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_sb.h"
25#include "xfs_ag.h"
26#include "xfs_mount.h"
27#include "xfs_da_format.h"
28#include "xfs_da_btree.h"
29#include "xfs_inode.h"
30#include "xfs_trans.h"
31#include "xfs_inode_item.h"
32#include "xfs_bmap.h"
33#include "xfs_bmap_util.h"
34#include "xfs_error.h"
35#include "xfs_dir2.h"
36#include "xfs_dir2_priv.h"
37#include "xfs_ioctl.h"
38#include "xfs_trace.h"
39#include "xfs_log.h"
40#include "xfs_dinode.h"
41
42#include <linux/aio.h>
43#include <linux/dcache.h>
44#include <linux/falloc.h>
45#include <linux/pagevec.h>
46
47static const struct vm_operations_struct xfs_file_vm_ops;
48
49/*
50 * Locking primitives for read and write IO paths to ensure we consistently use
51 * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
52 */
53static inline void
54xfs_rw_ilock(
55 struct xfs_inode *ip,
56 int type)
57{
58 if (type & XFS_IOLOCK_EXCL)
59 mutex_lock(&VFS_I(ip)->i_mutex);
60 xfs_ilock(ip, type);
61}
62
63static inline void
64xfs_rw_iunlock(
65 struct xfs_inode *ip,
66 int type)
67{
68 xfs_iunlock(ip, type);
69 if (type & XFS_IOLOCK_EXCL)
70 mutex_unlock(&VFS_I(ip)->i_mutex);
71}
72
73static inline void
74xfs_rw_ilock_demote(
75 struct xfs_inode *ip,
76 int type)
77{
78 xfs_ilock_demote(ip, type);
79 if (type & XFS_IOLOCK_EXCL)
80 mutex_unlock(&VFS_I(ip)->i_mutex);
81}
82
83/*
84 * xfs_iozero
85 *
86 * xfs_iozero clears the specified range of buffer supplied,
87 * and marks all the affected blocks as valid and modified. If
88 * an affected block is not allocated, it will be allocated. If
89 * an affected block is not completely overwritten, and is not
90 * valid before the operation, it will be read from disk before
91 * being partially zeroed.
92 */
93int
94xfs_iozero(
95 struct xfs_inode *ip, /* inode */
96 loff_t pos, /* offset in file */
97 size_t count) /* size of data to zero */
98{
99 struct page *page;
100 struct address_space *mapping;
101 int status;
102
103 mapping = VFS_I(ip)->i_mapping;
104 do {
105 unsigned offset, bytes;
106 void *fsdata;
107
108 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
109 bytes = PAGE_CACHE_SIZE - offset;
110 if (bytes > count)
111 bytes = count;
112
113 status = pagecache_write_begin(NULL, mapping, pos, bytes,
114 AOP_FLAG_UNINTERRUPTIBLE,
115 &page, &fsdata);
116 if (status)
117 break;
118
119 zero_user(page, offset, bytes);
120
121 status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
122 page, fsdata);
123 WARN_ON(status <= 0); /* can't return less than zero! */
124 pos += bytes;
125 count -= bytes;
126 status = 0;
127 } while (count);
128
129 return (-status);
130}
131
132/*
133 * Fsync operations on directories are much simpler than on regular files,
134 * as there is no file data to flush, and thus also no need for explicit
135 * cache flush operations, and there are no non-transaction metadata updates
136 * on directories either.
137 */
138STATIC int
139xfs_dir_fsync(
140 struct file *file,
141 loff_t start,
142 loff_t end,
143 int datasync)
144{
145 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
146 struct xfs_mount *mp = ip->i_mount;
147 xfs_lsn_t lsn = 0;
148
149 trace_xfs_dir_fsync(ip);
150
151 xfs_ilock(ip, XFS_ILOCK_SHARED);
152 if (xfs_ipincount(ip))
153 lsn = ip->i_itemp->ili_last_lsn;
154 xfs_iunlock(ip, XFS_ILOCK_SHARED);
155
156 if (!lsn)
157 return 0;
158 return -_xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, NULL);
159}
160
161STATIC int
162xfs_file_fsync(
163 struct file *file,
164 loff_t start,
165 loff_t end,
166 int datasync)
167{
168 struct inode *inode = file->f_mapping->host;
169 struct xfs_inode *ip = XFS_I(inode);
170 struct xfs_mount *mp = ip->i_mount;
171 int error = 0;
172 int log_flushed = 0;
173 xfs_lsn_t lsn = 0;
174
175 trace_xfs_file_fsync(ip);
176
177 error = filemap_write_and_wait_range(inode->i_mapping, start, end);
178 if (error)
179 return error;
180
181 if (XFS_FORCED_SHUTDOWN(mp))
182 return -XFS_ERROR(EIO);
183
184 xfs_iflags_clear(ip, XFS_ITRUNCATED);
185
186 if (mp->m_flags & XFS_MOUNT_BARRIER) {
187 /*
188 * If we have an RT and/or log subvolume we need to make sure
189 * to flush the write cache the device used for file data
190 * first. This is to ensure newly written file data make
191 * it to disk before logging the new inode size in case of
192 * an extending write.
193 */
194 if (XFS_IS_REALTIME_INODE(ip))
195 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
196 else if (mp->m_logdev_targp != mp->m_ddev_targp)
197 xfs_blkdev_issue_flush(mp->m_ddev_targp);
198 }
199
200 /*
201 * All metadata updates are logged, which means that we just have
202 * to flush the log up to the latest LSN that touched the inode.
203 */
204 xfs_ilock(ip, XFS_ILOCK_SHARED);
205 if (xfs_ipincount(ip)) {
206 if (!datasync ||
207 (ip->i_itemp->ili_fields & ~XFS_ILOG_TIMESTAMP))
208 lsn = ip->i_itemp->ili_last_lsn;
209 }
210 xfs_iunlock(ip, XFS_ILOCK_SHARED);
211
212 if (lsn)
213 error = _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
214
215 /*
216 * If we only have a single device, and the log force about was
217 * a no-op we might have to flush the data device cache here.
218 * This can only happen for fdatasync/O_DSYNC if we were overwriting
219 * an already allocated file and thus do not have any metadata to
220 * commit.
221 */
222 if ((mp->m_flags & XFS_MOUNT_BARRIER) &&
223 mp->m_logdev_targp == mp->m_ddev_targp &&
224 !XFS_IS_REALTIME_INODE(ip) &&
225 !log_flushed)
226 xfs_blkdev_issue_flush(mp->m_ddev_targp);
227
228 return -error;
229}
230
231STATIC ssize_t
232xfs_file_aio_read(
233 struct kiocb *iocb,
234 const struct iovec *iovp,
235 unsigned long nr_segs,
236 loff_t pos)
237{
238 struct file *file = iocb->ki_filp;
239 struct inode *inode = file->f_mapping->host;
240 struct xfs_inode *ip = XFS_I(inode);
241 struct xfs_mount *mp = ip->i_mount;
242 size_t size = 0;
243 ssize_t ret = 0;
244 int ioflags = 0;
245 xfs_fsize_t n;
246
247 XFS_STATS_INC(xs_read_calls);
248
249 BUG_ON(iocb->ki_pos != pos);
250
251 if (unlikely(file->f_flags & O_DIRECT))
252 ioflags |= IO_ISDIRECT;
253 if (file->f_mode & FMODE_NOCMTIME)
254 ioflags |= IO_INVIS;
255
256 ret = generic_segment_checks(iovp, &nr_segs, &size, VERIFY_WRITE);
257 if (ret < 0)
258 return ret;
259
260 if (unlikely(ioflags & IO_ISDIRECT)) {
261 xfs_buftarg_t *target =
262 XFS_IS_REALTIME_INODE(ip) ?
263 mp->m_rtdev_targp : mp->m_ddev_targp;
264 /* DIO must be aligned to device logical sector size */
265 if ((pos | size) & target->bt_logical_sectormask) {
266 if (pos == i_size_read(inode))
267 return 0;
268 return -XFS_ERROR(EINVAL);
269 }
270 }
271
272 n = mp->m_super->s_maxbytes - pos;
273 if (n <= 0 || size == 0)
274 return 0;
275
276 if (n < size)
277 size = n;
278
279 if (XFS_FORCED_SHUTDOWN(mp))
280 return -EIO;
281
282 /*
283 * Locking is a bit tricky here. If we take an exclusive lock
284 * for direct IO, we effectively serialise all new concurrent
285 * read IO to this file and block it behind IO that is currently in
286 * progress because IO in progress holds the IO lock shared. We only
287 * need to hold the lock exclusive to blow away the page cache, so
288 * only take lock exclusively if the page cache needs invalidation.
289 * This allows the normal direct IO case of no page cache pages to
290 * proceeed concurrently without serialisation.
291 */
292 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
293 if ((ioflags & IO_ISDIRECT) && inode->i_mapping->nrpages) {
294 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
295 xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
296
297 if (inode->i_mapping->nrpages) {
298 ret = filemap_write_and_wait_range(
299 VFS_I(ip)->i_mapping,
300 pos, -1);
301 if (ret) {
302 xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
303 return ret;
304 }
305 truncate_pagecache_range(VFS_I(ip), pos, -1);
306 }
307 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
308 }
309
310 trace_xfs_file_read(ip, size, pos, ioflags);
311
312 ret = generic_file_aio_read(iocb, iovp, nr_segs, pos);
313 if (ret > 0)
314 XFS_STATS_ADD(xs_read_bytes, ret);
315
316 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
317 return ret;
318}
319
320STATIC ssize_t
321xfs_file_splice_read(
322 struct file *infilp,
323 loff_t *ppos,
324 struct pipe_inode_info *pipe,
325 size_t count,
326 unsigned int flags)
327{
328 struct xfs_inode *ip = XFS_I(infilp->f_mapping->host);
329 int ioflags = 0;
330 ssize_t ret;
331
332 XFS_STATS_INC(xs_read_calls);
333
334 if (infilp->f_mode & FMODE_NOCMTIME)
335 ioflags |= IO_INVIS;
336
337 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
338 return -EIO;
339
340 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
341
342 trace_xfs_file_splice_read(ip, count, *ppos, ioflags);
343
344 ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
345 if (ret > 0)
346 XFS_STATS_ADD(xs_read_bytes, ret);
347
348 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
349 return ret;
350}
351
352/*
353 * xfs_file_splice_write() does not use xfs_rw_ilock() because
354 * generic_file_splice_write() takes the i_mutex itself. This, in theory,
355 * couuld cause lock inversions between the aio_write path and the splice path
356 * if someone is doing concurrent splice(2) based writes and write(2) based
357 * writes to the same inode. The only real way to fix this is to re-implement
358 * the generic code here with correct locking orders.
359 */
360STATIC ssize_t
361xfs_file_splice_write(
362 struct pipe_inode_info *pipe,
363 struct file *outfilp,
364 loff_t *ppos,
365 size_t count,
366 unsigned int flags)
367{
368 struct inode *inode = outfilp->f_mapping->host;
369 struct xfs_inode *ip = XFS_I(inode);
370 int ioflags = 0;
371 ssize_t ret;
372
373 XFS_STATS_INC(xs_write_calls);
374
375 if (outfilp->f_mode & FMODE_NOCMTIME)
376 ioflags |= IO_INVIS;
377
378 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
379 return -EIO;
380
381 xfs_ilock(ip, XFS_IOLOCK_EXCL);
382
383 trace_xfs_file_splice_write(ip, count, *ppos, ioflags);
384
385 ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
386 if (ret > 0)
387 XFS_STATS_ADD(xs_write_bytes, ret);
388
389 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
390 return ret;
391}
392
393/*
394 * This routine is called to handle zeroing any space in the last block of the
395 * file that is beyond the EOF. We do this since the size is being increased
396 * without writing anything to that block and we don't want to read the
397 * garbage on the disk.
398 */
399STATIC int /* error (positive) */
400xfs_zero_last_block(
401 struct xfs_inode *ip,
402 xfs_fsize_t offset,
403 xfs_fsize_t isize)
404{
405 struct xfs_mount *mp = ip->i_mount;
406 xfs_fileoff_t last_fsb = XFS_B_TO_FSBT(mp, isize);
407 int zero_offset = XFS_B_FSB_OFFSET(mp, isize);
408 int zero_len;
409 int nimaps = 1;
410 int error = 0;
411 struct xfs_bmbt_irec imap;
412
413 xfs_ilock(ip, XFS_ILOCK_EXCL);
414 error = xfs_bmapi_read(ip, last_fsb, 1, &imap, &nimaps, 0);
415 xfs_iunlock(ip, XFS_ILOCK_EXCL);
416 if (error)
417 return error;
418
419 ASSERT(nimaps > 0);
420
421 /*
422 * If the block underlying isize is just a hole, then there
423 * is nothing to zero.
424 */
425 if (imap.br_startblock == HOLESTARTBLOCK)
426 return 0;
427
428 zero_len = mp->m_sb.sb_blocksize - zero_offset;
429 if (isize + zero_len > offset)
430 zero_len = offset - isize;
431 return xfs_iozero(ip, isize, zero_len);
432}
433
434/*
435 * Zero any on disk space between the current EOF and the new, larger EOF.
436 *
437 * This handles the normal case of zeroing the remainder of the last block in
438 * the file and the unusual case of zeroing blocks out beyond the size of the
439 * file. This second case only happens with fixed size extents and when the
440 * system crashes before the inode size was updated but after blocks were
441 * allocated.
442 *
443 * Expects the iolock to be held exclusive, and will take the ilock internally.
444 */
445int /* error (positive) */
446xfs_zero_eof(
447 struct xfs_inode *ip,
448 xfs_off_t offset, /* starting I/O offset */
449 xfs_fsize_t isize) /* current inode size */
450{
451 struct xfs_mount *mp = ip->i_mount;
452 xfs_fileoff_t start_zero_fsb;
453 xfs_fileoff_t end_zero_fsb;
454 xfs_fileoff_t zero_count_fsb;
455 xfs_fileoff_t last_fsb;
456 xfs_fileoff_t zero_off;
457 xfs_fsize_t zero_len;
458 int nimaps;
459 int error = 0;
460 struct xfs_bmbt_irec imap;
461
462 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
463 ASSERT(offset > isize);
464
465 /*
466 * First handle zeroing the block on which isize resides.
467 *
468 * We only zero a part of that block so it is handled specially.
469 */
470 if (XFS_B_FSB_OFFSET(mp, isize) != 0) {
471 error = xfs_zero_last_block(ip, offset, isize);
472 if (error)
473 return error;
474 }
475
476 /*
477 * Calculate the range between the new size and the old where blocks
478 * needing to be zeroed may exist.
479 *
480 * To get the block where the last byte in the file currently resides,
481 * we need to subtract one from the size and truncate back to a block
482 * boundary. We subtract 1 in case the size is exactly on a block
483 * boundary.
484 */
485 last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
486 start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
487 end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
488 ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
489 if (last_fsb == end_zero_fsb) {
490 /*
491 * The size was only incremented on its last block.
492 * We took care of that above, so just return.
493 */
494 return 0;
495 }
496
497 ASSERT(start_zero_fsb <= end_zero_fsb);
498 while (start_zero_fsb <= end_zero_fsb) {
499 nimaps = 1;
500 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
501
502 xfs_ilock(ip, XFS_ILOCK_EXCL);
503 error = xfs_bmapi_read(ip, start_zero_fsb, zero_count_fsb,
504 &imap, &nimaps, 0);
505 xfs_iunlock(ip, XFS_ILOCK_EXCL);
506 if (error)
507 return error;
508
509 ASSERT(nimaps > 0);
510
511 if (imap.br_state == XFS_EXT_UNWRITTEN ||
512 imap.br_startblock == HOLESTARTBLOCK) {
513 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
514 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
515 continue;
516 }
517
518 /*
519 * There are blocks we need to zero.
520 */
521 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
522 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
523
524 if ((zero_off + zero_len) > offset)
525 zero_len = offset - zero_off;
526
527 error = xfs_iozero(ip, zero_off, zero_len);
528 if (error)
529 return error;
530
531 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
532 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
533 }
534
535 return 0;
536}
537
538/*
539 * Common pre-write limit and setup checks.
540 *
541 * Called with the iolocked held either shared and exclusive according to
542 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
543 * if called for a direct write beyond i_size.
544 */
545STATIC ssize_t
546xfs_file_aio_write_checks(
547 struct file *file,
548 loff_t *pos,
549 size_t *count,
550 int *iolock)
551{
552 struct inode *inode = file->f_mapping->host;
553 struct xfs_inode *ip = XFS_I(inode);
554 int error = 0;
555
556restart:
557 error = generic_write_checks(file, pos, count, S_ISBLK(inode->i_mode));
558 if (error)
559 return error;
560
561 /*
562 * If the offset is beyond the size of the file, we need to zero any
563 * blocks that fall between the existing EOF and the start of this
564 * write. If zeroing is needed and we are currently holding the
565 * iolock shared, we need to update it to exclusive which implies
566 * having to redo all checks before.
567 */
568 if (*pos > i_size_read(inode)) {
569 if (*iolock == XFS_IOLOCK_SHARED) {
570 xfs_rw_iunlock(ip, *iolock);
571 *iolock = XFS_IOLOCK_EXCL;
572 xfs_rw_ilock(ip, *iolock);
573 goto restart;
574 }
575 error = -xfs_zero_eof(ip, *pos, i_size_read(inode));
576 if (error)
577 return error;
578 }
579
580 /*
581 * Updating the timestamps will grab the ilock again from
582 * xfs_fs_dirty_inode, so we have to call it after dropping the
583 * lock above. Eventually we should look into a way to avoid
584 * the pointless lock roundtrip.
585 */
586 if (likely(!(file->f_mode & FMODE_NOCMTIME))) {
587 error = file_update_time(file);
588 if (error)
589 return error;
590 }
591
592 /*
593 * If we're writing the file then make sure to clear the setuid and
594 * setgid bits if the process is not being run by root. This keeps
595 * people from modifying setuid and setgid binaries.
596 */
597 return file_remove_suid(file);
598}
599
600/*
601 * xfs_file_dio_aio_write - handle direct IO writes
602 *
603 * Lock the inode appropriately to prepare for and issue a direct IO write.
604 * By separating it from the buffered write path we remove all the tricky to
605 * follow locking changes and looping.
606 *
607 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
608 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
609 * pages are flushed out.
610 *
611 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
612 * allowing them to be done in parallel with reads and other direct IO writes.
613 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
614 * needs to do sub-block zeroing and that requires serialisation against other
615 * direct IOs to the same block. In this case we need to serialise the
616 * submission of the unaligned IOs so that we don't get racing block zeroing in
617 * the dio layer. To avoid the problem with aio, we also need to wait for
618 * outstanding IOs to complete so that unwritten extent conversion is completed
619 * before we try to map the overlapping block. This is currently implemented by
620 * hitting it with a big hammer (i.e. inode_dio_wait()).
621 *
622 * Returns with locks held indicated by @iolock and errors indicated by
623 * negative return values.
624 */
625STATIC ssize_t
626xfs_file_dio_aio_write(
627 struct kiocb *iocb,
628 const struct iovec *iovp,
629 unsigned long nr_segs,
630 loff_t pos,
631 size_t ocount)
632{
633 struct file *file = iocb->ki_filp;
634 struct address_space *mapping = file->f_mapping;
635 struct inode *inode = mapping->host;
636 struct xfs_inode *ip = XFS_I(inode);
637 struct xfs_mount *mp = ip->i_mount;
638 ssize_t ret = 0;
639 size_t count = ocount;
640 int unaligned_io = 0;
641 int iolock;
642 struct xfs_buftarg *target = XFS_IS_REALTIME_INODE(ip) ?
643 mp->m_rtdev_targp : mp->m_ddev_targp;
644
645 /* DIO must be aligned to device logical sector size */
646 if ((pos | count) & target->bt_logical_sectormask)
647 return -XFS_ERROR(EINVAL);
648
649 /* "unaligned" here means not aligned to a filesystem block */
650 if ((pos & mp->m_blockmask) || ((pos + count) & mp->m_blockmask))
651 unaligned_io = 1;
652
653 /*
654 * We don't need to take an exclusive lock unless there page cache needs
655 * to be invalidated or unaligned IO is being executed. We don't need to
656 * consider the EOF extension case here because
657 * xfs_file_aio_write_checks() will relock the inode as necessary for
658 * EOF zeroing cases and fill out the new inode size as appropriate.
659 */
660 if (unaligned_io || mapping->nrpages)
661 iolock = XFS_IOLOCK_EXCL;
662 else
663 iolock = XFS_IOLOCK_SHARED;
664 xfs_rw_ilock(ip, iolock);
665
666 /*
667 * Recheck if there are cached pages that need invalidate after we got
668 * the iolock to protect against other threads adding new pages while
669 * we were waiting for the iolock.
670 */
671 if (mapping->nrpages && iolock == XFS_IOLOCK_SHARED) {
672 xfs_rw_iunlock(ip, iolock);
673 iolock = XFS_IOLOCK_EXCL;
674 xfs_rw_ilock(ip, iolock);
675 }
676
677 ret = xfs_file_aio_write_checks(file, &pos, &count, &iolock);
678 if (ret)
679 goto out;
680
681 if (mapping->nrpages) {
682 ret = filemap_write_and_wait_range(VFS_I(ip)->i_mapping,
683 pos, -1);
684 if (ret)
685 goto out;
686 truncate_pagecache_range(VFS_I(ip), pos, -1);
687 }
688
689 /*
690 * If we are doing unaligned IO, wait for all other IO to drain,
691 * otherwise demote the lock if we had to flush cached pages
692 */
693 if (unaligned_io)
694 inode_dio_wait(inode);
695 else if (iolock == XFS_IOLOCK_EXCL) {
696 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
697 iolock = XFS_IOLOCK_SHARED;
698 }
699
700 trace_xfs_file_direct_write(ip, count, iocb->ki_pos, 0);
701 ret = generic_file_direct_write(iocb, iovp,
702 &nr_segs, pos, count, ocount);
703
704out:
705 xfs_rw_iunlock(ip, iolock);
706
707 /* No fallback to buffered IO on errors for XFS. */
708 ASSERT(ret < 0 || ret == count);
709 return ret;
710}
711
712STATIC ssize_t
713xfs_file_buffered_aio_write(
714 struct kiocb *iocb,
715 const struct iovec *iovp,
716 unsigned long nr_segs,
717 loff_t pos,
718 size_t count)
719{
720 struct file *file = iocb->ki_filp;
721 struct address_space *mapping = file->f_mapping;
722 struct inode *inode = mapping->host;
723 struct xfs_inode *ip = XFS_I(inode);
724 ssize_t ret;
725 int enospc = 0;
726 int iolock = XFS_IOLOCK_EXCL;
727 struct iov_iter from;
728
729 xfs_rw_ilock(ip, iolock);
730
731 ret = xfs_file_aio_write_checks(file, &pos, &count, &iolock);
732 if (ret)
733 goto out;
734
735 iov_iter_init(&from, iovp, nr_segs, count, 0);
736 /* We can write back this queue in page reclaim */
737 current->backing_dev_info = mapping->backing_dev_info;
738
739write_retry:
740 trace_xfs_file_buffered_write(ip, count, iocb->ki_pos, 0);
741 ret = generic_perform_write(file, &from, pos);
742 if (likely(ret >= 0))
743 iocb->ki_pos = pos + ret;
744 /*
745 * If we just got an ENOSPC, try to write back all dirty inodes to
746 * convert delalloc space to free up some of the excess reserved
747 * metadata space.
748 */
749 if (ret == -ENOSPC && !enospc) {
750 enospc = 1;
751 xfs_flush_inodes(ip->i_mount);
752 goto write_retry;
753 }
754
755 current->backing_dev_info = NULL;
756out:
757 xfs_rw_iunlock(ip, iolock);
758 return ret;
759}
760
761STATIC ssize_t
762xfs_file_aio_write(
763 struct kiocb *iocb,
764 const struct iovec *iovp,
765 unsigned long nr_segs,
766 loff_t pos)
767{
768 struct file *file = iocb->ki_filp;
769 struct address_space *mapping = file->f_mapping;
770 struct inode *inode = mapping->host;
771 struct xfs_inode *ip = XFS_I(inode);
772 ssize_t ret;
773 size_t ocount = 0;
774
775 XFS_STATS_INC(xs_write_calls);
776
777 BUG_ON(iocb->ki_pos != pos);
778
779 ret = generic_segment_checks(iovp, &nr_segs, &ocount, VERIFY_READ);
780 if (ret)
781 return ret;
782
783 if (ocount == 0)
784 return 0;
785
786 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
787 ret = -EIO;
788 goto out;
789 }
790
791 if (unlikely(file->f_flags & O_DIRECT))
792 ret = xfs_file_dio_aio_write(iocb, iovp, nr_segs, pos, ocount);
793 else
794 ret = xfs_file_buffered_aio_write(iocb, iovp, nr_segs, pos,
795 ocount);
796
797 if (ret > 0) {
798 ssize_t err;
799
800 XFS_STATS_ADD(xs_write_bytes, ret);
801
802 /* Handle various SYNC-type writes */
803 err = generic_write_sync(file, iocb->ki_pos - ret, ret);
804 if (err < 0)
805 ret = err;
806 }
807
808out:
809 return ret;
810}
811
812STATIC long
813xfs_file_fallocate(
814 struct file *file,
815 int mode,
816 loff_t offset,
817 loff_t len)
818{
819 struct inode *inode = file_inode(file);
820 struct xfs_inode *ip = XFS_I(inode);
821 struct xfs_trans *tp;
822 long error;
823 loff_t new_size = 0;
824
825 if (!S_ISREG(inode->i_mode))
826 return -EINVAL;
827 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
828 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
829 return -EOPNOTSUPP;
830
831 xfs_ilock(ip, XFS_IOLOCK_EXCL);
832 if (mode & FALLOC_FL_PUNCH_HOLE) {
833 error = xfs_free_file_space(ip, offset, len);
834 if (error)
835 goto out_unlock;
836 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
837 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
838
839 if (offset & blksize_mask || len & blksize_mask) {
840 error = EINVAL;
841 goto out_unlock;
842 }
843
844 /*
845 * There is no need to overlap collapse range with EOF,
846 * in which case it is effectively a truncate operation
847 */
848 if (offset + len >= i_size_read(inode)) {
849 error = EINVAL;
850 goto out_unlock;
851 }
852
853 new_size = i_size_read(inode) - len;
854
855 error = xfs_collapse_file_space(ip, offset, len);
856 if (error)
857 goto out_unlock;
858 } else {
859 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
860 offset + len > i_size_read(inode)) {
861 new_size = offset + len;
862 error = -inode_newsize_ok(inode, new_size);
863 if (error)
864 goto out_unlock;
865 }
866
867 if (mode & FALLOC_FL_ZERO_RANGE)
868 error = xfs_zero_file_space(ip, offset, len);
869 else
870 error = xfs_alloc_file_space(ip, offset, len,
871 XFS_BMAPI_PREALLOC);
872 if (error)
873 goto out_unlock;
874 }
875
876 tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_WRITEID);
877 error = xfs_trans_reserve(tp, &M_RES(ip->i_mount)->tr_writeid, 0, 0);
878 if (error) {
879 xfs_trans_cancel(tp, 0);
880 goto out_unlock;
881 }
882
883 xfs_ilock(ip, XFS_ILOCK_EXCL);
884 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
885 ip->i_d.di_mode &= ~S_ISUID;
886 if (ip->i_d.di_mode & S_IXGRP)
887 ip->i_d.di_mode &= ~S_ISGID;
888
889 if (!(mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_COLLAPSE_RANGE)))
890 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
891
892 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
893 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
894
895 if (file->f_flags & O_DSYNC)
896 xfs_trans_set_sync(tp);
897 error = xfs_trans_commit(tp, 0);
898 if (error)
899 goto out_unlock;
900
901 /* Change file size if needed */
902 if (new_size) {
903 struct iattr iattr;
904
905 iattr.ia_valid = ATTR_SIZE;
906 iattr.ia_size = new_size;
907 error = xfs_setattr_size(ip, &iattr);
908 }
909
910out_unlock:
911 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
912 return -error;
913}
914
915
916STATIC int
917xfs_file_open(
918 struct inode *inode,
919 struct file *file)
920{
921 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
922 return -EFBIG;
923 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
924 return -EIO;
925 return 0;
926}
927
928STATIC int
929xfs_dir_open(
930 struct inode *inode,
931 struct file *file)
932{
933 struct xfs_inode *ip = XFS_I(inode);
934 int mode;
935 int error;
936
937 error = xfs_file_open(inode, file);
938 if (error)
939 return error;
940
941 /*
942 * If there are any blocks, read-ahead block 0 as we're almost
943 * certain to have the next operation be a read there.
944 */
945 mode = xfs_ilock_data_map_shared(ip);
946 if (ip->i_d.di_nextents > 0)
947 xfs_dir3_data_readahead(NULL, ip, 0, -1);
948 xfs_iunlock(ip, mode);
949 return 0;
950}
951
952STATIC int
953xfs_file_release(
954 struct inode *inode,
955 struct file *filp)
956{
957 return -xfs_release(XFS_I(inode));
958}
959
960STATIC int
961xfs_file_readdir(
962 struct file *file,
963 struct dir_context *ctx)
964{
965 struct inode *inode = file_inode(file);
966 xfs_inode_t *ip = XFS_I(inode);
967 int error;
968 size_t bufsize;
969
970 /*
971 * The Linux API doesn't pass down the total size of the buffer
972 * we read into down to the filesystem. With the filldir concept
973 * it's not needed for correct information, but the XFS dir2 leaf
974 * code wants an estimate of the buffer size to calculate it's
975 * readahead window and size the buffers used for mapping to
976 * physical blocks.
977 *
978 * Try to give it an estimate that's good enough, maybe at some
979 * point we can change the ->readdir prototype to include the
980 * buffer size. For now we use the current glibc buffer size.
981 */
982 bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
983
984 error = xfs_readdir(ip, ctx, bufsize);
985 if (error)
986 return -error;
987 return 0;
988}
989
990STATIC int
991xfs_file_mmap(
992 struct file *filp,
993 struct vm_area_struct *vma)
994{
995 vma->vm_ops = &xfs_file_vm_ops;
996
997 file_accessed(filp);
998 return 0;
999}
1000
1001/*
1002 * mmap()d file has taken write protection fault and is being made
1003 * writable. We can set the page state up correctly for a writable
1004 * page, which means we can do correct delalloc accounting (ENOSPC
1005 * checking!) and unwritten extent mapping.
1006 */
1007STATIC int
1008xfs_vm_page_mkwrite(
1009 struct vm_area_struct *vma,
1010 struct vm_fault *vmf)
1011{
1012 return block_page_mkwrite(vma, vmf, xfs_get_blocks);
1013}
1014
1015/*
1016 * This type is designed to indicate the type of offset we would like
1017 * to search from page cache for either xfs_seek_data() or xfs_seek_hole().
1018 */
1019enum {
1020 HOLE_OFF = 0,
1021 DATA_OFF,
1022};
1023
1024/*
1025 * Lookup the desired type of offset from the given page.
1026 *
1027 * On success, return true and the offset argument will point to the
1028 * start of the region that was found. Otherwise this function will
1029 * return false and keep the offset argument unchanged.
1030 */
1031STATIC bool
1032xfs_lookup_buffer_offset(
1033 struct page *page,
1034 loff_t *offset,
1035 unsigned int type)
1036{
1037 loff_t lastoff = page_offset(page);
1038 bool found = false;
1039 struct buffer_head *bh, *head;
1040
1041 bh = head = page_buffers(page);
1042 do {
1043 /*
1044 * Unwritten extents that have data in the page
1045 * cache covering them can be identified by the
1046 * BH_Unwritten state flag. Pages with multiple
1047 * buffers might have a mix of holes, data and
1048 * unwritten extents - any buffer with valid
1049 * data in it should have BH_Uptodate flag set
1050 * on it.
1051 */
1052 if (buffer_unwritten(bh) ||
1053 buffer_uptodate(bh)) {
1054 if (type == DATA_OFF)
1055 found = true;
1056 } else {
1057 if (type == HOLE_OFF)
1058 found = true;
1059 }
1060
1061 if (found) {
1062 *offset = lastoff;
1063 break;
1064 }
1065 lastoff += bh->b_size;
1066 } while ((bh = bh->b_this_page) != head);
1067
1068 return found;
1069}
1070
1071/*
1072 * This routine is called to find out and return a data or hole offset
1073 * from the page cache for unwritten extents according to the desired
1074 * type for xfs_seek_data() or xfs_seek_hole().
1075 *
1076 * The argument offset is used to tell where we start to search from the
1077 * page cache. Map is used to figure out the end points of the range to
1078 * lookup pages.
1079 *
1080 * Return true if the desired type of offset was found, and the argument
1081 * offset is filled with that address. Otherwise, return false and keep
1082 * offset unchanged.
1083 */
1084STATIC bool
1085xfs_find_get_desired_pgoff(
1086 struct inode *inode,
1087 struct xfs_bmbt_irec *map,
1088 unsigned int type,
1089 loff_t *offset)
1090{
1091 struct xfs_inode *ip = XFS_I(inode);
1092 struct xfs_mount *mp = ip->i_mount;
1093 struct pagevec pvec;
1094 pgoff_t index;
1095 pgoff_t end;
1096 loff_t endoff;
1097 loff_t startoff = *offset;
1098 loff_t lastoff = startoff;
1099 bool found = false;
1100
1101 pagevec_init(&pvec, 0);
1102
1103 index = startoff >> PAGE_CACHE_SHIFT;
1104 endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1105 end = endoff >> PAGE_CACHE_SHIFT;
1106 do {
1107 int want;
1108 unsigned nr_pages;
1109 unsigned int i;
1110
1111 want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1112 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1113 want);
1114 /*
1115 * No page mapped into given range. If we are searching holes
1116 * and if this is the first time we got into the loop, it means
1117 * that the given offset is landed in a hole, return it.
1118 *
1119 * If we have already stepped through some block buffers to find
1120 * holes but they all contains data. In this case, the last
1121 * offset is already updated and pointed to the end of the last
1122 * mapped page, if it does not reach the endpoint to search,
1123 * that means there should be a hole between them.
1124 */
1125 if (nr_pages == 0) {
1126 /* Data search found nothing */
1127 if (type == DATA_OFF)
1128 break;
1129
1130 ASSERT(type == HOLE_OFF);
1131 if (lastoff == startoff || lastoff < endoff) {
1132 found = true;
1133 *offset = lastoff;
1134 }
1135 break;
1136 }
1137
1138 /*
1139 * At lease we found one page. If this is the first time we
1140 * step into the loop, and if the first page index offset is
1141 * greater than the given search offset, a hole was found.
1142 */
1143 if (type == HOLE_OFF && lastoff == startoff &&
1144 lastoff < page_offset(pvec.pages[0])) {
1145 found = true;
1146 break;
1147 }
1148
1149 for (i = 0; i < nr_pages; i++) {
1150 struct page *page = pvec.pages[i];
1151 loff_t b_offset;
1152
1153 /*
1154 * At this point, the page may be truncated or
1155 * invalidated (changing page->mapping to NULL),
1156 * or even swizzled back from swapper_space to tmpfs
1157 * file mapping. However, page->index will not change
1158 * because we have a reference on the page.
1159 *
1160 * Searching done if the page index is out of range.
1161 * If the current offset is not reaches the end of
1162 * the specified search range, there should be a hole
1163 * between them.
1164 */
1165 if (page->index > end) {
1166 if (type == HOLE_OFF && lastoff < endoff) {
1167 *offset = lastoff;
1168 found = true;
1169 }
1170 goto out;
1171 }
1172
1173 lock_page(page);
1174 /*
1175 * Page truncated or invalidated(page->mapping == NULL).
1176 * We can freely skip it and proceed to check the next
1177 * page.
1178 */
1179 if (unlikely(page->mapping != inode->i_mapping)) {
1180 unlock_page(page);
1181 continue;
1182 }
1183
1184 if (!page_has_buffers(page)) {
1185 unlock_page(page);
1186 continue;
1187 }
1188
1189 found = xfs_lookup_buffer_offset(page, &b_offset, type);
1190 if (found) {
1191 /*
1192 * The found offset may be less than the start
1193 * point to search if this is the first time to
1194 * come here.
1195 */
1196 *offset = max_t(loff_t, startoff, b_offset);
1197 unlock_page(page);
1198 goto out;
1199 }
1200
1201 /*
1202 * We either searching data but nothing was found, or
1203 * searching hole but found a data buffer. In either
1204 * case, probably the next page contains the desired
1205 * things, update the last offset to it so.
1206 */
1207 lastoff = page_offset(page) + PAGE_SIZE;
1208 unlock_page(page);
1209 }
1210
1211 /*
1212 * The number of returned pages less than our desired, search
1213 * done. In this case, nothing was found for searching data,
1214 * but we found a hole behind the last offset.
1215 */
1216 if (nr_pages < want) {
1217 if (type == HOLE_OFF) {
1218 *offset = lastoff;
1219 found = true;
1220 }
1221 break;
1222 }
1223
1224 index = pvec.pages[i - 1]->index + 1;
1225 pagevec_release(&pvec);
1226 } while (index <= end);
1227
1228out:
1229 pagevec_release(&pvec);
1230 return found;
1231}
1232
1233STATIC loff_t
1234xfs_seek_data(
1235 struct file *file,
1236 loff_t start)
1237{
1238 struct inode *inode = file->f_mapping->host;
1239 struct xfs_inode *ip = XFS_I(inode);
1240 struct xfs_mount *mp = ip->i_mount;
1241 loff_t uninitialized_var(offset);
1242 xfs_fsize_t isize;
1243 xfs_fileoff_t fsbno;
1244 xfs_filblks_t end;
1245 uint lock;
1246 int error;
1247
1248 lock = xfs_ilock_data_map_shared(ip);
1249
1250 isize = i_size_read(inode);
1251 if (start >= isize) {
1252 error = ENXIO;
1253 goto out_unlock;
1254 }
1255
1256 /*
1257 * Try to read extents from the first block indicated
1258 * by fsbno to the end block of the file.
1259 */
1260 fsbno = XFS_B_TO_FSBT(mp, start);
1261 end = XFS_B_TO_FSB(mp, isize);
1262 for (;;) {
1263 struct xfs_bmbt_irec map[2];
1264 int nmap = 2;
1265 unsigned int i;
1266
1267 error = xfs_bmapi_read(ip, fsbno, end - fsbno, map, &nmap,
1268 XFS_BMAPI_ENTIRE);
1269 if (error)
1270 goto out_unlock;
1271
1272 /* No extents at given offset, must be beyond EOF */
1273 if (nmap == 0) {
1274 error = ENXIO;
1275 goto out_unlock;
1276 }
1277
1278 for (i = 0; i < nmap; i++) {
1279 offset = max_t(loff_t, start,
1280 XFS_FSB_TO_B(mp, map[i].br_startoff));
1281
1282 /* Landed in a data extent */
1283 if (map[i].br_startblock == DELAYSTARTBLOCK ||
1284 (map[i].br_state == XFS_EXT_NORM &&
1285 !isnullstartblock(map[i].br_startblock)))
1286 goto out;
1287
1288 /*
1289 * Landed in an unwritten extent, try to search data
1290 * from page cache.
1291 */
1292 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1293 if (xfs_find_get_desired_pgoff(inode, &map[i],
1294 DATA_OFF, &offset))
1295 goto out;
1296 }
1297 }
1298
1299 /*
1300 * map[0] is hole or its an unwritten extent but
1301 * without data in page cache. Probably means that
1302 * we are reading after EOF if nothing in map[1].
1303 */
1304 if (nmap == 1) {
1305 error = ENXIO;
1306 goto out_unlock;
1307 }
1308
1309 ASSERT(i > 1);
1310
1311 /*
1312 * Nothing was found, proceed to the next round of search
1313 * if reading offset not beyond or hit EOF.
1314 */
1315 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1316 start = XFS_FSB_TO_B(mp, fsbno);
1317 if (start >= isize) {
1318 error = ENXIO;
1319 goto out_unlock;
1320 }
1321 }
1322
1323out:
1324 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1325
1326out_unlock:
1327 xfs_iunlock(ip, lock);
1328
1329 if (error)
1330 return -error;
1331 return offset;
1332}
1333
1334STATIC loff_t
1335xfs_seek_hole(
1336 struct file *file,
1337 loff_t start)
1338{
1339 struct inode *inode = file->f_mapping->host;
1340 struct xfs_inode *ip = XFS_I(inode);
1341 struct xfs_mount *mp = ip->i_mount;
1342 loff_t uninitialized_var(offset);
1343 xfs_fsize_t isize;
1344 xfs_fileoff_t fsbno;
1345 xfs_filblks_t end;
1346 uint lock;
1347 int error;
1348
1349 if (XFS_FORCED_SHUTDOWN(mp))
1350 return -XFS_ERROR(EIO);
1351
1352 lock = xfs_ilock_data_map_shared(ip);
1353
1354 isize = i_size_read(inode);
1355 if (start >= isize) {
1356 error = ENXIO;
1357 goto out_unlock;
1358 }
1359
1360 fsbno = XFS_B_TO_FSBT(mp, start);
1361 end = XFS_B_TO_FSB(mp, isize);
1362
1363 for (;;) {
1364 struct xfs_bmbt_irec map[2];
1365 int nmap = 2;
1366 unsigned int i;
1367
1368 error = xfs_bmapi_read(ip, fsbno, end - fsbno, map, &nmap,
1369 XFS_BMAPI_ENTIRE);
1370 if (error)
1371 goto out_unlock;
1372
1373 /* No extents at given offset, must be beyond EOF */
1374 if (nmap == 0) {
1375 error = ENXIO;
1376 goto out_unlock;
1377 }
1378
1379 for (i = 0; i < nmap; i++) {
1380 offset = max_t(loff_t, start,
1381 XFS_FSB_TO_B(mp, map[i].br_startoff));
1382
1383 /* Landed in a hole */
1384 if (map[i].br_startblock == HOLESTARTBLOCK)
1385 goto out;
1386
1387 /*
1388 * Landed in an unwritten extent, try to search hole
1389 * from page cache.
1390 */
1391 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1392 if (xfs_find_get_desired_pgoff(inode, &map[i],
1393 HOLE_OFF, &offset))
1394 goto out;
1395 }
1396 }
1397
1398 /*
1399 * map[0] contains data or its unwritten but contains
1400 * data in page cache, probably means that we are
1401 * reading after EOF. We should fix offset to point
1402 * to the end of the file(i.e., there is an implicit
1403 * hole at the end of any file).
1404 */
1405 if (nmap == 1) {
1406 offset = isize;
1407 break;
1408 }
1409
1410 ASSERT(i > 1);
1411
1412 /*
1413 * Both mappings contains data, proceed to the next round of
1414 * search if the current reading offset not beyond or hit EOF.
1415 */
1416 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1417 start = XFS_FSB_TO_B(mp, fsbno);
1418 if (start >= isize) {
1419 offset = isize;
1420 break;
1421 }
1422 }
1423
1424out:
1425 /*
1426 * At this point, we must have found a hole. However, the returned
1427 * offset may be bigger than the file size as it may be aligned to
1428 * page boundary for unwritten extents, we need to deal with this
1429 * situation in particular.
1430 */
1431 offset = min_t(loff_t, offset, isize);
1432 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1433
1434out_unlock:
1435 xfs_iunlock(ip, lock);
1436
1437 if (error)
1438 return -error;
1439 return offset;
1440}
1441
1442STATIC loff_t
1443xfs_file_llseek(
1444 struct file *file,
1445 loff_t offset,
1446 int origin)
1447{
1448 switch (origin) {
1449 case SEEK_END:
1450 case SEEK_CUR:
1451 case SEEK_SET:
1452 return generic_file_llseek(file, offset, origin);
1453 case SEEK_DATA:
1454 return xfs_seek_data(file, offset);
1455 case SEEK_HOLE:
1456 return xfs_seek_hole(file, offset);
1457 default:
1458 return -EINVAL;
1459 }
1460}
1461
1462const struct file_operations xfs_file_operations = {
1463 .llseek = xfs_file_llseek,
1464 .read = do_sync_read,
1465 .write = do_sync_write,
1466 .aio_read = xfs_file_aio_read,
1467 .aio_write = xfs_file_aio_write,
1468 .splice_read = xfs_file_splice_read,
1469 .splice_write = xfs_file_splice_write,
1470 .unlocked_ioctl = xfs_file_ioctl,
1471#ifdef CONFIG_COMPAT
1472 .compat_ioctl = xfs_file_compat_ioctl,
1473#endif
1474 .mmap = xfs_file_mmap,
1475 .open = xfs_file_open,
1476 .release = xfs_file_release,
1477 .fsync = xfs_file_fsync,
1478 .fallocate = xfs_file_fallocate,
1479};
1480
1481const struct file_operations xfs_dir_file_operations = {
1482 .open = xfs_dir_open,
1483 .read = generic_read_dir,
1484 .iterate = xfs_file_readdir,
1485 .llseek = generic_file_llseek,
1486 .unlocked_ioctl = xfs_file_ioctl,
1487#ifdef CONFIG_COMPAT
1488 .compat_ioctl = xfs_file_compat_ioctl,
1489#endif
1490 .fsync = xfs_dir_fsync,
1491};
1492
1493static const struct vm_operations_struct xfs_file_vm_ops = {
1494 .fault = filemap_fault,
1495 .map_pages = filemap_map_pages,
1496 .page_mkwrite = xfs_vm_page_mkwrite,
1497 .remap_pages = generic_file_remap_pages,
1498};