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_bit.h"
21#include "xfs_log.h"
22#include "xfs_inum.h"
23#include "xfs_sb.h"
24#include "xfs_ag.h"
25#include "xfs_trans.h"
26#include "xfs_mount.h"
27#include "xfs_bmap_btree.h"
28#include "xfs_alloc.h"
29#include "xfs_dinode.h"
30#include "xfs_inode.h"
31#include "xfs_inode_item.h"
32#include "xfs_bmap.h"
33#include "xfs_error.h"
34#include "xfs_vnodeops.h"
35#include "xfs_da_btree.h"
36#include "xfs_ioctl.h"
37#include "xfs_trace.h"
38
39#include <linux/dcache.h>
40#include <linux/falloc.h>
41
42static const struct vm_operations_struct xfs_file_vm_ops;
43
44/*
45 * Locking primitives for read and write IO paths to ensure we consistently use
46 * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
47 */
48static inline void
49xfs_rw_ilock(
50 struct xfs_inode *ip,
51 int type)
52{
53 if (type & XFS_IOLOCK_EXCL)
54 mutex_lock(&VFS_I(ip)->i_mutex);
55 xfs_ilock(ip, type);
56}
57
58static inline void
59xfs_rw_iunlock(
60 struct xfs_inode *ip,
61 int type)
62{
63 xfs_iunlock(ip, type);
64 if (type & XFS_IOLOCK_EXCL)
65 mutex_unlock(&VFS_I(ip)->i_mutex);
66}
67
68static inline void
69xfs_rw_ilock_demote(
70 struct xfs_inode *ip,
71 int type)
72{
73 xfs_ilock_demote(ip, type);
74 if (type & XFS_IOLOCK_EXCL)
75 mutex_unlock(&VFS_I(ip)->i_mutex);
76}
77
78/*
79 * xfs_iozero
80 *
81 * xfs_iozero clears the specified range of buffer supplied,
82 * and marks all the affected blocks as valid and modified. If
83 * an affected block is not allocated, it will be allocated. If
84 * an affected block is not completely overwritten, and is not
85 * valid before the operation, it will be read from disk before
86 * being partially zeroed.
87 */
88STATIC int
89xfs_iozero(
90 struct xfs_inode *ip, /* inode */
91 loff_t pos, /* offset in file */
92 size_t count) /* size of data to zero */
93{
94 struct page *page;
95 struct address_space *mapping;
96 int status;
97
98 mapping = VFS_I(ip)->i_mapping;
99 do {
100 unsigned offset, bytes;
101 void *fsdata;
102
103 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
104 bytes = PAGE_CACHE_SIZE - offset;
105 if (bytes > count)
106 bytes = count;
107
108 status = pagecache_write_begin(NULL, mapping, pos, bytes,
109 AOP_FLAG_UNINTERRUPTIBLE,
110 &page, &fsdata);
111 if (status)
112 break;
113
114 zero_user(page, offset, bytes);
115
116 status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
117 page, fsdata);
118 WARN_ON(status <= 0); /* can't return less than zero! */
119 pos += bytes;
120 count -= bytes;
121 status = 0;
122 } while (count);
123
124 return (-status);
125}
126
127STATIC int
128xfs_file_fsync(
129 struct file *file,
130 loff_t start,
131 loff_t end,
132 int datasync)
133{
134 struct inode *inode = file->f_mapping->host;
135 struct xfs_inode *ip = XFS_I(inode);
136 struct xfs_mount *mp = ip->i_mount;
137 struct xfs_trans *tp;
138 int error = 0;
139 int log_flushed = 0;
140
141 trace_xfs_file_fsync(ip);
142
143 error = filemap_write_and_wait_range(inode->i_mapping, start, end);
144 if (error)
145 return error;
146
147 if (XFS_FORCED_SHUTDOWN(mp))
148 return -XFS_ERROR(EIO);
149
150 xfs_iflags_clear(ip, XFS_ITRUNCATED);
151
152 xfs_ilock(ip, XFS_IOLOCK_SHARED);
153 xfs_ioend_wait(ip);
154 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
155
156 if (mp->m_flags & XFS_MOUNT_BARRIER) {
157 /*
158 * If we have an RT and/or log subvolume we need to make sure
159 * to flush the write cache the device used for file data
160 * first. This is to ensure newly written file data make
161 * it to disk before logging the new inode size in case of
162 * an extending write.
163 */
164 if (XFS_IS_REALTIME_INODE(ip))
165 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
166 else if (mp->m_logdev_targp != mp->m_ddev_targp)
167 xfs_blkdev_issue_flush(mp->m_ddev_targp);
168 }
169
170 /*
171 * We always need to make sure that the required inode state is safe on
172 * disk. The inode might be clean but we still might need to force the
173 * log because of committed transactions that haven't hit the disk yet.
174 * Likewise, there could be unflushed non-transactional changes to the
175 * inode core that have to go to disk and this requires us to issue
176 * a synchronous transaction to capture these changes correctly.
177 *
178 * This code relies on the assumption that if the i_update_core field
179 * of the inode is clear and the inode is unpinned then it is clean
180 * and no action is required.
181 */
182 xfs_ilock(ip, XFS_ILOCK_SHARED);
183
184 /*
185 * First check if the VFS inode is marked dirty. All the dirtying
186 * of non-transactional updates no goes through mark_inode_dirty*,
187 * which allows us to distinguish beteeen pure timestamp updates
188 * and i_size updates which need to be caught for fdatasync.
189 * After that also theck for the dirty state in the XFS inode, which
190 * might gets cleared when the inode gets written out via the AIL
191 * or xfs_iflush_cluster.
192 */
193 if (((inode->i_state & I_DIRTY_DATASYNC) ||
194 ((inode->i_state & I_DIRTY_SYNC) && !datasync)) &&
195 ip->i_update_core) {
196 /*
197 * Kick off a transaction to log the inode core to get the
198 * updates. The sync transaction will also force the log.
199 */
200 xfs_iunlock(ip, XFS_ILOCK_SHARED);
201 tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS);
202 error = xfs_trans_reserve(tp, 0,
203 XFS_FSYNC_TS_LOG_RES(mp), 0, 0, 0);
204 if (error) {
205 xfs_trans_cancel(tp, 0);
206 return -error;
207 }
208 xfs_ilock(ip, XFS_ILOCK_EXCL);
209
210 /*
211 * Note - it's possible that we might have pushed ourselves out
212 * of the way during trans_reserve which would flush the inode.
213 * But there's no guarantee that the inode buffer has actually
214 * gone out yet (it's delwri). Plus the buffer could be pinned
215 * anyway if it's part of an inode in another recent
216 * transaction. So we play it safe and fire off the
217 * transaction anyway.
218 */
219 xfs_trans_ijoin(tp, ip);
220 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
221 xfs_trans_set_sync(tp);
222 error = _xfs_trans_commit(tp, 0, &log_flushed);
223
224 xfs_iunlock(ip, XFS_ILOCK_EXCL);
225 } else {
226 /*
227 * Timestamps/size haven't changed since last inode flush or
228 * inode transaction commit. That means either nothing got
229 * written or a transaction committed which caught the updates.
230 * If the latter happened and the transaction hasn't hit the
231 * disk yet, the inode will be still be pinned. If it is,
232 * force the log.
233 */
234 if (xfs_ipincount(ip)) {
235 error = _xfs_log_force_lsn(mp,
236 ip->i_itemp->ili_last_lsn,
237 XFS_LOG_SYNC, &log_flushed);
238 }
239 xfs_iunlock(ip, XFS_ILOCK_SHARED);
240 }
241
242 /*
243 * If we only have a single device, and the log force about was
244 * a no-op we might have to flush the data device cache here.
245 * This can only happen for fdatasync/O_DSYNC if we were overwriting
246 * an already allocated file and thus do not have any metadata to
247 * commit.
248 */
249 if ((mp->m_flags & XFS_MOUNT_BARRIER) &&
250 mp->m_logdev_targp == mp->m_ddev_targp &&
251 !XFS_IS_REALTIME_INODE(ip) &&
252 !log_flushed)
253 xfs_blkdev_issue_flush(mp->m_ddev_targp);
254
255 return -error;
256}
257
258STATIC ssize_t
259xfs_file_aio_read(
260 struct kiocb *iocb,
261 const struct iovec *iovp,
262 unsigned long nr_segs,
263 loff_t pos)
264{
265 struct file *file = iocb->ki_filp;
266 struct inode *inode = file->f_mapping->host;
267 struct xfs_inode *ip = XFS_I(inode);
268 struct xfs_mount *mp = ip->i_mount;
269 size_t size = 0;
270 ssize_t ret = 0;
271 int ioflags = 0;
272 xfs_fsize_t n;
273 unsigned long seg;
274
275 XFS_STATS_INC(xs_read_calls);
276
277 BUG_ON(iocb->ki_pos != pos);
278
279 if (unlikely(file->f_flags & O_DIRECT))
280 ioflags |= IO_ISDIRECT;
281 if (file->f_mode & FMODE_NOCMTIME)
282 ioflags |= IO_INVIS;
283
284 /* START copy & waste from filemap.c */
285 for (seg = 0; seg < nr_segs; seg++) {
286 const struct iovec *iv = &iovp[seg];
287
288 /*
289 * If any segment has a negative length, or the cumulative
290 * length ever wraps negative then return -EINVAL.
291 */
292 size += iv->iov_len;
293 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
294 return XFS_ERROR(-EINVAL);
295 }
296 /* END copy & waste from filemap.c */
297
298 if (unlikely(ioflags & IO_ISDIRECT)) {
299 xfs_buftarg_t *target =
300 XFS_IS_REALTIME_INODE(ip) ?
301 mp->m_rtdev_targp : mp->m_ddev_targp;
302 if ((iocb->ki_pos & target->bt_smask) ||
303 (size & target->bt_smask)) {
304 if (iocb->ki_pos == ip->i_size)
305 return 0;
306 return -XFS_ERROR(EINVAL);
307 }
308 }
309
310 n = XFS_MAXIOFFSET(mp) - iocb->ki_pos;
311 if (n <= 0 || size == 0)
312 return 0;
313
314 if (n < size)
315 size = n;
316
317 if (XFS_FORCED_SHUTDOWN(mp))
318 return -EIO;
319
320 if (unlikely(ioflags & IO_ISDIRECT)) {
321 xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
322
323 if (inode->i_mapping->nrpages) {
324 ret = -xfs_flushinval_pages(ip,
325 (iocb->ki_pos & PAGE_CACHE_MASK),
326 -1, FI_REMAPF_LOCKED);
327 if (ret) {
328 xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
329 return ret;
330 }
331 }
332 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
333 } else
334 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
335
336 trace_xfs_file_read(ip, size, iocb->ki_pos, ioflags);
337
338 ret = generic_file_aio_read(iocb, iovp, nr_segs, iocb->ki_pos);
339 if (ret > 0)
340 XFS_STATS_ADD(xs_read_bytes, ret);
341
342 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
343 return ret;
344}
345
346STATIC ssize_t
347xfs_file_splice_read(
348 struct file *infilp,
349 loff_t *ppos,
350 struct pipe_inode_info *pipe,
351 size_t count,
352 unsigned int flags)
353{
354 struct xfs_inode *ip = XFS_I(infilp->f_mapping->host);
355 int ioflags = 0;
356 ssize_t ret;
357
358 XFS_STATS_INC(xs_read_calls);
359
360 if (infilp->f_mode & FMODE_NOCMTIME)
361 ioflags |= IO_INVIS;
362
363 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
364 return -EIO;
365
366 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
367
368 trace_xfs_file_splice_read(ip, count, *ppos, ioflags);
369
370 ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
371 if (ret > 0)
372 XFS_STATS_ADD(xs_read_bytes, ret);
373
374 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
375 return ret;
376}
377
378STATIC void
379xfs_aio_write_isize_update(
380 struct inode *inode,
381 loff_t *ppos,
382 ssize_t bytes_written)
383{
384 struct xfs_inode *ip = XFS_I(inode);
385 xfs_fsize_t isize = i_size_read(inode);
386
387 if (bytes_written > 0)
388 XFS_STATS_ADD(xs_write_bytes, bytes_written);
389
390 if (unlikely(bytes_written < 0 && bytes_written != -EFAULT &&
391 *ppos > isize))
392 *ppos = isize;
393
394 if (*ppos > ip->i_size) {
395 xfs_rw_ilock(ip, XFS_ILOCK_EXCL);
396 if (*ppos > ip->i_size)
397 ip->i_size = *ppos;
398 xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
399 }
400}
401
402/*
403 * If this was a direct or synchronous I/O that failed (such as ENOSPC) then
404 * part of the I/O may have been written to disk before the error occurred. In
405 * this case the on-disk file size may have been adjusted beyond the in-memory
406 * file size and now needs to be truncated back.
407 */
408STATIC void
409xfs_aio_write_newsize_update(
410 struct xfs_inode *ip)
411{
412 if (ip->i_new_size) {
413 xfs_rw_ilock(ip, XFS_ILOCK_EXCL);
414 ip->i_new_size = 0;
415 if (ip->i_d.di_size > ip->i_size)
416 ip->i_d.di_size = ip->i_size;
417 xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
418 }
419}
420
421/*
422 * xfs_file_splice_write() does not use xfs_rw_ilock() because
423 * generic_file_splice_write() takes the i_mutex itself. This, in theory,
424 * couuld cause lock inversions between the aio_write path and the splice path
425 * if someone is doing concurrent splice(2) based writes and write(2) based
426 * writes to the same inode. The only real way to fix this is to re-implement
427 * the generic code here with correct locking orders.
428 */
429STATIC ssize_t
430xfs_file_splice_write(
431 struct pipe_inode_info *pipe,
432 struct file *outfilp,
433 loff_t *ppos,
434 size_t count,
435 unsigned int flags)
436{
437 struct inode *inode = outfilp->f_mapping->host;
438 struct xfs_inode *ip = XFS_I(inode);
439 xfs_fsize_t new_size;
440 int ioflags = 0;
441 ssize_t ret;
442
443 XFS_STATS_INC(xs_write_calls);
444
445 if (outfilp->f_mode & FMODE_NOCMTIME)
446 ioflags |= IO_INVIS;
447
448 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
449 return -EIO;
450
451 xfs_ilock(ip, XFS_IOLOCK_EXCL);
452
453 new_size = *ppos + count;
454
455 xfs_ilock(ip, XFS_ILOCK_EXCL);
456 if (new_size > ip->i_size)
457 ip->i_new_size = new_size;
458 xfs_iunlock(ip, XFS_ILOCK_EXCL);
459
460 trace_xfs_file_splice_write(ip, count, *ppos, ioflags);
461
462 ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
463
464 xfs_aio_write_isize_update(inode, ppos, ret);
465 xfs_aio_write_newsize_update(ip);
466 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
467 return ret;
468}
469
470/*
471 * This routine is called to handle zeroing any space in the last
472 * block of the file that is beyond the EOF. We do this since the
473 * size is being increased without writing anything to that block
474 * and we don't want anyone to read the garbage on the disk.
475 */
476STATIC int /* error (positive) */
477xfs_zero_last_block(
478 xfs_inode_t *ip,
479 xfs_fsize_t offset,
480 xfs_fsize_t isize)
481{
482 xfs_fileoff_t last_fsb;
483 xfs_mount_t *mp = ip->i_mount;
484 int nimaps;
485 int zero_offset;
486 int zero_len;
487 int error = 0;
488 xfs_bmbt_irec_t imap;
489
490 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
491
492 zero_offset = XFS_B_FSB_OFFSET(mp, isize);
493 if (zero_offset == 0) {
494 /*
495 * There are no extra bytes in the last block on disk to
496 * zero, so return.
497 */
498 return 0;
499 }
500
501 last_fsb = XFS_B_TO_FSBT(mp, isize);
502 nimaps = 1;
503 error = xfs_bmapi(NULL, ip, last_fsb, 1, 0, NULL, 0, &imap,
504 &nimaps, NULL);
505 if (error) {
506 return error;
507 }
508 ASSERT(nimaps > 0);
509 /*
510 * If the block underlying isize is just a hole, then there
511 * is nothing to zero.
512 */
513 if (imap.br_startblock == HOLESTARTBLOCK) {
514 return 0;
515 }
516 /*
517 * Zero the part of the last block beyond the EOF, and write it
518 * out sync. We need to drop the ilock while we do this so we
519 * don't deadlock when the buffer cache calls back to us.
520 */
521 xfs_iunlock(ip, XFS_ILOCK_EXCL);
522
523 zero_len = mp->m_sb.sb_blocksize - zero_offset;
524 if (isize + zero_len > offset)
525 zero_len = offset - isize;
526 error = xfs_iozero(ip, isize, zero_len);
527
528 xfs_ilock(ip, XFS_ILOCK_EXCL);
529 ASSERT(error >= 0);
530 return error;
531}
532
533/*
534 * Zero any on disk space between the current EOF and the new,
535 * larger EOF. This handles the normal case of zeroing the remainder
536 * of the last block in the file and the unusual case of zeroing blocks
537 * out beyond the size of the file. This second case only happens
538 * with fixed size extents and when the system crashes before the inode
539 * size was updated but after blocks were allocated. If fill is set,
540 * then any holes in the range are filled and zeroed. If not, the holes
541 * are left alone as holes.
542 */
543
544int /* error (positive) */
545xfs_zero_eof(
546 xfs_inode_t *ip,
547 xfs_off_t offset, /* starting I/O offset */
548 xfs_fsize_t isize) /* current inode size */
549{
550 xfs_mount_t *mp = ip->i_mount;
551 xfs_fileoff_t start_zero_fsb;
552 xfs_fileoff_t end_zero_fsb;
553 xfs_fileoff_t zero_count_fsb;
554 xfs_fileoff_t last_fsb;
555 xfs_fileoff_t zero_off;
556 xfs_fsize_t zero_len;
557 int nimaps;
558 int error = 0;
559 xfs_bmbt_irec_t imap;
560
561 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
562 ASSERT(offset > isize);
563
564 /*
565 * First handle zeroing the block on which isize resides.
566 * We only zero a part of that block so it is handled specially.
567 */
568 error = xfs_zero_last_block(ip, offset, isize);
569 if (error) {
570 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
571 return error;
572 }
573
574 /*
575 * Calculate the range between the new size and the old
576 * where blocks needing to be zeroed may exist. To get the
577 * block where the last byte in the file currently resides,
578 * we need to subtract one from the size and truncate back
579 * to a block boundary. We subtract 1 in case the size is
580 * exactly on a block boundary.
581 */
582 last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
583 start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
584 end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
585 ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
586 if (last_fsb == end_zero_fsb) {
587 /*
588 * The size was only incremented on its last block.
589 * We took care of that above, so just return.
590 */
591 return 0;
592 }
593
594 ASSERT(start_zero_fsb <= end_zero_fsb);
595 while (start_zero_fsb <= end_zero_fsb) {
596 nimaps = 1;
597 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
598 error = xfs_bmapi(NULL, ip, start_zero_fsb, zero_count_fsb,
599 0, NULL, 0, &imap, &nimaps, NULL);
600 if (error) {
601 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
602 return error;
603 }
604 ASSERT(nimaps > 0);
605
606 if (imap.br_state == XFS_EXT_UNWRITTEN ||
607 imap.br_startblock == HOLESTARTBLOCK) {
608 /*
609 * This loop handles initializing pages that were
610 * partially initialized by the code below this
611 * loop. It basically zeroes the part of the page
612 * that sits on a hole and sets the page as P_HOLE
613 * and calls remapf if it is a mapped file.
614 */
615 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
616 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
617 continue;
618 }
619
620 /*
621 * There are blocks we need to zero.
622 * Drop the inode lock while we're doing the I/O.
623 * We'll still have the iolock to protect us.
624 */
625 xfs_iunlock(ip, XFS_ILOCK_EXCL);
626
627 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
628 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
629
630 if ((zero_off + zero_len) > offset)
631 zero_len = offset - zero_off;
632
633 error = xfs_iozero(ip, zero_off, zero_len);
634 if (error) {
635 goto out_lock;
636 }
637
638 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
639 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
640
641 xfs_ilock(ip, XFS_ILOCK_EXCL);
642 }
643
644 return 0;
645
646out_lock:
647 xfs_ilock(ip, XFS_ILOCK_EXCL);
648 ASSERT(error >= 0);
649 return error;
650}
651
652/*
653 * Common pre-write limit and setup checks.
654 *
655 * Returns with iolock held according to @iolock.
656 */
657STATIC ssize_t
658xfs_file_aio_write_checks(
659 struct file *file,
660 loff_t *pos,
661 size_t *count,
662 int *iolock)
663{
664 struct inode *inode = file->f_mapping->host;
665 struct xfs_inode *ip = XFS_I(inode);
666 xfs_fsize_t new_size;
667 int error = 0;
668
669 error = generic_write_checks(file, pos, count, S_ISBLK(inode->i_mode));
670 if (error) {
671 xfs_rw_iunlock(ip, XFS_ILOCK_EXCL | *iolock);
672 *iolock = 0;
673 return error;
674 }
675
676 new_size = *pos + *count;
677 if (new_size > ip->i_size)
678 ip->i_new_size = new_size;
679
680 if (likely(!(file->f_mode & FMODE_NOCMTIME)))
681 file_update_time(file);
682
683 /*
684 * If the offset is beyond the size of the file, we need to zero any
685 * blocks that fall between the existing EOF and the start of this
686 * write.
687 */
688 if (*pos > ip->i_size)
689 error = -xfs_zero_eof(ip, *pos, ip->i_size);
690
691 xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
692 if (error)
693 return error;
694
695 /*
696 * If we're writing the file then make sure to clear the setuid and
697 * setgid bits if the process is not being run by root. This keeps
698 * people from modifying setuid and setgid binaries.
699 */
700 return file_remove_suid(file);
701
702}
703
704/*
705 * xfs_file_dio_aio_write - handle direct IO writes
706 *
707 * Lock the inode appropriately to prepare for and issue a direct IO write.
708 * By separating it from the buffered write path we remove all the tricky to
709 * follow locking changes and looping.
710 *
711 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
712 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
713 * pages are flushed out.
714 *
715 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
716 * allowing them to be done in parallel with reads and other direct IO writes.
717 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
718 * needs to do sub-block zeroing and that requires serialisation against other
719 * direct IOs to the same block. In this case we need to serialise the
720 * submission of the unaligned IOs so that we don't get racing block zeroing in
721 * the dio layer. To avoid the problem with aio, we also need to wait for
722 * outstanding IOs to complete so that unwritten extent conversion is completed
723 * before we try to map the overlapping block. This is currently implemented by
724 * hitting it with a big hammer (i.e. xfs_ioend_wait()).
725 *
726 * Returns with locks held indicated by @iolock and errors indicated by
727 * negative return values.
728 */
729STATIC ssize_t
730xfs_file_dio_aio_write(
731 struct kiocb *iocb,
732 const struct iovec *iovp,
733 unsigned long nr_segs,
734 loff_t pos,
735 size_t ocount,
736 int *iolock)
737{
738 struct file *file = iocb->ki_filp;
739 struct address_space *mapping = file->f_mapping;
740 struct inode *inode = mapping->host;
741 struct xfs_inode *ip = XFS_I(inode);
742 struct xfs_mount *mp = ip->i_mount;
743 ssize_t ret = 0;
744 size_t count = ocount;
745 int unaligned_io = 0;
746 struct xfs_buftarg *target = XFS_IS_REALTIME_INODE(ip) ?
747 mp->m_rtdev_targp : mp->m_ddev_targp;
748
749 *iolock = 0;
750 if ((pos & target->bt_smask) || (count & target->bt_smask))
751 return -XFS_ERROR(EINVAL);
752
753 if ((pos & mp->m_blockmask) || ((pos + count) & mp->m_blockmask))
754 unaligned_io = 1;
755
756 if (unaligned_io || mapping->nrpages || pos > ip->i_size)
757 *iolock = XFS_IOLOCK_EXCL;
758 else
759 *iolock = XFS_IOLOCK_SHARED;
760 xfs_rw_ilock(ip, XFS_ILOCK_EXCL | *iolock);
761
762 ret = xfs_file_aio_write_checks(file, &pos, &count, iolock);
763 if (ret)
764 return ret;
765
766 if (mapping->nrpages) {
767 WARN_ON(*iolock != XFS_IOLOCK_EXCL);
768 ret = -xfs_flushinval_pages(ip, (pos & PAGE_CACHE_MASK), -1,
769 FI_REMAPF_LOCKED);
770 if (ret)
771 return ret;
772 }
773
774 /*
775 * If we are doing unaligned IO, wait for all other IO to drain,
776 * otherwise demote the lock if we had to flush cached pages
777 */
778 if (unaligned_io)
779 xfs_ioend_wait(ip);
780 else if (*iolock == XFS_IOLOCK_EXCL) {
781 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
782 *iolock = XFS_IOLOCK_SHARED;
783 }
784
785 trace_xfs_file_direct_write(ip, count, iocb->ki_pos, 0);
786 ret = generic_file_direct_write(iocb, iovp,
787 &nr_segs, pos, &iocb->ki_pos, count, ocount);
788
789 /* No fallback to buffered IO on errors for XFS. */
790 ASSERT(ret < 0 || ret == count);
791 return ret;
792}
793
794STATIC ssize_t
795xfs_file_buffered_aio_write(
796 struct kiocb *iocb,
797 const struct iovec *iovp,
798 unsigned long nr_segs,
799 loff_t pos,
800 size_t ocount,
801 int *iolock)
802{
803 struct file *file = iocb->ki_filp;
804 struct address_space *mapping = file->f_mapping;
805 struct inode *inode = mapping->host;
806 struct xfs_inode *ip = XFS_I(inode);
807 ssize_t ret;
808 int enospc = 0;
809 size_t count = ocount;
810
811 *iolock = XFS_IOLOCK_EXCL;
812 xfs_rw_ilock(ip, XFS_ILOCK_EXCL | *iolock);
813
814 ret = xfs_file_aio_write_checks(file, &pos, &count, iolock);
815 if (ret)
816 return ret;
817
818 /* We can write back this queue in page reclaim */
819 current->backing_dev_info = mapping->backing_dev_info;
820
821write_retry:
822 trace_xfs_file_buffered_write(ip, count, iocb->ki_pos, 0);
823 ret = generic_file_buffered_write(iocb, iovp, nr_segs,
824 pos, &iocb->ki_pos, count, ret);
825 /*
826 * if we just got an ENOSPC, flush the inode now we aren't holding any
827 * page locks and retry *once*
828 */
829 if (ret == -ENOSPC && !enospc) {
830 ret = -xfs_flush_pages(ip, 0, -1, 0, FI_NONE);
831 if (ret)
832 return ret;
833 enospc = 1;
834 goto write_retry;
835 }
836 current->backing_dev_info = NULL;
837 return ret;
838}
839
840STATIC ssize_t
841xfs_file_aio_write(
842 struct kiocb *iocb,
843 const struct iovec *iovp,
844 unsigned long nr_segs,
845 loff_t pos)
846{
847 struct file *file = iocb->ki_filp;
848 struct address_space *mapping = file->f_mapping;
849 struct inode *inode = mapping->host;
850 struct xfs_inode *ip = XFS_I(inode);
851 ssize_t ret;
852 int iolock;
853 size_t ocount = 0;
854
855 XFS_STATS_INC(xs_write_calls);
856
857 BUG_ON(iocb->ki_pos != pos);
858
859 ret = generic_segment_checks(iovp, &nr_segs, &ocount, VERIFY_READ);
860 if (ret)
861 return ret;
862
863 if (ocount == 0)
864 return 0;
865
866 xfs_wait_for_freeze(ip->i_mount, SB_FREEZE_WRITE);
867
868 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
869 return -EIO;
870
871 if (unlikely(file->f_flags & O_DIRECT))
872 ret = xfs_file_dio_aio_write(iocb, iovp, nr_segs, pos,
873 ocount, &iolock);
874 else
875 ret = xfs_file_buffered_aio_write(iocb, iovp, nr_segs, pos,
876 ocount, &iolock);
877
878 xfs_aio_write_isize_update(inode, &iocb->ki_pos, ret);
879
880 if (ret <= 0)
881 goto out_unlock;
882
883 /* Handle various SYNC-type writes */
884 if ((file->f_flags & O_DSYNC) || IS_SYNC(inode)) {
885 loff_t end = pos + ret - 1;
886 int error;
887
888 xfs_rw_iunlock(ip, iolock);
889 error = xfs_file_fsync(file, pos, end,
890 (file->f_flags & __O_SYNC) ? 0 : 1);
891 xfs_rw_ilock(ip, iolock);
892 if (error)
893 ret = error;
894 }
895
896out_unlock:
897 xfs_aio_write_newsize_update(ip);
898 xfs_rw_iunlock(ip, iolock);
899 return ret;
900}
901
902STATIC long
903xfs_file_fallocate(
904 struct file *file,
905 int mode,
906 loff_t offset,
907 loff_t len)
908{
909 struct inode *inode = file->f_path.dentry->d_inode;
910 long error;
911 loff_t new_size = 0;
912 xfs_flock64_t bf;
913 xfs_inode_t *ip = XFS_I(inode);
914 int cmd = XFS_IOC_RESVSP;
915 int attr_flags = XFS_ATTR_NOLOCK;
916
917 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
918 return -EOPNOTSUPP;
919
920 bf.l_whence = 0;
921 bf.l_start = offset;
922 bf.l_len = len;
923
924 xfs_ilock(ip, XFS_IOLOCK_EXCL);
925
926 if (mode & FALLOC_FL_PUNCH_HOLE)
927 cmd = XFS_IOC_UNRESVSP;
928
929 /* check the new inode size is valid before allocating */
930 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
931 offset + len > i_size_read(inode)) {
932 new_size = offset + len;
933 error = inode_newsize_ok(inode, new_size);
934 if (error)
935 goto out_unlock;
936 }
937
938 if (file->f_flags & O_DSYNC)
939 attr_flags |= XFS_ATTR_SYNC;
940
941 error = -xfs_change_file_space(ip, cmd, &bf, 0, attr_flags);
942 if (error)
943 goto out_unlock;
944
945 /* Change file size if needed */
946 if (new_size) {
947 struct iattr iattr;
948
949 iattr.ia_valid = ATTR_SIZE;
950 iattr.ia_size = new_size;
951 error = -xfs_setattr_size(ip, &iattr, XFS_ATTR_NOLOCK);
952 }
953
954out_unlock:
955 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
956 return error;
957}
958
959
960STATIC int
961xfs_file_open(
962 struct inode *inode,
963 struct file *file)
964{
965 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
966 return -EFBIG;
967 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
968 return -EIO;
969 return 0;
970}
971
972STATIC int
973xfs_dir_open(
974 struct inode *inode,
975 struct file *file)
976{
977 struct xfs_inode *ip = XFS_I(inode);
978 int mode;
979 int error;
980
981 error = xfs_file_open(inode, file);
982 if (error)
983 return error;
984
985 /*
986 * If there are any blocks, read-ahead block 0 as we're almost
987 * certain to have the next operation be a read there.
988 */
989 mode = xfs_ilock_map_shared(ip);
990 if (ip->i_d.di_nextents > 0)
991 xfs_da_reada_buf(NULL, ip, 0, XFS_DATA_FORK);
992 xfs_iunlock(ip, mode);
993 return 0;
994}
995
996STATIC int
997xfs_file_release(
998 struct inode *inode,
999 struct file *filp)
1000{
1001 return -xfs_release(XFS_I(inode));
1002}
1003
1004STATIC int
1005xfs_file_readdir(
1006 struct file *filp,
1007 void *dirent,
1008 filldir_t filldir)
1009{
1010 struct inode *inode = filp->f_path.dentry->d_inode;
1011 xfs_inode_t *ip = XFS_I(inode);
1012 int error;
1013 size_t bufsize;
1014
1015 /*
1016 * The Linux API doesn't pass down the total size of the buffer
1017 * we read into down to the filesystem. With the filldir concept
1018 * it's not needed for correct information, but the XFS dir2 leaf
1019 * code wants an estimate of the buffer size to calculate it's
1020 * readahead window and size the buffers used for mapping to
1021 * physical blocks.
1022 *
1023 * Try to give it an estimate that's good enough, maybe at some
1024 * point we can change the ->readdir prototype to include the
1025 * buffer size. For now we use the current glibc buffer size.
1026 */
1027 bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
1028
1029 error = xfs_readdir(ip, dirent, bufsize,
1030 (xfs_off_t *)&filp->f_pos, filldir);
1031 if (error)
1032 return -error;
1033 return 0;
1034}
1035
1036STATIC int
1037xfs_file_mmap(
1038 struct file *filp,
1039 struct vm_area_struct *vma)
1040{
1041 vma->vm_ops = &xfs_file_vm_ops;
1042 vma->vm_flags |= VM_CAN_NONLINEAR;
1043
1044 file_accessed(filp);
1045 return 0;
1046}
1047
1048/*
1049 * mmap()d file has taken write protection fault and is being made
1050 * writable. We can set the page state up correctly for a writable
1051 * page, which means we can do correct delalloc accounting (ENOSPC
1052 * checking!) and unwritten extent mapping.
1053 */
1054STATIC int
1055xfs_vm_page_mkwrite(
1056 struct vm_area_struct *vma,
1057 struct vm_fault *vmf)
1058{
1059 return block_page_mkwrite(vma, vmf, xfs_get_blocks);
1060}
1061
1062const struct file_operations xfs_file_operations = {
1063 .llseek = generic_file_llseek,
1064 .read = do_sync_read,
1065 .write = do_sync_write,
1066 .aio_read = xfs_file_aio_read,
1067 .aio_write = xfs_file_aio_write,
1068 .splice_read = xfs_file_splice_read,
1069 .splice_write = xfs_file_splice_write,
1070 .unlocked_ioctl = xfs_file_ioctl,
1071#ifdef CONFIG_COMPAT
1072 .compat_ioctl = xfs_file_compat_ioctl,
1073#endif
1074 .mmap = xfs_file_mmap,
1075 .open = xfs_file_open,
1076 .release = xfs_file_release,
1077 .fsync = xfs_file_fsync,
1078 .fallocate = xfs_file_fallocate,
1079};
1080
1081const struct file_operations xfs_dir_file_operations = {
1082 .open = xfs_dir_open,
1083 .read = generic_read_dir,
1084 .readdir = xfs_file_readdir,
1085 .llseek = generic_file_llseek,
1086 .unlocked_ioctl = xfs_file_ioctl,
1087#ifdef CONFIG_COMPAT
1088 .compat_ioctl = xfs_file_compat_ioctl,
1089#endif
1090 .fsync = xfs_file_fsync,
1091};
1092
1093static const struct vm_operations_struct xfs_file_vm_ops = {
1094 .fault = filemap_fault,
1095 .page_mkwrite = xfs_vm_page_mkwrite,
1096};
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_inode.h"
14#include "xfs_trans.h"
15#include "xfs_inode_item.h"
16#include "xfs_bmap.h"
17#include "xfs_bmap_util.h"
18#include "xfs_dir2.h"
19#include "xfs_dir2_priv.h"
20#include "xfs_ioctl.h"
21#include "xfs_trace.h"
22#include "xfs_log.h"
23#include "xfs_icache.h"
24#include "xfs_pnfs.h"
25#include "xfs_iomap.h"
26#include "xfs_reflink.h"
27
28#include <linux/falloc.h>
29#include <linux/backing-dev.h>
30#include <linux/mman.h>
31#include <linux/fadvise.h>
32
33static const struct vm_operations_struct xfs_file_vm_ops;
34
35int
36xfs_update_prealloc_flags(
37 struct xfs_inode *ip,
38 enum xfs_prealloc_flags flags)
39{
40 struct xfs_trans *tp;
41 int error;
42
43 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_writeid,
44 0, 0, 0, &tp);
45 if (error)
46 return error;
47
48 xfs_ilock(ip, XFS_ILOCK_EXCL);
49 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
50
51 if (!(flags & XFS_PREALLOC_INVISIBLE)) {
52 VFS_I(ip)->i_mode &= ~S_ISUID;
53 if (VFS_I(ip)->i_mode & S_IXGRP)
54 VFS_I(ip)->i_mode &= ~S_ISGID;
55 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
56 }
57
58 if (flags & XFS_PREALLOC_SET)
59 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
60 if (flags & XFS_PREALLOC_CLEAR)
61 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
62
63 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
64 if (flags & XFS_PREALLOC_SYNC)
65 xfs_trans_set_sync(tp);
66 return xfs_trans_commit(tp);
67}
68
69/*
70 * Fsync operations on directories are much simpler than on regular files,
71 * as there is no file data to flush, and thus also no need for explicit
72 * cache flush operations, and there are no non-transaction metadata updates
73 * on directories either.
74 */
75STATIC int
76xfs_dir_fsync(
77 struct file *file,
78 loff_t start,
79 loff_t end,
80 int datasync)
81{
82 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
83
84 trace_xfs_dir_fsync(ip);
85 return xfs_log_force_inode(ip);
86}
87
88STATIC int
89xfs_file_fsync(
90 struct file *file,
91 loff_t start,
92 loff_t end,
93 int datasync)
94{
95 struct inode *inode = file->f_mapping->host;
96 struct xfs_inode *ip = XFS_I(inode);
97 struct xfs_inode_log_item *iip = ip->i_itemp;
98 struct xfs_mount *mp = ip->i_mount;
99 int error = 0;
100 int log_flushed = 0;
101 xfs_lsn_t lsn = 0;
102
103 trace_xfs_file_fsync(ip);
104
105 error = file_write_and_wait_range(file, start, end);
106 if (error)
107 return error;
108
109 if (XFS_FORCED_SHUTDOWN(mp))
110 return -EIO;
111
112 xfs_iflags_clear(ip, XFS_ITRUNCATED);
113
114 /*
115 * If we have an RT and/or log subvolume we need to make sure to flush
116 * the write cache the device used for file data first. This is to
117 * ensure newly written file data make it to disk before logging the new
118 * inode size in case of an extending write.
119 */
120 if (XFS_IS_REALTIME_INODE(ip))
121 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
122 else if (mp->m_logdev_targp != mp->m_ddev_targp)
123 xfs_blkdev_issue_flush(mp->m_ddev_targp);
124
125 /*
126 * All metadata updates are logged, which means that we just have to
127 * flush the log up to the latest LSN that touched the inode. If we have
128 * concurrent fsync/fdatasync() calls, we need them to all block on the
129 * log force before we clear the ili_fsync_fields field. This ensures
130 * that we don't get a racing sync operation that does not wait for the
131 * metadata to hit the journal before returning. If we race with
132 * clearing the ili_fsync_fields, then all that will happen is the log
133 * force will do nothing as the lsn will already be on disk. We can't
134 * race with setting ili_fsync_fields because that is done under
135 * XFS_ILOCK_EXCL, and that can't happen because we hold the lock shared
136 * until after the ili_fsync_fields is cleared.
137 */
138 xfs_ilock(ip, XFS_ILOCK_SHARED);
139 if (xfs_ipincount(ip)) {
140 if (!datasync ||
141 (iip->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
142 lsn = iip->ili_last_lsn;
143 }
144
145 if (lsn) {
146 error = xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
147 spin_lock(&iip->ili_lock);
148 iip->ili_fsync_fields = 0;
149 spin_unlock(&iip->ili_lock);
150 }
151 xfs_iunlock(ip, XFS_ILOCK_SHARED);
152
153 /*
154 * If we only have a single device, and the log force about was
155 * a no-op we might have to flush the data device cache here.
156 * This can only happen for fdatasync/O_DSYNC if we were overwriting
157 * an already allocated file and thus do not have any metadata to
158 * commit.
159 */
160 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
161 mp->m_logdev_targp == mp->m_ddev_targp)
162 xfs_blkdev_issue_flush(mp->m_ddev_targp);
163
164 return error;
165}
166
167STATIC ssize_t
168xfs_file_dio_aio_read(
169 struct kiocb *iocb,
170 struct iov_iter *to)
171{
172 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
173 size_t count = iov_iter_count(to);
174 ssize_t ret;
175
176 trace_xfs_file_direct_read(ip, count, iocb->ki_pos);
177
178 if (!count)
179 return 0; /* skip atime */
180
181 file_accessed(iocb->ki_filp);
182
183 if (iocb->ki_flags & IOCB_NOWAIT) {
184 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
185 return -EAGAIN;
186 } else {
187 xfs_ilock(ip, XFS_IOLOCK_SHARED);
188 }
189 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL,
190 is_sync_kiocb(iocb));
191 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
192
193 return ret;
194}
195
196static noinline ssize_t
197xfs_file_dax_read(
198 struct kiocb *iocb,
199 struct iov_iter *to)
200{
201 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
202 size_t count = iov_iter_count(to);
203 ssize_t ret = 0;
204
205 trace_xfs_file_dax_read(ip, count, iocb->ki_pos);
206
207 if (!count)
208 return 0; /* skip atime */
209
210 if (iocb->ki_flags & IOCB_NOWAIT) {
211 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
212 return -EAGAIN;
213 } else {
214 xfs_ilock(ip, XFS_IOLOCK_SHARED);
215 }
216
217 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
218 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
219
220 file_accessed(iocb->ki_filp);
221 return ret;
222}
223
224STATIC ssize_t
225xfs_file_buffered_aio_read(
226 struct kiocb *iocb,
227 struct iov_iter *to)
228{
229 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
230 ssize_t ret;
231
232 trace_xfs_file_buffered_read(ip, iov_iter_count(to), iocb->ki_pos);
233
234 if (iocb->ki_flags & IOCB_NOWAIT) {
235 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
236 return -EAGAIN;
237 } else {
238 xfs_ilock(ip, XFS_IOLOCK_SHARED);
239 }
240 ret = generic_file_read_iter(iocb, to);
241 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
242
243 return ret;
244}
245
246STATIC ssize_t
247xfs_file_read_iter(
248 struct kiocb *iocb,
249 struct iov_iter *to)
250{
251 struct inode *inode = file_inode(iocb->ki_filp);
252 struct xfs_mount *mp = XFS_I(inode)->i_mount;
253 ssize_t ret = 0;
254
255 XFS_STATS_INC(mp, xs_read_calls);
256
257 if (XFS_FORCED_SHUTDOWN(mp))
258 return -EIO;
259
260 if (IS_DAX(inode))
261 ret = xfs_file_dax_read(iocb, to);
262 else if (iocb->ki_flags & IOCB_DIRECT)
263 ret = xfs_file_dio_aio_read(iocb, to);
264 else
265 ret = xfs_file_buffered_aio_read(iocb, to);
266
267 if (ret > 0)
268 XFS_STATS_ADD(mp, xs_read_bytes, ret);
269 return ret;
270}
271
272/*
273 * Common pre-write limit and setup checks.
274 *
275 * Called with the iolocked held either shared and exclusive according to
276 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
277 * if called for a direct write beyond i_size.
278 */
279STATIC ssize_t
280xfs_file_aio_write_checks(
281 struct kiocb *iocb,
282 struct iov_iter *from,
283 int *iolock)
284{
285 struct file *file = iocb->ki_filp;
286 struct inode *inode = file->f_mapping->host;
287 struct xfs_inode *ip = XFS_I(inode);
288 ssize_t error = 0;
289 size_t count = iov_iter_count(from);
290 bool drained_dio = false;
291 loff_t isize;
292
293restart:
294 error = generic_write_checks(iocb, from);
295 if (error <= 0)
296 return error;
297
298 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
299 if (error)
300 return error;
301
302 /*
303 * For changing security info in file_remove_privs() we need i_rwsem
304 * exclusively.
305 */
306 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
307 xfs_iunlock(ip, *iolock);
308 *iolock = XFS_IOLOCK_EXCL;
309 xfs_ilock(ip, *iolock);
310 goto restart;
311 }
312 /*
313 * If the offset is beyond the size of the file, we need to zero any
314 * blocks that fall between the existing EOF and the start of this
315 * write. If zeroing is needed and we are currently holding the
316 * iolock shared, we need to update it to exclusive which implies
317 * having to redo all checks before.
318 *
319 * We need to serialise against EOF updates that occur in IO
320 * completions here. We want to make sure that nobody is changing the
321 * size while we do this check until we have placed an IO barrier (i.e.
322 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
323 * The spinlock effectively forms a memory barrier once we have the
324 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
325 * and hence be able to correctly determine if we need to run zeroing.
326 */
327 spin_lock(&ip->i_flags_lock);
328 isize = i_size_read(inode);
329 if (iocb->ki_pos > isize) {
330 spin_unlock(&ip->i_flags_lock);
331 if (!drained_dio) {
332 if (*iolock == XFS_IOLOCK_SHARED) {
333 xfs_iunlock(ip, *iolock);
334 *iolock = XFS_IOLOCK_EXCL;
335 xfs_ilock(ip, *iolock);
336 iov_iter_reexpand(from, count);
337 }
338 /*
339 * We now have an IO submission barrier in place, but
340 * AIO can do EOF updates during IO completion and hence
341 * we now need to wait for all of them to drain. Non-AIO
342 * DIO will have drained before we are given the
343 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
344 * no-op.
345 */
346 inode_dio_wait(inode);
347 drained_dio = true;
348 goto restart;
349 }
350
351 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
352 error = iomap_zero_range(inode, isize, iocb->ki_pos - isize,
353 NULL, &xfs_buffered_write_iomap_ops);
354 if (error)
355 return error;
356 } else
357 spin_unlock(&ip->i_flags_lock);
358
359 /*
360 * Updating the timestamps will grab the ilock again from
361 * xfs_fs_dirty_inode, so we have to call it after dropping the
362 * lock above. Eventually we should look into a way to avoid
363 * the pointless lock roundtrip.
364 */
365 return file_modified(file);
366}
367
368static int
369xfs_dio_write_end_io(
370 struct kiocb *iocb,
371 ssize_t size,
372 int error,
373 unsigned flags)
374{
375 struct inode *inode = file_inode(iocb->ki_filp);
376 struct xfs_inode *ip = XFS_I(inode);
377 loff_t offset = iocb->ki_pos;
378 unsigned int nofs_flag;
379
380 trace_xfs_end_io_direct_write(ip, offset, size);
381
382 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
383 return -EIO;
384
385 if (error)
386 return error;
387 if (!size)
388 return 0;
389
390 /*
391 * Capture amount written on completion as we can't reliably account
392 * for it on submission.
393 */
394 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
395
396 /*
397 * We can allocate memory here while doing writeback on behalf of
398 * memory reclaim. To avoid memory allocation deadlocks set the
399 * task-wide nofs context for the following operations.
400 */
401 nofs_flag = memalloc_nofs_save();
402
403 if (flags & IOMAP_DIO_COW) {
404 error = xfs_reflink_end_cow(ip, offset, size);
405 if (error)
406 goto out;
407 }
408
409 /*
410 * Unwritten conversion updates the in-core isize after extent
411 * conversion but before updating the on-disk size. Updating isize any
412 * earlier allows a racing dio read to find unwritten extents before
413 * they are converted.
414 */
415 if (flags & IOMAP_DIO_UNWRITTEN) {
416 error = xfs_iomap_write_unwritten(ip, offset, size, true);
417 goto out;
418 }
419
420 /*
421 * We need to update the in-core inode size here so that we don't end up
422 * with the on-disk inode size being outside the in-core inode size. We
423 * have no other method of updating EOF for AIO, so always do it here
424 * if necessary.
425 *
426 * We need to lock the test/set EOF update as we can be racing with
427 * other IO completions here to update the EOF. Failing to serialise
428 * here can result in EOF moving backwards and Bad Things Happen when
429 * that occurs.
430 */
431 spin_lock(&ip->i_flags_lock);
432 if (offset + size > i_size_read(inode)) {
433 i_size_write(inode, offset + size);
434 spin_unlock(&ip->i_flags_lock);
435 error = xfs_setfilesize(ip, offset, size);
436 } else {
437 spin_unlock(&ip->i_flags_lock);
438 }
439
440out:
441 memalloc_nofs_restore(nofs_flag);
442 return error;
443}
444
445static const struct iomap_dio_ops xfs_dio_write_ops = {
446 .end_io = xfs_dio_write_end_io,
447};
448
449/*
450 * xfs_file_dio_aio_write - handle direct IO writes
451 *
452 * Lock the inode appropriately to prepare for and issue a direct IO write.
453 * By separating it from the buffered write path we remove all the tricky to
454 * follow locking changes and looping.
455 *
456 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
457 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
458 * pages are flushed out.
459 *
460 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
461 * allowing them to be done in parallel with reads and other direct IO writes.
462 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
463 * needs to do sub-block zeroing and that requires serialisation against other
464 * direct IOs to the same block. In this case we need to serialise the
465 * submission of the unaligned IOs so that we don't get racing block zeroing in
466 * the dio layer. To avoid the problem with aio, we also need to wait for
467 * outstanding IOs to complete so that unwritten extent conversion is completed
468 * before we try to map the overlapping block. This is currently implemented by
469 * hitting it with a big hammer (i.e. inode_dio_wait()).
470 *
471 * Returns with locks held indicated by @iolock and errors indicated by
472 * negative return values.
473 */
474STATIC ssize_t
475xfs_file_dio_aio_write(
476 struct kiocb *iocb,
477 struct iov_iter *from)
478{
479 struct file *file = iocb->ki_filp;
480 struct address_space *mapping = file->f_mapping;
481 struct inode *inode = mapping->host;
482 struct xfs_inode *ip = XFS_I(inode);
483 struct xfs_mount *mp = ip->i_mount;
484 ssize_t ret = 0;
485 int unaligned_io = 0;
486 int iolock;
487 size_t count = iov_iter_count(from);
488 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
489
490 /* DIO must be aligned to device logical sector size */
491 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
492 return -EINVAL;
493
494 /*
495 * Don't take the exclusive iolock here unless the I/O is unaligned to
496 * the file system block size. We don't need to consider the EOF
497 * extension case here because xfs_file_aio_write_checks() will relock
498 * the inode as necessary for EOF zeroing cases and fill out the new
499 * inode size as appropriate.
500 */
501 if ((iocb->ki_pos & mp->m_blockmask) ||
502 ((iocb->ki_pos + count) & mp->m_blockmask)) {
503 unaligned_io = 1;
504
505 /*
506 * We can't properly handle unaligned direct I/O to reflink
507 * files yet, as we can't unshare a partial block.
508 */
509 if (xfs_is_cow_inode(ip)) {
510 trace_xfs_reflink_bounce_dio_write(ip, iocb->ki_pos, count);
511 return -ENOTBLK;
512 }
513 iolock = XFS_IOLOCK_EXCL;
514 } else {
515 iolock = XFS_IOLOCK_SHARED;
516 }
517
518 if (iocb->ki_flags & IOCB_NOWAIT) {
519 /* unaligned dio always waits, bail */
520 if (unaligned_io)
521 return -EAGAIN;
522 if (!xfs_ilock_nowait(ip, iolock))
523 return -EAGAIN;
524 } else {
525 xfs_ilock(ip, iolock);
526 }
527
528 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
529 if (ret)
530 goto out;
531 count = iov_iter_count(from);
532
533 /*
534 * If we are doing unaligned IO, we can't allow any other overlapping IO
535 * in-flight at the same time or we risk data corruption. Wait for all
536 * other IO to drain before we submit. If the IO is aligned, demote the
537 * iolock if we had to take the exclusive lock in
538 * xfs_file_aio_write_checks() for other reasons.
539 */
540 if (unaligned_io) {
541 inode_dio_wait(inode);
542 } else if (iolock == XFS_IOLOCK_EXCL) {
543 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
544 iolock = XFS_IOLOCK_SHARED;
545 }
546
547 trace_xfs_file_direct_write(ip, count, iocb->ki_pos);
548 /*
549 * If unaligned, this is the only IO in-flight. Wait on it before we
550 * release the iolock to prevent subsequent overlapping IO.
551 */
552 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
553 &xfs_dio_write_ops,
554 is_sync_kiocb(iocb) || unaligned_io);
555out:
556 xfs_iunlock(ip, iolock);
557
558 /*
559 * No fallback to buffered IO after short writes for XFS, direct I/O
560 * will either complete fully or return an error.
561 */
562 ASSERT(ret < 0 || ret == count);
563 return ret;
564}
565
566static noinline ssize_t
567xfs_file_dax_write(
568 struct kiocb *iocb,
569 struct iov_iter *from)
570{
571 struct inode *inode = iocb->ki_filp->f_mapping->host;
572 struct xfs_inode *ip = XFS_I(inode);
573 int iolock = XFS_IOLOCK_EXCL;
574 ssize_t ret, error = 0;
575 size_t count;
576 loff_t pos;
577
578 if (iocb->ki_flags & IOCB_NOWAIT) {
579 if (!xfs_ilock_nowait(ip, iolock))
580 return -EAGAIN;
581 } else {
582 xfs_ilock(ip, iolock);
583 }
584
585 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
586 if (ret)
587 goto out;
588
589 pos = iocb->ki_pos;
590 count = iov_iter_count(from);
591
592 trace_xfs_file_dax_write(ip, count, pos);
593 ret = dax_iomap_rw(iocb, from, &xfs_direct_write_iomap_ops);
594 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
595 i_size_write(inode, iocb->ki_pos);
596 error = xfs_setfilesize(ip, pos, ret);
597 }
598out:
599 xfs_iunlock(ip, iolock);
600 if (error)
601 return error;
602
603 if (ret > 0) {
604 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
605
606 /* Handle various SYNC-type writes */
607 ret = generic_write_sync(iocb, ret);
608 }
609 return ret;
610}
611
612STATIC ssize_t
613xfs_file_buffered_aio_write(
614 struct kiocb *iocb,
615 struct iov_iter *from)
616{
617 struct file *file = iocb->ki_filp;
618 struct address_space *mapping = file->f_mapping;
619 struct inode *inode = mapping->host;
620 struct xfs_inode *ip = XFS_I(inode);
621 ssize_t ret;
622 int enospc = 0;
623 int iolock;
624
625 if (iocb->ki_flags & IOCB_NOWAIT)
626 return -EOPNOTSUPP;
627
628write_retry:
629 iolock = XFS_IOLOCK_EXCL;
630 xfs_ilock(ip, iolock);
631
632 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
633 if (ret)
634 goto out;
635
636 /* We can write back this queue in page reclaim */
637 current->backing_dev_info = inode_to_bdi(inode);
638
639 trace_xfs_file_buffered_write(ip, iov_iter_count(from), iocb->ki_pos);
640 ret = iomap_file_buffered_write(iocb, from,
641 &xfs_buffered_write_iomap_ops);
642 if (likely(ret >= 0))
643 iocb->ki_pos += ret;
644
645 /*
646 * If we hit a space limit, try to free up some lingering preallocated
647 * space before returning an error. In the case of ENOSPC, first try to
648 * write back all dirty inodes to free up some of the excess reserved
649 * metadata space. This reduces the chances that the eofblocks scan
650 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
651 * also behaves as a filter to prevent too many eofblocks scans from
652 * running at the same time.
653 */
654 if (ret == -EDQUOT && !enospc) {
655 xfs_iunlock(ip, iolock);
656 enospc = xfs_inode_free_quota_eofblocks(ip);
657 if (enospc)
658 goto write_retry;
659 enospc = xfs_inode_free_quota_cowblocks(ip);
660 if (enospc)
661 goto write_retry;
662 iolock = 0;
663 } else if (ret == -ENOSPC && !enospc) {
664 struct xfs_eofblocks eofb = {0};
665
666 enospc = 1;
667 xfs_flush_inodes(ip->i_mount);
668
669 xfs_iunlock(ip, iolock);
670 eofb.eof_flags = XFS_EOF_FLAGS_SYNC;
671 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
672 xfs_icache_free_cowblocks(ip->i_mount, &eofb);
673 goto write_retry;
674 }
675
676 current->backing_dev_info = NULL;
677out:
678 if (iolock)
679 xfs_iunlock(ip, iolock);
680
681 if (ret > 0) {
682 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
683 /* Handle various SYNC-type writes */
684 ret = generic_write_sync(iocb, ret);
685 }
686 return ret;
687}
688
689STATIC ssize_t
690xfs_file_write_iter(
691 struct kiocb *iocb,
692 struct iov_iter *from)
693{
694 struct file *file = iocb->ki_filp;
695 struct address_space *mapping = file->f_mapping;
696 struct inode *inode = mapping->host;
697 struct xfs_inode *ip = XFS_I(inode);
698 ssize_t ret;
699 size_t ocount = iov_iter_count(from);
700
701 XFS_STATS_INC(ip->i_mount, xs_write_calls);
702
703 if (ocount == 0)
704 return 0;
705
706 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
707 return -EIO;
708
709 if (IS_DAX(inode))
710 return xfs_file_dax_write(iocb, from);
711
712 if (iocb->ki_flags & IOCB_DIRECT) {
713 /*
714 * Allow a directio write to fall back to a buffered
715 * write *only* in the case that we're doing a reflink
716 * CoW. In all other directio scenarios we do not
717 * allow an operation to fall back to buffered mode.
718 */
719 ret = xfs_file_dio_aio_write(iocb, from);
720 if (ret != -ENOTBLK)
721 return ret;
722 }
723
724 return xfs_file_buffered_aio_write(iocb, from);
725}
726
727static void
728xfs_wait_dax_page(
729 struct inode *inode)
730{
731 struct xfs_inode *ip = XFS_I(inode);
732
733 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
734 schedule();
735 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
736}
737
738static int
739xfs_break_dax_layouts(
740 struct inode *inode,
741 bool *retry)
742{
743 struct page *page;
744
745 ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
746
747 page = dax_layout_busy_page(inode->i_mapping);
748 if (!page)
749 return 0;
750
751 *retry = true;
752 return ___wait_var_event(&page->_refcount,
753 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
754 0, 0, xfs_wait_dax_page(inode));
755}
756
757int
758xfs_break_layouts(
759 struct inode *inode,
760 uint *iolock,
761 enum layout_break_reason reason)
762{
763 bool retry;
764 int error;
765
766 ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
767
768 do {
769 retry = false;
770 switch (reason) {
771 case BREAK_UNMAP:
772 error = xfs_break_dax_layouts(inode, &retry);
773 if (error || retry)
774 break;
775 /* fall through */
776 case BREAK_WRITE:
777 error = xfs_break_leased_layouts(inode, iolock, &retry);
778 break;
779 default:
780 WARN_ON_ONCE(1);
781 error = -EINVAL;
782 }
783 } while (error == 0 && retry);
784
785 return error;
786}
787
788#define XFS_FALLOC_FL_SUPPORTED \
789 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
790 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
791 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
792
793STATIC long
794xfs_file_fallocate(
795 struct file *file,
796 int mode,
797 loff_t offset,
798 loff_t len)
799{
800 struct inode *inode = file_inode(file);
801 struct xfs_inode *ip = XFS_I(inode);
802 long error;
803 enum xfs_prealloc_flags flags = 0;
804 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
805 loff_t new_size = 0;
806 bool do_file_insert = false;
807
808 if (!S_ISREG(inode->i_mode))
809 return -EINVAL;
810 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
811 return -EOPNOTSUPP;
812
813 xfs_ilock(ip, iolock);
814 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
815 if (error)
816 goto out_unlock;
817
818 /*
819 * Must wait for all AIO to complete before we continue as AIO can
820 * change the file size on completion without holding any locks we
821 * currently hold. We must do this first because AIO can update both
822 * the on disk and in memory inode sizes, and the operations that follow
823 * require the in-memory size to be fully up-to-date.
824 */
825 inode_dio_wait(inode);
826
827 /*
828 * Now AIO and DIO has drained we flush and (if necessary) invalidate
829 * the cached range over the first operation we are about to run.
830 *
831 * We care about zero and collapse here because they both run a hole
832 * punch over the range first. Because that can zero data, and the range
833 * of invalidation for the shift operations is much larger, we still do
834 * the required flush for collapse in xfs_prepare_shift().
835 *
836 * Insert has the same range requirements as collapse, and we extend the
837 * file first which can zero data. Hence insert has the same
838 * flush/invalidate requirements as collapse and so they are both
839 * handled at the right time by xfs_prepare_shift().
840 */
841 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
842 FALLOC_FL_COLLAPSE_RANGE)) {
843 error = xfs_flush_unmap_range(ip, offset, len);
844 if (error)
845 goto out_unlock;
846 }
847
848 if (mode & FALLOC_FL_PUNCH_HOLE) {
849 error = xfs_free_file_space(ip, offset, len);
850 if (error)
851 goto out_unlock;
852 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
853 unsigned int blksize_mask = i_blocksize(inode) - 1;
854
855 if (offset & blksize_mask || len & blksize_mask) {
856 error = -EINVAL;
857 goto out_unlock;
858 }
859
860 /*
861 * There is no need to overlap collapse range with EOF,
862 * in which case it is effectively a truncate operation
863 */
864 if (offset + len >= i_size_read(inode)) {
865 error = -EINVAL;
866 goto out_unlock;
867 }
868
869 new_size = i_size_read(inode) - len;
870
871 error = xfs_collapse_file_space(ip, offset, len);
872 if (error)
873 goto out_unlock;
874 } else if (mode & FALLOC_FL_INSERT_RANGE) {
875 unsigned int blksize_mask = i_blocksize(inode) - 1;
876 loff_t isize = i_size_read(inode);
877
878 if (offset & blksize_mask || len & blksize_mask) {
879 error = -EINVAL;
880 goto out_unlock;
881 }
882
883 /*
884 * New inode size must not exceed ->s_maxbytes, accounting for
885 * possible signed overflow.
886 */
887 if (inode->i_sb->s_maxbytes - isize < len) {
888 error = -EFBIG;
889 goto out_unlock;
890 }
891 new_size = isize + len;
892
893 /* Offset should be less than i_size */
894 if (offset >= isize) {
895 error = -EINVAL;
896 goto out_unlock;
897 }
898 do_file_insert = true;
899 } else {
900 flags |= XFS_PREALLOC_SET;
901
902 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
903 offset + len > i_size_read(inode)) {
904 new_size = offset + len;
905 error = inode_newsize_ok(inode, new_size);
906 if (error)
907 goto out_unlock;
908 }
909
910 if (mode & FALLOC_FL_ZERO_RANGE) {
911 /*
912 * Punch a hole and prealloc the range. We use a hole
913 * punch rather than unwritten extent conversion for two
914 * reasons:
915 *
916 * 1.) Hole punch handles partial block zeroing for us.
917 * 2.) If prealloc returns ENOSPC, the file range is
918 * still zero-valued by virtue of the hole punch.
919 */
920 unsigned int blksize = i_blocksize(inode);
921
922 trace_xfs_zero_file_space(ip);
923
924 error = xfs_free_file_space(ip, offset, len);
925 if (error)
926 goto out_unlock;
927
928 len = round_up(offset + len, blksize) -
929 round_down(offset, blksize);
930 offset = round_down(offset, blksize);
931 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
932 error = xfs_reflink_unshare(ip, offset, len);
933 if (error)
934 goto out_unlock;
935 } else {
936 /*
937 * If always_cow mode we can't use preallocations and
938 * thus should not create them.
939 */
940 if (xfs_is_always_cow_inode(ip)) {
941 error = -EOPNOTSUPP;
942 goto out_unlock;
943 }
944 }
945
946 if (!xfs_is_always_cow_inode(ip)) {
947 error = xfs_alloc_file_space(ip, offset, len,
948 XFS_BMAPI_PREALLOC);
949 if (error)
950 goto out_unlock;
951 }
952 }
953
954 if (file->f_flags & O_DSYNC)
955 flags |= XFS_PREALLOC_SYNC;
956
957 error = xfs_update_prealloc_flags(ip, flags);
958 if (error)
959 goto out_unlock;
960
961 /* Change file size if needed */
962 if (new_size) {
963 struct iattr iattr;
964
965 iattr.ia_valid = ATTR_SIZE;
966 iattr.ia_size = new_size;
967 error = xfs_vn_setattr_size(file_dentry(file), &iattr);
968 if (error)
969 goto out_unlock;
970 }
971
972 /*
973 * Perform hole insertion now that the file size has been
974 * updated so that if we crash during the operation we don't
975 * leave shifted extents past EOF and hence losing access to
976 * the data that is contained within them.
977 */
978 if (do_file_insert)
979 error = xfs_insert_file_space(ip, offset, len);
980
981out_unlock:
982 xfs_iunlock(ip, iolock);
983 return error;
984}
985
986STATIC int
987xfs_file_fadvise(
988 struct file *file,
989 loff_t start,
990 loff_t end,
991 int advice)
992{
993 struct xfs_inode *ip = XFS_I(file_inode(file));
994 int ret;
995 int lockflags = 0;
996
997 /*
998 * Operations creating pages in page cache need protection from hole
999 * punching and similar ops
1000 */
1001 if (advice == POSIX_FADV_WILLNEED) {
1002 lockflags = XFS_IOLOCK_SHARED;
1003 xfs_ilock(ip, lockflags);
1004 }
1005 ret = generic_fadvise(file, start, end, advice);
1006 if (lockflags)
1007 xfs_iunlock(ip, lockflags);
1008 return ret;
1009}
1010
1011STATIC loff_t
1012xfs_file_remap_range(
1013 struct file *file_in,
1014 loff_t pos_in,
1015 struct file *file_out,
1016 loff_t pos_out,
1017 loff_t len,
1018 unsigned int remap_flags)
1019{
1020 struct inode *inode_in = file_inode(file_in);
1021 struct xfs_inode *src = XFS_I(inode_in);
1022 struct inode *inode_out = file_inode(file_out);
1023 struct xfs_inode *dest = XFS_I(inode_out);
1024 struct xfs_mount *mp = src->i_mount;
1025 loff_t remapped = 0;
1026 xfs_extlen_t cowextsize;
1027 int ret;
1028
1029 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1030 return -EINVAL;
1031
1032 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1033 return -EOPNOTSUPP;
1034
1035 if (XFS_FORCED_SHUTDOWN(mp))
1036 return -EIO;
1037
1038 /* Prepare and then clone file data. */
1039 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1040 &len, remap_flags);
1041 if (ret || len == 0)
1042 return ret;
1043
1044 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1045
1046 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1047 &remapped);
1048 if (ret)
1049 goto out_unlock;
1050
1051 /*
1052 * Carry the cowextsize hint from src to dest if we're sharing the
1053 * entire source file to the entire destination file, the source file
1054 * has a cowextsize hint, and the destination file does not.
1055 */
1056 cowextsize = 0;
1057 if (pos_in == 0 && len == i_size_read(inode_in) &&
1058 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1059 pos_out == 0 && len >= i_size_read(inode_out) &&
1060 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1061 cowextsize = src->i_d.di_cowextsize;
1062
1063 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1064 remap_flags);
1065 if (ret)
1066 goto out_unlock;
1067
1068 if (mp->m_flags & XFS_MOUNT_WSYNC)
1069 xfs_log_force_inode(dest);
1070out_unlock:
1071 xfs_iunlock2_io_mmap(src, dest);
1072 if (ret)
1073 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1074 return remapped > 0 ? remapped : ret;
1075}
1076
1077STATIC int
1078xfs_file_open(
1079 struct inode *inode,
1080 struct file *file)
1081{
1082 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
1083 return -EFBIG;
1084 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
1085 return -EIO;
1086 file->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
1087 return 0;
1088}
1089
1090STATIC int
1091xfs_dir_open(
1092 struct inode *inode,
1093 struct file *file)
1094{
1095 struct xfs_inode *ip = XFS_I(inode);
1096 int mode;
1097 int error;
1098
1099 error = xfs_file_open(inode, file);
1100 if (error)
1101 return error;
1102
1103 /*
1104 * If there are any blocks, read-ahead block 0 as we're almost
1105 * certain to have the next operation be a read there.
1106 */
1107 mode = xfs_ilock_data_map_shared(ip);
1108 if (ip->i_df.if_nextents > 0)
1109 error = xfs_dir3_data_readahead(ip, 0, 0);
1110 xfs_iunlock(ip, mode);
1111 return error;
1112}
1113
1114STATIC int
1115xfs_file_release(
1116 struct inode *inode,
1117 struct file *filp)
1118{
1119 return xfs_release(XFS_I(inode));
1120}
1121
1122STATIC int
1123xfs_file_readdir(
1124 struct file *file,
1125 struct dir_context *ctx)
1126{
1127 struct inode *inode = file_inode(file);
1128 xfs_inode_t *ip = XFS_I(inode);
1129 size_t bufsize;
1130
1131 /*
1132 * The Linux API doesn't pass down the total size of the buffer
1133 * we read into down to the filesystem. With the filldir concept
1134 * it's not needed for correct information, but the XFS dir2 leaf
1135 * code wants an estimate of the buffer size to calculate it's
1136 * readahead window and size the buffers used for mapping to
1137 * physical blocks.
1138 *
1139 * Try to give it an estimate that's good enough, maybe at some
1140 * point we can change the ->readdir prototype to include the
1141 * buffer size. For now we use the current glibc buffer size.
1142 */
1143 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_d.di_size);
1144
1145 return xfs_readdir(NULL, ip, ctx, bufsize);
1146}
1147
1148STATIC loff_t
1149xfs_file_llseek(
1150 struct file *file,
1151 loff_t offset,
1152 int whence)
1153{
1154 struct inode *inode = file->f_mapping->host;
1155
1156 if (XFS_FORCED_SHUTDOWN(XFS_I(inode)->i_mount))
1157 return -EIO;
1158
1159 switch (whence) {
1160 default:
1161 return generic_file_llseek(file, offset, whence);
1162 case SEEK_HOLE:
1163 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1164 break;
1165 case SEEK_DATA:
1166 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1167 break;
1168 }
1169
1170 if (offset < 0)
1171 return offset;
1172 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1173}
1174
1175/*
1176 * Locking for serialisation of IO during page faults. This results in a lock
1177 * ordering of:
1178 *
1179 * mmap_lock (MM)
1180 * sb_start_pagefault(vfs, freeze)
1181 * i_mmaplock (XFS - truncate serialisation)
1182 * page_lock (MM)
1183 * i_lock (XFS - extent map serialisation)
1184 */
1185static vm_fault_t
1186__xfs_filemap_fault(
1187 struct vm_fault *vmf,
1188 enum page_entry_size pe_size,
1189 bool write_fault)
1190{
1191 struct inode *inode = file_inode(vmf->vma->vm_file);
1192 struct xfs_inode *ip = XFS_I(inode);
1193 vm_fault_t ret;
1194
1195 trace_xfs_filemap_fault(ip, pe_size, write_fault);
1196
1197 if (write_fault) {
1198 sb_start_pagefault(inode->i_sb);
1199 file_update_time(vmf->vma->vm_file);
1200 }
1201
1202 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1203 if (IS_DAX(inode)) {
1204 pfn_t pfn;
1205
1206 ret = dax_iomap_fault(vmf, pe_size, &pfn, NULL,
1207 (write_fault && !vmf->cow_page) ?
1208 &xfs_direct_write_iomap_ops :
1209 &xfs_read_iomap_ops);
1210 if (ret & VM_FAULT_NEEDDSYNC)
1211 ret = dax_finish_sync_fault(vmf, pe_size, pfn);
1212 } else {
1213 if (write_fault)
1214 ret = iomap_page_mkwrite(vmf,
1215 &xfs_buffered_write_iomap_ops);
1216 else
1217 ret = filemap_fault(vmf);
1218 }
1219 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1220
1221 if (write_fault)
1222 sb_end_pagefault(inode->i_sb);
1223 return ret;
1224}
1225
1226static inline bool
1227xfs_is_write_fault(
1228 struct vm_fault *vmf)
1229{
1230 return (vmf->flags & FAULT_FLAG_WRITE) &&
1231 (vmf->vma->vm_flags & VM_SHARED);
1232}
1233
1234static vm_fault_t
1235xfs_filemap_fault(
1236 struct vm_fault *vmf)
1237{
1238 /* DAX can shortcut the normal fault path on write faults! */
1239 return __xfs_filemap_fault(vmf, PE_SIZE_PTE,
1240 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1241 xfs_is_write_fault(vmf));
1242}
1243
1244static vm_fault_t
1245xfs_filemap_huge_fault(
1246 struct vm_fault *vmf,
1247 enum page_entry_size pe_size)
1248{
1249 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1250 return VM_FAULT_FALLBACK;
1251
1252 /* DAX can shortcut the normal fault path on write faults! */
1253 return __xfs_filemap_fault(vmf, pe_size,
1254 xfs_is_write_fault(vmf));
1255}
1256
1257static vm_fault_t
1258xfs_filemap_page_mkwrite(
1259 struct vm_fault *vmf)
1260{
1261 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1262}
1263
1264/*
1265 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1266 * on write faults. In reality, it needs to serialise against truncate and
1267 * prepare memory for writing so handle is as standard write fault.
1268 */
1269static vm_fault_t
1270xfs_filemap_pfn_mkwrite(
1271 struct vm_fault *vmf)
1272{
1273
1274 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1275}
1276
1277static void
1278xfs_filemap_map_pages(
1279 struct vm_fault *vmf,
1280 pgoff_t start_pgoff,
1281 pgoff_t end_pgoff)
1282{
1283 struct inode *inode = file_inode(vmf->vma->vm_file);
1284
1285 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1286 filemap_map_pages(vmf, start_pgoff, end_pgoff);
1287 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1288}
1289
1290static const struct vm_operations_struct xfs_file_vm_ops = {
1291 .fault = xfs_filemap_fault,
1292 .huge_fault = xfs_filemap_huge_fault,
1293 .map_pages = xfs_filemap_map_pages,
1294 .page_mkwrite = xfs_filemap_page_mkwrite,
1295 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1296};
1297
1298STATIC int
1299xfs_file_mmap(
1300 struct file *file,
1301 struct vm_area_struct *vma)
1302{
1303 struct inode *inode = file_inode(file);
1304 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1305
1306 /*
1307 * We don't support synchronous mappings for non-DAX files and
1308 * for DAX files if underneath dax_device is not synchronous.
1309 */
1310 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1311 return -EOPNOTSUPP;
1312
1313 file_accessed(file);
1314 vma->vm_ops = &xfs_file_vm_ops;
1315 if (IS_DAX(inode))
1316 vma->vm_flags |= VM_HUGEPAGE;
1317 return 0;
1318}
1319
1320const struct file_operations xfs_file_operations = {
1321 .llseek = xfs_file_llseek,
1322 .read_iter = xfs_file_read_iter,
1323 .write_iter = xfs_file_write_iter,
1324 .splice_read = generic_file_splice_read,
1325 .splice_write = iter_file_splice_write,
1326 .iopoll = iomap_dio_iopoll,
1327 .unlocked_ioctl = xfs_file_ioctl,
1328#ifdef CONFIG_COMPAT
1329 .compat_ioctl = xfs_file_compat_ioctl,
1330#endif
1331 .mmap = xfs_file_mmap,
1332 .mmap_supported_flags = MAP_SYNC,
1333 .open = xfs_file_open,
1334 .release = xfs_file_release,
1335 .fsync = xfs_file_fsync,
1336 .get_unmapped_area = thp_get_unmapped_area,
1337 .fallocate = xfs_file_fallocate,
1338 .fadvise = xfs_file_fadvise,
1339 .remap_file_range = xfs_file_remap_range,
1340};
1341
1342const struct file_operations xfs_dir_file_operations = {
1343 .open = xfs_dir_open,
1344 .read = generic_read_dir,
1345 .iterate_shared = xfs_file_readdir,
1346 .llseek = generic_file_llseek,
1347 .unlocked_ioctl = xfs_file_ioctl,
1348#ifdef CONFIG_COMPAT
1349 .compat_ioctl = xfs_file_compat_ioctl,
1350#endif
1351 .fsync = xfs_dir_fsync,
1352};