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_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};
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_inode.h"
14#include "xfs_trans.h"
15#include "xfs_inode_item.h"
16#include "xfs_bmap.h"
17#include "xfs_bmap_util.h"
18#include "xfs_dir2.h"
19#include "xfs_dir2_priv.h"
20#include "xfs_ioctl.h"
21#include "xfs_trace.h"
22#include "xfs_log.h"
23#include "xfs_icache.h"
24#include "xfs_pnfs.h"
25#include "xfs_iomap.h"
26#include "xfs_reflink.h"
27#include "xfs_file.h"
28
29#include <linux/dax.h>
30#include <linux/falloc.h>
31#include <linux/backing-dev.h>
32#include <linux/mman.h>
33#include <linux/fadvise.h>
34#include <linux/mount.h>
35
36static const struct vm_operations_struct xfs_file_vm_ops;
37
38/*
39 * Decide if the given file range is aligned to the size of the fundamental
40 * allocation unit for the file.
41 */
42bool
43xfs_is_falloc_aligned(
44 struct xfs_inode *ip,
45 loff_t pos,
46 long long int len)
47{
48 unsigned int alloc_unit = xfs_inode_alloc_unitsize(ip);
49
50 if (!is_power_of_2(alloc_unit))
51 return isaligned_64(pos, alloc_unit) &&
52 isaligned_64(len, alloc_unit);
53
54 return !((pos | len) & (alloc_unit - 1));
55}
56
57/*
58 * Fsync operations on directories are much simpler than on regular files,
59 * as there is no file data to flush, and thus also no need for explicit
60 * cache flush operations, and there are no non-transaction metadata updates
61 * on directories either.
62 */
63STATIC int
64xfs_dir_fsync(
65 struct file *file,
66 loff_t start,
67 loff_t end,
68 int datasync)
69{
70 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
71
72 trace_xfs_dir_fsync(ip);
73 return xfs_log_force_inode(ip);
74}
75
76static xfs_csn_t
77xfs_fsync_seq(
78 struct xfs_inode *ip,
79 bool datasync)
80{
81 if (!xfs_ipincount(ip))
82 return 0;
83 if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
84 return 0;
85 return ip->i_itemp->ili_commit_seq;
86}
87
88/*
89 * All metadata updates are logged, which means that we just have to flush the
90 * log up to the latest LSN that touched the inode.
91 *
92 * If we have concurrent fsync/fdatasync() calls, we need them to all block on
93 * the log force before we clear the ili_fsync_fields field. This ensures that
94 * we don't get a racing sync operation that does not wait for the metadata to
95 * hit the journal before returning. If we race with clearing ili_fsync_fields,
96 * then all that will happen is the log force will do nothing as the lsn will
97 * already be on disk. We can't race with setting ili_fsync_fields because that
98 * is done under XFS_ILOCK_EXCL, and that can't happen because we hold the lock
99 * shared until after the ili_fsync_fields is cleared.
100 */
101static int
102xfs_fsync_flush_log(
103 struct xfs_inode *ip,
104 bool datasync,
105 int *log_flushed)
106{
107 int error = 0;
108 xfs_csn_t seq;
109
110 xfs_ilock(ip, XFS_ILOCK_SHARED);
111 seq = xfs_fsync_seq(ip, datasync);
112 if (seq) {
113 error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
114 log_flushed);
115
116 spin_lock(&ip->i_itemp->ili_lock);
117 ip->i_itemp->ili_fsync_fields = 0;
118 spin_unlock(&ip->i_itemp->ili_lock);
119 }
120 xfs_iunlock(ip, XFS_ILOCK_SHARED);
121 return error;
122}
123
124STATIC int
125xfs_file_fsync(
126 struct file *file,
127 loff_t start,
128 loff_t end,
129 int datasync)
130{
131 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
132 struct xfs_mount *mp = ip->i_mount;
133 int error, err2;
134 int log_flushed = 0;
135
136 trace_xfs_file_fsync(ip);
137
138 error = file_write_and_wait_range(file, start, end);
139 if (error)
140 return error;
141
142 if (xfs_is_shutdown(mp))
143 return -EIO;
144
145 xfs_iflags_clear(ip, XFS_ITRUNCATED);
146
147 /*
148 * If we have an RT and/or log subvolume we need to make sure to flush
149 * the write cache the device used for file data first. This is to
150 * ensure newly written file data make it to disk before logging the new
151 * inode size in case of an extending write.
152 */
153 if (XFS_IS_REALTIME_INODE(ip))
154 error = blkdev_issue_flush(mp->m_rtdev_targp->bt_bdev);
155 else if (mp->m_logdev_targp != mp->m_ddev_targp)
156 error = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
157
158 /*
159 * Any inode that has dirty modifications in the log is pinned. The
160 * racy check here for a pinned inode will not catch modifications
161 * that happen concurrently to the fsync call, but fsync semantics
162 * only require to sync previously completed I/O.
163 */
164 if (xfs_ipincount(ip)) {
165 err2 = xfs_fsync_flush_log(ip, datasync, &log_flushed);
166 if (err2 && !error)
167 error = err2;
168 }
169
170 /*
171 * If we only have a single device, and the log force about was
172 * a no-op we might have to flush the data device cache here.
173 * This can only happen for fdatasync/O_DSYNC if we were overwriting
174 * an already allocated file and thus do not have any metadata to
175 * commit.
176 */
177 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
178 mp->m_logdev_targp == mp->m_ddev_targp) {
179 err2 = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
180 if (err2 && !error)
181 error = err2;
182 }
183
184 return error;
185}
186
187static int
188xfs_ilock_iocb(
189 struct kiocb *iocb,
190 unsigned int lock_mode)
191{
192 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
193
194 if (iocb->ki_flags & IOCB_NOWAIT) {
195 if (!xfs_ilock_nowait(ip, lock_mode))
196 return -EAGAIN;
197 } else {
198 xfs_ilock(ip, lock_mode);
199 }
200
201 return 0;
202}
203
204static int
205xfs_ilock_iocb_for_write(
206 struct kiocb *iocb,
207 unsigned int *lock_mode)
208{
209 ssize_t ret;
210 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
211
212 ret = xfs_ilock_iocb(iocb, *lock_mode);
213 if (ret)
214 return ret;
215
216 /*
217 * If a reflink remap is in progress we always need to take the iolock
218 * exclusively to wait for it to finish.
219 */
220 if (*lock_mode == XFS_IOLOCK_SHARED &&
221 xfs_iflags_test(ip, XFS_IREMAPPING)) {
222 xfs_iunlock(ip, *lock_mode);
223 *lock_mode = XFS_IOLOCK_EXCL;
224 return xfs_ilock_iocb(iocb, *lock_mode);
225 }
226
227 return 0;
228}
229
230STATIC ssize_t
231xfs_file_dio_read(
232 struct kiocb *iocb,
233 struct iov_iter *to)
234{
235 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
236 ssize_t ret;
237
238 trace_xfs_file_direct_read(iocb, to);
239
240 if (!iov_iter_count(to))
241 return 0; /* skip atime */
242
243 file_accessed(iocb->ki_filp);
244
245 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
246 if (ret)
247 return ret;
248 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
249 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
250
251 return ret;
252}
253
254static noinline ssize_t
255xfs_file_dax_read(
256 struct kiocb *iocb,
257 struct iov_iter *to)
258{
259 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
260 ssize_t ret = 0;
261
262 trace_xfs_file_dax_read(iocb, to);
263
264 if (!iov_iter_count(to))
265 return 0; /* skip atime */
266
267 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
268 if (ret)
269 return ret;
270 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
271 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
272
273 file_accessed(iocb->ki_filp);
274 return ret;
275}
276
277STATIC ssize_t
278xfs_file_buffered_read(
279 struct kiocb *iocb,
280 struct iov_iter *to)
281{
282 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
283 ssize_t ret;
284
285 trace_xfs_file_buffered_read(iocb, to);
286
287 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
288 if (ret)
289 return ret;
290 ret = generic_file_read_iter(iocb, to);
291 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
292
293 return ret;
294}
295
296STATIC ssize_t
297xfs_file_read_iter(
298 struct kiocb *iocb,
299 struct iov_iter *to)
300{
301 struct inode *inode = file_inode(iocb->ki_filp);
302 struct xfs_mount *mp = XFS_I(inode)->i_mount;
303 ssize_t ret = 0;
304
305 XFS_STATS_INC(mp, xs_read_calls);
306
307 if (xfs_is_shutdown(mp))
308 return -EIO;
309
310 if (IS_DAX(inode))
311 ret = xfs_file_dax_read(iocb, to);
312 else if (iocb->ki_flags & IOCB_DIRECT)
313 ret = xfs_file_dio_read(iocb, to);
314 else
315 ret = xfs_file_buffered_read(iocb, to);
316
317 if (ret > 0)
318 XFS_STATS_ADD(mp, xs_read_bytes, ret);
319 return ret;
320}
321
322STATIC ssize_t
323xfs_file_splice_read(
324 struct file *in,
325 loff_t *ppos,
326 struct pipe_inode_info *pipe,
327 size_t len,
328 unsigned int flags)
329{
330 struct inode *inode = file_inode(in);
331 struct xfs_inode *ip = XFS_I(inode);
332 struct xfs_mount *mp = ip->i_mount;
333 ssize_t ret = 0;
334
335 XFS_STATS_INC(mp, xs_read_calls);
336
337 if (xfs_is_shutdown(mp))
338 return -EIO;
339
340 trace_xfs_file_splice_read(ip, *ppos, len);
341
342 xfs_ilock(ip, XFS_IOLOCK_SHARED);
343 ret = filemap_splice_read(in, ppos, pipe, len, flags);
344 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
345 if (ret > 0)
346 XFS_STATS_ADD(mp, xs_read_bytes, ret);
347 return ret;
348}
349
350/*
351 * Take care of zeroing post-EOF blocks when they might exist.
352 *
353 * Returns 0 if successfully, a negative error for a failure, or 1 if this
354 * function dropped the iolock and reacquired it exclusively and the caller
355 * needs to restart the write sanity checks.
356 */
357static ssize_t
358xfs_file_write_zero_eof(
359 struct kiocb *iocb,
360 struct iov_iter *from,
361 unsigned int *iolock,
362 size_t count,
363 bool *drained_dio)
364{
365 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
366 loff_t isize;
367 int error;
368
369 /*
370 * We need to serialise against EOF updates that occur in IO completions
371 * here. We want to make sure that nobody is changing the size while
372 * we do this check until we have placed an IO barrier (i.e. hold
373 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
374 * spinlock effectively forms a memory barrier once we have
375 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
376 * hence be able to correctly determine if we need to run zeroing.
377 */
378 spin_lock(&ip->i_flags_lock);
379 isize = i_size_read(VFS_I(ip));
380 if (iocb->ki_pos <= isize) {
381 spin_unlock(&ip->i_flags_lock);
382 return 0;
383 }
384 spin_unlock(&ip->i_flags_lock);
385
386 if (iocb->ki_flags & IOCB_NOWAIT)
387 return -EAGAIN;
388
389 if (!*drained_dio) {
390 /*
391 * If zeroing is needed and we are currently holding the iolock
392 * shared, we need to update it to exclusive which implies
393 * having to redo all checks before.
394 */
395 if (*iolock == XFS_IOLOCK_SHARED) {
396 xfs_iunlock(ip, *iolock);
397 *iolock = XFS_IOLOCK_EXCL;
398 xfs_ilock(ip, *iolock);
399 iov_iter_reexpand(from, count);
400 }
401
402 /*
403 * We now have an IO submission barrier in place, but AIO can do
404 * EOF updates during IO completion and hence we now need to
405 * wait for all of them to drain. Non-AIO DIO will have drained
406 * before we are given the XFS_IOLOCK_EXCL, and so for most
407 * cases this wait is a no-op.
408 */
409 inode_dio_wait(VFS_I(ip));
410 *drained_dio = true;
411 return 1;
412 }
413
414 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
415
416 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
417 error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
418 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
419
420 return error;
421}
422
423/*
424 * Common pre-write limit and setup checks.
425 *
426 * Called with the iolock held either shared and exclusive according to
427 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
428 * if called for a direct write beyond i_size.
429 */
430STATIC ssize_t
431xfs_file_write_checks(
432 struct kiocb *iocb,
433 struct iov_iter *from,
434 unsigned int *iolock)
435{
436 struct inode *inode = iocb->ki_filp->f_mapping->host;
437 size_t count = iov_iter_count(from);
438 bool drained_dio = false;
439 ssize_t error;
440
441restart:
442 error = generic_write_checks(iocb, from);
443 if (error <= 0)
444 return error;
445
446 if (iocb->ki_flags & IOCB_NOWAIT) {
447 error = break_layout(inode, false);
448 if (error == -EWOULDBLOCK)
449 error = -EAGAIN;
450 } else {
451 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
452 }
453
454 if (error)
455 return error;
456
457 /*
458 * For changing security info in file_remove_privs() we need i_rwsem
459 * exclusively.
460 */
461 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
462 xfs_iunlock(XFS_I(inode), *iolock);
463 *iolock = XFS_IOLOCK_EXCL;
464 error = xfs_ilock_iocb(iocb, *iolock);
465 if (error) {
466 *iolock = 0;
467 return error;
468 }
469 goto restart;
470 }
471
472 /*
473 * If the offset is beyond the size of the file, we need to zero all
474 * blocks that fall between the existing EOF and the start of this
475 * write.
476 *
477 * We can do an unlocked check for i_size here safely as I/O completion
478 * can only extend EOF. Truncate is locked out at this point, so the
479 * EOF can not move backwards, only forwards. Hence we only need to take
480 * the slow path when we are at or beyond the current EOF.
481 */
482 if (iocb->ki_pos > i_size_read(inode)) {
483 error = xfs_file_write_zero_eof(iocb, from, iolock, count,
484 &drained_dio);
485 if (error == 1)
486 goto restart;
487 if (error)
488 return error;
489 }
490
491 return kiocb_modified(iocb);
492}
493
494static int
495xfs_dio_write_end_io(
496 struct kiocb *iocb,
497 ssize_t size,
498 int error,
499 unsigned flags)
500{
501 struct inode *inode = file_inode(iocb->ki_filp);
502 struct xfs_inode *ip = XFS_I(inode);
503 loff_t offset = iocb->ki_pos;
504 unsigned int nofs_flag;
505
506 trace_xfs_end_io_direct_write(ip, offset, size);
507
508 if (xfs_is_shutdown(ip->i_mount))
509 return -EIO;
510
511 if (error)
512 return error;
513 if (!size)
514 return 0;
515
516 /*
517 * Capture amount written on completion as we can't reliably account
518 * for it on submission.
519 */
520 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
521
522 /*
523 * We can allocate memory here while doing writeback on behalf of
524 * memory reclaim. To avoid memory allocation deadlocks set the
525 * task-wide nofs context for the following operations.
526 */
527 nofs_flag = memalloc_nofs_save();
528
529 if (flags & IOMAP_DIO_COW) {
530 error = xfs_reflink_end_cow(ip, offset, size);
531 if (error)
532 goto out;
533 }
534
535 /*
536 * Unwritten conversion updates the in-core isize after extent
537 * conversion but before updating the on-disk size. Updating isize any
538 * earlier allows a racing dio read to find unwritten extents before
539 * they are converted.
540 */
541 if (flags & IOMAP_DIO_UNWRITTEN) {
542 error = xfs_iomap_write_unwritten(ip, offset, size, true);
543 goto out;
544 }
545
546 /*
547 * We need to update the in-core inode size here so that we don't end up
548 * with the on-disk inode size being outside the in-core inode size. We
549 * have no other method of updating EOF for AIO, so always do it here
550 * if necessary.
551 *
552 * We need to lock the test/set EOF update as we can be racing with
553 * other IO completions here to update the EOF. Failing to serialise
554 * here can result in EOF moving backwards and Bad Things Happen when
555 * that occurs.
556 *
557 * As IO completion only ever extends EOF, we can do an unlocked check
558 * here to avoid taking the spinlock. If we land within the current EOF,
559 * then we do not need to do an extending update at all, and we don't
560 * need to take the lock to check this. If we race with an update moving
561 * EOF, then we'll either still be beyond EOF and need to take the lock,
562 * or we'll be within EOF and we don't need to take it at all.
563 */
564 if (offset + size <= i_size_read(inode))
565 goto out;
566
567 spin_lock(&ip->i_flags_lock);
568 if (offset + size > i_size_read(inode)) {
569 i_size_write(inode, offset + size);
570 spin_unlock(&ip->i_flags_lock);
571 error = xfs_setfilesize(ip, offset, size);
572 } else {
573 spin_unlock(&ip->i_flags_lock);
574 }
575
576out:
577 memalloc_nofs_restore(nofs_flag);
578 return error;
579}
580
581static const struct iomap_dio_ops xfs_dio_write_ops = {
582 .end_io = xfs_dio_write_end_io,
583};
584
585/*
586 * Handle block aligned direct I/O writes
587 */
588static noinline ssize_t
589xfs_file_dio_write_aligned(
590 struct xfs_inode *ip,
591 struct kiocb *iocb,
592 struct iov_iter *from)
593{
594 unsigned int iolock = XFS_IOLOCK_SHARED;
595 ssize_t ret;
596
597 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
598 if (ret)
599 return ret;
600 ret = xfs_file_write_checks(iocb, from, &iolock);
601 if (ret)
602 goto out_unlock;
603
604 /*
605 * We don't need to hold the IOLOCK exclusively across the IO, so demote
606 * the iolock back to shared if we had to take the exclusive lock in
607 * xfs_file_write_checks() for other reasons.
608 */
609 if (iolock == XFS_IOLOCK_EXCL) {
610 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
611 iolock = XFS_IOLOCK_SHARED;
612 }
613 trace_xfs_file_direct_write(iocb, from);
614 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
615 &xfs_dio_write_ops, 0, NULL, 0);
616out_unlock:
617 if (iolock)
618 xfs_iunlock(ip, iolock);
619 return ret;
620}
621
622/*
623 * Handle block unaligned direct I/O writes
624 *
625 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
626 * them to be done in parallel with reads and other direct I/O writes. However,
627 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
628 * to do sub-block zeroing and that requires serialisation against other direct
629 * I/O to the same block. In this case we need to serialise the submission of
630 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
631 * In the case where sub-block zeroing is not required, we can do concurrent
632 * sub-block dios to the same block successfully.
633 *
634 * Optimistically submit the I/O using the shared lock first, but use the
635 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
636 * if block allocation or partial block zeroing would be required. In that case
637 * we try again with the exclusive lock.
638 */
639static noinline ssize_t
640xfs_file_dio_write_unaligned(
641 struct xfs_inode *ip,
642 struct kiocb *iocb,
643 struct iov_iter *from)
644{
645 size_t isize = i_size_read(VFS_I(ip));
646 size_t count = iov_iter_count(from);
647 unsigned int iolock = XFS_IOLOCK_SHARED;
648 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
649 ssize_t ret;
650
651 /*
652 * Extending writes need exclusivity because of the sub-block zeroing
653 * that the DIO code always does for partial tail blocks beyond EOF, so
654 * don't even bother trying the fast path in this case.
655 */
656 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
657 if (iocb->ki_flags & IOCB_NOWAIT)
658 return -EAGAIN;
659retry_exclusive:
660 iolock = XFS_IOLOCK_EXCL;
661 flags = IOMAP_DIO_FORCE_WAIT;
662 }
663
664 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
665 if (ret)
666 return ret;
667
668 /*
669 * We can't properly handle unaligned direct I/O to reflink files yet,
670 * as we can't unshare a partial block.
671 */
672 if (xfs_is_cow_inode(ip)) {
673 trace_xfs_reflink_bounce_dio_write(iocb, from);
674 ret = -ENOTBLK;
675 goto out_unlock;
676 }
677
678 ret = xfs_file_write_checks(iocb, from, &iolock);
679 if (ret)
680 goto out_unlock;
681
682 /*
683 * If we are doing exclusive unaligned I/O, this must be the only I/O
684 * in-flight. Otherwise we risk data corruption due to unwritten extent
685 * conversions from the AIO end_io handler. Wait for all other I/O to
686 * drain first.
687 */
688 if (flags & IOMAP_DIO_FORCE_WAIT)
689 inode_dio_wait(VFS_I(ip));
690
691 trace_xfs_file_direct_write(iocb, from);
692 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
693 &xfs_dio_write_ops, flags, NULL, 0);
694
695 /*
696 * Retry unaligned I/O with exclusive blocking semantics if the DIO
697 * layer rejected it for mapping or locking reasons. If we are doing
698 * nonblocking user I/O, propagate the error.
699 */
700 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
701 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
702 xfs_iunlock(ip, iolock);
703 goto retry_exclusive;
704 }
705
706out_unlock:
707 if (iolock)
708 xfs_iunlock(ip, iolock);
709 return ret;
710}
711
712static ssize_t
713xfs_file_dio_write(
714 struct kiocb *iocb,
715 struct iov_iter *from)
716{
717 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
718 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
719 size_t count = iov_iter_count(from);
720
721 /* direct I/O must be aligned to device logical sector size */
722 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
723 return -EINVAL;
724 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
725 return xfs_file_dio_write_unaligned(ip, iocb, from);
726 return xfs_file_dio_write_aligned(ip, iocb, from);
727}
728
729static noinline ssize_t
730xfs_file_dax_write(
731 struct kiocb *iocb,
732 struct iov_iter *from)
733{
734 struct inode *inode = iocb->ki_filp->f_mapping->host;
735 struct xfs_inode *ip = XFS_I(inode);
736 unsigned int iolock = XFS_IOLOCK_EXCL;
737 ssize_t ret, error = 0;
738 loff_t pos;
739
740 ret = xfs_ilock_iocb(iocb, iolock);
741 if (ret)
742 return ret;
743 ret = xfs_file_write_checks(iocb, from, &iolock);
744 if (ret)
745 goto out;
746
747 pos = iocb->ki_pos;
748
749 trace_xfs_file_dax_write(iocb, from);
750 ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
751 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
752 i_size_write(inode, iocb->ki_pos);
753 error = xfs_setfilesize(ip, pos, ret);
754 }
755out:
756 if (iolock)
757 xfs_iunlock(ip, iolock);
758 if (error)
759 return error;
760
761 if (ret > 0) {
762 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
763
764 /* Handle various SYNC-type writes */
765 ret = generic_write_sync(iocb, ret);
766 }
767 return ret;
768}
769
770STATIC ssize_t
771xfs_file_buffered_write(
772 struct kiocb *iocb,
773 struct iov_iter *from)
774{
775 struct inode *inode = iocb->ki_filp->f_mapping->host;
776 struct xfs_inode *ip = XFS_I(inode);
777 ssize_t ret;
778 bool cleared_space = false;
779 unsigned int iolock;
780
781write_retry:
782 iolock = XFS_IOLOCK_EXCL;
783 ret = xfs_ilock_iocb(iocb, iolock);
784 if (ret)
785 return ret;
786
787 ret = xfs_file_write_checks(iocb, from, &iolock);
788 if (ret)
789 goto out;
790
791 trace_xfs_file_buffered_write(iocb, from);
792 ret = iomap_file_buffered_write(iocb, from,
793 &xfs_buffered_write_iomap_ops, NULL);
794
795 /*
796 * If we hit a space limit, try to free up some lingering preallocated
797 * space before returning an error. In the case of ENOSPC, first try to
798 * write back all dirty inodes to free up some of the excess reserved
799 * metadata space. This reduces the chances that the eofblocks scan
800 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
801 * also behaves as a filter to prevent too many eofblocks scans from
802 * running at the same time. Use a synchronous scan to increase the
803 * effectiveness of the scan.
804 */
805 if (ret == -EDQUOT && !cleared_space) {
806 xfs_iunlock(ip, iolock);
807 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
808 cleared_space = true;
809 goto write_retry;
810 } else if (ret == -ENOSPC && !cleared_space) {
811 struct xfs_icwalk icw = {0};
812
813 cleared_space = true;
814 xfs_flush_inodes(ip->i_mount);
815
816 xfs_iunlock(ip, iolock);
817 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
818 xfs_blockgc_free_space(ip->i_mount, &icw);
819 goto write_retry;
820 }
821
822out:
823 if (iolock)
824 xfs_iunlock(ip, iolock);
825
826 if (ret > 0) {
827 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
828 /* Handle various SYNC-type writes */
829 ret = generic_write_sync(iocb, ret);
830 }
831 return ret;
832}
833
834STATIC ssize_t
835xfs_file_write_iter(
836 struct kiocb *iocb,
837 struct iov_iter *from)
838{
839 struct inode *inode = iocb->ki_filp->f_mapping->host;
840 struct xfs_inode *ip = XFS_I(inode);
841 ssize_t ret;
842 size_t ocount = iov_iter_count(from);
843
844 XFS_STATS_INC(ip->i_mount, xs_write_calls);
845
846 if (ocount == 0)
847 return 0;
848
849 if (xfs_is_shutdown(ip->i_mount))
850 return -EIO;
851
852 if (IS_DAX(inode))
853 return xfs_file_dax_write(iocb, from);
854
855 if (iocb->ki_flags & IOCB_ATOMIC) {
856 /*
857 * Currently only atomic writing of a single FS block is
858 * supported. It would be possible to atomic write smaller than
859 * a FS block, but there is no requirement to support this.
860 * Note that iomap also does not support this yet.
861 */
862 if (ocount != ip->i_mount->m_sb.sb_blocksize)
863 return -EINVAL;
864 ret = generic_atomic_write_valid(iocb, from);
865 if (ret)
866 return ret;
867 }
868
869 if (iocb->ki_flags & IOCB_DIRECT) {
870 /*
871 * Allow a directio write to fall back to a buffered
872 * write *only* in the case that we're doing a reflink
873 * CoW. In all other directio scenarios we do not
874 * allow an operation to fall back to buffered mode.
875 */
876 ret = xfs_file_dio_write(iocb, from);
877 if (ret != -ENOTBLK)
878 return ret;
879 }
880
881 return xfs_file_buffered_write(iocb, from);
882}
883
884/* Does this file, inode, or mount want synchronous writes? */
885static inline bool xfs_file_sync_writes(struct file *filp)
886{
887 struct xfs_inode *ip = XFS_I(file_inode(filp));
888
889 if (xfs_has_wsync(ip->i_mount))
890 return true;
891 if (filp->f_flags & (__O_SYNC | O_DSYNC))
892 return true;
893 if (IS_SYNC(file_inode(filp)))
894 return true;
895
896 return false;
897}
898
899static int
900xfs_falloc_newsize(
901 struct file *file,
902 int mode,
903 loff_t offset,
904 loff_t len,
905 loff_t *new_size)
906{
907 struct inode *inode = file_inode(file);
908
909 if ((mode & FALLOC_FL_KEEP_SIZE) || offset + len <= i_size_read(inode))
910 return 0;
911 *new_size = offset + len;
912 return inode_newsize_ok(inode, *new_size);
913}
914
915static int
916xfs_falloc_setsize(
917 struct file *file,
918 loff_t new_size)
919{
920 struct iattr iattr = {
921 .ia_valid = ATTR_SIZE,
922 .ia_size = new_size,
923 };
924
925 if (!new_size)
926 return 0;
927 return xfs_vn_setattr_size(file_mnt_idmap(file), file_dentry(file),
928 &iattr);
929}
930
931static int
932xfs_falloc_collapse_range(
933 struct file *file,
934 loff_t offset,
935 loff_t len)
936{
937 struct inode *inode = file_inode(file);
938 loff_t new_size = i_size_read(inode) - len;
939 int error;
940
941 if (!xfs_is_falloc_aligned(XFS_I(inode), offset, len))
942 return -EINVAL;
943
944 /*
945 * There is no need to overlap collapse range with EOF, in which case it
946 * is effectively a truncate operation
947 */
948 if (offset + len >= i_size_read(inode))
949 return -EINVAL;
950
951 error = xfs_collapse_file_space(XFS_I(inode), offset, len);
952 if (error)
953 return error;
954 return xfs_falloc_setsize(file, new_size);
955}
956
957static int
958xfs_falloc_insert_range(
959 struct file *file,
960 loff_t offset,
961 loff_t len)
962{
963 struct inode *inode = file_inode(file);
964 loff_t isize = i_size_read(inode);
965 int error;
966
967 if (!xfs_is_falloc_aligned(XFS_I(inode), offset, len))
968 return -EINVAL;
969
970 /*
971 * New inode size must not exceed ->s_maxbytes, accounting for
972 * possible signed overflow.
973 */
974 if (inode->i_sb->s_maxbytes - isize < len)
975 return -EFBIG;
976
977 /* Offset should be less than i_size */
978 if (offset >= isize)
979 return -EINVAL;
980
981 error = xfs_falloc_setsize(file, isize + len);
982 if (error)
983 return error;
984
985 /*
986 * Perform hole insertion now that the file size has been updated so
987 * that if we crash during the operation we don't leave shifted extents
988 * past EOF and hence losing access to the data that is contained within
989 * them.
990 */
991 return xfs_insert_file_space(XFS_I(inode), offset, len);
992}
993
994/*
995 * Punch a hole and prealloc the range. We use a hole punch rather than
996 * unwritten extent conversion for two reasons:
997 *
998 * 1.) Hole punch handles partial block zeroing for us.
999 * 2.) If prealloc returns ENOSPC, the file range is still zero-valued by
1000 * virtue of the hole punch.
1001 */
1002static int
1003xfs_falloc_zero_range(
1004 struct file *file,
1005 int mode,
1006 loff_t offset,
1007 loff_t len)
1008{
1009 struct inode *inode = file_inode(file);
1010 unsigned int blksize = i_blocksize(inode);
1011 loff_t new_size = 0;
1012 int error;
1013
1014 trace_xfs_zero_file_space(XFS_I(inode));
1015
1016 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1017 if (error)
1018 return error;
1019
1020 error = xfs_free_file_space(XFS_I(inode), offset, len);
1021 if (error)
1022 return error;
1023
1024 len = round_up(offset + len, blksize) - round_down(offset, blksize);
1025 offset = round_down(offset, blksize);
1026 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1027 if (error)
1028 return error;
1029 return xfs_falloc_setsize(file, new_size);
1030}
1031
1032static int
1033xfs_falloc_unshare_range(
1034 struct file *file,
1035 int mode,
1036 loff_t offset,
1037 loff_t len)
1038{
1039 struct inode *inode = file_inode(file);
1040 loff_t new_size = 0;
1041 int error;
1042
1043 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1044 if (error)
1045 return error;
1046
1047 error = xfs_reflink_unshare(XFS_I(inode), offset, len);
1048 if (error)
1049 return error;
1050
1051 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1052 if (error)
1053 return error;
1054 return xfs_falloc_setsize(file, new_size);
1055}
1056
1057static int
1058xfs_falloc_allocate_range(
1059 struct file *file,
1060 int mode,
1061 loff_t offset,
1062 loff_t len)
1063{
1064 struct inode *inode = file_inode(file);
1065 loff_t new_size = 0;
1066 int error;
1067
1068 /*
1069 * If always_cow mode we can't use preallocations and thus should not
1070 * create them.
1071 */
1072 if (xfs_is_always_cow_inode(XFS_I(inode)))
1073 return -EOPNOTSUPP;
1074
1075 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1076 if (error)
1077 return error;
1078
1079 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1080 if (error)
1081 return error;
1082 return xfs_falloc_setsize(file, new_size);
1083}
1084
1085#define XFS_FALLOC_FL_SUPPORTED \
1086 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
1087 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
1088 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
1089
1090STATIC long
1091xfs_file_fallocate(
1092 struct file *file,
1093 int mode,
1094 loff_t offset,
1095 loff_t len)
1096{
1097 struct inode *inode = file_inode(file);
1098 struct xfs_inode *ip = XFS_I(inode);
1099 long error;
1100 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
1101
1102 if (!S_ISREG(inode->i_mode))
1103 return -EINVAL;
1104 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
1105 return -EOPNOTSUPP;
1106
1107 xfs_ilock(ip, iolock);
1108 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
1109 if (error)
1110 goto out_unlock;
1111
1112 /*
1113 * Must wait for all AIO to complete before we continue as AIO can
1114 * change the file size on completion without holding any locks we
1115 * currently hold. We must do this first because AIO can update both
1116 * the on disk and in memory inode sizes, and the operations that follow
1117 * require the in-memory size to be fully up-to-date.
1118 */
1119 inode_dio_wait(inode);
1120
1121 error = file_modified(file);
1122 if (error)
1123 goto out_unlock;
1124
1125 switch (mode & FALLOC_FL_MODE_MASK) {
1126 case FALLOC_FL_PUNCH_HOLE:
1127 error = xfs_free_file_space(ip, offset, len);
1128 break;
1129 case FALLOC_FL_COLLAPSE_RANGE:
1130 error = xfs_falloc_collapse_range(file, offset, len);
1131 break;
1132 case FALLOC_FL_INSERT_RANGE:
1133 error = xfs_falloc_insert_range(file, offset, len);
1134 break;
1135 case FALLOC_FL_ZERO_RANGE:
1136 error = xfs_falloc_zero_range(file, mode, offset, len);
1137 break;
1138 case FALLOC_FL_UNSHARE_RANGE:
1139 error = xfs_falloc_unshare_range(file, mode, offset, len);
1140 break;
1141 case FALLOC_FL_ALLOCATE_RANGE:
1142 error = xfs_falloc_allocate_range(file, mode, offset, len);
1143 break;
1144 default:
1145 error = -EOPNOTSUPP;
1146 break;
1147 }
1148
1149 if (!error && xfs_file_sync_writes(file))
1150 error = xfs_log_force_inode(ip);
1151
1152out_unlock:
1153 xfs_iunlock(ip, iolock);
1154 return error;
1155}
1156
1157STATIC int
1158xfs_file_fadvise(
1159 struct file *file,
1160 loff_t start,
1161 loff_t end,
1162 int advice)
1163{
1164 struct xfs_inode *ip = XFS_I(file_inode(file));
1165 int ret;
1166 int lockflags = 0;
1167
1168 /*
1169 * Operations creating pages in page cache need protection from hole
1170 * punching and similar ops
1171 */
1172 if (advice == POSIX_FADV_WILLNEED) {
1173 lockflags = XFS_IOLOCK_SHARED;
1174 xfs_ilock(ip, lockflags);
1175 }
1176 ret = generic_fadvise(file, start, end, advice);
1177 if (lockflags)
1178 xfs_iunlock(ip, lockflags);
1179 return ret;
1180}
1181
1182STATIC loff_t
1183xfs_file_remap_range(
1184 struct file *file_in,
1185 loff_t pos_in,
1186 struct file *file_out,
1187 loff_t pos_out,
1188 loff_t len,
1189 unsigned int remap_flags)
1190{
1191 struct inode *inode_in = file_inode(file_in);
1192 struct xfs_inode *src = XFS_I(inode_in);
1193 struct inode *inode_out = file_inode(file_out);
1194 struct xfs_inode *dest = XFS_I(inode_out);
1195 struct xfs_mount *mp = src->i_mount;
1196 loff_t remapped = 0;
1197 xfs_extlen_t cowextsize;
1198 int ret;
1199
1200 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1201 return -EINVAL;
1202
1203 if (!xfs_has_reflink(mp))
1204 return -EOPNOTSUPP;
1205
1206 if (xfs_is_shutdown(mp))
1207 return -EIO;
1208
1209 /* Prepare and then clone file data. */
1210 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1211 &len, remap_flags);
1212 if (ret || len == 0)
1213 return ret;
1214
1215 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1216
1217 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1218 &remapped);
1219 if (ret)
1220 goto out_unlock;
1221
1222 /*
1223 * Carry the cowextsize hint from src to dest if we're sharing the
1224 * entire source file to the entire destination file, the source file
1225 * has a cowextsize hint, and the destination file does not.
1226 */
1227 cowextsize = 0;
1228 if (pos_in == 0 && len == i_size_read(inode_in) &&
1229 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1230 pos_out == 0 && len >= i_size_read(inode_out) &&
1231 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1232 cowextsize = src->i_cowextsize;
1233
1234 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1235 remap_flags);
1236 if (ret)
1237 goto out_unlock;
1238
1239 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1240 xfs_log_force_inode(dest);
1241out_unlock:
1242 xfs_iunlock2_remapping(src, dest);
1243 if (ret)
1244 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1245 /*
1246 * If the caller did not set CAN_SHORTEN, then it is not prepared to
1247 * handle partial results -- either the whole remap succeeds, or we
1248 * must say why it did not. In this case, any error should be returned
1249 * to the caller.
1250 */
1251 if (ret && remapped < len && !(remap_flags & REMAP_FILE_CAN_SHORTEN))
1252 return ret;
1253 return remapped > 0 ? remapped : ret;
1254}
1255
1256STATIC int
1257xfs_file_open(
1258 struct inode *inode,
1259 struct file *file)
1260{
1261 if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1262 return -EIO;
1263 file->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
1264 if (xfs_inode_can_atomicwrite(XFS_I(inode)))
1265 file->f_mode |= FMODE_CAN_ATOMIC_WRITE;
1266 return generic_file_open(inode, file);
1267}
1268
1269STATIC int
1270xfs_dir_open(
1271 struct inode *inode,
1272 struct file *file)
1273{
1274 struct xfs_inode *ip = XFS_I(inode);
1275 unsigned int mode;
1276 int error;
1277
1278 if (xfs_is_shutdown(ip->i_mount))
1279 return -EIO;
1280 error = generic_file_open(inode, file);
1281 if (error)
1282 return error;
1283
1284 /*
1285 * If there are any blocks, read-ahead block 0 as we're almost
1286 * certain to have the next operation be a read there.
1287 */
1288 mode = xfs_ilock_data_map_shared(ip);
1289 if (ip->i_df.if_nextents > 0)
1290 error = xfs_dir3_data_readahead(ip, 0, 0);
1291 xfs_iunlock(ip, mode);
1292 return error;
1293}
1294
1295/*
1296 * Don't bother propagating errors. We're just doing cleanup, and the caller
1297 * ignores the return value anyway.
1298 */
1299STATIC int
1300xfs_file_release(
1301 struct inode *inode,
1302 struct file *file)
1303{
1304 struct xfs_inode *ip = XFS_I(inode);
1305 struct xfs_mount *mp = ip->i_mount;
1306
1307 /*
1308 * If this is a read-only mount or the file system has been shut down,
1309 * don't generate I/O.
1310 */
1311 if (xfs_is_readonly(mp) || xfs_is_shutdown(mp))
1312 return 0;
1313
1314 /*
1315 * If we previously truncated this file and removed old data in the
1316 * process, we want to initiate "early" writeout on the last close.
1317 * This is an attempt to combat the notorious NULL files problem which
1318 * is particularly noticeable from a truncate down, buffered (re-)write
1319 * (delalloc), followed by a crash. What we are effectively doing here
1320 * is significantly reducing the time window where we'd otherwise be
1321 * exposed to that problem.
1322 */
1323 if (xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED)) {
1324 xfs_iflags_clear(ip, XFS_EOFBLOCKS_RELEASED);
1325 if (ip->i_delayed_blks > 0)
1326 filemap_flush(inode->i_mapping);
1327 }
1328
1329 /*
1330 * XFS aggressively preallocates post-EOF space to generate contiguous
1331 * allocations for writers that append to the end of the file.
1332 *
1333 * To support workloads that close and reopen the file frequently, these
1334 * preallocations usually persist after a close unless it is the first
1335 * close for the inode. This is a tradeoff to generate tightly packed
1336 * data layouts for unpacking tarballs or similar archives that write
1337 * one file after another without going back to it while keeping the
1338 * preallocation for files that have recurring open/write/close cycles.
1339 *
1340 * This heuristic is skipped for inodes with the append-only flag as
1341 * that flag is rather pointless for inodes written only once.
1342 *
1343 * There is no point in freeing blocks here for open but unlinked files
1344 * as they will be taken care of by the inactivation path soon.
1345 *
1346 * When releasing a read-only context, don't flush data or trim post-EOF
1347 * blocks. This avoids open/read/close workloads from removing EOF
1348 * blocks that other writers depend upon to reduce fragmentation.
1349 *
1350 * If we can't get the iolock just skip truncating the blocks past EOF
1351 * because we could deadlock with the mmap_lock otherwise. We'll get
1352 * another chance to drop them once the last reference to the inode is
1353 * dropped, so we'll never leak blocks permanently.
1354 */
1355 if (inode->i_nlink &&
1356 (file->f_mode & FMODE_WRITE) &&
1357 !(ip->i_diflags & XFS_DIFLAG_APPEND) &&
1358 !xfs_iflags_test(ip, XFS_EOFBLOCKS_RELEASED) &&
1359 xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1360 if (xfs_can_free_eofblocks(ip) &&
1361 !xfs_iflags_test_and_set(ip, XFS_EOFBLOCKS_RELEASED))
1362 xfs_free_eofblocks(ip);
1363 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1364 }
1365
1366 return 0;
1367}
1368
1369STATIC int
1370xfs_file_readdir(
1371 struct file *file,
1372 struct dir_context *ctx)
1373{
1374 struct inode *inode = file_inode(file);
1375 xfs_inode_t *ip = XFS_I(inode);
1376 size_t bufsize;
1377
1378 /*
1379 * The Linux API doesn't pass down the total size of the buffer
1380 * we read into down to the filesystem. With the filldir concept
1381 * it's not needed for correct information, but the XFS dir2 leaf
1382 * code wants an estimate of the buffer size to calculate it's
1383 * readahead window and size the buffers used for mapping to
1384 * physical blocks.
1385 *
1386 * Try to give it an estimate that's good enough, maybe at some
1387 * point we can change the ->readdir prototype to include the
1388 * buffer size. For now we use the current glibc buffer size.
1389 */
1390 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1391
1392 return xfs_readdir(NULL, ip, ctx, bufsize);
1393}
1394
1395STATIC loff_t
1396xfs_file_llseek(
1397 struct file *file,
1398 loff_t offset,
1399 int whence)
1400{
1401 struct inode *inode = file->f_mapping->host;
1402
1403 if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1404 return -EIO;
1405
1406 switch (whence) {
1407 default:
1408 return generic_file_llseek(file, offset, whence);
1409 case SEEK_HOLE:
1410 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1411 break;
1412 case SEEK_DATA:
1413 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1414 break;
1415 }
1416
1417 if (offset < 0)
1418 return offset;
1419 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1420}
1421
1422static inline vm_fault_t
1423xfs_dax_fault_locked(
1424 struct vm_fault *vmf,
1425 unsigned int order,
1426 bool write_fault)
1427{
1428 vm_fault_t ret;
1429 pfn_t pfn;
1430
1431 if (!IS_ENABLED(CONFIG_FS_DAX)) {
1432 ASSERT(0);
1433 return VM_FAULT_SIGBUS;
1434 }
1435 ret = dax_iomap_fault(vmf, order, &pfn, NULL,
1436 (write_fault && !vmf->cow_page) ?
1437 &xfs_dax_write_iomap_ops :
1438 &xfs_read_iomap_ops);
1439 if (ret & VM_FAULT_NEEDDSYNC)
1440 ret = dax_finish_sync_fault(vmf, order, pfn);
1441 return ret;
1442}
1443
1444static vm_fault_t
1445xfs_dax_read_fault(
1446 struct vm_fault *vmf,
1447 unsigned int order)
1448{
1449 struct xfs_inode *ip = XFS_I(file_inode(vmf->vma->vm_file));
1450 vm_fault_t ret;
1451
1452 trace_xfs_read_fault(ip, order);
1453
1454 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1455 ret = xfs_dax_fault_locked(vmf, order, false);
1456 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1457
1458 return ret;
1459}
1460
1461/*
1462 * Locking for serialisation of IO during page faults. This results in a lock
1463 * ordering of:
1464 *
1465 * mmap_lock (MM)
1466 * sb_start_pagefault(vfs, freeze)
1467 * invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1468 * page_lock (MM)
1469 * i_lock (XFS - extent map serialisation)
1470 */
1471static vm_fault_t
1472xfs_write_fault(
1473 struct vm_fault *vmf,
1474 unsigned int order)
1475{
1476 struct inode *inode = file_inode(vmf->vma->vm_file);
1477 struct xfs_inode *ip = XFS_I(inode);
1478 unsigned int lock_mode = XFS_MMAPLOCK_SHARED;
1479 vm_fault_t ret;
1480
1481 trace_xfs_write_fault(ip, order);
1482
1483 sb_start_pagefault(inode->i_sb);
1484 file_update_time(vmf->vma->vm_file);
1485
1486 /*
1487 * Normally we only need the shared mmaplock, but if a reflink remap is
1488 * in progress we take the exclusive lock to wait for the remap to
1489 * finish before taking a write fault.
1490 */
1491 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1492 if (xfs_iflags_test(ip, XFS_IREMAPPING)) {
1493 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1494 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1495 lock_mode = XFS_MMAPLOCK_EXCL;
1496 }
1497
1498 if (IS_DAX(inode))
1499 ret = xfs_dax_fault_locked(vmf, order, true);
1500 else
1501 ret = iomap_page_mkwrite(vmf, &xfs_buffered_write_iomap_ops);
1502 xfs_iunlock(ip, lock_mode);
1503
1504 sb_end_pagefault(inode->i_sb);
1505 return ret;
1506}
1507
1508static inline bool
1509xfs_is_write_fault(
1510 struct vm_fault *vmf)
1511{
1512 return (vmf->flags & FAULT_FLAG_WRITE) &&
1513 (vmf->vma->vm_flags & VM_SHARED);
1514}
1515
1516static vm_fault_t
1517xfs_filemap_fault(
1518 struct vm_fault *vmf)
1519{
1520 struct inode *inode = file_inode(vmf->vma->vm_file);
1521
1522 /* DAX can shortcut the normal fault path on write faults! */
1523 if (IS_DAX(inode)) {
1524 if (xfs_is_write_fault(vmf))
1525 return xfs_write_fault(vmf, 0);
1526 return xfs_dax_read_fault(vmf, 0);
1527 }
1528
1529 trace_xfs_read_fault(XFS_I(inode), 0);
1530 return filemap_fault(vmf);
1531}
1532
1533static vm_fault_t
1534xfs_filemap_huge_fault(
1535 struct vm_fault *vmf,
1536 unsigned int order)
1537{
1538 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1539 return VM_FAULT_FALLBACK;
1540
1541 /* DAX can shortcut the normal fault path on write faults! */
1542 if (xfs_is_write_fault(vmf))
1543 return xfs_write_fault(vmf, order);
1544 return xfs_dax_read_fault(vmf, order);
1545}
1546
1547static vm_fault_t
1548xfs_filemap_page_mkwrite(
1549 struct vm_fault *vmf)
1550{
1551 return xfs_write_fault(vmf, 0);
1552}
1553
1554/*
1555 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1556 * on write faults. In reality, it needs to serialise against truncate and
1557 * prepare memory for writing so handle is as standard write fault.
1558 */
1559static vm_fault_t
1560xfs_filemap_pfn_mkwrite(
1561 struct vm_fault *vmf)
1562{
1563 return xfs_write_fault(vmf, 0);
1564}
1565
1566static const struct vm_operations_struct xfs_file_vm_ops = {
1567 .fault = xfs_filemap_fault,
1568 .huge_fault = xfs_filemap_huge_fault,
1569 .map_pages = filemap_map_pages,
1570 .page_mkwrite = xfs_filemap_page_mkwrite,
1571 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1572};
1573
1574STATIC int
1575xfs_file_mmap(
1576 struct file *file,
1577 struct vm_area_struct *vma)
1578{
1579 struct inode *inode = file_inode(file);
1580 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1581
1582 /*
1583 * We don't support synchronous mappings for non-DAX files and
1584 * for DAX files if underneath dax_device is not synchronous.
1585 */
1586 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1587 return -EOPNOTSUPP;
1588
1589 file_accessed(file);
1590 vma->vm_ops = &xfs_file_vm_ops;
1591 if (IS_DAX(inode))
1592 vm_flags_set(vma, VM_HUGEPAGE);
1593 return 0;
1594}
1595
1596const struct file_operations xfs_file_operations = {
1597 .llseek = xfs_file_llseek,
1598 .read_iter = xfs_file_read_iter,
1599 .write_iter = xfs_file_write_iter,
1600 .splice_read = xfs_file_splice_read,
1601 .splice_write = iter_file_splice_write,
1602 .iopoll = iocb_bio_iopoll,
1603 .unlocked_ioctl = xfs_file_ioctl,
1604#ifdef CONFIG_COMPAT
1605 .compat_ioctl = xfs_file_compat_ioctl,
1606#endif
1607 .mmap = xfs_file_mmap,
1608 .open = xfs_file_open,
1609 .release = xfs_file_release,
1610 .fsync = xfs_file_fsync,
1611 .get_unmapped_area = thp_get_unmapped_area,
1612 .fallocate = xfs_file_fallocate,
1613 .fadvise = xfs_file_fadvise,
1614 .remap_file_range = xfs_file_remap_range,
1615 .fop_flags = FOP_MMAP_SYNC | FOP_BUFFER_RASYNC |
1616 FOP_BUFFER_WASYNC | FOP_DIO_PARALLEL_WRITE,
1617};
1618
1619const struct file_operations xfs_dir_file_operations = {
1620 .open = xfs_dir_open,
1621 .read = generic_read_dir,
1622 .iterate_shared = xfs_file_readdir,
1623 .llseek = generic_file_llseek,
1624 .unlocked_ioctl = xfs_file_ioctl,
1625#ifdef CONFIG_COMPAT
1626 .compat_ioctl = xfs_file_compat_ioctl,
1627#endif
1628 .fsync = xfs_dir_fsync,
1629};