_blocksize(inode) - 1;
849 loff_t isize = i_size_read(inode);
850
851 if (offset & blksize_mask || len & blksize_mask) {
852 error = -EINVAL;
853 goto out_unlock;
854 }
855
856 /*
857 * New inode size must not exceed ->s_maxbytes, accounting for
858 * possible signed overflow.
859 */
860 if (inode->i_sb->s_maxbytes - isize < len) {
861 error = -EFBIG;
862 goto out_unlock;
863 }
864 new_size = isize + len;
865
866 /* Offset should be less than i_size */
867 if (offset >= isize) {
868 error = -EINVAL;
869 goto out_unlock;
870 }
871 do_file_insert = true;
872 } else {
873 flags |= XFS_PREALLOC_SET;
874
875 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
876 offset + len > i_size_read(inode)) {
877 new_size = offset + len;
878 error = inode_newsize_ok(inode, new_size);
879 if (error)
880 goto out_unlock;
881 }
882
883 if (mode & FALLOC_FL_ZERO_RANGE) {
884 error = xfs_zero_file_space(ip, offset, len);
885 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
886 error = xfs_reflink_unshare(ip, offset, len);
887 if (error)
888 goto out_unlock;
889
890 if (!xfs_is_always_cow_inode(ip)) {
891 error = xfs_alloc_file_space(ip, offset, len,
892 XFS_BMAPI_PREALLOC);
893 }
894 } else {
895 /*
896 * If always_cow mode we can't use preallocations and
897 * thus should not create them.
898 */
899 if (xfs_is_always_cow_inode(ip)) {
900 error = -EOPNOTSUPP;
901 goto out_unlock;
902 }
903
904 error = xfs_alloc_file_space(ip, offset, len,
905 XFS_BMAPI_PREALLOC);
906 }
907 if (error)
908 goto out_unlock;
909 }
910
911 if (file->f_flags & O_DSYNC)
912 flags |= XFS_PREALLOC_SYNC;
913
914 error = xfs_update_prealloc_flags(ip, flags);
915 if (error)
916 goto out_unlock;
917
918 /* Change file size if needed */
919 if (new_size) {
920 struct iattr iattr;
921
922 iattr.ia_valid = ATTR_SIZE;
923 iattr.ia_size = new_size;
924 error = xfs_vn_setattr_size(file_dentry(file), &iattr);
925 if (error)
926 goto out_unlock;
927 }
928
929 /*
930 * Perform hole insertion now that the file size has been
931 * updated so that if we crash during the operation we don't
932 * leave shifted extents past EOF and hence losing access to
933 * the data that is contained within them.
934 */
935 if (do_file_insert)
936 error = xfs_insert_file_space(ip, offset, len);
937
938out_unlock:
939 xfs_iunlock(ip, iolock);
940 return error;
941}
942
943STATIC int
944xfs_file_fadvise(
945 struct file *file,
946 loff_t start,
947 loff_t end,
948 int advice)
949{
950 struct xfs_inode *ip = XFS_I(file_inode(file));
951 int ret;
952 int lockflags = 0;
953
954 /*
955 * Operations creating pages in page cache need protection from hole
956 * punching and similar ops
957 */
958 if (advice == POSIX_FADV_WILLNEED) {
959 lockflags = XFS_IOLOCK_SHARED;
960 xfs_ilock(ip, lockflags);
961 }
962 ret = generic_fadvise(file, start, end, advice);
963 if (lockflags)
964 xfs_iunlock(ip, lockflags);
965 return ret;
966}
967
968STATIC loff_t
969xfs_file_remap_range(
970 struct file *file_in,
971 loff_t pos_in,
972 struct file *file_out,
973 loff_t pos_out,
974 loff_t len,
975 unsigned int remap_flags)
976{
977 struct inode *inode_in = file_inode(file_in);
978 struct xfs_inode *src = XFS_I(inode_in);
979 struct inode *inode_out = file_inode(file_out);
980 struct xfs_inode *dest = XFS_I(inode_out);
981 struct xfs_mount *mp = src->i_mount;
982 loff_t remapped = 0;
983 xfs_extlen_t cowextsize;
984 int ret;
985
986 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
987 return -EINVAL;
988
989 if (!xfs_sb_version_hasreflink(&mp->m_sb))
990 return -EOPNOTSUPP;
991
992 if (XFS_FORCED_SHUTDOWN(mp))
993 return -EIO;
994
995 /* Prepare and then clone file data. */
996 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
997 &len, remap_flags);
998 if (ret < 0 || len == 0)
999 return ret;
1000
1001 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1002
1003 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1004 &remapped);
1005 if (ret)
1006 goto out_unlock;
1007
1008 /*
1009 * Carry the cowextsize hint from src to dest if we're sharing the
1010 * entire source file to the entire destination file, the source file
1011 * has a cowextsize hint, and the destination file does not.
1012 */
1013 cowextsize = 0;
1014 if (pos_in == 0 && len == i_size_read(inode_in) &&
1015 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1016 pos_out == 0 && len >= i_size_read(inode_out) &&
1017 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1018 cowextsize = src->i_d.di_cowextsize;
1019
1020 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1021 remap_flags);
1022
1023out_unlock:
1024 xfs_reflink_remap_unlock(file_in, file_out);
1025 if (ret)
1026 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1027 return remapped > 0 ? remapped : ret;
1028}
1029
1030STATIC int
1031xfs_file_open(
1032 struct inode *inode,
1033 struct file *file)
1034{
1035 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
1036 return -EFBIG;
1037 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
1038 return -EIO;
1039 file->f_mode |= FMODE_NOWAIT;
1040 return 0;
1041}
1042
1043STATIC int
1044xfs_dir_open(
1045 struct inode *inode,
1046 struct file *file)
1047{
1048 struct xfs_inode *ip = XFS_I(inode);
1049 int mode;
1050 int error;
1051
1052 error = xfs_file_open(inode, file);
1053 if (error)
1054 return error;
1055
1056 /*
1057 * If there are any blocks, read-ahead block 0 as we're almost
1058 * certain to have the next operation be a read there.
1059 */
1060 mode = xfs_ilock_data_map_shared(ip);
1061 if (ip->i_d.di_nextents > 0)
1062 error = xfs_dir3_data_readahead(ip, 0, -1);
1063 xfs_iunlock(ip, mode);
1064 return error;
1065}
1066
1067STATIC int
1068xfs_file_release(
1069 struct inode *inode,
1070 struct file *filp)
1071{
1072 return xfs_release(XFS_I(inode));
1073}
1074
1075STATIC int
1076xfs_file_readdir(
1077 struct file *file,
1078 struct dir_context *ctx)
1079{
1080 struct inode *inode = file_inode(file);
1081 xfs_inode_t *ip = XFS_I(inode);
1082 size_t bufsize;
1083
1084 /*
1085 * The Linux API doesn't pass down the total size of the buffer
1086 * we read into down to the filesystem. With the filldir concept
1087 * it's not needed for correct information, but the XFS dir2 leaf
1088 * code wants an estimate of the buffer size to calculate it's
1089 * readahead window and size the buffers used for mapping to
1090 * physical blocks.
1091 *
1092 * Try to give it an estimate that's good enough, maybe at some
1093 * point we can change the ->readdir prototype to include the
1094 * buffer size. For now we use the current glibc buffer size.
1095 */
1096 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_d.di_size);
1097
1098 return xfs_readdir(NULL, ip, ctx, bufsize);
1099}
1100
1101STATIC loff_t
1102xfs_file_llseek(
1103 struct file *file,
1104 loff_t offset,
1105 int whence)
1106{
1107 struct inode *inode = file->f_mapping->host;
1108
1109 if (XFS_FORCED_SHUTDOWN(XFS_I(inode)->i_mount))
1110 return -EIO;
1111
1112 switch (whence) {
1113 default:
1114 return generic_file_llseek(file, offset, whence);
1115 case SEEK_HOLE:
1116 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1117 break;
1118 case SEEK_DATA:
1119 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1120 break;
1121 }
1122
1123 if (offset < 0)
1124 return offset;
1125 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1126}
1127
1128/*
1129 * Locking for serialisation of IO during page faults. This results in a lock
1130 * ordering of:
1131 *
1132 * mmap_sem (MM)
1133 * sb_start_pagefault(vfs, freeze)
1134 * i_mmaplock (XFS - truncate serialisation)
1135 * page_lock (MM)
1136 * i_lock (XFS - extent map serialisation)
1137 */
1138static vm_fault_t
1139__xfs_filemap_fault(
1140 struct vm_fault *vmf,
1141 enum page_entry_size pe_size,
1142 bool write_fault)
1143{
1144 struct inode *inode = file_inode(vmf->vma->vm_file);
1145 struct xfs_inode *ip = XFS_I(inode);
1146 vm_fault_t ret;
1147
1148 trace_xfs_filemap_fault(ip, pe_size, write_fault);
1149
1150 if (write_fault) {
1151 sb_start_pagefault(inode->i_sb);
1152 file_update_time(vmf->vma->vm_file);
1153 }
1154
1155 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1156 if (IS_DAX(inode)) {
1157 pfn_t pfn;
1158
1159 ret = dax_iomap_fault(vmf, pe_size, &pfn, NULL, &xfs_iomap_ops);
1160 if (ret & VM_FAULT_NEEDDSYNC)
1161 ret = dax_finish_sync_fault(vmf, pe_size, pfn);
1162 } else {
1163 if (write_fault)
1164 ret = iomap_page_mkwrite(vmf, &xfs_iomap_ops);
1165 else
1166 ret = filemap_fault(vmf);
1167 }
1168 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1169
1170 if (write_fault)
1171 sb_end_pagefault(inode->i_sb);
1172 return ret;
1173}
1174
1175static vm_fault_t
1176xfs_filemap_fault(
1177 struct vm_fault *vmf)
1178{
1179 /* DAX can shortcut the normal fault path on write faults! */
1180 return __xfs_filemap_fault(vmf, PE_SIZE_PTE,
1181 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1182 (vmf->flags & FAULT_FLAG_WRITE));
1183}
1184
1185static vm_fault_t
1186xfs_filemap_huge_fault(
1187 struct vm_fault *vmf,
1188 enum page_entry_size pe_size)
1189{
1190 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1191 return VM_FAULT_FALLBACK;
1192
1193 /* DAX can shortcut the normal fault path on write faults! */
1194 return __xfs_filemap_fault(vmf, pe_size,
1195 (vmf->flags & FAULT_FLAG_WRITE));
1196}
1197
1198static vm_fault_t
1199xfs_filemap_page_mkwrite(
1200 struct vm_fault *vmf)
1201{
1202 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1203}
1204
1205/*
1206 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1207 * on write faults. In reality, it needs to serialise against truncate and
1208 * prepare memory for writing so handle is as standard write fault.
1209 */
1210static vm_fault_t
1211xfs_filemap_pfn_mkwrite(
1212 struct vm_fault *vmf)
1213{
1214
1215 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1216}
1217
1218static const struct vm_operations_struct xfs_file_vm_ops = {
1219 .fault = xfs_filemap_fault,
1220 .huge_fault = xfs_filemap_huge_fault,
1221 .map_pages = filemap_map_pages,
1222 .page_mkwrite = xfs_filemap_page_mkwrite,
1223 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1224};
1225
1226STATIC int
1227xfs_file_mmap(
1228 struct file *filp,
1229 struct vm_area_struct *vma)
1230{
1231 struct dax_device *dax_dev;
1232
1233 dax_dev = xfs_find_daxdev_for_inode(file_inode(filp));
1234 /*
1235 * We don't support synchronous mappings for non-DAX files and
1236 * for DAX files if underneath dax_device is not synchronous.
1237 */
1238 if (!daxdev_mapping_supported(vma, dax_dev))
1239 return -EOPNOTSUPP;
1240
1241 file_accessed(filp);
1242 vma->vm_ops = &xfs_file_vm_ops;
1243 if (IS_DAX(file_inode(filp)))
1244 vma->vm_flags |= VM_HUGEPAGE;
1245 return 0;
1246}
1247
1248const struct file_operations xfs_file_operations = {
1249 .llseek = xfs_file_llseek,
1250 .read_iter = xfs_file_read_iter,
1251 .write_iter = xfs_file_write_iter,
1252 .splice_read = generic_file_splice_read,
1253 .splice_write = iter_file_splice_write,
1254 .iopoll = iomap_dio_iopoll,
1255 .unlocked_ioctl = xfs_file_ioctl,
1256#ifdef CONFIG_COMPAT
1257 .compat_ioctl = xfs_file_compat_ioctl,
1258#endif
1259 .mmap = xfs_file_mmap,
1260 .mmap_supported_flags = MAP_SYNC,
1261 .open = xfs_file_open,
1262 .release = xfs_file_release,
1263 .fsync = xfs_file_fsync,
1264 .get_unmapped_area = thp_get_unmapped_area,
1265 .fallocate = xfs_file_fallocate,
1266 .fadvise = xfs_file_fadvise,
1267 .remap_file_range = xfs_file_remap_range,
1268};
1269
1270const struct file_operations xfs_dir_file_operations = {
1271 .open = xfs_dir_open,
1272 .read = generic_read_dir,
1273 .iterate_shared = xfs_file_readdir,
1274 .llseek = generic_file_llseek,
1275 .unlocked_ioctl = xfs_file_ioctl,
1276#ifdef CONFIG_COMPAT
1277 .compat_ioctl = xfs_file_compat_ioctl,
1278#endif
1279 .fsync = xfs_dir_fsync,
1280};