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/*
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
48static const struct vm_operations_struct xfs_file_vm_ops;
49
50/*
51 * Clear the specified ranges to zero through either the pagecache or DAX.
52 * Holes and unwritten extents will be left as-is as they already are zeroed.
53 */
54int
55xfs_zero_range(
56 struct xfs_inode *ip,
57 xfs_off_t pos,
58 xfs_off_t count,
59 bool *did_zero)
60{
61 return iomap_zero_range(VFS_I(ip), pos, count, NULL, &xfs_iomap_ops);
62}
63
64int
65xfs_update_prealloc_flags(
66 struct xfs_inode *ip,
67 enum xfs_prealloc_flags flags)
68{
69 struct xfs_trans *tp;
70 int error;
71
72 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_writeid,
73 0, 0, 0, &tp);
74 if (error)
75 return error;
76
77 xfs_ilock(ip, XFS_ILOCK_EXCL);
78 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
79
80 if (!(flags & XFS_PREALLOC_INVISIBLE)) {
81 VFS_I(ip)->i_mode &= ~S_ISUID;
82 if (VFS_I(ip)->i_mode & S_IXGRP)
83 VFS_I(ip)->i_mode &= ~S_ISGID;
84 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
85 }
86
87 if (flags & XFS_PREALLOC_SET)
88 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
89 if (flags & XFS_PREALLOC_CLEAR)
90 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
91
92 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
93 if (flags & XFS_PREALLOC_SYNC)
94 xfs_trans_set_sync(tp);
95 return xfs_trans_commit(tp);
96}
97
98/*
99 * Fsync operations on directories are much simpler than on regular files,
100 * as there is no file data to flush, and thus also no need for explicit
101 * cache flush operations, and there are no non-transaction metadata updates
102 * on directories either.
103 */
104STATIC int
105xfs_dir_fsync(
106 struct file *file,
107 loff_t start,
108 loff_t end,
109 int datasync)
110{
111 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
112 struct xfs_mount *mp = ip->i_mount;
113 xfs_lsn_t lsn = 0;
114
115 trace_xfs_dir_fsync(ip);
116
117 xfs_ilock(ip, XFS_ILOCK_SHARED);
118 if (xfs_ipincount(ip))
119 lsn = ip->i_itemp->ili_last_lsn;
120 xfs_iunlock(ip, XFS_ILOCK_SHARED);
121
122 if (!lsn)
123 return 0;
124 return _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, NULL);
125}
126
127STATIC int
128xfs_file_fsync(
129 struct file *file,
130 loff_t start,
131 loff_t end,
132 int datasync)
133{
134 struct inode *inode = file->f_mapping->host;
135 struct xfs_inode *ip = XFS_I(inode);
136 struct xfs_mount *mp = ip->i_mount;
137 int error = 0;
138 int log_flushed = 0;
139 xfs_lsn_t lsn = 0;
140
141 trace_xfs_file_fsync(ip);
142
143 error = filemap_write_and_wait_range(inode->i_mapping, start, end);
144 if (error)
145 return error;
146
147 if (XFS_FORCED_SHUTDOWN(mp))
148 return -EIO;
149
150 xfs_iflags_clear(ip, XFS_ITRUNCATED);
151
152 /*
153 * If we have an RT and/or log subvolume we need to make sure to flush
154 * the write cache the device used for file data first. This is to
155 * ensure newly written file data make it to disk before logging the new
156 * inode size in case of an extending write.
157 */
158 if (XFS_IS_REALTIME_INODE(ip))
159 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
160 else if (mp->m_logdev_targp != mp->m_ddev_targp)
161 xfs_blkdev_issue_flush(mp->m_ddev_targp);
162
163 /*
164 * All metadata updates are logged, which means that we just have to
165 * flush the log up to the latest LSN that touched the inode. If we have
166 * concurrent fsync/fdatasync() calls, we need them to all block on the
167 * log force before we clear the ili_fsync_fields field. This ensures
168 * that we don't get a racing sync operation that does not wait for the
169 * metadata to hit the journal before returning. If we race with
170 * clearing the ili_fsync_fields, then all that will happen is the log
171 * force will do nothing as the lsn will already be on disk. We can't
172 * race with setting ili_fsync_fields because that is done under
173 * XFS_ILOCK_EXCL, and that can't happen because we hold the lock shared
174 * until after the ili_fsync_fields is cleared.
175 */
176 xfs_ilock(ip, XFS_ILOCK_SHARED);
177 if (xfs_ipincount(ip)) {
178 if (!datasync ||
179 (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
180 lsn = ip->i_itemp->ili_last_lsn;
181 }
182
183 if (lsn) {
184 error = _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
185 ip->i_itemp->ili_fsync_fields = 0;
186 }
187 xfs_iunlock(ip, XFS_ILOCK_SHARED);
188
189 /*
190 * If we only have a single device, and the log force about was
191 * a no-op we might have to flush the data device cache here.
192 * This can only happen for fdatasync/O_DSYNC if we were overwriting
193 * an already allocated file and thus do not have any metadata to
194 * commit.
195 */
196 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
197 mp->m_logdev_targp == mp->m_ddev_targp)
198 xfs_blkdev_issue_flush(mp->m_ddev_targp);
199
200 return error;
201}
202
203STATIC ssize_t
204xfs_file_dio_aio_read(
205 struct kiocb *iocb,
206 struct iov_iter *to)
207{
208 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
209 size_t count = iov_iter_count(to);
210 ssize_t ret;
211
212 trace_xfs_file_direct_read(ip, count, iocb->ki_pos);
213
214 if (!count)
215 return 0; /* skip atime */
216
217 file_accessed(iocb->ki_filp);
218
219 xfs_ilock(ip, XFS_IOLOCK_SHARED);
220 ret = iomap_dio_rw(iocb, to, &xfs_iomap_ops, NULL);
221 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
222
223 return ret;
224}
225
226static noinline ssize_t
227xfs_file_dax_read(
228 struct kiocb *iocb,
229 struct iov_iter *to)
230{
231 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
232 size_t count = iov_iter_count(to);
233 ssize_t ret = 0;
234
235 trace_xfs_file_dax_read(ip, count, iocb->ki_pos);
236
237 if (!count)
238 return 0; /* skip atime */
239
240 xfs_ilock(ip, XFS_IOLOCK_SHARED);
241 ret = dax_iomap_rw(iocb, to, &xfs_iomap_ops);
242 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
243
244 file_accessed(iocb->ki_filp);
245 return ret;
246}
247
248STATIC ssize_t
249xfs_file_buffered_aio_read(
250 struct kiocb *iocb,
251 struct iov_iter *to)
252{
253 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
254 ssize_t ret;
255
256 trace_xfs_file_buffered_read(ip, iov_iter_count(to), iocb->ki_pos);
257
258 xfs_ilock(ip, XFS_IOLOCK_SHARED);
259 ret = generic_file_read_iter(iocb, to);
260 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
261
262 return ret;
263}
264
265STATIC ssize_t
266xfs_file_read_iter(
267 struct kiocb *iocb,
268 struct iov_iter *to)
269{
270 struct inode *inode = file_inode(iocb->ki_filp);
271 struct xfs_mount *mp = XFS_I(inode)->i_mount;
272 ssize_t ret = 0;
273
274 XFS_STATS_INC(mp, xs_read_calls);
275
276 if (XFS_FORCED_SHUTDOWN(mp))
277 return -EIO;
278
279 if (IS_DAX(inode))
280 ret = xfs_file_dax_read(iocb, to);
281 else if (iocb->ki_flags & IOCB_DIRECT)
282 ret = xfs_file_dio_aio_read(iocb, to);
283 else
284 ret = xfs_file_buffered_aio_read(iocb, to);
285
286 if (ret > 0)
287 XFS_STATS_ADD(mp, xs_read_bytes, ret);
288 return ret;
289}
290
291/*
292 * Zero any on disk space between the current EOF and the new, larger EOF.
293 *
294 * This handles the normal case of zeroing the remainder of the last block in
295 * the file and the unusual case of zeroing blocks out beyond the size of the
296 * file. This second case only happens with fixed size extents and when the
297 * system crashes before the inode size was updated but after blocks were
298 * allocated.
299 *
300 * Expects the iolock to be held exclusive, and will take the ilock internally.
301 */
302int /* error (positive) */
303xfs_zero_eof(
304 struct xfs_inode *ip,
305 xfs_off_t offset, /* starting I/O offset */
306 xfs_fsize_t isize, /* current inode size */
307 bool *did_zeroing)
308{
309 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
310 ASSERT(offset > isize);
311
312 trace_xfs_zero_eof(ip, isize, offset - isize);
313 return xfs_zero_range(ip, isize, offset - isize, did_zeroing);
314}
315
316/*
317 * Common pre-write limit and setup checks.
318 *
319 * Called with the iolocked held either shared and exclusive according to
320 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
321 * if called for a direct write beyond i_size.
322 */
323STATIC ssize_t
324xfs_file_aio_write_checks(
325 struct kiocb *iocb,
326 struct iov_iter *from,
327 int *iolock)
328{
329 struct file *file = iocb->ki_filp;
330 struct inode *inode = file->f_mapping->host;
331 struct xfs_inode *ip = XFS_I(inode);
332 ssize_t error = 0;
333 size_t count = iov_iter_count(from);
334 bool drained_dio = false;
335
336restart:
337 error = generic_write_checks(iocb, from);
338 if (error <= 0)
339 return error;
340
341 error = xfs_break_layouts(inode, iolock);
342 if (error)
343 return error;
344
345 /*
346 * For changing security info in file_remove_privs() we need i_rwsem
347 * exclusively.
348 */
349 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
350 xfs_iunlock(ip, *iolock);
351 *iolock = XFS_IOLOCK_EXCL;
352 xfs_ilock(ip, *iolock);
353 goto restart;
354 }
355 /*
356 * If the offset is beyond the size of the file, we need to zero any
357 * blocks that fall between the existing EOF and the start of this
358 * write. If zeroing is needed and we are currently holding the
359 * iolock shared, we need to update it to exclusive which implies
360 * having to redo all checks before.
361 *
362 * We need to serialise against EOF updates that occur in IO
363 * completions here. We want to make sure that nobody is changing the
364 * size while we do this check until we have placed an IO barrier (i.e.
365 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
366 * The spinlock effectively forms a memory barrier once we have the
367 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
368 * and hence be able to correctly determine if we need to run zeroing.
369 */
370 spin_lock(&ip->i_flags_lock);
371 if (iocb->ki_pos > i_size_read(inode)) {
372 bool zero = false;
373
374 spin_unlock(&ip->i_flags_lock);
375 if (!drained_dio) {
376 if (*iolock == XFS_IOLOCK_SHARED) {
377 xfs_iunlock(ip, *iolock);
378 *iolock = XFS_IOLOCK_EXCL;
379 xfs_ilock(ip, *iolock);
380 iov_iter_reexpand(from, count);
381 }
382 /*
383 * We now have an IO submission barrier in place, but
384 * AIO can do EOF updates during IO completion and hence
385 * we now need to wait for all of them to drain. Non-AIO
386 * DIO will have drained before we are given the
387 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
388 * no-op.
389 */
390 inode_dio_wait(inode);
391 drained_dio = true;
392 goto restart;
393 }
394 error = xfs_zero_eof(ip, iocb->ki_pos, i_size_read(inode), &zero);
395 if (error)
396 return error;
397 } else
398 spin_unlock(&ip->i_flags_lock);
399
400 /*
401 * Updating the timestamps will grab the ilock again from
402 * xfs_fs_dirty_inode, so we have to call it after dropping the
403 * lock above. Eventually we should look into a way to avoid
404 * the pointless lock roundtrip.
405 */
406 if (likely(!(file->f_mode & FMODE_NOCMTIME))) {
407 error = file_update_time(file);
408 if (error)
409 return error;
410 }
411
412 /*
413 * If we're writing the file then make sure to clear the setuid and
414 * setgid bits if the process is not being run by root. This keeps
415 * people from modifying setuid and setgid binaries.
416 */
417 if (!IS_NOSEC(inode))
418 return file_remove_privs(file);
419 return 0;
420}
421
422static int
423xfs_dio_write_end_io(
424 struct kiocb *iocb,
425 ssize_t size,
426 unsigned flags)
427{
428 struct inode *inode = file_inode(iocb->ki_filp);
429 struct xfs_inode *ip = XFS_I(inode);
430 loff_t offset = iocb->ki_pos;
431 bool update_size = false;
432 int error = 0;
433
434 trace_xfs_end_io_direct_write(ip, offset, size);
435
436 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
437 return -EIO;
438
439 if (size <= 0)
440 return size;
441
442 /*
443 * We need to update the in-core inode size here so that we don't end up
444 * with the on-disk inode size being outside the in-core inode size. We
445 * have no other method of updating EOF for AIO, so always do it here
446 * if necessary.
447 *
448 * We need to lock the test/set EOF update as we can be racing with
449 * other IO completions here to update the EOF. Failing to serialise
450 * here can result in EOF moving backwards and Bad Things Happen when
451 * that occurs.
452 */
453 spin_lock(&ip->i_flags_lock);
454 if (offset + size > i_size_read(inode)) {
455 i_size_write(inode, offset + size);
456 update_size = true;
457 }
458 spin_unlock(&ip->i_flags_lock);
459
460 if (flags & IOMAP_DIO_COW) {
461 error = xfs_reflink_end_cow(ip, offset, size);
462 if (error)
463 return error;
464 }
465
466 if (flags & IOMAP_DIO_UNWRITTEN)
467 error = xfs_iomap_write_unwritten(ip, offset, size);
468 else if (update_size)
469 error = xfs_setfilesize(ip, offset, size);
470
471 return error;
472}
473
474/*
475 * xfs_file_dio_aio_write - handle direct IO writes
476 *
477 * Lock the inode appropriately to prepare for and issue a direct IO write.
478 * By separating it from the buffered write path we remove all the tricky to
479 * follow locking changes and looping.
480 *
481 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
482 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
483 * pages are flushed out.
484 *
485 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
486 * allowing them to be done in parallel with reads and other direct IO writes.
487 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
488 * needs to do sub-block zeroing and that requires serialisation against other
489 * direct IOs to the same block. In this case we need to serialise the
490 * submission of the unaligned IOs so that we don't get racing block zeroing in
491 * the dio layer. To avoid the problem with aio, we also need to wait for
492 * outstanding IOs to complete so that unwritten extent conversion is completed
493 * before we try to map the overlapping block. This is currently implemented by
494 * hitting it with a big hammer (i.e. inode_dio_wait()).
495 *
496 * Returns with locks held indicated by @iolock and errors indicated by
497 * negative return values.
498 */
499STATIC ssize_t
500xfs_file_dio_aio_write(
501 struct kiocb *iocb,
502 struct iov_iter *from)
503{
504 struct file *file = iocb->ki_filp;
505 struct address_space *mapping = file->f_mapping;
506 struct inode *inode = mapping->host;
507 struct xfs_inode *ip = XFS_I(inode);
508 struct xfs_mount *mp = ip->i_mount;
509 ssize_t ret = 0;
510 int unaligned_io = 0;
511 int iolock;
512 size_t count = iov_iter_count(from);
513 struct xfs_buftarg *target = XFS_IS_REALTIME_INODE(ip) ?
514 mp->m_rtdev_targp : mp->m_ddev_targp;
515
516 /* DIO must be aligned to device logical sector size */
517 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
518 return -EINVAL;
519
520 /*
521 * Don't take the exclusive iolock here unless the I/O is unaligned to
522 * the file system block size. We don't need to consider the EOF
523 * extension case here because xfs_file_aio_write_checks() will relock
524 * the inode as necessary for EOF zeroing cases and fill out the new
525 * inode size as appropriate.
526 */
527 if ((iocb->ki_pos & mp->m_blockmask) ||
528 ((iocb->ki_pos + count) & mp->m_blockmask)) {
529 unaligned_io = 1;
530
531 /*
532 * We can't properly handle unaligned direct I/O to reflink
533 * files yet, as we can't unshare a partial block.
534 */
535 if (xfs_is_reflink_inode(ip)) {
536 trace_xfs_reflink_bounce_dio_write(ip, iocb->ki_pos, count);
537 return -EREMCHG;
538 }
539 iolock = XFS_IOLOCK_EXCL;
540 } else {
541 iolock = XFS_IOLOCK_SHARED;
542 }
543
544 xfs_ilock(ip, iolock);
545
546 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
547 if (ret)
548 goto out;
549 count = iov_iter_count(from);
550
551 /*
552 * If we are doing unaligned IO, wait for all other IO to drain,
553 * otherwise demote the lock if we had to take the exclusive lock
554 * for other reasons in xfs_file_aio_write_checks.
555 */
556 if (unaligned_io)
557 inode_dio_wait(inode);
558 else if (iolock == XFS_IOLOCK_EXCL) {
559 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
560 iolock = XFS_IOLOCK_SHARED;
561 }
562
563 trace_xfs_file_direct_write(ip, count, iocb->ki_pos);
564
565 /* If this is a block-aligned directio CoW, remap immediately. */
566 if (xfs_is_reflink_inode(ip) && !unaligned_io) {
567 ret = xfs_reflink_allocate_cow_range(ip, iocb->ki_pos, count);
568 if (ret)
569 goto out;
570 }
571
572 ret = iomap_dio_rw(iocb, from, &xfs_iomap_ops, xfs_dio_write_end_io);
573out:
574 xfs_iunlock(ip, iolock);
575
576 /*
577 * No fallback to buffered IO on errors for XFS, direct IO will either
578 * complete fully or fail.
579 */
580 ASSERT(ret < 0 || ret == count);
581 return ret;
582}
583
584static noinline ssize_t
585xfs_file_dax_write(
586 struct kiocb *iocb,
587 struct iov_iter *from)
588{
589 struct inode *inode = iocb->ki_filp->f_mapping->host;
590 struct xfs_inode *ip = XFS_I(inode);
591 int iolock = XFS_IOLOCK_EXCL;
592 ssize_t ret, error = 0;
593 size_t count;
594 loff_t pos;
595
596 xfs_ilock(ip, iolock);
597 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
598 if (ret)
599 goto out;
600
601 pos = iocb->ki_pos;
602 count = iov_iter_count(from);
603
604 trace_xfs_file_dax_write(ip, count, pos);
605 ret = dax_iomap_rw(iocb, from, &xfs_iomap_ops);
606 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
607 i_size_write(inode, iocb->ki_pos);
608 error = xfs_setfilesize(ip, pos, ret);
609 }
610out:
611 xfs_iunlock(ip, iolock);
612 return error ? error : ret;
613}
614
615STATIC ssize_t
616xfs_file_buffered_aio_write(
617 struct kiocb *iocb,
618 struct iov_iter *from)
619{
620 struct file *file = iocb->ki_filp;
621 struct address_space *mapping = file->f_mapping;
622 struct inode *inode = mapping->host;
623 struct xfs_inode *ip = XFS_I(inode);
624 ssize_t ret;
625 int enospc = 0;
626 int iolock;
627
628write_retry:
629 iolock = XFS_IOLOCK_EXCL;
630 xfs_ilock(ip, iolock);
631
632 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
633 if (ret)
634 goto out;
635
636 /* We can write back this queue in page reclaim */
637 current->backing_dev_info = inode_to_bdi(inode);
638
639 trace_xfs_file_buffered_write(ip, iov_iter_count(from), iocb->ki_pos);
640 ret = iomap_file_buffered_write(iocb, from, &xfs_iomap_ops);
641 if (likely(ret >= 0))
642 iocb->ki_pos += ret;
643
644 /*
645 * If we hit a space limit, try to free up some lingering preallocated
646 * space before returning an error. In the case of ENOSPC, first try to
647 * write back all dirty inodes to free up some of the excess reserved
648 * metadata space. This reduces the chances that the eofblocks scan
649 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
650 * also behaves as a filter to prevent too many eofblocks scans from
651 * running at the same time.
652 */
653 if (ret == -EDQUOT && !enospc) {
654 xfs_iunlock(ip, iolock);
655 enospc = xfs_inode_free_quota_eofblocks(ip);
656 if (enospc)
657 goto write_retry;
658 enospc = xfs_inode_free_quota_cowblocks(ip);
659 if (enospc)
660 goto write_retry;
661 iolock = 0;
662 } else if (ret == -ENOSPC && !enospc) {
663 struct xfs_eofblocks eofb = {0};
664
665 enospc = 1;
666 xfs_flush_inodes(ip->i_mount);
667
668 xfs_iunlock(ip, iolock);
669 eofb.eof_flags = XFS_EOF_FLAGS_SYNC;
670 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
671 goto write_retry;
672 }
673
674 current->backing_dev_info = NULL;
675out:
676 if (iolock)
677 xfs_iunlock(ip, iolock);
678 return ret;
679}
680
681STATIC ssize_t
682xfs_file_write_iter(
683 struct kiocb *iocb,
684 struct iov_iter *from)
685{
686 struct file *file = iocb->ki_filp;
687 struct address_space *mapping = file->f_mapping;
688 struct inode *inode = mapping->host;
689 struct xfs_inode *ip = XFS_I(inode);
690 ssize_t ret;
691 size_t ocount = iov_iter_count(from);
692
693 XFS_STATS_INC(ip->i_mount, xs_write_calls);
694
695 if (ocount == 0)
696 return 0;
697
698 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
699 return -EIO;
700
701 if (IS_DAX(inode))
702 ret = xfs_file_dax_write(iocb, from);
703 else if (iocb->ki_flags & IOCB_DIRECT) {
704 /*
705 * Allow a directio write to fall back to a buffered
706 * write *only* in the case that we're doing a reflink
707 * CoW. In all other directio scenarios we do not
708 * allow an operation to fall back to buffered mode.
709 */
710 ret = xfs_file_dio_aio_write(iocb, from);
711 if (ret == -EREMCHG)
712 goto buffered;
713 } else {
714buffered:
715 ret = xfs_file_buffered_aio_write(iocb, from);
716 }
717
718 if (ret > 0) {
719 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
720
721 /* Handle various SYNC-type writes */
722 ret = generic_write_sync(iocb, ret);
723 }
724 return ret;
725}
726
727#define XFS_FALLOC_FL_SUPPORTED \
728 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
729 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
730 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
731
732STATIC long
733xfs_file_fallocate(
734 struct file *file,
735 int mode,
736 loff_t offset,
737 loff_t len)
738{
739 struct inode *inode = file_inode(file);
740 struct xfs_inode *ip = XFS_I(inode);
741 long error;
742 enum xfs_prealloc_flags flags = 0;
743 uint iolock = XFS_IOLOCK_EXCL;
744 loff_t new_size = 0;
745 bool do_file_insert = 0;
746
747 if (!S_ISREG(inode->i_mode))
748 return -EINVAL;
749 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
750 return -EOPNOTSUPP;
751
752 xfs_ilock(ip, iolock);
753 error = xfs_break_layouts(inode, &iolock);
754 if (error)
755 goto out_unlock;
756
757 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
758 iolock |= XFS_MMAPLOCK_EXCL;
759
760 if (mode & FALLOC_FL_PUNCH_HOLE) {
761 error = xfs_free_file_space(ip, offset, len);
762 if (error)
763 goto out_unlock;
764 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
765 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
766
767 if (offset & blksize_mask || len & blksize_mask) {
768 error = -EINVAL;
769 goto out_unlock;
770 }
771
772 /*
773 * There is no need to overlap collapse range with EOF,
774 * in which case it is effectively a truncate operation
775 */
776 if (offset + len >= i_size_read(inode)) {
777 error = -EINVAL;
778 goto out_unlock;
779 }
780
781 new_size = i_size_read(inode) - len;
782
783 error = xfs_collapse_file_space(ip, offset, len);
784 if (error)
785 goto out_unlock;
786 } else if (mode & FALLOC_FL_INSERT_RANGE) {
787 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
788
789 new_size = i_size_read(inode) + len;
790 if (offset & blksize_mask || len & blksize_mask) {
791 error = -EINVAL;
792 goto out_unlock;
793 }
794
795 /* check the new inode size does not wrap through zero */
796 if (new_size > inode->i_sb->s_maxbytes) {
797 error = -EFBIG;
798 goto out_unlock;
799 }
800
801 /* Offset should be less than i_size */
802 if (offset >= i_size_read(inode)) {
803 error = -EINVAL;
804 goto out_unlock;
805 }
806 do_file_insert = 1;
807 } else {
808 flags |= XFS_PREALLOC_SET;
809
810 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
811 offset + len > i_size_read(inode)) {
812 new_size = offset + len;
813 error = inode_newsize_ok(inode, new_size);
814 if (error)
815 goto out_unlock;
816 }
817
818 if (mode & FALLOC_FL_ZERO_RANGE)
819 error = xfs_zero_file_space(ip, offset, len);
820 else {
821 if (mode & FALLOC_FL_UNSHARE_RANGE) {
822 error = xfs_reflink_unshare(ip, offset, len);
823 if (error)
824 goto out_unlock;
825 }
826 error = xfs_alloc_file_space(ip, offset, len,
827 XFS_BMAPI_PREALLOC);
828 }
829 if (error)
830 goto out_unlock;
831 }
832
833 if (file->f_flags & O_DSYNC)
834 flags |= XFS_PREALLOC_SYNC;
835
836 error = xfs_update_prealloc_flags(ip, flags);
837 if (error)
838 goto out_unlock;
839
840 /* Change file size if needed */
841 if (new_size) {
842 struct iattr iattr;
843
844 iattr.ia_valid = ATTR_SIZE;
845 iattr.ia_size = new_size;
846 error = xfs_vn_setattr_size(file_dentry(file), &iattr);
847 if (error)
848 goto out_unlock;
849 }
850
851 /*
852 * Perform hole insertion now that the file size has been
853 * updated so that if we crash during the operation we don't
854 * leave shifted extents past EOF and hence losing access to
855 * the data that is contained within them.
856 */
857 if (do_file_insert)
858 error = xfs_insert_file_space(ip, offset, len);
859
860out_unlock:
861 xfs_iunlock(ip, iolock);
862 return error;
863}
864
865STATIC int
866xfs_file_clone_range(
867 struct file *file_in,
868 loff_t pos_in,
869 struct file *file_out,
870 loff_t pos_out,
871 u64 len)
872{
873 return xfs_reflink_remap_range(file_in, pos_in, file_out, pos_out,
874 len, false);
875}
876
877STATIC ssize_t
878xfs_file_dedupe_range(
879 struct file *src_file,
880 u64 loff,
881 u64 len,
882 struct file *dst_file,
883 u64 dst_loff)
884{
885 int error;
886
887 error = xfs_reflink_remap_range(src_file, loff, dst_file, dst_loff,
888 len, true);
889 if (error)
890 return error;
891 return len;
892}
893
894STATIC int
895xfs_file_open(
896 struct inode *inode,
897 struct file *file)
898{
899 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
900 return -EFBIG;
901 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
902 return -EIO;
903 return 0;
904}
905
906STATIC int
907xfs_dir_open(
908 struct inode *inode,
909 struct file *file)
910{
911 struct xfs_inode *ip = XFS_I(inode);
912 int mode;
913 int error;
914
915 error = xfs_file_open(inode, file);
916 if (error)
917 return error;
918
919 /*
920 * If there are any blocks, read-ahead block 0 as we're almost
921 * certain to have the next operation be a read there.
922 */
923 mode = xfs_ilock_data_map_shared(ip);
924 if (ip->i_d.di_nextents > 0)
925 error = xfs_dir3_data_readahead(ip, 0, -1);
926 xfs_iunlock(ip, mode);
927 return error;
928}
929
930STATIC int
931xfs_file_release(
932 struct inode *inode,
933 struct file *filp)
934{
935 return xfs_release(XFS_I(inode));
936}
937
938STATIC int
939xfs_file_readdir(
940 struct file *file,
941 struct dir_context *ctx)
942{
943 struct inode *inode = file_inode(file);
944 xfs_inode_t *ip = XFS_I(inode);
945 size_t bufsize;
946
947 /*
948 * The Linux API doesn't pass down the total size of the buffer
949 * we read into down to the filesystem. With the filldir concept
950 * it's not needed for correct information, but the XFS dir2 leaf
951 * code wants an estimate of the buffer size to calculate it's
952 * readahead window and size the buffers used for mapping to
953 * physical blocks.
954 *
955 * Try to give it an estimate that's good enough, maybe at some
956 * point we can change the ->readdir prototype to include the
957 * buffer size. For now we use the current glibc buffer size.
958 */
959 bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
960
961 return xfs_readdir(ip, ctx, bufsize);
962}
963
964/*
965 * This type is designed to indicate the type of offset we would like
966 * to search from page cache for xfs_seek_hole_data().
967 */
968enum {
969 HOLE_OFF = 0,
970 DATA_OFF,
971};
972
973/*
974 * Lookup the desired type of offset from the given page.
975 *
976 * On success, return true and the offset argument will point to the
977 * start of the region that was found. Otherwise this function will
978 * return false and keep the offset argument unchanged.
979 */
980STATIC bool
981xfs_lookup_buffer_offset(
982 struct page *page,
983 loff_t *offset,
984 unsigned int type)
985{
986 loff_t lastoff = page_offset(page);
987 bool found = false;
988 struct buffer_head *bh, *head;
989
990 bh = head = page_buffers(page);
991 do {
992 /*
993 * Unwritten extents that have data in the page
994 * cache covering them can be identified by the
995 * BH_Unwritten state flag. Pages with multiple
996 * buffers might have a mix of holes, data and
997 * unwritten extents - any buffer with valid
998 * data in it should have BH_Uptodate flag set
999 * on it.
1000 */
1001 if (buffer_unwritten(bh) ||
1002 buffer_uptodate(bh)) {
1003 if (type == DATA_OFF)
1004 found = true;
1005 } else {
1006 if (type == HOLE_OFF)
1007 found = true;
1008 }
1009
1010 if (found) {
1011 *offset = lastoff;
1012 break;
1013 }
1014 lastoff += bh->b_size;
1015 } while ((bh = bh->b_this_page) != head);
1016
1017 return found;
1018}
1019
1020/*
1021 * This routine is called to find out and return a data or hole offset
1022 * from the page cache for unwritten extents according to the desired
1023 * type for xfs_seek_hole_data().
1024 *
1025 * The argument offset is used to tell where we start to search from the
1026 * page cache. Map is used to figure out the end points of the range to
1027 * lookup pages.
1028 *
1029 * Return true if the desired type of offset was found, and the argument
1030 * offset is filled with that address. Otherwise, return false and keep
1031 * offset unchanged.
1032 */
1033STATIC bool
1034xfs_find_get_desired_pgoff(
1035 struct inode *inode,
1036 struct xfs_bmbt_irec *map,
1037 unsigned int type,
1038 loff_t *offset)
1039{
1040 struct xfs_inode *ip = XFS_I(inode);
1041 struct xfs_mount *mp = ip->i_mount;
1042 struct pagevec pvec;
1043 pgoff_t index;
1044 pgoff_t end;
1045 loff_t endoff;
1046 loff_t startoff = *offset;
1047 loff_t lastoff = startoff;
1048 bool found = false;
1049
1050 pagevec_init(&pvec, 0);
1051
1052 index = startoff >> PAGE_SHIFT;
1053 endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1054 end = endoff >> PAGE_SHIFT;
1055 do {
1056 int want;
1057 unsigned nr_pages;
1058 unsigned int i;
1059
1060 want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1061 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1062 want);
1063 /*
1064 * No page mapped into given range. If we are searching holes
1065 * and if this is the first time we got into the loop, it means
1066 * that the given offset is landed in a hole, return it.
1067 *
1068 * If we have already stepped through some block buffers to find
1069 * holes but they all contains data. In this case, the last
1070 * offset is already updated and pointed to the end of the last
1071 * mapped page, if it does not reach the endpoint to search,
1072 * that means there should be a hole between them.
1073 */
1074 if (nr_pages == 0) {
1075 /* Data search found nothing */
1076 if (type == DATA_OFF)
1077 break;
1078
1079 ASSERT(type == HOLE_OFF);
1080 if (lastoff == startoff || lastoff < endoff) {
1081 found = true;
1082 *offset = lastoff;
1083 }
1084 break;
1085 }
1086
1087 /*
1088 * At lease we found one page. If this is the first time we
1089 * step into the loop, and if the first page index offset is
1090 * greater than the given search offset, a hole was found.
1091 */
1092 if (type == HOLE_OFF && lastoff == startoff &&
1093 lastoff < page_offset(pvec.pages[0])) {
1094 found = true;
1095 break;
1096 }
1097
1098 for (i = 0; i < nr_pages; i++) {
1099 struct page *page = pvec.pages[i];
1100 loff_t b_offset;
1101
1102 /*
1103 * At this point, the page may be truncated or
1104 * invalidated (changing page->mapping to NULL),
1105 * or even swizzled back from swapper_space to tmpfs
1106 * file mapping. However, page->index will not change
1107 * because we have a reference on the page.
1108 *
1109 * Searching done if the page index is out of range.
1110 * If the current offset is not reaches the end of
1111 * the specified search range, there should be a hole
1112 * between them.
1113 */
1114 if (page->index > end) {
1115 if (type == HOLE_OFF && lastoff < endoff) {
1116 *offset = lastoff;
1117 found = true;
1118 }
1119 goto out;
1120 }
1121
1122 lock_page(page);
1123 /*
1124 * Page truncated or invalidated(page->mapping == NULL).
1125 * We can freely skip it and proceed to check the next
1126 * page.
1127 */
1128 if (unlikely(page->mapping != inode->i_mapping)) {
1129 unlock_page(page);
1130 continue;
1131 }
1132
1133 if (!page_has_buffers(page)) {
1134 unlock_page(page);
1135 continue;
1136 }
1137
1138 found = xfs_lookup_buffer_offset(page, &b_offset, type);
1139 if (found) {
1140 /*
1141 * The found offset may be less than the start
1142 * point to search if this is the first time to
1143 * come here.
1144 */
1145 *offset = max_t(loff_t, startoff, b_offset);
1146 unlock_page(page);
1147 goto out;
1148 }
1149
1150 /*
1151 * We either searching data but nothing was found, or
1152 * searching hole but found a data buffer. In either
1153 * case, probably the next page contains the desired
1154 * things, update the last offset to it so.
1155 */
1156 lastoff = page_offset(page) + PAGE_SIZE;
1157 unlock_page(page);
1158 }
1159
1160 /*
1161 * The number of returned pages less than our desired, search
1162 * done. In this case, nothing was found for searching data,
1163 * but we found a hole behind the last offset.
1164 */
1165 if (nr_pages < want) {
1166 if (type == HOLE_OFF) {
1167 *offset = lastoff;
1168 found = true;
1169 }
1170 break;
1171 }
1172
1173 index = pvec.pages[i - 1]->index + 1;
1174 pagevec_release(&pvec);
1175 } while (index <= end);
1176
1177out:
1178 pagevec_release(&pvec);
1179 return found;
1180}
1181
1182/*
1183 * caller must lock inode with xfs_ilock_data_map_shared,
1184 * can we craft an appropriate ASSERT?
1185 *
1186 * end is because the VFS-level lseek interface is defined such that any
1187 * offset past i_size shall return -ENXIO, but we use this for quota code
1188 * which does not maintain i_size, and we want to SEEK_DATA past i_size.
1189 */
1190loff_t
1191__xfs_seek_hole_data(
1192 struct inode *inode,
1193 loff_t start,
1194 loff_t end,
1195 int whence)
1196{
1197 struct xfs_inode *ip = XFS_I(inode);
1198 struct xfs_mount *mp = ip->i_mount;
1199 loff_t uninitialized_var(offset);
1200 xfs_fileoff_t fsbno;
1201 xfs_filblks_t lastbno;
1202 int error;
1203
1204 if (start >= end) {
1205 error = -ENXIO;
1206 goto out_error;
1207 }
1208
1209 /*
1210 * Try to read extents from the first block indicated
1211 * by fsbno to the end block of the file.
1212 */
1213 fsbno = XFS_B_TO_FSBT(mp, start);
1214 lastbno = XFS_B_TO_FSB(mp, end);
1215
1216 for (;;) {
1217 struct xfs_bmbt_irec map[2];
1218 int nmap = 2;
1219 unsigned int i;
1220
1221 error = xfs_bmapi_read(ip, fsbno, lastbno - fsbno, map, &nmap,
1222 XFS_BMAPI_ENTIRE);
1223 if (error)
1224 goto out_error;
1225
1226 /* No extents at given offset, must be beyond EOF */
1227 if (nmap == 0) {
1228 error = -ENXIO;
1229 goto out_error;
1230 }
1231
1232 for (i = 0; i < nmap; i++) {
1233 offset = max_t(loff_t, start,
1234 XFS_FSB_TO_B(mp, map[i].br_startoff));
1235
1236 /* Landed in the hole we wanted? */
1237 if (whence == SEEK_HOLE &&
1238 map[i].br_startblock == HOLESTARTBLOCK)
1239 goto out;
1240
1241 /* Landed in the data extent we wanted? */
1242 if (whence == SEEK_DATA &&
1243 (map[i].br_startblock == DELAYSTARTBLOCK ||
1244 (map[i].br_state == XFS_EXT_NORM &&
1245 !isnullstartblock(map[i].br_startblock))))
1246 goto out;
1247
1248 /*
1249 * Landed in an unwritten extent, try to search
1250 * for hole or data from page cache.
1251 */
1252 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1253 if (xfs_find_get_desired_pgoff(inode, &map[i],
1254 whence == SEEK_HOLE ? HOLE_OFF : DATA_OFF,
1255 &offset))
1256 goto out;
1257 }
1258 }
1259
1260 /*
1261 * We only received one extent out of the two requested. This
1262 * means we've hit EOF and didn't find what we are looking for.
1263 */
1264 if (nmap == 1) {
1265 /*
1266 * If we were looking for a hole, set offset to
1267 * the end of the file (i.e., there is an implicit
1268 * hole at the end of any file).
1269 */
1270 if (whence == SEEK_HOLE) {
1271 offset = end;
1272 break;
1273 }
1274 /*
1275 * If we were looking for data, it's nowhere to be found
1276 */
1277 ASSERT(whence == SEEK_DATA);
1278 error = -ENXIO;
1279 goto out_error;
1280 }
1281
1282 ASSERT(i > 1);
1283
1284 /*
1285 * Nothing was found, proceed to the next round of search
1286 * if the next reading offset is not at or beyond EOF.
1287 */
1288 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1289 start = XFS_FSB_TO_B(mp, fsbno);
1290 if (start >= end) {
1291 if (whence == SEEK_HOLE) {
1292 offset = end;
1293 break;
1294 }
1295 ASSERT(whence == SEEK_DATA);
1296 error = -ENXIO;
1297 goto out_error;
1298 }
1299 }
1300
1301out:
1302 /*
1303 * If at this point we have found the hole we wanted, the returned
1304 * offset may be bigger than the file size as it may be aligned to
1305 * page boundary for unwritten extents. We need to deal with this
1306 * situation in particular.
1307 */
1308 if (whence == SEEK_HOLE)
1309 offset = min_t(loff_t, offset, end);
1310
1311 return offset;
1312
1313out_error:
1314 return error;
1315}
1316
1317STATIC loff_t
1318xfs_seek_hole_data(
1319 struct file *file,
1320 loff_t start,
1321 int whence)
1322{
1323 struct inode *inode = file->f_mapping->host;
1324 struct xfs_inode *ip = XFS_I(inode);
1325 struct xfs_mount *mp = ip->i_mount;
1326 uint lock;
1327 loff_t offset, end;
1328 int error = 0;
1329
1330 if (XFS_FORCED_SHUTDOWN(mp))
1331 return -EIO;
1332
1333 lock = xfs_ilock_data_map_shared(ip);
1334
1335 end = i_size_read(inode);
1336 offset = __xfs_seek_hole_data(inode, start, end, whence);
1337 if (offset < 0) {
1338 error = offset;
1339 goto out_unlock;
1340 }
1341
1342 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1343
1344out_unlock:
1345 xfs_iunlock(ip, lock);
1346
1347 if (error)
1348 return error;
1349 return offset;
1350}
1351
1352STATIC loff_t
1353xfs_file_llseek(
1354 struct file *file,
1355 loff_t offset,
1356 int whence)
1357{
1358 switch (whence) {
1359 case SEEK_END:
1360 case SEEK_CUR:
1361 case SEEK_SET:
1362 return generic_file_llseek(file, offset, whence);
1363 case SEEK_HOLE:
1364 case SEEK_DATA:
1365 return xfs_seek_hole_data(file, offset, whence);
1366 default:
1367 return -EINVAL;
1368 }
1369}
1370
1371/*
1372 * Locking for serialisation of IO during page faults. This results in a lock
1373 * ordering of:
1374 *
1375 * mmap_sem (MM)
1376 * sb_start_pagefault(vfs, freeze)
1377 * i_mmaplock (XFS - truncate serialisation)
1378 * page_lock (MM)
1379 * i_lock (XFS - extent map serialisation)
1380 */
1381
1382/*
1383 * mmap()d file has taken write protection fault and is being made writable. We
1384 * can set the page state up correctly for a writable page, which means we can
1385 * do correct delalloc accounting (ENOSPC checking!) and unwritten extent
1386 * mapping.
1387 */
1388STATIC int
1389xfs_filemap_page_mkwrite(
1390 struct vm_area_struct *vma,
1391 struct vm_fault *vmf)
1392{
1393 struct inode *inode = file_inode(vma->vm_file);
1394 int ret;
1395
1396 trace_xfs_filemap_page_mkwrite(XFS_I(inode));
1397
1398 sb_start_pagefault(inode->i_sb);
1399 file_update_time(vma->vm_file);
1400 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1401
1402 if (IS_DAX(inode)) {
1403 ret = dax_iomap_fault(vma, vmf, &xfs_iomap_ops);
1404 } else {
1405 ret = iomap_page_mkwrite(vma, vmf, &xfs_iomap_ops);
1406 ret = block_page_mkwrite_return(ret);
1407 }
1408
1409 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1410 sb_end_pagefault(inode->i_sb);
1411
1412 return ret;
1413}
1414
1415STATIC int
1416xfs_filemap_fault(
1417 struct vm_area_struct *vma,
1418 struct vm_fault *vmf)
1419{
1420 struct inode *inode = file_inode(vma->vm_file);
1421 int ret;
1422
1423 trace_xfs_filemap_fault(XFS_I(inode));
1424
1425 /* DAX can shortcut the normal fault path on write faults! */
1426 if ((vmf->flags & FAULT_FLAG_WRITE) && IS_DAX(inode))
1427 return xfs_filemap_page_mkwrite(vma, vmf);
1428
1429 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1430 if (IS_DAX(inode))
1431 ret = dax_iomap_fault(vma, vmf, &xfs_iomap_ops);
1432 else
1433 ret = filemap_fault(vma, vmf);
1434 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1435
1436 return ret;
1437}
1438
1439/*
1440 * Similar to xfs_filemap_fault(), the DAX fault path can call into here on
1441 * both read and write faults. Hence we need to handle both cases. There is no
1442 * ->pmd_mkwrite callout for huge pages, so we have a single function here to
1443 * handle both cases here. @flags carries the information on the type of fault
1444 * occuring.
1445 */
1446STATIC int
1447xfs_filemap_pmd_fault(
1448 struct vm_area_struct *vma,
1449 unsigned long addr,
1450 pmd_t *pmd,
1451 unsigned int flags)
1452{
1453 struct inode *inode = file_inode(vma->vm_file);
1454 struct xfs_inode *ip = XFS_I(inode);
1455 int ret;
1456
1457 if (!IS_DAX(inode))
1458 return VM_FAULT_FALLBACK;
1459
1460 trace_xfs_filemap_pmd_fault(ip);
1461
1462 if (flags & FAULT_FLAG_WRITE) {
1463 sb_start_pagefault(inode->i_sb);
1464 file_update_time(vma->vm_file);
1465 }
1466
1467 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1468 ret = dax_iomap_pmd_fault(vma, addr, pmd, flags, &xfs_iomap_ops);
1469 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1470
1471 if (flags & FAULT_FLAG_WRITE)
1472 sb_end_pagefault(inode->i_sb);
1473
1474 return ret;
1475}
1476
1477/*
1478 * pfn_mkwrite was originally inteneded to ensure we capture time stamp
1479 * updates on write faults. In reality, it's need to serialise against
1480 * truncate similar to page_mkwrite. Hence we cycle the XFS_MMAPLOCK_SHARED
1481 * to ensure we serialise the fault barrier in place.
1482 */
1483static int
1484xfs_filemap_pfn_mkwrite(
1485 struct vm_area_struct *vma,
1486 struct vm_fault *vmf)
1487{
1488
1489 struct inode *inode = file_inode(vma->vm_file);
1490 struct xfs_inode *ip = XFS_I(inode);
1491 int ret = VM_FAULT_NOPAGE;
1492 loff_t size;
1493
1494 trace_xfs_filemap_pfn_mkwrite(ip);
1495
1496 sb_start_pagefault(inode->i_sb);
1497 file_update_time(vma->vm_file);
1498
1499 /* check if the faulting page hasn't raced with truncate */
1500 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1501 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1502 if (vmf->pgoff >= size)
1503 ret = VM_FAULT_SIGBUS;
1504 else if (IS_DAX(inode))
1505 ret = dax_pfn_mkwrite(vma, vmf);
1506 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1507 sb_end_pagefault(inode->i_sb);
1508 return ret;
1509
1510}
1511
1512static const struct vm_operations_struct xfs_file_vm_ops = {
1513 .fault = xfs_filemap_fault,
1514 .pmd_fault = xfs_filemap_pmd_fault,
1515 .map_pages = filemap_map_pages,
1516 .page_mkwrite = xfs_filemap_page_mkwrite,
1517 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1518};
1519
1520STATIC int
1521xfs_file_mmap(
1522 struct file *filp,
1523 struct vm_area_struct *vma)
1524{
1525 file_accessed(filp);
1526 vma->vm_ops = &xfs_file_vm_ops;
1527 if (IS_DAX(file_inode(filp)))
1528 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
1529 return 0;
1530}
1531
1532const struct file_operations xfs_file_operations = {
1533 .llseek = xfs_file_llseek,
1534 .read_iter = xfs_file_read_iter,
1535 .write_iter = xfs_file_write_iter,
1536 .splice_read = generic_file_splice_read,
1537 .splice_write = iter_file_splice_write,
1538 .unlocked_ioctl = xfs_file_ioctl,
1539#ifdef CONFIG_COMPAT
1540 .compat_ioctl = xfs_file_compat_ioctl,
1541#endif
1542 .mmap = xfs_file_mmap,
1543 .open = xfs_file_open,
1544 .release = xfs_file_release,
1545 .fsync = xfs_file_fsync,
1546 .get_unmapped_area = thp_get_unmapped_area,
1547 .fallocate = xfs_file_fallocate,
1548 .clone_file_range = xfs_file_clone_range,
1549 .dedupe_file_range = xfs_file_dedupe_range,
1550};
1551
1552const struct file_operations xfs_dir_file_operations = {
1553 .open = xfs_dir_open,
1554 .read = generic_read_dir,
1555 .iterate_shared = xfs_file_readdir,
1556 .llseek = generic_file_llseek,
1557 .unlocked_ioctl = xfs_file_ioctl,
1558#ifdef CONFIG_COMPAT
1559 .compat_ioctl = xfs_file_compat_ioctl,
1560#endif
1561 .fsync = xfs_dir_fsync,
1562};