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/*
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};